Understanding how canopy structure and plant nutritional status jointly regulate crop productivity remains a central challenge for precision agriculture, particularly when observations are limited to single growth stages. This study examines whether three-dimensional canopy information derived from unmanned aerial vehicle (UAV) LiDAR can be integrated with multispectral observations to improve spatial characterization of potato yield potential and nitrogen status under irrigated Prairie conditions.

Multispectral imagery and high-density UAV-LiDAR data were acquired at row closure across two growing seasons in southwestern Manitoba, Canada, spanning a controlled gradient of nitrogen availability. Rather than treating yield and nitrogen status as independent targets, we evaluated a joint learning framework in which both variables were estimated simultaneously from the same fused feature space. Multiple neural network architectures were compared under identical data partitions to isolate the effects of shared representation learning. Model interpretation was performed using attribution analysis to distinguish spectral versus structural feature dependence.

Joint learning substantially altered model behaviour. Yield estimation, which proved weak when optimized in isolation, improved markedly when trained alongside nitrogen status, indicating that shared canopy representations capture integrative growth signals not accessible through yield-only optimization. In contrast, nitrogen prediction exhibited limited or inconsistent benefit from joint learning, remaining primarily governed by chlorophyll-sensitive spectral information. Attribution results revealed that yield relied on a broader combination of spectral responses and LiDAR-derived structural descriptors, including canopy height distribution, volumetric development, and spatial heterogeneity, whereas nitrogen status remained physiologically localized within the spectral domain.

These results demonstrate that canopy structure provides complementary information for cumulative traits such as yield, even from single-date acquisitions, while offering limited leverage for physiologically proximal indicators like nitrogen concentration. More broadly, the study shows that multi-task learning does not uniformly enhance prediction accuracy but instead exposes how different agronomic traits are encoded across spectral and structural dimensions. This has direct implications for designing UAV-based decision support systems, where aligning sensing modalities, learning strategy, and crop physiology is critical for meaningful inference.

How to cite: Chatraei Azizabadi, E. and Badreldin, N.: Joint estimation of potato yield and nitrogen status using UAV-derived spectral and structural data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4124, https://doi.org/10.5194/egusphere-egu26-4124, 2026.

EGU26-4124

Download Close

Reliable agricultural statistics support food security monitoring and evidence-based decision making. In Mozambique, official agricultural statistics are primarily derived from the Integrated Agricultural Survey (IAI), an enumerator-based field survey that provides essential contextual information on agricultural production but remains labour-intensive, costly and spatially and temporally constrained, particularly in remote rural areas. While satellite remote sensing offers complementary, wall-to-wall coverage, its spatial resolution is often insufficient to directly capture the fragmented fields, mixed and intercropping patterns, shifting cultivation and strong sub-field variability typical of smallholder farming systems. Consequently, consistent estimation of crop area and crop type derived from enumerator-based crop cover assessments remains challenging in these landscapes.

This study investigates the potential of high-resolution multispectral data acquired with Uncrewed Aerial Vehicles (UAVs) to complement field surveys by providing spatially explicit and internally consistent crop cover and crop fraction estimates at the field and sub-field scale. By resolving individual crops and dominant intercropping systems, UAV-based observations support the interpretation of farmer-reported crop cover proportions, improve consistency across enumerators, and enable post-survey correction of crop area estimates, while providing a basis for future integration with coarser-resolution satellite remote sensing. High-resolution RGB and multispectral imagery (green, red, red edge, and near-infrared; ≤5 cm ground sampling distance) was collected using a DJI Mavic 3M with RTK over 30 sampling areas of 500 × 500 m in Manica Province during the 2025 agricultural season. In parallel, a field survey recorded standardized observations of agricultural activity, including crop type (of most field and tree crops), intercropping combinations and enumerator-based estimates of fractional crop cover. UAV images were processed using a workflow tailored to heterogeneous smallholder landscapes to produce orthomosaics, digital surface models (DSMs), and vegetation indices. These products were linked to field observations through segments representing relatively homogeneous land units, enabling direct comparison between UAV-derived and survey-based crop cover estimates.

