AlphaEarth Foundations crop mapping

The recently released AlphaEarth Foundations Model was generated using multi-source satellite data (optical, radar, LiDAR, etc.) and consists of a 64-dimensional normalized vector for each 10m by 10m pixel on the whole globe for each year since 2017. This representation enables the use of cosine distance to quantify similarity between pixels in the parameter space. The model has been made available in Google Earth Engine, providing new opportunities for environmental monitoring. In this study, we revisited previous analyses of agricultural areas and bee pastures to evaluate the applicability of the new methodology. We compared its performance with earlier approaches based on machine learning classifiers and spectral indices calculations. Our results show that, in most cases, computations were significantly faster and required less code, while accuracy improved slightly compared to traditional methods. We also examined the spatial and temporal transferability of the approach, emphasizing its potential for broader applications in land monitoring.

Accurate mapping of tea plantations in fragmented, mountainous landscapes faces challenges from spectral confusion, cloud-induced data gaps, and limited model transferability. To address these issues, this study proposes a data-driven approach leveraging 64-dimensional Google AlphaEarth Foundations (AEF) satellite embeddings as core predictive features, integrated with Sentinel-2 spectral, textural, and topographic variables. Prior to feature optimization, comparative experiments confirmed that Random Forest outperformed Gradient Boosting Trees, Classification and Regression Trees, and Support Vector Machines in stability and accuracy, serving as the core classifier. Leveraging a robust sampling strategy, this study evaluated 12 classification scenarios. Results showed that the AEF-augmented scenario achieved the best performance in Rizhao (Overall Accuracy 92.69%, Kappa 0.90), with a high Producer’s Accuracy of 97.47% that effectively minimized omission errors. SHapley Additive exPlanations (SHAP) analysis revealed the model’s physically interpretable logic: utilizing embeddings as “exclusion filters” to separate tea from non-target classes by encoding latent phenological patterns, while relying on original spectral bands to capture canopy biological signals. Crucially, the model demonstrated exceptional generalizability when transferred to the unseen Qingdao region without retraining. This study validates AEF embeddings as a robust, scalable feature representation for regional crop monitoring in label-scarce and heterogeneous environments, offering a transferable data foundation for precise agricultural inventory and sustainable development planning.

No abstract available

Satellite imagery is now widely available at scale, yet reliable ground-truth labels remain scarce. Monitoring urban–forest boundaries over time is crucial for climate research and for improving wildfire risk assessment. Large-scale pretrained satellite foundational model embeddings provide a promising alternative to supervised learning for land-cover analysis. This work proposes an unsupervised land-use/land-cover analysis method using Google Earth Engine (GEE)’s AlphaEarth Foundations satellite embeddings and tests the effectiveness of unsupervised clustering applied to forest and urban segmentation. In our pipeline, we assess clustering algorithms on the 64-dimensional annual embeddings with global coverage over a regional study area and analyze the impact of cluster count on semantic separability. Cluster labels are assigned to land-cover classes using post-hoc majority overlap with ESA WorldCover, enabling quantitative evaluation without supervised training. Performance is assessed using Intersection over Union (IoU), F1-score, and accuracy. Results indicate that moderate cluster counts yield stable and semantically meaningful partitions, achieving up to 0.66 mean IoU and 81% accuracy for forest–urban discrimination. The proposed framework is fully and globally scalable within GEE and is designed to support additional clustering methods implemented in external machine learning libraries to provide efficient land-use/land-cover mapping in unsupervised environments.

El estudio compara la clasificación de nueve tipos de bosque en el Sistema Ibérico Norte mediante dos fuentes de datos de teledetección: compuestos estacionales Landsat 8 y 9 y embeddings de AlphaEarth Foundations (AEF) de Google. Los tipos de bosque se distinguen por la morfología foliar (aciculiforme o planifolia) y el hábito foliar (perennifolio, marcescente o caducifolio), incluyendo formaciones puras y mixtas. Se generó una máscara forestal a partir del Mapa Forestal de España y se seleccionaron muestras de entrenamiento. En el caso de Landsat, se calcularon índices espectrales y compuestos estacionales como variables predictoras. Los embeddings de AEF integran información multifuente (óptica, radar, LiDAR y climática) en vectores latentes que sintetizan la dinámica anual y la estructura de la vegetación. Ambas bases de datos se clasificaron con Random Forest y los resultados se compararon píxel a píxel mediante una matriz de confusión. Los resultados muestran mayor estabilidad en las clases puras, especialmente coníferas y caducifolias, y mayor confusión en las clases mixtas. Landsat captó mejor los patrones de hábito foliar, mientras que AEF proporcionó una delimitación espacial más precisa y mejor discriminación estructural.

Large-scale vegetation height mapping remains a key challenge in remote sensing, with implications for harvest planning, carbon accounting, and habitat assessment. Despite recent methodological advances, limitations persist, including signal saturation in high biomass stands and high preprocessing demands for multisensor harmonization. Advances in artificial intelligence, particularly geospatial foundation models, offer new opportunities to address these limitations. AlphaEarth Foundations (AEF; Google DeepMind, 2025) integrates multisource Earth observation data into compact representations that may alleviate some of these constraints. However, its performance relative to traditional approaches remains unclear. Here, we confronted models trained on AEF embeddings with models trained on conventional optical and SAR features derived from Sentinel-1 and Sentinel-2 data. We tested two modeling approaches–random forest (RF) and a deep learning U-Net architecture–resulting in four data-model configurations: AEF+RF, AEF+U-Net, Sentinel+RF, and Sentinel+U-Net. Vegetation height reference data were derived from uncrewed aerial vehicle-based canopy height models (40 sites; 478 ha). AEF+U-Net significantly outperformed all other configurations with R2 = 0.83, root-mean-square error (RMSE) = 1.30 m, and relative RMSE (rRMSE) = 54.7% (p<0.001, Diebold–Mariano test, K = 40 spatial blocks). Importantly, AEF+U-Net showed greater performance stability across the vegetation height gradient than all other configurations. Beyond predictive accuracy, AEF+U-Net required 29%–40% fewer manual preprocessing and feature-engineering tasks relative to conventional configurations, further supporting the operational viability of foundation model embedding-based approaches for vegetation structure mapping.

Abstract. Accurate mapping of forest types and vegetation characteristics is essential for monitoring biodiversity and forest dynamics. Traditional Deep Learning (DL) models trained on Sentinel-2 time series achieve high performance, but require extensive preprocessing and sensor-related fine-tuning. In this study, we evaluate the recently introduced AlphaEarth Foundations (AEF) embeddings, which is a global, multi-modal feature representation of the earths surface, for forest mapping in Italy. We compare a) a Random Forest model trained on Sentinel-2 and climate time series features, b) a Multi-Layer Perceptron trained on AEF, c) a Time-Series Transformer trained on Sentinel-2 and climate annual time series, and d) a Cross-Attention fusion model combining both feature sets. Using 5-fold cross-validation in a regression and a classification task on two datasets (evergreen broad-leaved tree cover ETC, forest vegetation type FVT) we find that the combined model consistently outperforms the single-source approaches (RMSE = 0.161, Acc = 0.757). AEF-based models achieve comparable accuracy to the Sentinel-2-based models, while reducing extensive time series preprocessing and training time by an order of magnitude. Feature attribution using integrated gradients reveals that AEF provides stable baseline representations, while Sentinel-2 inputs add phenology-related detail. The results show, that integrating generalized embeddings with specialized spectral-temporal features improves predictive performance for forest mapping.

Mangrove ecosystems are critical for climate change mitigation; however, monitoring these environments in high-precipitation regions, such as the Colombian Pacific coast, is often hindered by persistent cloud cover and complex terrain. This study addresses these challenges by implementing the novel Google AlphaEarth Foundations (AEF) technology, leveraging 64-dimensional embeddings integrated with Digital Elevation Models (DEM) and Slope data. For classification, a Random Forest (RF) algorithm was deployed using a subset of only 7 embedding dimensions alongside topographical variables. The model estimated a mangrove extent of 209,262 ha, compared to a reference baseline of 137,732 ha. This discrepancy is hypothesized to stem from the model’s ability to map changes in the mangrove forest due to anthropogenic and natural factors and species migrating upstream, areas frequently overlooked in traditional inventories. The classification performance, evaluated on a spatially independent hold-out validation set, yielded an Overall Accuracy of 0.9844 and a Kappa Index of 0.9747. Regarding biomass estimation, the RF algorithm utilized 4 embedding dimensions plus DEM and Slope to achieve a Coefficient of Determination (R2) of 0.5844, a Root Mean Square Error (RMSE) of 18.99 Mg/ha, and a Mean Absolute Error (MAE) of 14.88 Mg/ha within a range of 100–300 Mg/ha. These metrics represent a notable advancement, successfully mitigating the physical signal saturation that typically constrains traditional single-sensor remote sensing models at high biomass thresholds. Significant advantages of this methodology include the complete elimination of cloud interference and a drastic reduction in processing time. These findings demonstrate that the synergy between foundational models and machine learning provides a robust, scalable, and efficient framework for managing blue carbon stocks in critical tropical regions.

Litchi is an important cash crop in China. Accurate litchi orchard extraction is crucial for agricultural management and economic development. However, litchi orchards are spectrally very similar to evergreen broad-leaved forests and are poorly resolved in medium- and low-resolution satellite imagery, so large-scale extraction remains challenging. A cross-resolution fusion network via semantic anchor-graph enhancement (SAGE-Net) is therefore proposed, using a dual-branch encoder–decoder to aggregate complementary features of AlphaEarth Foundations (AEF) embeddings and GF-1 imagery. To enable semantic interaction across resolutions, the SAGE module is designed to gather the AEF embeddings into a few semantic anchors, reason over the anchor graph, and scatter them back under GF-1 guidance, so that semantic content and spatial extent are supplied by the source better suited to each. To refine the coarse prediction to full resolution without discarding what is already reliable, the reliability-adaptive correction (RAC) head is designed to gate a detail-guided correction by the confidence and local consistency of that prediction, concentrating refinement where it is uncertain. In Shenzhen, SAGE-Net attains the highest IoU (0.5660) and F1-score (0.7229) among eight comparison methods, with a Precision of 0.7162 and a Recall of 0.7297. Under stricter protocols that separate the training and the evaluation patches by a 256 m buffer or by whole spatial blocks, the IoU falls to 0.5224 and 0.5000 and SAGE-Net remains the most accurate of the networks retested, by a smaller margin. The results demonstrate the value of role-aware enhancement and reliability-adaptive correction for fine-grained orchard mapping, and confirm AEF embeddings as an effective supplementary data source for large-scale mapping.

No abstract available

No abstract available

No abstract available

Accurate irrigated cropland mapping is critical for agricultural water management and food security. Existing image-based irrigation mapping workflows primarily rely on vegetation indices and synthetic aperture radar (SAR) backscatter features, which have limited capacity to characterize the temporal evolution of irrigation processes and crop growth conditions. The AlphaEarth Foundation (AEF) model developed by Google DeepMind provides compact embeddings with temporal semantic information learned via self-supervision, yet their utility for irrigation mapping has not been systematically assessed. In this study, a comprehensive assessment of AEF embeddings for irrigated cropland mapping was performed in terms of feature separability, classification performance, and spatiotemporal transferability. Experiments were conducted in two representative irrigated regions: the Guanzhong Plain in China and Kansas in the USA. Class separability of the 64 embedding dimensions was quantified using the Jeffries–Matusita (JM) distance. Then, the AEF embeddings were compared with the Sentinel feature set (Sentinel-2 bands, normalized difference vegetation index(NDVI), enhanced vegetation index(EVI), normalized difference water index(NDWI) and Sentinel-1 vertical transmit vertical receive(VV), vertical transmit horizontal receive(VH)) using K-means clustering and supervised classifiers, including Decision Tree (DT), Random Forest (RF), Gradient Boosting Decision Trees (GBDT), Support Vector Machine (SVM), and Multi-layer Perceptron (MLP). Finally, transfer experiments across 2022 and 2024 in the Guanzhong Plain and Kansas were conducted to examine cross-year and cross-region performance. The results showed that AEF embeddings consistently provide stronger class separability in both study areas, with a maximum JM distance of 1.58 (A29). Using AEF embeddings, RF achieved overall accuracies (OA) of 0.95 in the Guanzhong Plain and 0.93 in Kansas, outperforming models based on Sentinel-1/2 bands and indices. Notably, unsupervised K-means clustering on AEF embeddings yielded OA > 0.85, indicating high intrinsic separability between irrigated and rainfed croplands. Transfer experiments further demonstrate stable temporal transfer (cross-year OA > 0.87), whereas cross-region transfer is constrained by differences in irrigation regimes, crop phenology and management practices, resulting in limited spatial generalization (OA~0.3). Overall, this study demonstrates the potential of high-information-density representations from geospatial foundation models for irrigated cropland mapping and provides methodological and technical insights to support transfer learning and operational mapping over large areas.

Accurate, timely crop yield estimates are critical for food security and agricultural planning, yet most operational pipelines still rely on handcrafted indices or weather-driven regressors. We assess whether Google DeepMind's AlphaEarth Foundations (AEF), a multimodal Geo-Foundation Model that provides annual 64-D embeddings at 10 m resolution, are sufficient for county-level U.S. corn yield estimation. Using one embedding per county-year and a light RBF-SVR learner, we evaluate leave-one-year-out generalization. Despite the absence of time-resolved inputs, the embeddings yield strong predictive accuracy across years. For corn, mean test performance is $\bar{R}^{2}=0.825$ with RMSE 14.8 bu/ac, reaching $R^{2}=0.861$ in the best year. The same workflow transfers to soybean with $\bar{R}^{2}=0.814$. Spatial diagnostics show strong positive autocorrelation of residuals every year, indicating missing slow-varying drivers rather than stochastic noise. Test-fold permutation importance shows feature importance, with several directions shared across crops, while a few flip sign, suggesting crop-specific semantics. These results demonstrate that annual AEF embeddings serve as analysis-ready features for county-scale yield estimation, enabling a lightweight, reproducible pipeline (mask, mean, regress) that is competitive with more complex alternatives.