For crop classification, training polygons were delineated on RGB orthomosaics for single-crop fields (e.g. maize, beans, sorghum and cassava) and common intercropping combinations (e.g. maize–beans). Annotated mosaics were tiled and augmented and used to train convolutional neural network models (e.g. UNet++), incorporating multispectral vegetation indices and DSM-derived height information as additional input channels. Model performance was evaluated using Intersection over Union, Dice coefficients, and regression metrics for fractional cover accuracy.

A comparison framework was implemented to relate UAV-derived crop type, crop combinations and fractional cover to field survey observations while explicitly accounting for measurement uncertainty. Model II regression quantified systematic bias and proportional differences between the two methods. Initial results indicate that UAV-derived estimates provide spatially consistent crop cover information in fields with complex intercropping structures. Ongoing work focuses on refining segmentation accuracy, analysing residual discrepancies and assessing how UAV-derived crop cover information can be integrated to expand the spatial coverage and reliability of agricultural statistics in smallholder landscapes.

How to cite: Ellsäßer, F. J., Paris, C., Mananze, S., Manuel, L., and Nelson, A.: Supporting agricultural statistics through multispectral UAV-based crop cover mapping in complex smallholder farming systems in Mozambique, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2159, https://doi.org/10.5194/egusphere-egu26-2159, 2026.

EGU26-2159

Download Close

Mango inflorescence malformation, caused by Fusarium mangiferae, represents a major constraint to sustainable mango production worldwide, leading to severe yield losses, reduced fruit set, and long-term reinfection of orchards. Current mitigation strategies rely on labor-intensive sanitation and fungicide applications, which are costly, environmentally burdensome, and often only partially effective and insufficiently timed. There is therefore a critical need for scalable, data-driven tools that enable early, accurate, and spatially explicit detection of disease hotspots within orchards.

In this study, we develop and evaluate an automated detection framework that integrates high-resolution Earth observation data with deep learning to identify malformed mango inflorescences at the canopy and tree level. RGB imagery was collected across multiple seasons (2022–2025) using complementary sensing platforms, including UAVs and ground-based imaging, covering three commercially important cultivars (‘Keitt’, ‘Lilly’, and ‘Kent’) and multiple phenological stages. Semantic segmentation models based on an enhanced U-Net architecture with a ResNet decoder were trained to discriminate healthy and malformed inflorescences at the pixel level, enabling fine-scale disease mapping under heterogeneous field conditions.

Results from the 2025 season demonstrate that millimetric ground-based imagery (0.19–0.68 mm pixel size) enables highly accurate detection of malformation at peak flowering, with average precision exceeding 90% and F1-scores above 0.85 for the disease-sensitive ‘Keitt’ and ‘Lilly’ cultivars. Importantly, incorporating multi-year data and balancing validation datasets significantly improved model robustness and generalization. For the first time, meaningful detection performance from UAV imagery was achieved (up to 71% and 87% precision for malformed and healthy inflorescences, respectively), indicating strong potential for operational orchard-scale monitoring. Cross-cultivar evaluation further revealed partial generalization to ‘Kent’, a cultivar unseen during training, highlighting both the promise and current limits of model transferability.

Beyond detection accuracy, this work delivers key operational insights: disease recognition is highly sensitive to spatial resolution and phenological timing, and segmentation-based approaches provide a strong foundation for precision sanitation, infestation quantification, and decision support. Future work will focus on instance segmentation for whole-inflorescence detection, early-stage disease identification prior to peak bloom, improved cross-cultivar generalization, and integration with UAV- and robot-assisted sanitation workflows. Overall, the study demonstrates how AI-driven Earth observation can support sustainable agroecosystem management by directing sanitation efforts to affected orchard zones, verifying their effectiveness, and enabling disease monitoring during periods of limited field activity, ultimately reducing chemical inputs, labor demands, and pathogen spread.