Accurate mapping of agricultural land use activity, particularly long-term transition from cropland to pasture and short-term transition between cropland to fallow land, is essential for sustainable water management, drought response, and food-system resilience which directly supports United Nations Sustainable Development Goals (SDG-2 and SDG-8). However, reliable agricultural land use activity mapping is challenging due to spectral ambiguity, temporal irregularities, severe class imbalance, and limited generalization across agricultural regions. In this work, we propose AG-STELLA, a knowledge guided spatio-temporal model that (i) captures temporal changes of agricultural lands using pretrained spatio-temporal transformers; (ii) integrates geospatial context using AlphaEarth embedding; (iii) introduces a temporal transition latent space with temporal consistency constraints; (iv) employs guidance through hydroclimatic consistency; and (v) uses a land use-aware gated decoder to improve robustness across regions. Through experimentation across three water-stressed U.S. states, we show consistent gains over baseline vision and foundation models, achieving up to 27% F1-score improvement for pasture (minority class) and 16% overall. We further show the robustness across heterogeneous regions through cross-state transfer learning, where AG-STELLA consistently outperforms foundation model baselines and achieve up to 82.3% F1-score for fallow land with a 9.6% improvement over the best foundation model. Code: https://github.com/Nibir088/AG-STELLA

Land cover mapping from satellite imagery is a fundamental task for agricultural monitoring, especially in heterogeneous regions. Despite advances in space-based earth observation and artificial intelligence, accurately representing complex spatial patterns in agricultural landscapes remains challenging when mapping solely depends on traditional spectral features. Recently, foundation models such as Google’s AlphaEarth have introduced high-dimensional embedding representations derived from large-scale self-supervised learning on satellite imagery. In this study, we benchmark AlphaEarth embeddings for land cover classification in the Nile Delta, Egypt, and compare their performance against conventional Sentinel-2 spectral bands, widely used spectral indices, and their combination. Experimental results demonstrate that AlphaEarth embeddings exhibit improved classification performance in heterogeneous agricultural environments, highlighting their potential as an effective alternative to well-known spectral indices and raw spectral reflectance information for land cover classification.

Abstract. Mapping complex cropping patterns and temporal dynamics is of great significance for addressing the challenges faced by agricultural systems. However, in China, annual nationwide maps depicting multiple crops and rotation sequences are still lacking. In this study, we developed a data-driven crop mapping framework by integrating multiple existing crop products with the Google Satellite Embeddings derived from the AlphaEarth foundation model, and produced 10 m resolution mapping of complex cropping patterns across China from 2018 to 2021. Firstly, we integrated multiple publicly available crop mapping products within a harmonized framework that applies a unified cropland extent and cropping intensity, providing a systematic assessment of their consistency at pixel level. Consistency analysis results classify the study area into areas with consistency and areas with confusion, the latter serving as the mapping focus. Then, by combining harmonized crop data layers with random forest classifiers trained on foundation-derived embedding features, our framework effectively enhanced the spatial coherence and temporal stability, achieving an overall accuracy of 92.60 % and an F1 score of 0.7584. Compared with second-level administrative unit (ADM-2) statistics, the mapped cropping areas achieved high consistency (R2= 0.849, RMSE = 4.61 × 104 ha, MAE = 2.07 × 104 ha). The resulting datasets provide an integrated depiction of China's complex cropping systems, enabling consistent interannual assessments of changes in crop types, cropping sequences, and spatial patterns at 10 m resolution, thereby establishing a robust foundation for refined agricultural management and policy decisions, while supporting climate-smart and sustainable land-use planning.

This study introduces Hawaii Cropland Data Layers (HCDL) V2.0, the first operational crop mapping product leveraging Google DeepMind’s AlphaEarth Foundations Satellite Embedding dataset. The workflow replaces earlier multi-sensor image composites with 64-dimensional satellite embeddings to generate annual 10-meter resolution cropland maps for Hawaii in 2023 and 2024. The approach addresses challenges posed by persistent cloud cover, fragmented agricultural landscapes, and limited ground-reference data, while improving classification accuracy and consistency across 27 land-cover classes. Using Random Forest classifiers and a mixed sampling strategy, the embedding-based method achieved superior performance compared to previous composite-based models, with macro-average F1-score gains and reduced variability across classes. HCDL V2.0 provides a scalable, analysis-ready solution for agricultural monitoring, land-use planning, and disaster resilience in tropical regions. The datasets are publicly available and visualized through the AgriWatch web application, with HCDL 2025 currently under development.

The era of generative artificial intelligence is reshaping multiple scientific fields, and Earth Observation is no exception. One promising direction is the use of satellite embeddings as a virtual abstraction layer, enabling models to operate on semantically rich representations instead of heterogeneous, multi-source, and multi-sensor satellite bands. Here, the study evaluates AlphaEarth Embeddings for two LULC retrieval tasks in Morocco: (i) irrigation type discrimination and (ii) crop type classification. Ground-truth data were collected through field surveys conducted between 2023 and 2025 across Morocco's main agricultural regions, including information on field boundaries, crop type, and agricultural management practices. To build the supervised dataset, the 64-dimensional embedding was downloaded and matched with corresponding fields. Then, two classifiers, namely XGBoost and CatBoost, were compared against TabPFN, a state-of-the-art foundation model (FM) for tabular data learning. Experimental results demonstrate that TabPFN outperforms all trained models across both tasks. For irrigation-type retrieval, TabPFN achieves 0.85 accuracy under leave-one-out validation (LOO), compared to 0.81 for CatBoost and XGBoost; its 5 -fold cross-validation accuracy reaches 0.85, exceeding XGBoost (0.82) and CatBoost (0.81). Similarly, for crop-type classification, TabPFN attains 0.91 accuracy in LOO and in 5-fold cross-validation, outperforming both CatBoost and XGBoost. For training time, for a dataset of 5868 and 6888 rows, XGBoost was faster than CatBoost using 1000 iterations. However, TabPFN produced state-of-the-art performance with no training. This represents a shift and changing paradigm in the use of classic statistical machine learning approaches towards training-free models and may offer substantial computational advantages for large-scale operational deployments of geospatial FM. While the findings also highlight the strong retrieval capacity of satellite embeddings, an evaluation against satellite bands needs to be investigated in future work.

This article presents a spatially explicit persistent cropland mask for the Canadian prairies covering Alberta, Saskatchewan, Manitoba. Data was generated using 64-band embedding data from AlphaEarth with Agri-Food Canada (AAFC) Annual Crop Inventory used for label generation. Stratified-random points from across the prairies were used with a Random Forest classifier to determine cropland and noncropland areas. The trained models were applied across the prairies from 2017 to 2024 in an annual wall-to-wall classification framework at 10 m resolution for each year. Annual classifications were then combined to create a multi-year frequency layer where pixels with more than two years of continuous cropping were labelled as cropland, creating a stable mask layer. The dataset contains a 10 m resolution binary raster mask layer in GeoTIFF format to support a wide range of applications including cropland mapping, land-use change assessment, agricultural monitoring, yield modelling, soil and climate studies, and machine-learning-based geospatial applications. • Annual Prairie-wide cropland mask cloud-optimized GeoTIFFs generated using AlphaEarth data embeddings and Random Forest models trained on stratified random reference samples. • Multi-year cropland frequency and stable cropland mask layers derived from aggregated annual predictions, enabling consistent identification of persistent cropland across the Canadian Prairies.

Existing frameworks using Landsat/Sentinel imagery have shown limited classification accuracy, whereas the AlphaEarth Foundation (AEF) embedding model generates high-resolution, multi-source spatial layers with superior performance in sparse data regimes. This study used a 2017–2022 field survey from two Kansas counties to classify tillage intensity with Harmonized Landsat-Sentinel 2 (HLS) and AEF datasets. Random Forest (RF) and Tabular Prior-Data Fitted Network (TabPFN) models were employed for field-level classification, with results showing AEF embeddings improved accuracy compared to HLS only for RF (3% gain: 0.66 validation/0.67 test vs. 0.64/0.64 for HLS), while TabPFN exhibited negligible 1% accuracy difference between the two datasets (0.67 validation/0.67 test for AEF vs. 0.66/0.66 for HLS). This framework provides a scalable solution for sustainable agricultural management. This framework provides a scalable solution for sustainable agricultural management. Cross-year and crop-wise analyses further verified AEF’s robustness, with the model outperforming HLS across most years and crop types (soybeans and wheat) while performing comparably in corn and sorghum fields. Overall, this work evaluates the efficacy of cutting-edge satellite embedding technology against conventional remote sensing datasets, offering actionable guidance for advancing large-scale agricultural monitoring systems.

Abstract. Accurate large-scale monitoring of wheat cultivation and its crop rotation patterns is essential for food security and agricultural management. However, traditional remote sensing classification approaches typically rely on long-term multi-source imagery, complex feature engineering, and extensive labelled samples, limiting their scalability and spatiotemporal generalisation. To address these challenges, this study explores the potential of the AlphaEarth Foundation (AEF) embeddings—a global, annual, analysis-ready satellite embedding dataset—for winter wheat and crop rotation mapping. Firstly, we analyze AEF embeddings for intra-class consistency and inter-class separability, assessing their effectiveness in representing wheat. Subsequently, we compare multiple lightweight classifiers to identify an optimal model and conduct spatiotemporal generalization experiments across Henan Province from 2018 to 2024 using only limited labelled samples from 2020. Based on the resulting wheat maps, crop rotation patterns are further identified. Experimental results demonstrate that AEF embeddings exhibit strong semantic coherence and discriminative capability. Acceptable accuracy (OA=0.85) can already be achieved with simple models like cosine similarity and linear regression. More advanced lightweight classifiers further improve performance (OA=0.86–0.93) while maintaining stable results across different years and regions (spatial consistency=0.82). In addition, the crop rotation maps show high spatial agreement with existing products, while producing more spatially contiguous field patterns. Overall, AEF embeddings can serve as effective, ready-to-use features for large-scale agricultural remote sensing applications. By substantially reducing reliance on complex feature engineering and extensive training samples, they provide a practical and scalable solution for mapping winter wheat and its crop rotation patterns.

Accurate monitoring of rice cultivation areas is of paramount importance for global food security and economic stability. With the rapid advancement of Artificial Intelligence in remote sensing, feature representations derived from pre-trained foundation models are progressively reshaping traditional paradigms of crop mapping. However, the applicability of these high-dimensional abstract features to fine-scale paddy rice identification, along with their physical interpretability, remains insufficiently explored. To address this gap, this study proposes a novel rice extraction framework that integrates Google Satellite Embeddings with interpretable machine learning techniques. Taking Changxing County, Zhejiang Province, China, which has a complex terrain, as the study area, we coupled 64-dimensional embedding features with four mainstream classifiers (Random Forest (RF), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Gradient Boosted Trees (GBT).). Furthermore, the SHAP method was employed to deconstruct the decision-making mechanisms of the models. The results indicate that the embedding dataset encapsulates rich semantic information regarding land features, with all models achieving OA exceeding 94%. Notably, the SVM demonstrated optimal generalization capability and robustness, achieving an impressive OA of 96.69% and a Kappa coefficient of 93.37%. SHAP analysis further revealed the sparsity of the embedding feature contributions, identifying that dimension A58 (accounting for 11.8% of the contribution) alongside a minority of key channels encapsulates the primary discriminative information for paddy rice, exhibiting distinct non-linear response patterns. By bypassing laborious data preprocessing workflows, this framework offers a computationally efficient and interpretable new pathway for food security monitoring at regional and even global scales.

Cassava is a critical staple crop for food security and rural livelihoods in Sub-Saharan Africa, yet high-resolution maps of its distribution remain scarce, particularly for smallholder systems. In this study, we generated a 10 m resolution cassava presence map for Uganda (CM24) by fine-tuning a Random Forest classifier on TESSERA foundation model embeddings derived from Sentinel-1 and Sentinel-2 time series. Using field survey data from the Copernicus4GEOGLAM campaign for training and validation, the model achieved excellent discriminative ability (validation AUC = 0.9532, test AUC = 0.9524). Visual validation against high-resolution satellite imagery confirmed good spatial agreement, capturing both large contiguous fields and small fragmented plots. Comparison with two existing global products (CassavaMap and SPAM2020) and two seasons of national survey data conducted by the Uganda Bureau of Statistics showed that CM24 produced national harvested area estimates that fell between the two survey totals, whereas CassavaMap and SPAM2020 systematically overestimated harvested area by factors of two to three. Our results demonstrate that foundation-model embeddings offer a robust and scalable approach for mapping cassava in heterogeneous smallholder landscapes. The resulting CM24 map provides a spatially explicit tool to support disease surveillance, agricultural monitoring, and food security planning in Uganda and beyond.