How to cite: Chen, A., Nagar, Y., and Dafny-Yelin, M.: Data-Driven Detection of Mango Inflorescence Malformation Using Remote Sensing and Deep Learning for Precision Agroecosystem Management, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20230, https://doi.org/10.5194/egusphere-egu26-20230, 2026.

EGU26-20230

Download Close

Vision–language models (VLMs) are increasingly used for the semantic interpretation of visual data, enabling flexible, open-vocabulary analysis of images based on natural language descriptions. These capabilities offer new opportunities for large-scale semantic mapping, particularly in domains where comprehensive labeled training data are scarce or difficult to obtain, such as agricultural and horticultural environments.

Recent research has explored the transfer of semantic information from 2D imagery into three-dimensional representations, a process often called semantic lifting. This approach is attractive for outdoor scene understanding, as training native 3D vision–language models that generalize across landscapes and management regimes remains challenging and tools for 3D data are therefore not developed as far as in the 2D domain. However, most existing studies on semantic lifting focus on indoor environments or urban outdoor scenes, while agricultural landscapes—with their distinct structural characteristics, vegetation dynamics, and management patterns—remain underexplored.

In this contribution, we investigate the applicability of open-vocabulary, VLM-based semantic lifting for large-scale 3D semantic mapping in agricultural settings. Building on insights from urban-scale benchmarks, we analyze how vision–language-driven semantic segmentation transfers to outdoor agricultural and horticultural scenes reconstructed from multi-view UAV imagery. Our results highlight both the potential of these models to generate spatially consistent semantic representations and their limitations, which are strongly dependent on land cover type and semantic classes.

We discuss how such preliminary semantic 3D representations can support large-scale agroecosystem mapping and serve as an initial layer for downstream applications, including spatial analysis and the deployment of agricultural robotic systems. The findings provide guidance on the opportunities and current constraints of foundation-model-based semantic mapping for sustainable agricultural monitoring.

How to cite: Schütte, T., Kontetzki, S., and Hänel, T.: Potentials and Limitations of Vision-Language Models for Large-Scale 3D Semantic Mapping in Agricultural Environments, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21351, https://doi.org/10.5194/egusphere-egu26-21351, 2026.

EGU26-21351

Download Close

Unmanned aerial vehicle (UAV) remote sensing plays an increasingly important role in crop phenotyping and precision agriculture. As GAI (Green Area Index) is one of the main crop characteristics desired for crop management or plant selection, several retrieval algorithms have been proposed from multispectral observations. The inputs of these retrieval algorithms could be the spectral radiance, or the spectral reflectance that is based either on the calibration over a reference panel (PanelCal) or on the use of a Downwelling Light Sensor (DLS) aboard the UAV. However, variability in illumination conditions during UAV flights introduces pronounced artifacts, leading to unreliable inputs of the retrieval algorithms that degrade the accuracy of GAI estimates.

In this study, we propose a Spectral Normalization for Illumination Invariant Calibration (SNIC) method that aims at eliminating the artefacts introduced in the retrieval algorithms when the illumination conditions are changing during the flight of a multispectral camera aboard a UAV.

A Digital Plant Phenotyping Platform (D3P) coupled with a three-dimensional radiative transfer model was employed to simulate wheat canopy reflectance and GAI across a wide range of illumination scenarios. The simulated datasets provide a physically consistent benchmark for evaluating the robustness of different radiometric calibration strategies under varying illumination conditions during the UAV flight. Our model driven GAI retrieval approach is based on XGBoost (eXtreme Gradient Boosting) regression. Four calibration strategies—Radiance, PanelCal, DLS, and SNIC—were then systematically assessed in terms of GAI retrieval performance.

This in-silico experiment demonstrates that SNIC substantially minimizes the sensitivity of GAI retrieval to illumination variability, whereas PanelCal exhibits pronounced degradation under fluctuating illumination conditions. Validation against 4,000 in situ measurements collected under diverse weather conditions further confirms that SNIC is resistant to changes in illumination conditions. The radiance-based method performs also nicely. Conversely, the reflectance-based methods suffer from severe limitations under such conditions (PanelCal) or from the artefacts introduced by the DLS sensor.