No abstract available

Crop mapping using satellite imagery is crucial for agriculture applications. However, a fundamental challenge that hinders crop mapping progress is the scarcity of samples. The latest foundation model, Segment Anything Model (SAM), provides an opportunity to address this issue, yet few studies have been conducted in this area. This study investigated the parcel segmentation performance of SAM on commonly used medium-resolution satellite imagery (i.e., Sentinel-2 and Landsat-8) and proposed a novel automated sample generation framework based on SAM. The framework comprises three steps. First, an image optimization automatically selects high-quality images as the inputs for SAM. Then, potential samples are generated based on the masks produced by SAM. Finally, the potential samples are subsequently subjected to a sample cleaning procedure to acquire the most reliable samples. Experiments were conducted in Henan Province, China, and southern Ontario, Canada, using six proven effective classifiers. The effectiveness of our method is demonstrated through the combination of field-survey-collected samples and differently proportioned generated samples. Our results indicated that directly using SAM for parcel segmentation remains challenging, unless the parcels are large, regular in shape, and have distinct color differences from surroundings. Additionally, the proposed approach significantly improved the performance of classifiers and alleviated the sample scarcity problem. Compared to classifiers trained only by field-survey-collected samples, our method resulted in an average improvement of 16% and 78.5% in Henan and Ontario, respectively. The random forest achieved relatively good performance, with weighted-average F1 of 0.97 and 0.996 obtained using Sentinel-2 imagery in the two study areas, respectively. Our study contributes insights into solutions for sample scarcity in crop mapping and highlights the promising application of foundation models like SAM.

No abstract available

Recently, deep learning methods have achieved promising crop mapping results. Yet, their classification performance is constrained by the scarcity of labeled samples. Therefore, the development of methods capable of exploiting label-rich environments to classify crops in label-scarce environments using only a few labeled samples per class is required. Few-shot learning (FSL) methods have achieved this goal in computer vision for natural images, but they remain largely unexplored in crop mapping from time series data. In order to address this gap, we adapted eight FSL methods to map infrequent crops cultivated in the selected study areas from France and a large diversity of crops from a complex agricultural area situated in Ghana. The FSL methods are commonly evaluated using class-balanced unlabeled sets from the target domain data (query sets), leading to overestimated classification results. This is unrealistic since these sets can have an arbitrary number of samples per class. In our work, we used the Dirichlet distribution to model the class proportions in few-shot query sets as random variables. We demonstrated that transductive information maximization based on α-divergence (α-TIM) performs better than the competing methods, including dynamic time warping (DTW), which is commonly used to tackle the lack of labeled samples. α-TIM achieved, for example, a macro F1-score of 59.6% in Ghana in a 24-way 20-shot setting (i.e., 20 labeled samples from each of the 24 crop types) and a macro F1-score of 75.9% in a seven-way 20-shot setting in France, outperforming the second best-performing methods by 2.7% and 5.7%, respectively. Moreover, α-TIM outperformed a baseline deep learning model, highlighting the benefits of effectively integrating the query sets into the learning process.

Recent studies demonstrate the effectiveness of deep learning (DL) methods in crop mapping using Satellite Image Time Series (SITS). Despite this progress, the existing crop mapping methods often neglect prior knowledge, leading to unsatisfactory performance, especially in scenarios with limited data. To address this limitation, this paper introduces the Knowledge-Guided Crop Mapping Network (KGCMNet) as an innovative solution for crop mapping. Beyond a spatiotemporal encoder and a crop type prediction decoder, KGCMNet incorporates a knowledge reconstruction (KR) module for introducing essential crop-related prior knowlege. The KR module, utilizing NDVI as the chosen knowledge, aids the encoder in learning discriminative phenology from the SITS data. Experimental results demonstrate the outperformance of KGCMNet compared to other methods, showcasing the effectiveness of KR module in improving the accuracy across various crop types.

The accurate mapping of crop types and rotation patterns is essential for promoting sustainable agricultural development, particularly in ecologically fragile regions such as the farming–pastoral ecotone of China (FPEC). This study focuses on Zhangjiakou, a representative area of the FPEC, to develop a multi-sensor remote sensing framework for monitoring crop distribution and analyzing rotation dynamics. After cloud removal and Savitzky–Golay filtering were applied to correct noise, we selected vegetation index features with maximum inter-class separability during the optimal classification window (June 15–August 20) and generated quarterly Sentinel-1 SAR composites. A Random Forest classifier was employed to perform crop classification based on these optimized features, enabling 10 m resolution crop mapping from 2019 to 2023. The proposed method achieved high classification accuracy (overall accuracy and Kappa > 0.90), with strong agreement between mapped and statistical crop areas (R2: 0.85–0.88; RMSE: 0.42–0.58 × 104 ha). Spatial analysis revealed distinct distribution patterns: oats, potato, sesame, and vegetables were predominantly cultivated in northern Zhangjiakou, while maize dominated southern regions. We observed significant annual variations in crop area proportions and identified specific altitudinal preferences: maize, potato, and sesame were mainly grown at 480–520 m, while oats and other crops at 520–600 m. Slope analysis showed that most crops were cultivated on gentle slopes of 0–5°, with sesame extending to 4–10° slopes. Temporal analysis from 2019 to 2023 indicated that sesame, oats, and potato predominantly followed rotation patterns, while maize cultivation was primarily monoculture. Key drivers of rotation change included water scarcity, economic incentives, and continuous cropping constraints. These findings provide critical insights for optimizing crop rotation strategies, enhancing agricultural sustainability, and improving land-use efficiency in ecologically fragile regions.

Accurate crop classification using satellite imagery is critical for agricultural monitoring, yield estimation, and land-use planning. However, this task remains challenging due to the spectral similarity among crops. Although crops differ in physiological characteristics, including chlorophyll content, they often exhibit only subtle differences in their spectral reflectance, which make their precise discrimination challenging. To address this, this study uses the high temporal and spectral resolution of Sentinel-2 imagery, including its red-edge bands and derived vegetation indices, which are particularly sensitive to vegetation health and structural differences. This study presents a hybrid deep learning framework for crop classification, conducted through a case study in a complex agricultural region of Northern Italy. We investigated the combined use of spectral bands and NDVI & red-edge-based vegetation indices as inputs to hybrid deep learning models. Previous studies have applied 1D CNN, 2D CNN, LSTM, and GRU, often standalone, but their capacity to jointly process spectral and vegetative features through integrated CNN-RNN structures remains underexplored in mixed agricultural regions. To fill this gap, we developed and assessed four hybrid architectures: (1) 1D CNN-LSTM, (2) 1D CNN-GRU, (3) 2D CNN-LSTM, and (4) 2D CNN-GRU. These models were trained using optimized hyperparameters on combined spectral and vegetative input features. The 2D CNN-GRU model achieved the highest overall accuracy (99.12%) and F1-macro (99.14%), followed by 2D CNN-LSTM (98.51%), while 1D CNN-GRU and 1D CNN-LSTM performed slightly lower (93.46% and 92.54%), respectively.

Accurate and timely information regarding the locations and types of crops cultivated is essential for sustainable agriculture and ensuring food security. However, accurately mapping season-specific crop types in tropical and subtropical regions is challenging due to smallholder farms, fragmented fields, predominant clouds, and limited seasonal reference data. To address these limitations, this study employed optical and radar satellite data in conjunction with machine learning algorithms, including Random Forest (RF), Support Vector Machine (SVM), and Gradient Tree Boosting (GBoost), utilizing a large number of reference datasets across crop seasons. To validate the results, extensive field visits were undertaken throughout the year. Our focus centered on two regions in Thailand recognized for their small fields and frequent overcast conditions. Utilizing over 8000 reference points, we mapped 12 crop types in Chaiyaphum province and 13 crop types in Suphan Buri province for three cropping seasons in 2023. The RF algorithm proved to be the most effective, demonstrating superior performance across all seasons in comparison to the other models, achieving an overall accuracy exceeding 85%, with classifications for sugarcane and rice exceeding 90%. The resultant maps identified sugarcane, rice, and cassava as the principal crops in the region. This research exemplifies a methodology for producing highly accurate seasonal crop maps, providing valuable tools for making informed decisions for crop sustainable management, thereby supporting sustainable agriculture practices. Our findings underscore the potential of Earth observation satellites and machine learning algorithms in addressing significant agricultural challenges and facilitating the development of more resilient strategies for food security.

No abstract available

Mapping nationwide in-season crop-type data is a significant and challenging task in agriculture remote sensing. The existing data product for U.S. crop-type planting, such as the Cropland Data Layer (CDL), falls short in facilitating near-real-time applications. This article designed a workflow aimed at automating the generation of in-season CDL-like products for USA. We methodically extracted trusted pixels as land cover labels from historical CDL datasets, employing Sentinel-2, Landsat 8, and Landsat-9 as sources for spectrum data, using the random forest classifier to conduct nationwide crop-type classifications. These classifications were integrated into the In-Season Crop Data Layer (ICDL) covering the entire Conterminous United States (CONUS). This approach facilitated the efficient generation of ICDLs for May, June, and July 2022, achieving satisfactory accuracy in July. Compared to Nebraska and Iowa ground truth data, ICDL achieved F1 scores of (0.911, 0.845) for corn and (0.959, 0.969) for soybean. Furthermore, ICDL’s regional acreage estimates for major crops (corn, soybean, spring wheat, cotton, winter wheat, and rice) closely align with the U.S. Department of Agriculture (USDA) National Agricultural Statistics Service (NASS) figures, showing minimal variances as low as (0.01%, −0.68%, 0.19%, −4.39%, −5.78%, −1.28%). Notably, ICDL outperforms CDL in most assessments. This research consistently produces annual ICDLs from May to July that are readily accessible to the public in the iCrop system. Simultaneously, it presents an alternative technique for nationwide, in-season mapping of crop types.

Abstract Crop mapping is essential for agricultural management, economic development planning, and ecological conservation. Remote sensing with a large field of view provides us with a potential technique for large-scale crop mapping. However, most of the previous studies have focused on multi-temporal crop mapping, requiring multiple imaging over a period of time, which is impossible in the cloudy season due to the absence of clear atmospheric windows. Recently, with the progress of spaceborne hyperspectral imaging technology, wide-width hyperspectral satellite images, which provide abundant spectral and spatial structure information, have made the precision crop mapping of large areas possible. This paper focuses on deep learning based crop mapping using one-shot hyperspectral satellite imagery, where three convolutional neural network (CNN) models, i.e., 1D-CNN, 2D-CNN, and 3D-CNN models, are applied for end-to-end crop mapping. In addition, a manifold learning based visualization approach, i.e., t-distributed stochastic neighbor embedding (t-SNE), is introduced to illustrate the discriminative ability of the deep semantic features extracted by the different CNN models. To demonstrate the advantages of one-shot hyperspectral satellite images, an experiment was designed to compare the crop mapping performance of different remote sensing data sources, where both mono-temporal and multi-temporal multispectral images (MSIs) of the same research area were introduced for a systematic comparison. The classification accuracy when using hyperspectral satellite images was found to reach more than 94%, which was much better than that when using mono-temporal MSIs, and was comparable to the result when using multi-temporal MSIs. These findings will be important for the application of hyperspectral data when mapping large-area crop landscapes, and they confirm the potential of CNN models, particularly 3D-CNN models, for crop recognition.

Parcel-level cropland maps are an essential data source for crop yield estimation, precision agriculture, and many other agronomy applications. Here, we proposed a rice field mapping approach that combines agricultural field boundary extraction with fine-resolution satellite images and pixel-wise cropland classification with Sentinel-1 time series SAR (Synthetic Aperture Radar) imagery. The agricultural field boundaries were delineated by image segmentation using U-net-based fully convolutional network (FCN) models. Meanwhile, a simple decision-tree classifier was developed based on rice phenology traits to extract rice pixels with time series SAR imagery. Agricultural fields were then classified as rice or non-rice by majority voting from pixel-wise classification results. The evaluation indicated that SeresNet34, as the backbone of the U-net model, had the best performance in agricultural field extraction with an IoU (Intersection over Union) of 0.801 compared to the simple U-net and ResNet-based U-net. The combination of agricultural field maps with the rice pixel detection model showed promising improvement in the accuracy and resolution of rice mapping. The produced rice field map had an IoU score of 0.953, while the User‘s Accuracy and Producer‘s Accuracy of pixel-wise rice field mapping were 0.824 and 0.816, respectively. The proposed model combination scheme merely requires a simple pixel-wise cropland classification model that incorporates the agricultural field mapping results to produce high-accuracy and high-resolution cropland maps.

No abstract available

This study aims to automate sample data acquisition and identify the optimal feature combination from remote sensing imagery for delineating cropland boundaries. The objective is to enable real-time, automated extraction of cropland extent, thereby providing critical data for precision agricultural management and crop pest and disease monitoring. Utilizing high-resolution Sentinel-2 time-series data, we established 14 experimental groups based on distinct feature sets: original spectral features, vegetation index (VI) features, VI-derived texture features, and their combinations. Leveraging initial samples automatically procured from a comprehensive database, experiments were conducted using the LightGBM ensemble classifier to pinpoint the most effective experimental configuration for cropland extraction. Key findings include: (1) Temporal feature combinations derived from statistical domains outperform the direct stacking of multitemporal data features. (2) The importance of VIs for cropland extraction, ranked in descending order, is: NDTI > NDVI > LSWI > REP > NDBI > EVI. (3) Experimental groups incorporating texture features demonstrated superior accuracy compared to those relying solely on VI features, with average improvements in Accuracy, F1-score, and ROC AUC of $3.46 \%, 3.48 \%$, and 3.47 %, respectively.