How to cite: Dong, M., Baret, F., Weiss, M., Ding, Y., Li, L., and Liu, S.: SNIC: A spectral normalization resistant to illumination conditions for robust estimates of GAI from UAV multispectral measurements , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-17039, https://doi.org/10.5194/egusphere-egu26-17039, 2026.

EGU26-17039

Download Close

Monitoring crop conditions is crucial for effective crop management and provides valuable insights into soil-plant-atmosphere interactions. While some studies have used unmanned aerial vehicle (UAV)-based light detection and ranging (LiDAR) data for mapping plant area index (PAI) in orchards, LiDAR-based time-series analysis to assess PAI variations with phenology throughout the growing season represents a significant gap in knowledge. Tracking PAI dynamics across phenological stages reflects canopy development and leaf expansion, which are directly linked to yield formation. Furthermore, the optimal spatial resolution for mapping biophysical variables of tree crops from LiDAR point clouds is yet to be determined. This study aimed to demonstrate the potential of UAV-derived LiDAR time-series to monitor the PAI and tree vertical profiles at high spatial resolution throughout the growing season of a cherry orchard located in southeastern France. A time series of 14 point cloud acquisitions with a density of 3300 points/m² was collected between February and December 2022, with at least one acquisition per month, covering all phenological stages of the cherry orchard. Field measurements were collected on May 30, and October 6, to measure the PAI at twilight using an LAI-2200C Plant Canopy Analyzer (LI-COR Biosciences, Lincoln, NE, USA), with 248 trees sampled. A voxel-based method was applied on the LiDAR point cloud data to create a three-dimensional grid within which PAI was estimated for each voxel. The results showed that a voxel size of at least 70 cm is required to retrieve reliable PAI estimates, while a voxel size of 100 cm produced the most accurate PAI estimates (RMSE = 0.5 m².m^-2, bias = 0.07, R² = 0.59), when assessed against in-situ PAI measurements. The temporal variation of canopy PAI illustrated the progression of the phenological stages, including flowering, leaf development, ripening and senescence, and the response of the canopy to drought stress (reduction in PAI due to leaf rolling) during the summer. The maps of PAI successfully described the variations in leaf canopy density for different cherry varieties and allowed assessment of the vertical PAI profile at the individual tree level. The LiDAR-derived PAI maps and vertical profiles were able to detect trees exhibiting poor leaf development, which is an important health indicator for effective crop management in orchard settings. Future work should focus on applying UAV-derived observations to optimize crop models to enhancing decision-making tools for effective orchard management.

How to cite: El Hajj, M., Johansen, K., Camargo, F., Lopez Valencia, O., Tu, Y.-H., Angulo Morales, V., López Camargo, O. A., Al Mashaharawi, S. K., Courault, D., and McCabe, M. F.: High-Resolution Plant Area Index Estimation in Cherry Orchards Using UAV LiDAR for Agroecosystem Monitoring, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10454, https://doi.org/10.5194/egusphere-egu26-10454, 2026.

EGU26-10454

Download Close

Pest and pathogen pressure in potato cultivation is increasingly affecting the potato quality and yield. The Netherlands, as the largest seed potato producer around the world, is particularly threatened by blackleg disease and potato virus Y (PVY). Uncrewed aerial vehicle (UAV)-based imaging combined with machine- and deep-learning methods have shown clear potential for potato disease identification, offering advantages over conventional human inspections, which are labor-intensive, expertise-demanding, and often subjective. Most existing studies focused on RGB data and pixel-level classification, producing maps that have limited practical value for targeted removal of infected plants. Earlier work demonstrated the potential of plant-level disease detection approaches. For example, Jia[1] employed hyperspectral data (specifically the first three principal component analysis (PCA) bands) with a YOLOv5s model to distinguish the blackleg- and PVY-infected plants from healthy ones, yielding average mAP@.50 scores of 0.85 for blackleg detection and 0.82 for PVY detection. Gibson-Poole[2] applied object-based image analysis (OBIA) to detect blackleg disease with RGB imagery, achieving a total accuracy of 87%. The findings suggest that multi-modal data (combining hyperspectral and RGB imagery) hold strong potential for plant-level disease detection. We aim to identify the most informative features derived from hyperspectral data and to investigate their integration with RGB data to enhance potato disease detection performance.

We proposed early fusion (E), where data were concatenated channel-wise before network input, and middle fusion (M) architectures, where features were extracted separately within a two-branch network and then merged at an intermediate stage, to integrate hyperspectral features and RGB imagery for potato disease detection. To reduce hyperspectral dimensionality, two feature sets were extracted: (i) the first three PCA bands, and (ii) 10 vegetation indices (VIs) selected from 64 candidates using variance inflation factor analysis to mitigate multicollinearity. Consequently, four models were developed and evaluated: E-PCA-RGB, E-VI-RGB, M-PCA-RGB, and M-VI-RGB. Unlike previous studies that focused on a single disease, our models detected blackleg-infected, PVY-infected, and healthy plants simultaneously. E-VI-RGB achieved the highest mAP@.50 value of 86.65±1.53, followed by M-VI-RGB (85.74±1.75). E-PCA-RGB and M-PCA-RGB yielded mAP@.50 scores of 83.00±2.52 and 83.11±2.20, respectively. These results demonstrate that combining hyperspectral features with RGB imagery improves detection performance compared with single-modality approaches (RGB 83.21±1.31, PCA 79.71±1.30, VIs 85.31±2.11). Our findings highlight the potential of multimodal fusion for potato disease detection in practice. The methods could enable automated systems not only to identify infected plants but also to support timely removal with machinery, mitigating the spread of disease in potato fields. The generalizability of our approach will be further tested and analyzed in future work.

How to cite: Jia, T., Smigaj, M., Kootstra, G., and Kooistra, L.: Fusing UAV-based hyperspectral and RGB imagery for potato plant disease detection, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10845, https://doi.org/10.5194/egusphere-egu26-10845, 2026.

EGU26-10845

Download Close

Monitoring soybean growth provides critical insights for farmers, enabling them to closely track crop development and implement proactive management practices that ultimately enhance yields. Inefficient management and excessive chemical use not only reduce efficiency but also result in significant environmental consequences, including water contamination and increased greenhouse gas (GHG) emissions. These environmental impacts degrade soil health, disrupt weather patterns, and contribute to issues such as soil nutrient depletion and irregular precipitation, all of which have direct, adverse effects on agricultural productivity. Integrating various sensor data, such as satellite and small Unmanned Aerial System (sUAS) data, with machine learning (ML) offers a pathway to precise soybean growth monitoring. This pathway enables farmers to make data-driven decisions that reduce the need for field scouting while improving resource efficiency. Though recent studies have begun to explore field-level, precise growth monitoring using sUAS and satellite imagery, in-depth research on their integration strategies is necessary to develop practical, cost-effective methods for accurately estimating soybean phenological stages. In this study, a comprehensive analysis of soybean growth is conducted across early vegetative to reproductive stages using ML and multi-sensor methods. The selected soybean fields are located at three Ohio State Agricultural Research Stations, which are geographically dispersed across Ohio, USA. Using fixed-wing Wingtra sUAS, high-resolution optical images of soybean fields were collected from 2023 to 2025. To determine whether simple machine learning or complex deep learning methods perform better, multiple combinations of these models with sUAS and satellite are trained, and their performance is evaluated. Best model performance was observed with the Vision Transformer (ViT) model on sUAS images, which detected soybean growth stages with an average Root Mean Squared Error (RMSE) of 0.7. Poor performance was observed with the Random Forest model on open-source Sentinel-2 images, with an RMSE of 3.1. Upon closer investigation of good-performing and poor-performing growth stages through sUAS and satellite, it was observed that early growth stages performed really well only with sUAS data (RMSE<1), while for later reproductive stages (>R2), satellite performed relatively well with RMSE<1. This indicates that using sUAS during the early growth phase and satellites during the late growth phase can be a promising approach for the future.  This strategy enables farmers to make data-driven decisions that optimize growth monitoring and resource use, reduce waste, and minimize environmental impacts.