Sustainable agriculture, particularly in semi-arid areas, depends on precise crop mapping and yield forecasting. In this study, Random Forest (RF) classification and regression implemented in Google Earth Engine (GEE) are combined with GeoAI and machine learning techniques. It evaluates crop distribution and predicts yields in Sokoto State, Nigeria, using Sentinel-2 surface reflectance imagery. 3,167.57 hectares of cropland were analysed using 690 cloud-filtered Sentinel-2 scenes that were gathered between January and December of 2023.  Crop classification demonstrated good reliability in identifying rice, pepper, and onion crops with an overall accuracy of 99.7% and a Kappa coefficient of 0.82. With a coefficient of determination (R²) of 0.84 and a Pearson correlation of 0.92, the Random Forest yield model predicted crop yields, demonstrating a very significant positive link between observed and anticipated yields. With a little positive bias of +9,244 kg/ha, the model's root mean squared error (RMSE) and mean absolute error (MAE) were 10,870 kg/ha and 9,244 kg/ha, respectively.  According to yield figures, the average yield was 20,971 kg/ha, with a range of 7,000 to 40,000 kg/ha. These findings show how precise crop mapping and yield estimation can be achieved by combining GeoAI, multispectral satellite data, and machine-learning approaches. The results assist better decision-making, sustainable farming methods, and increased food security in semi-arid regions by providing agricultural scientists and policymakers with practical insights.

The accurate extraction of cropland distribution is an important issue for precision agriculture and food security worldwide. The complex characteristics in southern China pose great challenges to the extraction. In this study, for the objective of accurate extraction and mapping of cropland parcels in multiple crop growth stages in southern China, we explored a method based on unmanned aerial vehicle (UAV) data and deep learning algorithms. Our method considered cropland size, cultivation patterns, spectral characteristics, and the terrain of the study area. From two aspects—model architecture of deep learning and the data form of UAV—four groups of experiments are performed to explore the optimal method for the extraction of cropland parcels in southern China. The optimal result obtained in October 2021 demonstrated an overall accuracy (OA) of 95.9%, a Kappa coefficient of 89.2%, and an Intersection-over-Union (IoU) of 95.7%. The optimal method also showed remarkable results in the maps of cropland distribution in multiple crop growth stages, with an average OA of 96.9%, an average Kappa coefficient of 89.5%, and an average IoU of 96.7% in August, November, and December of the same year. This study provides a valuable reference for the extraction of cropland parcels in multiple crop growth stages in southern China or regions with similar characteristics.

Cropland abandonment has multifaceted and controversial impacts on the natural environment and socioeconomic development. Utilizing remote sensing data offers the potential for comprehensive coverage and large-scale insights into automated abandoned cropland identification. However, accurately capturing small abandoned cropland, particularly in regions, such as southern China, with fragmentized farmland, poses a significant challenge using the traditional optical image-based mapping methods due to their coarse spatial resolution. In addition, irregular and chaotic textures of abandoned cropland further complicate the accurate prediction using very high resolution (VHR) data. In this article, we propose a novel deep learning network termed pixel attention contrastive network (PACnet) to map fine-grained abandoned cropland based on VHR data. Cross-image pixel contrast learning is introduced to discern distinctive features distinguishing abandoned cropland from other land types across various interimages. Moreover, a criss-cross attention module is embedded to enhance the contrasting characteristics within individual intraimages. Experimental outcomes validate the efficacy of PACnet, showcasing the highest accuracy (OA = 93.8% and mIOU = 71.7%) when compared with classical semantic segmentation networks. Our proposal not only underscores the potency of VHR remote sensing data in finely delineating abandoned cropland but also carries significant implications for cropland abandonment impact analysis and informed policy formulation.

No abstract available

Cropland abandonment is a worldwide problem that threatens food security and has significant consequences for the sustainable growth of the economy, society, and the natural ecosystem. However, detecting and mapping abandoned lands is challenging due to their diverse characteristics, like varying vegetation cover, spectral reflectance, and spatial patterns. To overcome these challenges, we employed Gaofen-6 (GF-6) imagery in conjunction with a Vision Transformer (ViT) model, harnessing self-attention and multi-scale feature learning to significantly enhance our ability to accurately and efficiently classify land covers. In Mianchi County, China, the study reveals that approximately 385 hectares of cropland (about 2.2% of the total cropland) were abandoned between 2019 and 2023. The highest annual abandonment occurred in 2021, with 214 hectares, followed by 170 hectares in 2023. The primary reason for the abandonment was the transformation of cropland into excavation activities, barren lands, and roadside greenways. The ViT’s performance peaked when multiple vegetation indices (VIs) were integrated into the GF-6 bands, resulting in the highest achieved results (F1 score = 0.89 and OA = 0.94). Our study represents an innovative approach by integrating ViT with 8 m multiband composite GF-6 imagery for precise identification and analysis of short-term cropland abandonment patterns, marking a distinct contribution compared to previous research. Moreover, our findings have broader implications for effective land use management, resource optimization, and addressing complex challenges in the field.

Mapping agricultural fields using high-resolution satellite imagery and deep learning (DL) models has advanced significantly, even in regions with small, irregularly shaped fields. However, effective DL models often require large, expensive labeled datasets, which are typically limited to specific years or regions. This restricts the ability to create annual maps needed for agricultural monitoring, as changes in farming practices and environmental conditions cause domain shifts between years and locations. To address this, we focused on improving model generalization without relying on yearly labels through a holistic approach that integrates several techniques, including an area-based loss function, Tversky-focal loss (TFL), data augmentation, and the use of regularization techniques like dropout. Photometric augmentations helped encode invariance to brightness changes but also increased the incidence of false positives. The best results were achieved by combining photometric augmentation, TFL, and Monte Carlo dropout, although dropout alone led to more false negatives. Input normalization also played a key role, with the best results obtained when normalization statistics were calculated locally (per chip) across all bands. Our U-Net-based workflow successfully generated multi-year crop maps over large areas, outperforming the base model without photometric augmentation or MC-dropout by 17 IoU points.

Timely and accurate cropland information at large spatial scales can improve crop management and support the government in decision making. Mapping the spatial extent and distribution of crops on a large spatial scale is challenging work due to the spatial variability. A multi-task spatiotemporal deep learning model, named LSTM-MTL, was developed in this study for large-scale rice mapping by utilizing time-series Sentinel-1 SAR data. The model showed a reasonable rice classification accuracy in the major rice production areas of the U.S. (OA = 98.3%, F1 score = 0.804), even when it only utilized SAR data. The model learned region-specific and common features simultaneously, and yielded a significant improved performance compared with RF and AtBiLSTM in both global and local training scenarios. We found that the LSTM-MTL model achieved a regional F1 score up to 10% higher than both global and local baseline models. The results demonstrated that the consideration of spatial variability via LSTM-MTL approach yielded an improved crop classification performance at a large spatial scale. We analyzed the input-output relationship through gradient backpropagation and found that low VH value in the early period and high VH value in the latter period were critical for rice classification. The results of in-season analysis showed that the model was able to yield a high accuracy (F1 score = 0.746) two months before rice maturity. The integration between multi-task learning and multi-temporal deep learning approach provides a promising approach for crop mapping at large spatial scales.

Abstract. Irrigation accounts for the major form of human water consumption and plays a pivotal role in enhancing crop yields and mitigating the effects of drought. Accurate mapping of irrigation distribution is essential for effective water resource management and the assessment of food security. However, the resolution of the global irrigated cropland map is coarse, typically approximately 10 km, and it lacks regular updates. In our study, we present a robust methodology that leverages irrigation performance during drought stress as an indicator of crop productivity and water consumption to identify global irrigated cropland. Within each irrigation mapping zone (IMZ), we identified the dry months of the growing season from 2017 to 2019 or the driest months from 2010 to 2019. To delineate irrigated cropland, we utilized the collected samples to calculate normalized difference vegetation index (NDVI) thresholds for the dry months of 2017 to 2019 and the NDVI deviation from the 10-year average for the driest month. By integrating the most accurate results from these two methods, we generated the Global Maximum Irrigation Extent dataset at 100 m resolution (GMIE-100), achieving an overall accuracy of 83.6 % ± 0.6 %. The GMIE-100 reveals that the maximum extent of irrigated cropland encompasses 403.17 ± 9.82 Mha, accounting for 23.4 % ± 0.6 % of the global cropland. Concentrated in fertile plains and regions adjacent to major rivers, the largest irrigated cropland areas are found in India, China, the United States, and Pakistan, which rank first to fourth, respectively. Importantly, the spatial resolution of GMIE-100 surpasses that of the dominant irrigation map, offering more detailed information essential to support estimates of agricultural water use and regional food security assessments. Furthermore, with the help of the deep learning (DL) method, the global central pivot irrigation system (CPIS) was identified using Pivot-Net, a novel convolutional neural network built on the U-net architecture. We found that there is 11.5 ± 0.01 Mha of CPIS, accounting for approximately 2.90 % ± 0.03 % of the total irrigated cropland. In Namibia, the United States, Saudi Arabia, South Africa, Canada, and Zambia, the CPIS proportion was greater than 10 %. To our knowledge, this is the inaugural study to undertake a global identification of specific irrigation methods, with a focus on the CPIS. The GMIE-100 dataset containing both the irrigated extent and CPIS distribution is publicly available on Harvard Dataverse at https://doi.org/10.7910/DVN/HKBAQQ (Tian et al., 2023a).

The study conducted experiments with different training strategies for scaling up U-Net convolutional neural network (CNN) models for regional land cover mapping with 2m multispectral WorldView (WV)-2 and –3, imagery in three distinct regions of Senegal which has complex seasonal wet/dry conditions and cropland-savanna mosaics over a large geographic extent. 106 images were selected across 25 10×10 km locations to serve as representative training and testing, which were validated by three independent observers (~16,000 points). Within each study region, our U-Net models rapidly achieved acceptable overall accuracies from using just 10-20 training images, after which additional training images did not improve accuracy beyond 70-85%. The two main target classes, Trees & Shrubs and Croplands had low commission and omission errors (7-18%). A model trained on 60 images from all three regions was applied to the full extent (> 65,000 km2) of 1,349 WV strips and achieved an accuracy of > 80%. We found that training data accuracy and diversity, as opposed to training data volume, determined model performance. Our experiments demonstrate tremendous potential for scaling up operational U-Net-based mapping in sub-Saharan African and highlighted practical challenges related to fundamental remote sensing issues which set an upper limit to AI’s performance.

Abstract Accurate crop mapping provides important and timely information for decision support on the estimation of crop production at large scale. Most existing crop-specific cover products based on remote sensing data and machine learning algorithms cannot serve large agriculture production areas as a result of poor model transfer capabilities. Developing a generalizable crop classification model for spatial transfer across regions is greatly needed. A deep learning approach, named DeepCropMapping (DCM), has been developed based on long short-term memory structure with attention mechanisms through integrating multi-temporal and multi-spectral remote sensing data for large-scale dynamic corn and soybean mapping. Full cross validation of classification experiments were conducted in six sites each covering 2,890,000 pixels at 30 m resolution in the U.S. corn belt from Year 2015 to 2018. Landsat Analysis Ready Data (ARD) and Cropland Data Layer (CDL) were adopted as the input satellite observations and ground reference, respectively. Transformer, Random Forest (RF), and Multilayer Perceptron (MLP) models were built for comparison. The DCM model produced a mean kappa score of 85.8% in base sites and a mean average kappa score of 82.0% in transfer sites at the end of the growing season. It yielded a comparable performance to Transformer and better than RF and MLP at the local test. The DCM model significantly outperformed other three models with a 95% confidence interval in the spatial transfer analysis. The results demonstrated the capability of learning generalizable features by the DCM model from ARD time series. The computational complexity analysis suggested that the DCM model required a shorter training time than Transformer but longer than MLP and RF. The results of the in-season classification experiment indicated the DCM model captured critical information from key growth phases and achieved higher accuracy than other models after the beginning of July. By monitoring the classification confidence in each time step, the results showed that the increased length of seasonal remote sensing time series would reduce the classification uncertainty in all sites. This study provided a viable option toward large-scale dynamic crop mapping through the integration of deep learning and remote sensing time series.

Mixed pixels often hinder accurate cropland mapping from remote sensing images with coarse spatial resolution. Image spatial super-resolution reconstruction technology is widely applied to address this issue, typically transforming coarse-resolution remote sensing images into fine spatial resolution images, which are then used to generate fine-resolution land cover maps using classification techniques. Deep learning has been widely used for image spatial super-resolution reconstruction; however, collecting training samples is often difficult for cropland mapping. Given that the quality of spatial super-resolution reconstruction directly impacts classification accuracy, this study aims to assess the impact of different types of training samples on image spatial super-resolution reconstruction and cropland mapping results by employing a Residual Channel Attention Network (RCAN) model combined with a spatial attention mechanism. Four types of samples were used for spatial super-resolution reconstruction model training, namely fine-resolution images and their corresponding coarse-resolution images, including original Sentinel-2 and degraded Sentinel-2 images, original GF-2 and degraded GF-2 images, histogram-matched GF-2 and degraded GF-2 images, and registered original GF-2 and Sentinel-2 images. The results indicate that the samples acquired by the histogram-matched GF-2 and degraded GF-2 images can resolve spectral band mismatches when simulating training samples from fine spatial resolution imagery, while the other three methods have limitations in their inability to fully address spectral and spatial mismatches. The histogram-matched method yielded the best image quality with PSNR, SSIM, and QNR values of 42.2813, 0.9778, and 0.9872, respectively, and produced the best mapping results, achieving an overall accuracy of 0.9306. By assessing the impact of training samples on image spatial super-resolution reconstruction and classification, this study addresses data limitations and contributes to improving the accuracy of cropland mapping, which is crucial for agricultural management and decision-making.

Prompt and precise cropland mapping is indispensable for safeguarding food security, enhancing land resource utilization, and advancing sustainable agricultural practices. Conventional approaches faced difficulties in complex terrain marked by fragmented plots, pronounced elevation differences, and non-uniform field borders. To address these challenges, we propose DAENet, a novel deep learning framework designed for accurate cropland extraction from high-resolution GaoFen-1 (GF-1) satellite imagery. DAENet employs a novel Geometric-Optimized and Boundary-Restrained (GOBR) Block, which combines channel attention, multi-scale spatial attention, and boundary supervision mechanisms to effectively mitigate challenges arising from disjointed cropland parcels, topography-cast shadows, and indistinct edges. We conducted comparative experiments using 8 mainstream semantic segmentation models. The results demonstrate that DAENet achieves superior performance, with an Intersection over Union (IoU) of 0.9636, representing a 4% improvement over the best-performing baseline, and an F1-score of 0.9811, marking a 2% increase. Ablation analysis further validated the indispensable contribution of GOBR modules in improving segmentation precision. Using our approach, we successfully extracted 25,556.98 hectares of cropland within the study area, encompassing a total of 67,850 individual blocks. Additionally, the proposed method exhibits robust generalization across varying spatial resolutions, underscoring its effectiveness as a high-accuracy solution for agricultural monitoring and sustainable land management in complex terrain.

Introduction Climate extremes increasingly threaten agricultural production, yet many artificial intelligence systems in agriculture remain local, reactive and narrowly trained for one crop, region or sensing modality. Methods We present AgriFM, a multimodal geospatial foundation model that combines satellite image time series, radar, thermal observations, weather trajectories, soil properties, topography and sparse management variables to estimate crop-stress and yield-failure risk across crops and regions. AgriFM was pretrained using self-supervised objectives on 2.4 million field-season sequences and evaluated on a curated benchmark spanning maize, wheat, soybean, rice and sorghum across five agroclimatic regions. Results In held-out geography and time-split evaluations, AgriFM improved early stress detection and yield-failure prediction over statistical, crop-model and deep-learning baselines. The largest gains occurred during compound drought and heat events, for which AgriFM produced alerts 18 to 24 days earlier than the satellite-only baseline while maintaining improved calibration. Phenology-conditioned fusion improved transfer across planting calendars, and uncertainty calibration reduced false alerts at fixed recall. Discussion Because the study is based on retrospective datasets, these findings establish cross-region retrospective performance rather than prospective field efficacy. The results support further field-based evaluation of multimodal foundation models for climate-resilient crop monitoring.

No abstract available

Fine-grained remote sensing classification using geospatial foundation model embeddings remains challenging due to high inter-class similarity and limited training data. We introduce IT-Flow (Interpolant Transport Flow), combining continuous normalizing flows with prototype-based learning for improved feature refinement in satellite imagery classification. Our method iteratively transforms embeddings from pre-trained Prithvi foundation model toward class prototypes through learned velocity fields, creating semantically meaningful trajectories in feature space. On the challenging EuroSAT Annual vs. Permanent Crop classification task, IT-Flow achieves 76.36% ± 1.67% F1 score across 10 random seeds, significantly outperforming standard adapter baseline (67.80% ± 4.12%) with p < 0.001. Our approach demonstrates both superior performance and 2.5× lower variance, making it particularly valuable for operational remote sensing applications requiring reliability for agricultural monitoring and food security.

The rapid advancement of foundation models in geoscience and remote sensing communities has presented great opportunities to unlock intelligent crop monitoring through Earth Observation (EO). However, existing general-purpose foundation models often fail to achieve satisfactory results in addressing the specific tasks especially for operational agricultural applications. This paper envisions the concept of an Agro-Geoinformation Foundation Model (AGFM) that is tailored to the unique characteristics of geospatial data and downstream tasks in agriculture. By leveraging diverse data sources such as crop type, vegetation dynamics, soil moisture, and meteorological data, AGFM aims to provide robust embeddings that can be fine-tuned for various downstream agricultural applications. Specifically, we outline a roadmap for the model development, including the plan for crop type mapping enhancement, the design of foundation model architectures, and examples of downstream tasks. The AGFM seeks to not only overcome current limitations in data silos, but also enhance the interpretability and accuracy of crop monitoring systems, advancing a more unified and scalable EO infrastructure for intelligent crop monitoring.

In this paper a new approach to the challenge of mapping integrated crop-livestock-forest systems is presented. A geospatial foundation model is leveraged to produce embeddings for images in examples of remote sensing time series. The embeddings, initially in high dimensional space, undergo a principal component analysis. Embedding trajectories are assembled considering the two principal components of each embedding from an image series, resulting in sequences of points in the plane. Considering time series from different cases of integrated systems, embedding trajectories are presented and the patterns for each case are analysed in terms of visual inspection and Euclidean distance based metrics, suggesting that the framework shows potential to be further explored in classification systems.

Recent geospatial foundation models (GFMs) have shown strong performance for crop yield prediction across regions and spatial scales, but operational deployment requires reliable performance under distribution drift. In agricultural settings, distribution drift may arise through temporal, label, concept, domain, or representation shifts, each of which can alter the relationship between incoming data and model behavior. This paper extends the previously published Fine-Tuning Agricultural Regression Models (FARM) framework, which adapts the Prithvi-EO-2.0-600M Vision Transformer for dense canola yield estimation over the Canadian Prairies, by introducing a systematic drift monitoring and reliability analysis pipeline. The proposed framework measures Mahalanobis, cosine, and Euclidean distances in both input and latent spaces. These distances characterize distributional shifts and allow us to evaluate their relationship with prediction error. Experiments on county-level and 30 m precision-agriculture datasets, spanning a normal growing season (2022) and a drift growing season (2021), show that both input-space and latent-space distances provide a useful signal of drift with latent-space metrics offering the clearest separation of anomalous conditions. At the same time, the results show that large drift scores do not necessarily correspond to large prediction errors. Some out-of-distribution samples are predicted accurately, while some in-distribution samples incur high error. These findings identify a critical gap between drift detection and uncertainty estimation and demonstrate that distributional distance alone is insufficient for operational reliability assessment. The results therefore motivate reliability-aware monitoring strategies for agricultural foundation models that combine drift detection with explicit uncertainty or error estimation in non-stationary environments.

Crop type mapping is a critical lever for decision making in smallholder food systems, yet high cost and expertise requirements of developing remote sensing-driven maps as well as need for recurrent data are persistent barriers to incorporating geospatial tools into decision making. To address this gap, we establish a three-part framework to evaluate crop type mapping: (1) performance, (2) temporal transferability, and (3) accessibility. We evaluate geospatial foundation model embeddings (TESSERA and AlphaEarth) against traditional baselines in Senegal’s groundnut basin. We find that the embedding-based approaches, in particular TESSERA, best fulfill these criteria. The TESSERA-based method shows a 78% reduction in both compute compared to a timeseries approach, a reduction in need for custom feature generation, and 28% higher relative accuracy (an absolute improvement of approximately 0.10 weighted F1) in one crop type classification temporal transfer test. By reducing the reliance on training data, computational resource use, and expertise for data processing, this approach shows potential to contribute to improved smallholder crop type classification.

Soil organic carbon (SOC) is a key indicator of soil health and climate resilience, yet large-scale spatial prediction remains challenging due to sparse field data and the complex dynamics of soil processes. We present FoundationSoil, a digital soil mapping framework that integrates a geospatial foundation model (Prithvi-EO-2.0) with environmental covariates to predict the spatial variability of SOC concentrations. Prithvi-EO-2.0 encodes multi-temporal satellite image patches via a transformer encoder, producing spatiotemporal representations that capture land-surface dynamics across seasons. These features are combined with 109 environmental covariates (e.g., climate, soil moisture) to train Random Forest and neural network regressors. We evaluated FoundationSoil using 12,692 georeferenced topsoil samples from the WoSIS database across the continental United States. Stratified grid-based partitions were applied to mitigate the influence of spatial autocorrelation on model performance evaluation. Results show that FoundationSoil consistently outperformed a pixel-based baseline, achieving best performance at R2 = 0.72 and RMSE = 38.62. Feature attribution shows that Prithvi-derived components capture semantically rich, non-redundant signals relevant to SOC processes. These results demonstrate the promise of pretrained vision foundation models for digital soil mapping and highlight their potential for data-efficient, transferable learning in Earth science applications.

The application of Artificial Intelligence (AI) in Earth Science and Remote Sensing has transformed environmental monitoring. Foundation models, pre-trained on massive unlabeled datasets, are particularly effective for geospatial tasks, offering adaptability and efficiency. In this study, we fine-tune Prithvi 100M, a transformer-based geospatial foundation model pre-trained on multispectral imagery from the Harmonized Landsat-Sentinel 2 (HLS) dataset, for burn scar mapping in the challenging Himalayan terrain.Burn scar mapping is crucial for monitoring post-fire landscapes, assessing ecosystem recovery, and disaster risk management. However, the complex topography and environmental conditions of the Himalayas pose significant challenges for traditional remote sensing methods. To address these challenges, A burn scar model based on Prithvi 100M was calibrated with region-specific burn scar labels and analyzed its performance using 2-band, 3-band, and 6-band spectral inputs. Our results revealed that 2-band and 3-band configurations achieved accuracy comparable to the 6-band setup, demonstrating the potential for reducing spectral inputs. Additionally, training on only 1/8th of the original data volume retained high accuracy, indicating the model’s robustness even with limited data.This study highlights the flexibility and effectiveness of geospatial foundation models for tasks in regions with limited training data. The ability to utilize fewer spectral bands without sacrificing performance is particularly advantageous for resource-constrained environments. Our findings demonstrate the promise of foundation models in remote sensing and pave the way for further applications in environmental monitoring and disaster management.

Geospatial foundation models are a new frontier in artificial intelligence, designed to understand and analyze spatial data at scale. Trained on huge sets of EO data, these models can support a wide range of applications—from monitoring natural disasters to guiding urban development and tracking climate change. To this aim, researchers and practitioners need to fine-tune the foundation models for specific tasks, utilizing a relatively small amount of additional data. As a result, geospatial foundation models are reshaping how we observe, manage, and protect our planet. Fine-tuning a geospatial foundation model requires carefully curated training datasets that reflect specific regions, time periods, or tasks—such as detecting deforestation or mapping urban growth. Yet preparing these datasets is often labor-intensive, involving steps like selecting relevant imagery, aligning spatial formats, and generating accurate labels. In practice, this means that the effectiveness of GFMs hinges on the availability of AI-ready data. This bottleneck limits the accessibility and scalability of GFMs for scientific and operational applications. In this work, we introduce a software library designed to automate these preparatory steps, streamlining the transformation of geospatial datasets into consistent, high-quality inputs for GFM fine-tuning. By reducing technical overhead and ensuring data readiness, the library enables faster, more reliable, and more inclusive adaptation of foundation models to local environmental challenges and specialized domain needs.

No abstract available

Frozen geospatial foundation model (GeoFM) embeddings are increasingly used as general-purpose features, yet their reliability under cross-city transfer remains unclear. We evaluate AlphaEarth and TESSERA for building-function and local-climate-zone mapping across 31 cities using leave-one-city-out (LOCO) transfer, comparing them with handcrafted Harmonized Landsat–Sentinel-2 (HLS) and Sentinel-1 features and annual band medians. When all representations are summarized over the same 90-m spatial support, both GeoFMs are competitive with the handcrafted baseline, although their average gains are modest: across both tasks and all three classifiers, the mean city-paired differences in macro-F1 are +0.011 for AlphaEarth and +0.012 for TESSERA, with both 95% confidence intervals spanning zero. Relative GeoFM performance also varies with label availability and spatial-support setting. Despite these small average differences, performance varies substantially from city to city. For AlphaEarth, this variation shows a clear contrast between Global North and Global South cities across both tasks; in building-function mapping, the association persists after adjustment for measured differences in urban form and reference-data characteristics, whereas the corresponding pattern is weaker for TESSERA. These findings show that competitive average accuracy does not necessarily imply reliable transfer across cities and highlight the importance of evaluating GeoFM embeddings across diverse deployment geographies.

Self-supervised learning techniques and geospatial foundation models are garnering increasing attention within the remote sensing community, marking a pivotal shift toward more adaptable, generalizable and data efficient methodologies. Despite their growing popularity, comparative analyses between these emergent models and traditional, task-specific supervised deep learning models for tasks such as semantic segmentation remain largely unexplored. This study is formulated as a comparative analysis to scrutinize the efficacy of the finetuned Prithvi geospatial foundation models against a supervised convolutional neural network trained from scratch for semantic segmentation of mangrove forests and pond aquaculture under varying levels of data availability and domain shift. Mangrove ecosystems are vital for coastal protection, biodiversity, and climate change mitigation. However, their conversion for economic activities like pond aquaculture requires timely and accurate land use land cover maps for effective monitoring. We introduce a novel labeled dataset tailored for this purpose and divide the training data into subsets (100%, 50%, 25%, 12.5%, and 5% sample size) to evaluate model performance under varying degrees of data availability. Our findings reveal that the U-Net model exhibits superior performance on all experiments except the 12.5% and 5% subsets. However, considering the difference between the pretraining and finetuning strategies, the results highlight the potential of foundation models to overcome the constraints of annotated data availability in remote sensing applications, offering a promising avenue for the scalable and efficient monitoring of critical coastal ecosystems.