How to cite: Katari, S. and Khanal, S.: AI-Driven Insights from Multimodal Data for Optimized Soybean Growth Monitoring , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-22080, https://doi.org/10.5194/egusphere-egu26-22080, 2026.

EGU26-22080

Download Close

Accurate, timely, and spatially consistent crop maps are a cornerstone of sustainable agricultural management, climate adaptation, and evidence-based policy. Yet national-scale crop mapping remains challenging in heterogeneous agroecosystems due to fragmented field structures, dynamic land use, and the contrasting spatiotemporal characteristics of annual crops and perennial orchards. Addressing these pressures requires scalable, data-driven approaches that translate advances in Earth observation (EO), data science, and modelling into actionable tools for climate-resilient agroecosystem management. Here we present an integrated, end-to-end crop-mapping framework that synthesizes complementary methodological advances to enable robust, operational monitoring of agricultural systems across space, time, and crop types. Using Israel as a national-scale case study representative of heterogeneous, intensively and extensively managed agroecosystems, the framework links fine-scale field structure, crop phenology, and multi-year dynamics to support decision-making under climatic variability. First, national cadastral parcel layers are refined into agronomically homogeneous field units using deep learning-based semantic segmentation (U-Net, DeepLabV3, and SegFormer) and foundation models (SAM), addressing a critical limitation of registry-based agricultural databases. A U-Net architecture outperformed SegFormer and DeepLabV3, achieving a mean Intersection-over-Union (IoU) of 0.76 with balanced precision-recall. At the national scale, polygon correctness improved from 75.16% to 86.37%, resulting in tens of thousands of fields segmented into homogeneous management units. This step substantially improves geometric consistency and the reliability of downstream crop classification and agroecosystem analysis. Second, a hierarchical, multi-sensor classification strategy integrates Sentinel-1 SAR and Sentinel-2 multispectral time series with phenological metrics and expert-driven feature selection to map agricultural land use and dynamically classify annual field crops across multiple growing seasons. XGBoost achieved the highest land-cover accuracy (OA = 0.909), driven primarily by vegetation, moisture, and chlorophyll-sensitive indices (NDVI, MCARI, NDMI, PGHI). For detailed row-crop mapping, deep learning models outperformed traditional machine learning (TabM OA = 0.861). Multi-satellite fusion ensured robustness and transferability, yielding an average leave-one-year-out accuracy of 0.833. This integration captures crop rotations, seasonal shifts, and climate-driven phenological gradients, enabling consistent multi-year monitoring in dryland and Mediterranean environments. Third, perennial orchard systems, often underrepresented in national crop statistics, are mapped using a multimodal fusion approach that combines very-high-resolution (VHR) aerial imagery with multi-temporal Sentinel-1/2 data. Deep learning architectures jointly exploit fine-scale spatial structure and phenological dynamics, achieving the highest performance across all evaluation settings (same-year OA = 0.890 ± 0.009; cross-year OA = 0.881 ± 0.014), with particularly strong gains for early-stage and sparsely vegetated orchards. Overall, the framework is designed for scalability, interpretability, and operational deployment, demonstrating how multi-modal remote sensing, deep learning, and hierarchical modelling can bridge scientific innovation and real-world agricultural decision-making under climate change.

How to cite: Paz-Kagan, T., Fine, L., Edri, A., Atanelov, A., Brickner, N. Z., and Rozenstein, O.: An Integrated Multi-Sensor Framework for National-Scale Crop Mapping and Climate-Resilient Agricultural Monitoring, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2432, https://doi.org/10.5194/egusphere-egu26-2432, 2026.