No abstract available

No abstract available

Accurate crop yield modeling (CYM) is inherently challenging due to the complex, nonlinear, and temporally dynamic interactions of biotic and abiotic factors. Crop traits, which historically capture the cumulative effect of these factors, exhibit functional relationships critical for optimizing productivity. This underscores the necessity of multi-trait-based CYM approaches. Crop growth models enable trait dynamics with reflectance data and spectral indices as proxies for crop health and traits, respectively, to have real-time, spatially explicit monitoring. The Agricultural Production Systems sIMulator was calibrated to simulate multiple traits across the growth season based on geo-tagged wheat field ground information. Reflectance and spectral indices were processed for the geo-tagged fields across temporal observations to enable real-time, spatially explicit monitoring. Based on these parameters, this study addresses a critical gap in existing CYM frameworks by proposing a machine learning-based model that synergized multiple crop traits with reflectance and spectral indices to generate site-specific yield estimates. The performance evaluation revealed that the Long Short-Term Memory (LSTM) model achieved superior accuracy for the integrated parameters (RMSE = 250.68 kg/ha, MAE = 193.76 kg/ha, and R2 = 0.84), followed by traits alone. The Random Forest model followed the LSTM model, with an RMSE = 293.56 kg/ha, MAE = 230.68 kg/ha, and R2 = 0.78 for integrated parameters, and an RMSE = 291.73 kg/ha, MAE = 223.17 kg/ha, and R2 = 0.78 for crop traits. The superior prediction demonstrated the dominant role of multiple crop traits with satellite-derived reflectance metrics to develop robust CYM frameworks capable of capturing intra- and inter-field yield variability.

This study underscores the critical importance of accurate crop yield information for national food security and export considerations, with a specific focus on wheat yield estimation at the Gram Panchayat (GP) level in Bareilly district, Uttar Pradesh, using technologies such as machine learning algorithms (ML), the Decision Support System for Agrotechnology Transfer (DSSAT) crop model and semi-physical models (SPMs). The research integrates Sentinel-2 time-series data and ground data to generate comprehensive crop type maps. These maps offer insights into spatial variations in crop extent, growth stages and the leaf area index (LAI), serving as essential components for precise yield assessment. The classification of crops employed spectral matching techniques (SMTs) on Sentinel-2 time-series data, complemented by field surveys and ground data on crop management. The strategic identification of crop-cutting experiment (CCE) locations, based on a combination of crop type maps, soil data and weather parameters, further enhanced the precision of the study. A systematic comparison of three major crop yield estimation models revealed distinctive gaps in each approach. Machine learning models exhibit effectiveness in homogenous areas with similar cultivars, while the accuracy of a semi-physical model depends upon the resolution of the utilized data. The DSSAT model is effective in predicting yields at specific locations but faces difficulties when trying to extend these predictions to cover a larger study area. This research provides valuable insights for policymakers by providing near-real-time, high-resolution crop yield estimates at the local level, facilitating informed decision making in attaining food security.

Accurate detection of crop green-up is essential for estimating planting progress, constraining crop models, and supporting in-season decisions across the U.S. Corn Belt. Conventional approaches apply threshold- or inflection-based methods to vegetation-index time series, which are sensitive to cloud contamination and computationally heavy. Geospatial foundation models (GFMs) pre-trained on large Earthobservation archives offer a data-driven alternative for capturing the spectral-temporal dynamics of phenological transitions, yet their effectiveness for field-scale green-up detection is largely unexplored. We present a controlled benchmark of two GFM paradigms for pixel-level green-up day-of-year (DOY) estimation in corn and soybean: AlphaEarth Foundations (AEF; precomputed 64-D annual embeddings) and Prithvi-EO-2.0 (a 300M vision transformer fine-tuned on multi-temporal imagery), and we evaluate fusing Sentinel-1 SAR with Harmonized Landsat-Sentinel-2 (HLS) optical inputs. Using verified Beck's Hybrids planting dates and an NDVI-based green-up reference, with leaveone-year-out (LOYO) evaluation over Nebraska (2022-2024), we find that: (i) SAR fusion is decisive for cross-year generalization, raising every model to $r \approx 0.75-0.81$; (ii) the dense decoder dominates Prithvi's performance, more than fine-tuning the backbone; and (iii) lightweight AEF rivals the 300 M model on matched data (MAE 5.1 vs. 5.6 d) at orders-of-magnitude lower compute, while Prithvi attains the best absolute accuracy (4.2 $\mathbf{d}, r=0.81)$ given additional training years. This is the first systematic comparison of embedding-based and fine-tunable GFM paradigms for in-season crop phenology.

Cropland monitoring is important for ensuring food security in the context of global climate change and population growth. Freely available satellite data allow for the monitoring of large areas, while cloud-processing platforms enable a wide user community to apply remote sensing techniques. Remote sensing-based estimates of cropped area and crop types can thus assist sustainable land management in developing countries such as Ethiopia. In this study, we developed a method for cropland and crop type classification based on Sentinel-1 and Sentinel-2 time-series data using Google Earth Engine. Field data on 18 different crop types from three study areas in Ethiopia were available as reference for the years 2021 and 2022. First, a land use/land cover classification was performed to identify cropland areas. We then evaluated different input parameters derived from Sentinel-2 and Sentinel-1, and combinations thereof, for crop type classification. We assessed the accuracy and robustness of 33 supervised random forest models for classifying crop types for three study areas and two years. Our results showed that classification accuracies were highest when Sentinel-2 spectral bands were included. The addition of Sentinel-1 parameters only slightly improved the accuracy compared to Sentinel-2 parameters alone. The variant including S2 bands, EVI2, and NDRe2 from Sentinel-2 and VV, VH, and Diff from Sentinel-1 was finally applied for crop type classification. Investigation results of class-specific accuracies reinforced the importance of sufficient reference sample availability. The developed methods and classification results can assist regional experts in Ethiopia to support agricultural monitoring and land management.

No abstract available

No abstract available

ABSTRACT Crop type mapping visualizes the spatial distribution pattern and proportion of planting areas of different crop types, which is the basis for subsequent agricultural applications. Although optical remote sensing has been widely used to monitor crop dynamics, data are not always available due to cloud and other atmospheric effects on optical sensors. Satellite microwave systems such as Synthetic Aperture Radar (SAR) have all-time and all-weather advantages in monitoring ground and crop conditions, combining optical imagery and SAR imagery for crop type classification is of great significance. Our study mainly proposes seven feature combination schemes based on the combination of multi-temporal spectral features and texture features of Sentinel-2 (S2), and radar backscattering features of Sentinel-1 (S1) evaluate the influence of different data sources and different features on classification accuracy, obtains the optimal classification strategy and analyses the contribution of different features to classification result, in the aim of providing a new technical approach for the fine identification of crops from multi-source remote-sensing data. Results show that the crop classification accuracy of combined multi-time series spectral, texture, and radar features is higher than that of combining two types of features. The features subset selected by multi-period spectral, texture, and radar features have the best classification result, the overall accuracy (OA) and kappa coefficients reach 96.40% and 0.93, respectively. The study provides a method reference for future research on larger-scale remote-sensing crop precise extraction.

This article presents TimeSen2Crop, a pixel-based dataset made up of more than 1 million samples of Sentinel 2 time series (TSs) associated to 16 crop types. This dataset, publicly available, aims to contribute to the worldwide research related to the supervised classification of TSs of Sentinel 2 data for crop type mapping. TimeSen2Crop includes atmospherically corrected images and reports the snow, shadows, and clouds information per labeled unit. The provided TSs represent an agronomic year (i.e., period from one year's harvest to the next one for agricultural commodity) ranging from September 2017 to August 2018. To generate the dataset, the publicly available Austrian crop type map based on farmer's declarations has been considered. To ensure the selection of reliable labeled units from the map (i.e., pure pixels correctly associated to their labels), an automatic procedure for the extraction of the training set based on a multitemporal deep learning model has been defined. TimeSen2Crop also includes a TS of Sentinel 2 images acquired in the following agronomic year (i.e., from September 2018 to August 2019). These data are provided with the aim of attract more research activities for solving a typical challenge of the crop type mapping task: adapting multitemporal deep learning models to different year (domain adaptation). The design of the dataset is described along with a benchmark comparison of deep learning models for crop type mapping.

Large-scale crop type mapping often requires prediction beyond the environmental settings of the training sites. Shifts in crop phenology, field characteristics, or ecological site conditions in the previously unseen area, may reduce the classification performance of machine learning classifiers that often overfit to the training sites. This study aims to assess the spatial transferability of Random Forest models for crop type classification across Germany. The effects of different input datasets, i.e., only optical, only Synthetic Aperture Radar (SAR), and optical-SAR data combination, and the impact of spatial feature selection were systematically tested to identify the optimal approach that shows the highest accuracy in the transfer region. The spatial feature selection, a feature selection approach combined with spatial cross-validation, should remove features that carry site-specific information in the training data, which in turn can reduce the accuracy of the classification model in previously unseen areas. Seven study sites distributed over Germany were analyzed using reference data for the major 11 crops grown in the year 2018. Sentinel-1 and Sentinel-2 data from October 2017 to October 2018 were used as input. The accuracy estimation was performed using the spatially independent sample sets. The results of the optical-SAR combination outperformed those of single sensors in the training sites (maximum F1-score–0.85), and likewise in the areas not covered by training data (maximum F1-score–0.79). Random forest models based on only SAR features showed the lowest accuracy losses when transferred to unseen regions (average F1loss–0.04). In contrast to using the entire feature set, spatial feature selection substantially reduces the number of input features while preserving good predictive performance on unseen sites. Altogether, applying spatial feature selection to a combination of optical-SAR features or using SAR-only features is beneficial for large-scale crop type classification where training data is not evenly distributed over the complete study region.

Abstract The Sentinel-2 mission of the ESA’s Copernicus programme is generating unprecedented volumes of data at high spatial, spectral and temporal resolutions. The objective of this short communication is to assess the value of multi-temporal information for crop type classification using Sentinel-2 data. The analysis is carried out in an agricultural region in Austria and considers nine crop types during two years (2016 and 2017). To assess the impact of multi-temporal information, we applied a Random Forest (RF) classifier and analysed the results by using the RF out-of-bag error to calculate the overall accuracy (OA) and F1 score. The models were also validated using an independent reference dataset. Results show how the addition of multi-temporal information increases the crop type classification accuracy with similar trends for 2016 and 2017. At the very beginning of the crop growing season (March–April), the classification achieves relatively low accuracies (OA: ∼0.50). Significant increases in OA can be obtained between May and June, until the OA reaches its highest value in July. The final RF model was able to predict with very high confidence nine crop types for both years (OA: 0.95–0.96 and F1 score: 0.83–1.00). The independent validation dataset with more than 5000 reference plots showed comparable results (OA: 91–95% and F1 score: 0.74–0.99). We conclude that the multi-temporal crop type classification efficiently mitigates negative effects observed when using single-date acquisition within sub-optimal temporal windows.

Maintaining productive farmland necessitates precise crop mapping and identification. While satellite remote sensing makes it possible to generate such maps, there are still issues to resolve, such as how to choose input data and the best classifier algorithm, especially in areas with scarce field data. Accurate assessments of the land used for farming are a crucial part of national food supply and production accounting in many African countries, and to this end, remote sensing tools are being increasingly put to use. The aim of this study was to assess the potentiality of Sentinel-2 to distinguish and discriminate crop species in the study area and constraints on accurately mapping cropping patterns in the winter season in River Nile and Northern State, Sudan. The research utilized Sentinel-2 Normalized Different Vegetation Index (NDVI) at 10 m resolution, unsupervised and supervised classification method with ground sample and accuracy assessment. The results of the study found that the signatures of grain sorghum, wheat, okra, Vicia faba, alfalfa, corn, haricot, onion, potato, tomato, lupine, tree cover, and garlic have clear distinctions, permitting an overall accuracy of 87.38%, with trees cover, onion, wheat, potato, garlic, alfalfa, tomato, lupine and Vicia faba achieving more than 87% accuracy. Major mislabeling problems occurred primarily in irrigated areas for grain sorghum, okra, corn, and haricot, in wooded areas comprised of small parcels of land. The research found that high-resolution temporal images combined with ground data had potential and utility for mapping cropland at the field scale in the winter.

Crop type classification using Earth Observation (EO) data is challenging, particularly for crop types with similar phenological growth stages. In this regard, the synergy of optical and Synthetic-Aperture Radar (SAR) data enables a broad representation of biophysical and structural information on target objects, enhancing crop type mapping. However, the fusion of multi-sensor dense time-series data often comes with the challenge of high dimensional feature space. In this study, we (1) evaluate how the usage of only optical, only SAR, and their fusion affect the classification accuracy; (2) identify the combination of which time-steps and feature-sets lead to peak accuracy; (3) analyze misclassifications based on the parcel size, optical data availability, and crops’ temporal profiles. Two fusion approaches were considered and compared in this study: feature stacking and decision fusion. To distinguish the most relevant feature subsets time- and variable-wise, grouped forward feature selection (gFFS) was used. gFFS allows focusing analysis and interpretation on feature sets of interest like spectral bands, vegetation indices (VIs), or data sensing time rather than on single features. This feature selection strategy leads to better interpretability of results while substantially reducing computational expenses. The results showed that, in contrast to most other studies, SAR datasets outperform optical datasets. Similar to most other studies, the optical-SAR combination outperformed single sensor predictions. No significant difference was recorded between feature stacking and decision fusion. Random Forest (RF) appears to be robust to high feature space dimensionality. The feature selection did not improve the accuracies even for the optical-SAR feature stack with 320 features. Nevertheless, the combination of RF feature importance and time- and variable-wise gFFS rankings in one visualization enhances interpretability and understanding of the features’ relevance for specific classification tasks. For example, by enabling the identification of features that have high RF feature importance values but are, in their information content, correlated with other features. This study contributes to the growing domain of interpretable machine learning.

No abstract available

No abstract available

In-season crop type mapping can assist in early yield estimation, however, such data are not widely available. Currently available crop type maps mostly rely on either optical imagery or synthetic aperture radar (SAR), but there is a growing number of research that demonstrates the potential of synergistic optical and SAR data fusion. This research investigates the performance of machine learning approaches that account for both optical and SAR features to generate in-season crop type maps. Classification performance of three supervised machine learning algorithms: Random Forest (RF), Support Vector Machine (SVM) and Extreme Gradient Boosting (XGBoost), were tested. Experimental results demonstrate that the best performance for the in-season classification of corn and soybeans was obtained four months after the sowing (April – July) from the fusion of optical (Sentinel-2) and SAR (Sentinel-1) images. The in-season classification from SVM and RF demonstrated 81.2 % (overall accuracy) agreement with ground truth.

Crop Type classification using Semantic Segmentation and remote sensing data is an important tool for decision-making related to precision agriculture. Such classification remains an unsolved challenge due to the choice of landscape, processing methodology and selected satellite imagery and its optical features, and most importantly the availability and usage of such datasets in a developing country like Pakistan. State-of-the-art semantic segmentation models lack in processing the temporal dimension of time series imagery and evident solution to process multi-spectral bands available in the satellite imagery. We propose a methodology to overcome these shortcomings by selecting appropriate band combinations for crop type classification and treating time series visual data as a single image. The proposed methodology is evaluated on the data set of six different crops collected from National Agriculture Research Center (NARC) Islamabad. The experimental results yield 85% accuracy for classifying various crop types based on the evaluation of five different semantic segmentation models. The code and the trained models are available at https://github.com/asimniazi63/crop-type-narc for other researchers working in the same domain.

Crop-type classification has been attracting a lot of attention in recent years. In particular since the launch of the Sentinel-2 (S2) satellite which combines a large amount of spectral and spatial information, compared to previous satellite generations. In the literature, several methods exist that perform crop classification in time series, but most of them: i) work at pixel level; ii) perform single-data analysis; and/or iii) consider a single feature. This results in low performance of state-of-the-art methods. This paper presents an approach that works at object-level and exploits both spatial and temporal information coded in NDVI time series and phenological parameters and takes advantage of a semi-supervised paradigm by combining a new hierarchical correlation clustering with an artificial neural network. The effectiveness of the proposed approach was corroborated over an intensive cultivated area located in Barrax, Spain. Crop-type classification was compared to state-of-the-art methods.

This paper presents the results of a sub-pixel classification of crop types in Bulgaria from PROBA-V 100 m normalized difference vegetation index (NDVI) time series. Two sub-pixel classification methods, artificial neural network (ANN) and support vector regression (SVR) were used where the output was a set of area fraction images (AFIs) at 100 m resolution with pixels containing estimated area fractions of each class. High-resolution maps of two test sites derived from Sentinel-2 classifications were used to obtain training data for the sub-pixel classifications. The estimated area fractions have a good correspondence with the true area fractions when aggregated to regions of 10 × 10 km2, especially when the SVR method was used. For the five dominant classes in the test sites the R2 obtained after the aggregation was 86% (winter cereals), 81% (sunflower), 92% (broad-leaved forest), 89% (maize), and 67% (grasslands) when the SVR method was used.

The aim of this study is to predict and map winter wheat yield in the Parvomay municipality, situated in the Upper Thracian Lowland of Bulgaria, utilizing satellite data from Sentinel-2. The main crops grown in the research area are winter wheat, rapeseed, sunflower, and maize. To distinguish winter wheat fields accurately, we evaluated classification methods such as Support Vector Machines (SVM) and Random Forest (RF). These methods were applied to satellite multispectral data acquired by the Sentinel-2 satellites during the growing season of 2020–2021. In accordance with their development cycles, temporal image composites were developed to identify suitable moments when each crop is most accurately distinguished from others. Ground truth data obtained from the integrated administration and control system (IACS) were used for training the classifiers and assessing the accuracy of the final maps. Winter wheat fields were masked using the crop mask created from the best-performing classification algorithm. Yields were predicted with regression models calibrated with in situ data collected in the Parvomay study area. Both SVM and RF algorithms performed well in classifying winter wheat fields, with SVM slightly outperforming RF. The produced crop maps enable the application of crop-specific yield models on a regional scale. The best predictor of yield was the green NDVI index (GNDVI) from the April monthly composite image.

No abstract available

The object of the study is zero-shot crop-type classification in a data-poor target region (Karabakh, Azerbaijan) using a single-date Sentinel-2 composite, with the classifier trained on labeled parcels from a data-rich source region (central France). Cross-regional deployment of crop classifiers is impeded by domain shift differences in phenology, management, and sensor-band responses and by the absence of local labels, which together degrade accuracy and trust in operational maps. Cloud-free July-2021 median composites were produced in Google Earth Engine, a fourteen-band stack (core optical bands plus NDVI, NDRE, NDWI, NDMI) was assembled, four supervised algorithms were trained on balanced French parcels, validated using overall accuracy and Cohen’s κ, and then applied zero-shot to Karabakh. Random Forest yielded 94.6% accuracy on French validation and, after instance reweighting and feature normalization, delivered spatially coherent predictions in Karabakh. The pipeline’s combination of harmonized inputs, index-augmented spectra, and lightweight domain correction enabled transfer without target-region labels, generating confidence-aware maps suitable for rapid decision support. Growth-stage mismatch and spectral sensitivity are the main causes of performance differences, red-edge information was essential for distinguishing structurally similar crops, and moisture indices helped with irrigation-induced discrimination. The approach is most effective under peak-season, cloud-free conditions with comparable agro-ecological settings and a harmonized crop taxonomy, it requires only open Sentinel-2 data, a cropland mask, and standard ML tools in GEE, supporting scalable, repeatable assessments where ground truth is scarce

Abstract. The demand for food production is increasing rapidly with a surge in the population. To cope with this increasing food demand, precise agricultural management is essential. The existing techniques involve extensive field surveys for agricultural land discrimination. To minimize the man-hour efforts and time required by these techniques, automated techniques for precise crop type mapping and monitoring have been used. These techniques utilize satellite imagery and advanced machine learning techniques for crop type mapping and monitoring. However, the performance of such techniques is affected by factors such as fragmented land parcels, seasonal variability, and inconsistent field-level observations. To overcome these issues, this study attempts to classify grape and non-grape crops and monitor their phenological stages in the study area in Pune district, India, using Sentinel-2 satellite imagery and deep learning (DL) segmentation techniques: U-Net and DeepLabV3. Further, Sentinel- 1C SAR imagery (VV and VH polarization) for the years 2016 to 2024 was utilized to train and evaluate a long short-term memory network (LSTM) model with an aim to analyze the temporal behavior of the grape crop from pruning to harvesting stage with emphasis on growth stages like leaf set, fruit set, and ripening. The experimental results demonstrate that U-Net outperforms DeepLabV3 (F1-score: 0.96; mAP: 0.95) in grape crop classification. The LSTM model showed performance (F1-score 0.82) for phenological stage identification. This study can help agricultural stakeholders in effective and large-scale crop discrimination with minimum human intervention. It has the potential to reveal grape distribution and development stages in a faster time.

In-season crop type mapping is essential to agriculture management applications, including yield estimates, crop planting acreage statistics, food market predictions, and land use change analysis that support relevant decision-making, pushing economic development in certain agricultural export nations like Keny. This study employed a supervised machine learning method to produce three Kenya counties’ in-season crop-type maps in September 2023. We used surveyed growing crop ground truth data at the end of August 2023 and European Space Agency (ESA) WorldCover data serving as training labels, including nine crop types (Maize, Coffee, Grassland, Tea, Sugarcane, Exotic tree, Legumes, Vegetable, Native tree). The 15-day composite Sentinel-2 time series data was generated, incorporating training labels to assemble into training samples. They engaged in training a random forest classifier, conducting crop-type classifying in Nandi, Vihiga, and Kisumu Counties of Kenya. Moreover, the majority filter served to refine the classification. The validation results confirmed that grassland, sugarcane, tree, and tea possess high classification accuracy $(0.80-0.91)$, and coffee and maize showcase low accuracy (0.67 0.73) due to the massive mix pixels. This study attempted to produce in-season crop-type maps in an African nation with fragmented crop fields.

Recently, the advancements of foundation model (FM) have profoundly transformed the paradigm of data processing and analysis. Typically, FMs learn from large-scale data and can be efficiently adapted to various downstream tasks. Inspired by the prosperity of FMs developed for natural images, the field of remote sensing (RS) has also seen the emergence of foundation models (RSFMs). Although RSFMs have achieved remarkable performance in certain downstream tasks, these tasks are only limited to computer vision-related applications, leaving the specific needs of agriculture mostly unaddressed. In this paper, we briefly review recent works on RSFMs from three key perspectives, including model architecture, pretraining methods, and agricultural applications. Specifically, the gap between general computer vision-involved applications and agricultural applications are highlighted. Furthermore, we discuss the existing challenges faced by current RSFMs, and hope this work to provide valuable insights into how RSFMs can empower future development of agricultural community.

Precision agriculture increasingly relies on robust visual understanding of crops and fields. However, agricultural scenes remain challenging due to large appearance variability, background clutter, and fine-grained multiscale structures. Although remote sensing foundation models (RSFMs) have demonstrated strong generalization in generic Earth observation tasks, adapting them effectively to domain-specific agricultural applications is still nontrivial. Existing parameter-efficient fine-tuning (PEFT) strategies often struggle to balance semantic adaptability across diverse crop phenotypes and domain shifts with spatial fidelity in preserving fine-grained structures and boundaries, which limits their effectiveness on dense prediction tasks. To this end, we propose token-driven residual adaptation (TORA), an efficient adaptation paradigm tailored to multiscale agricultural vision. TORA augments frozen RSFMs with hierarchical learnable semantic prototypes to explicitly model domain-specific semantic patterns and to interact with intermediate representations through tokenwise attention. Meanwhile, we design an instance-aware residual path to inject adaptation signals into backbone features, thereby strengthening domain alignment while effectively mitigating the degradation of spatial details. Furthermore, we establish a comprehensive evaluation benchmark spanning three representative tasks—semantic segmentation, object detection, and image classification—across nine datasets. Extensive experiments demonstrate that TORA trains only approximately 1.5%–3.8% of the backbone parameters, substantially fewer than full fine-tuning (FFT), yet achieves performance that is competitive with or superior to strong PEFT baselines and approaches FFT on multiple datasets.

The modernization of agriculture is a cornerstone of national economic growth and sustainable development. Accurate delineation of agricultural field boundaries plays a vital role in field protection, agricultural yield forecasting, and land use regulation. However, existing approaches often rely on large-scale human annotations and exhibit poor generalization ability. With the increasing availability of data from autonomous agricultural operations, particularly the movement patterns of field machinery, a new paradigm for field boundary extraction emerges. In this article, we introduce FieldMapper, a novel multimodal framework designed to automatically delineate agricultural fields by synergistically leveraging remote sensing imagery and machinery trajectory data. The framework first employs a powerful vision foundation model to generate fine-grained field boundary proposals from satellite images. These proposals are then weakly annotated by aligning them with coverage patterns derived from machinery trajectories, eliminating the need for dense manual labeling. Finally, a discriminative model is trained through weakly supervised learning to distinguish true field instances from background candidates. Extensive experiments demonstrate FieldMapper outperforms the best baseline by 23.0% in terms of IoU, highlighting its robustness and effectiveness in real-world field delineation scenarios.

Earth monitoring has important applications across major social sectors including urban planning, agriculture, ecology, etc. Many of these applications rely on semantic segmentation, where the goal is to generate pixel-level classifications (e.g., buildings and crops) to inform decision-making. Developing a general-purpose model for the problem is often challenging at large scale, given the often limited ground truth labels (e.g., many tasks require field surveys), cross-region variability, and different sensor characteristics. Recent progress in industrial general-purpose vision foundation models (VFMs) has shown encouraging potential on a wide range of downstream tasks in the natural image domain. However, existing evaluations have not characterized the effectiveness of these off-the-shelf industrial VFMs on remote sensing segmentation tasks to understand the status and suitability for different types of Earth monitoring tasks. This is particularly important considering that the end goal of the general-purpose VFMs is to reach zero-shot generalizability to broad tasks like in the language domain. In this work, we present preliminary results on the performance of DINOv2, a state-of-the-art industrial VFM, for remote sensing image segmentation. We hypothesize that off-the-shelf DINOv2 performs well on tasks where targets have clear texture patterns and sharp boundaries, such as building footprint mapping, but may struggle with targets having fuzzy boundaries or a lack of texture signature. We consider DINOv2 to be effective if it can achieve comparable or better results compared to a regular-sized supervised U-Net model, and ineffective if there exists a large gap. The preliminary results align with our expectations, highlighting both the potential and current limitations of industrial VFMs in remote sensing segmentation tasks when there are limited textures and the boundaries are softer.

Precision agriculture increasingly relies on high-resolution, long-term remote sensing to delineate sub-field management zones. However, traditional spatial zonation assumes temporal stationarity, utilizing seasonal aggregates that obscure transient, intra-annual stress signals. This study develops a data-driven framework to characterize both persistent and non-stationary crop water use dynamics by integrating monthly, 30-m evapotranspiration (ET) data from OpenET (2000–2025) with zero-shot temporal anomaly detection. A pre-trained time-series foundation model (Chronos-T5-Small) generated counterfactual expectations for sub-field ET, quantifying deviations using a mean absolute error-based anomaly score. Unsupervised clustering of these anomaly scores with longitudinal ET metrics partitioned the landscape into dynamic biophysical regimes. Cross-registered against legacy persistence mapping based on seasonal totals, the foundation model showed strong directional agreement (86.1%, Cohen’s Kappa = 0.716) in identifying chronically constrained zones across 869 shared active pixels. Crucially, the framework identified 966 historically persistent pixels undergoing stability decay, of which 95.3% were statistically verified via paired t-tests to have collapsed into the field’s baseline variance pool. Furthermore, counterfactual anomaly detection isolated zones of recent acute divergence, differentiating enduring edaphic constraints from sudden system disruptions. This approach demonstrates how foundation models can transition from purely predictive engines to diagnostic instruments, advancing operational precision agriculture.

Abstract. Land cover classification plays a pivotal role in understanding and managing Earth's resources, influencing decisions in agriculture, forestry, urban planning, and environmental conservation. This study evaluates the performance of fine-tuning the Prithvi Foundation Model, Clay Foundation Model, and U-Net++ with Sentinel-2 imagery for land cover classification in the regions of Kalimantan and Sumatra, Indonesia. Using a dataset of Sentinel-2 image tiles labelled with 12 land cover classes, the models were trained and assessed using Intersection over Union (IoU) metrics. Results demonstrate the superior performance of the Clay Foundation Model, achieving a mean IoU of 0.4819. This research paper highlights the potential of Vision Transformer-based foundation models in distinguishing complex land cover categories and suggests directions for future research.

Remote sensing for land use and land cover (LULC) classification using satellite imagery provides valuable insights to the prediction of the dynamical change of land use, the risk assessments due to climate change by identifying areas prone to natural disasters, the protection of the environment, as well as the land management and planning. In recent years, deep learning-based image analysis approaches, such as geospatial foundation models for earth observation data that are trained through self-supervised learning, are increasingly used for earth observation tasks. A typical downstream task for such models is LULC classification, where the foundation model is fine-tuned using a set of labeled data. Depending on the dataset, these labels can be very specific to certain targets (only agriculture-focused labels), or to a certain region (only covering the US), and usually vary a lot in spatial resolution. To improve the generalizability and accuracy, there is a need for LULC models that better capture diverse land use patterns across different geographical areas. Here, we introduce an approach leveraging techniques from domain adaptation to integrate information of an existing LULC application into a new model more easily and improve the model’s performance. We were able to show that our approach increased the accuracy by 32% and 7% compared to fine-tuning the model with different labels individually, while also reducing the training time by over 65%.

Advances in remote sensing with higher-resolution images provide new opportunities for monitoring tasks in forestry, ecology, agriculture, energy, etc. However, applications at large scale often face the challenge of spatial variability, where patterns vary over geographic regions. Existing methods often assume a fixed, known spatial partitioning as input, which is mostly unavailable in practice, or exhaust models over locations that are computationally prohibitive for large-scale Earth monitoring. We propose an adaptive router approach, which leverages large, pretrained multimodal foundation models as a model router to automatically select scenario-specialized segmentation models for each input image. We evaluate our method using high-resolution tree mortality mapping, which has been recognized as a critical gap in large-scale carbon monitoring. The new method shows over 30 % improvement in polygon F1 score over baselines including vision foundation models.

Accurate segmentation of orchard plants in high-resolution remote sensing images is crucial for precision agriculture and orchard monitoring. Traditional methods rely on manual features, thresholds, or labeled training data, and have poor generalization in complex scenarios such as weeds and terraces. This paper proposes a zero-sample semantic segmentation framework based on the image segmentation foundation model Segment Anything Model (SAM), combined with the sliding window strategy to process ultra-high-resolution images, without the need for any specific task training or labeling. The study uses the unmanned aerial vehicle orthophoto images of the carambola plantation in Luodian County, Guizhou Province as test data. This framework divides the images into overlapping image blocks ($1024 \times 1024$ pixels, with an overlap of 128 pixels), and uses SAM to automatically generate segmentation masks. Key parameters were finely tuned: the number of points on each side was set to 56, the prediction IoU threshold was 0.70, the stability score threshold was 0.80, and the minimum mask area was 200 pixels. Through coordinate transformation correction, the geometric offset between the vector output and the original image was eliminated, and the pseudo-polygons generated from the background area were filtered out using the effective region filtering mechanism (with an effective pixel ratio threshold of 0.5). Experimental results on carambola images show that this method can accurately segment individual carambola plants, generate clear boundaries, and directly export GIS-compatible shapefile files. In the comparison experiment in the apple orchard, by adjusting parameters, the same framework can adapt to different crop forms and achieve block-level segmentation of the tree canopy overlapping area. This study provides a flexible, zero-sample solution for high-resolution image crop mapping without the need for labeled training data. Future work will integrate multimodal data (unmanned aerial vehicle, near-infrared, SAR) and lightweight fine-tuning strategies to further improve the segmentation granularity and adaptability.

ABSTRACT Land Use and Land Cover (LULC) classification is critical for environmental monitoring and sustainable resource management, but faces challenges in accurately capturing complex spatial-spectral features and long-range dependencies in remote sensing imagery. To address this, we introduce a Hybrid Wide Residual Network-Dual Spatial Positional Embedding Vision Transformer (WRN-DSPViT) framework enhanced with Squeeze-and-Excitation (SE) blocks and dual spatial positional embeddings. This model integrates a Wide Residual Network (WRN) for local spatial feature extraction and a Vision Transformer (ViT) with novel dual spatial encoding to capture global context via multi-head self-attention, where SE blocks dynamically recalibrate channel-wise features. Attention pooling is employed to fuse spatial features, allowing for adaptive weighting of important regions in the image, further enhancing classification accuracy. For multimodal hyperspectral-LiDAR data (Houston 2013), we extend this framework with parallel WRN-SE streams and cross-modal transformers, preserving spatial relationships through dual encodings. Evaluated on three benchmarks—EuroSAT (Multispectral), Houston 2013 (hyperspectral-LiDAR), and DeepGlobe (road extraction)—the hybrid WRN-DSPViT achieves state-of-the-art performance: 98.80% accuracy on EuroSAT (3.26 M parameters, 201.68 MFLOPS), 91.24% overall accuracy and 92.18% average accuracy on Houston 2013, and 0.802 F1-score/0.660 IoU on DeepGlobe.

Land cover classification has been of great interest as one of the most prominent applications of remote sensing images. The emergence of convolutional neural networks has largely promoted the development of land cover classification, but it ignores the positional relationship between pixels. When remotely sensed features have both large intraclass scale differences and interclass similarities, it will result in the problems of fuzzy class boundaries of classification results and misclassification of small samples, which are difficult to be solved by existing methods. Inspired by the recent Transformer network, we propose a self-attentive bilateral network SABNet to alleviate these problems. Its backbone consists of a modified multiscale vision transformer and a stacked convolutional layer for extracting global spatial information and local contextual information. A local embedding module and a coordinate attention fusion module are further proposed in the feature fusion stage to reduce attention distraction and efficiently fuse the high and low features. A stepwise feature fusion module is proposed in the decoder to fully fuse the features extracted from the two branches. Experiments show that our method achieves the best results in mIoU on both Landcover.ai and GID-15 datasets with a similar number of parameters, 91.49% for the Landcover.ai dataset and 64.23% for the GID-15 dataset, compared with existing methods.

Abstract. Land cover classification is essential basic information and key parameters for environmental change research, geographical and national monitoring, and sustainable development planning. Deep learning can automatically and multi-level extract the features of complex features, which has been proven to be an effective method for information extraction. However, one of the major challenges of deep learning is its poor interpret-ability, which makes it difficult to understand and explain the reasoning behind its classification results. This paper proposes a deep cross-modal coupling model (CMCM) for integrating semantic features and visual features. The representation of knowledge map is indicatively introduced into remote sensing image classification. Compared to previous studies, the proposed method provides accurate descriptions of the complex semantic objects within a complex land cover environment. The results showed that the integration of semantic knowledge improved the accuracy and interpret-ability of land cover classification.

No abstract available

Remote sensing land-cover classification can provide valuable data support for natural resource management. Existing classification methods based on graph-neural networks rely mainly on the global features and non-Euclidean structural features of image objects without considering the local features that describe their internal structures and the raster-depth features in the form of Euclidean structures. To this end, this paper presents a multi-scale and deep-feature, remote sensing???image land cover???classification method that embeds raster-depth features into node features and captures multi-scale graph-embedding information from global graphs and subgraphs to fully express image information. The depth-feature map of the image is obtained through a visual geometry Group 16???layer network and integrated into the feature space. The fractal network evolution algorithm is adopted to obtain multi-scale image objects. Global-scale features such as spectral, texture, index, and raster-depth features of the image objects are extracted, and local-scale features (e.g., average degree, average path length, graph diameter, average clustering coefficient, small-world effect) of the subgraphs are extracted to construct multi-scale depth features. The composite global graph structure is constructed by adopting adaptive weights, the graph embeddings are extracted via the graph convolutional network, and the node categories are predicted via SoftMax. For the Gaofen Image Dataset (GID‐15, a public benchmark dataset for land cover classification) and the 2017 China Computer Federation Remote Sensing Image Classification Dataset (CCF 2017, released in the 2017 China Computer Federation Big Data and Computational Intelligence Contest), as compared with the traditional method that considers only the global scale and the single-graph structure, this method improves the overall accuracy by 3.83% and 3.46%, respectively, and increases the kappa coefficient by 0.0681 and 0.0637, respectively, which indicates its effectiveness.

Multispectral land cover classification using Sentinel-2 imagery has traditionally relied on physically interpretable spectral bands and indices such as NDVI, NDWI, and NDBI. More recently, satellite foundation models have introduced pretrained embedding representations that capture spatial and semantic information learned from large-scale Earth observation data. While such embeddings have shown promising performance across remote sensing tasks, their behavior relative to classical spectral representations under realistic sensing conditions remains insufficiently characterized.In this work, we conduct a systematic empirical comparison of three representation families for multispectral land cover classification on the EuroSAT benchmark: (i) handcrafted spectral features derived from Sentinel-2 bands and indices, (ii) CNN-based multispectral models trained directly on the bands, and (iii) frozen satellite embeddings extracted from a pretrained foundation model. Beyond standard clean-set evaluation, we analyze performance under several practical scenarios, including additive spectral noise, band corruption and removal, limited labeled training data, and cross-region transfer.Our results show that all representations achieve strong performance under clean conditions, with overall accuracy exceeding 97%. However, differences become more apparent under degraded sensing and data-constrained settings. In particular, embedding-based representations tend to maintain higher accuracy under spectral noise and band removal, require fewer labeled samples to reach competitive performance, and exhibit smaller performance drops in cross-region evaluation. These observations suggest that pretrained satellite embeddings can provide a useful and robust representation for multispectral land cover classification, while classical spectral features remain attractive when interpretability and simplicity are priorities.

No abstract available

No abstract available

No abstract available

Most supervised classification methods for polarimetric synthetic aperture radar (PolSAR) data rely on abundant labeled samples, and cannot tackle the problem that categorizes or infers unseen land cover classes without training samples. Aiming to categorize instances from both seen and unseen classes simultaneously, a generalized zero-shot learning (GZSL)-based PolSAR land cover classification framework is proposed. The semantic attributes are first collected to describe characteristics of typical land cover types in PolSAR images, and semantic relevance between attributes is established to relate unseen and seen classes. Via latent embedding, the projection between mid-level polarimetric features and semantic attributes for each land cover class can be obtained during the training stage. The GZSL model for PolSAR data is constructed by mid-level polarimetric features, the projection relationship, and the semantic relevance. Finally, the labels of the test instances can be predicted, even for some unseen classes. Experiments on three real RadarSAT-2 PolSAR datasets show that the proposed framework can classify both seen and unseen land cover classes with limited kinds of training classes, which reduces the requirement for labeled samples. The classification accuracy of the unseen land cover class reaches about 73% if semantic relevance exists during the training stage.

No abstract available

In the field of remote sensing, data acquired from a single sensor usually can't meet the needs of some special applications, because the information extracted from the data are often incomplete and limited. Data fusing can solve this problem, but it will lead to the redundant information. In this paper, we proposed a novel manifold learning approach to perform dimensionality reduction for the fusing optical and SAR data. And three typical manifold learning models, namely, ISOMAP, local linear embedding (LLE) and principle component analysis (PCA), were utilized to test the robustness of our method by comparing with the land cover classification results. Our experimental results showed that our proposed method obtained the best land cover classification results among these approaches for the fusing optical and SAR data.

No abstract available

Abstract. Accurate land-cover maps are essential, but medium-resolution imagery (e.g., Sentinel-2 at 10 m) often contains mixed pixels that include multiple land-cover types. Standard “hard” classification assigns one class per pixel, hiding minority classes and reducing map usefulness. Compositional classification instead estimates the proportion of each class within a pixel, preserving sub-pixel detail, but requires outputs that are non-negative and sum to one, constraints not naturally handled by typical ML/DL losses. This study proposed and evaluated a deep-learning framework for compositional land-cover estimation at 10 m resolution. It compared two input feature sets: (1) reflectance from 10 Sentinel-2 multispectral bands (B2–B8, B8A, B11, B12) and (2) Embedding V1, a 64- dimensional representation from the AlphaEarth Foundations model that integrates multi-source, multi-temporal Earth observation signals. Ground-truth composition vectors were derived from OpenEarthMap by aggregating 0.25–0.5 m labels to 10 m pixels for eight classes. Three architectures (MLP, 2D-CNN, 3D-CNN) used Softmax outputs to enforce the constant-sum constraint, and two losses (MAE vs Aitchison distance) were tested. Embedding V1 improved estimation accuracy across all model architectures compared to Sentinel-2 spectral bands alone. While 3D-CNN achieved the best performance with Sentinel-2 input (MAE: 0.1126), MLP outperformed all other architectures when Embedding V1 was used (MAE: 0.0989). Comparison of fraction maps revealed that MAE produced spatially smoothed outputs, whereas Aitchison distance yielded sharper and more realistic compositions. The combination of Embedding V1, MLP, and Aitchison distance loss achieved the best overall performance, suggesting that foundation model embeddings combined with compositional losses can improve sub-pixel land-cover estimation.