EGU26-2432

Download Close

Plant density at the wheat emergence stage is a fundamental structural attribute of agroecosystems, exerting strong control on early competition, resource use efficiency, and yield formation. While UAV-based counting approaches have been widely explored for visually distinct crops such as maize and cotton, accurate and scalable estimation of wheat seedlings remains challenging due to their small size, high spatial density, and spectral similarity to soil and residue backgrounds. Moreover, existing RGB-based UAV and ground imaging approaches face an inherent trade-off between spatial resolution, spectral sensitivity, and operational efficiency.

Here, we propose MS²‑Net (Multi-altitude, Multispectral Seedling Network), a high-throughput Earth-observation framework that integrates multi-altitude multispectral UAV observations with deep learning to enable robust estimation of wheat plant density at the emergence stage. Field experiments were conducted across three major wheat-growing regions in China (Henan, Hebei, and Shaanxi), covering approximately 1,500 plots spanning large variability in sowing density, genotype, and early growth conditions. Multispectral UAV imagery (blue, green, red, red-edge, and near-infrared) was acquired at four flight altitudes (12, 15, 20, and 40 m), enabling systematic evaluation of the trade-off between spatial detail and mapping efficiency. High-resolution smartphone images collected synchronously at plot level provided accurate reference plant counts for model training and validation.

All UAV data were radiometrically calibrated to surface reflectance and used to derive conventional vegetation indices (NDVI, GNDVI, NDRE, OSAVI, and a red-edge chlorophyll index) for spectral interpretability. Wheat plant density was estimated using a deep regression framework built on an EfficientNet-B6 backbone and enhanced with spectral-aware adaptation, spatial attention, and scale-consistent feature learning, allowing MS²-Net to exploit both multispectral information and multi-scale spatial patterns. Across five-fold cross-validation over regions and flight altitudes, MS²-Net achieved robust density estimation (R² = 0.86, RMSE = 37.20 plants m⁻², averaged across sites and flight altitudes), with red-edge and near-infrared bands contributing substantially to model stability across observation scales.

Results demonstrate that multi-altitude multispectral UAV observations provide a practical balance between spatial resolution, spectral sensitivity, and survey efficiency, outperforming both ground-based imaging and RGB-only UAV approaches for early wheat stand assessment. By enabling rapid, field-scale and spectrally informed plant density mapping, MS²-Net provides a scalable pathway for operational agroecosystem monitoring, high-throughput phenotyping, and precision crop management under real field conditions.

How to cite: Wang, M., Yao, L., Chen, B., Xia, Z., Pang, B., He, Z., Wei, Y., Wu, G., Yu, Q., and Zhao, G.: MS²-Net: A deep learning framework for high-throughput assessment of wheat emergence-stage plant density using multi-altitude multispectral UAV imagery, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5216, https://doi.org/10.5194/egusphere-egu26-5216, 2026.

EGU26-5216

Download Close

This study presents a novel framework for monitoring crop phenology using Sentinel-1 (S1) time series data. The proposed approach establishes explicit links between landscape-scale vegetation patterns and field-level phenological developments to address three key objectives: evaluating the agreement between field and landscape phenological signals, identifying dominant phenological tendencies at the field scale, and detecting phenological outliers. Two core indicators were developed—Average Agreement (AVA), which quantifies the correspondence between individual field dynamics and overall landscape development, and Dominance of Tendency (DoT), which characterizes whether fields are phenologically ahead or behind the broader landscape trend, while assessing the consistency of these tendencies across multiple S1 features and orbits.

Environmental descriptors, including soil organic carbon, topographic wetness index, and elevation, were found to shape the spatial and temporal variability of both indicators. Although no single dominant driver was identified, random forest analyses achieved an R² of 0.8, highlighting the complex, multifactorial nature of phenological processes. By integrating growing degree day (GDD) information and S1 time series metrics, the framework reduces reliance on extensive in situ measurements while enabling robust field-scale characterization of phenological progression.

Results show that combining outlier detection with cross-scale comparisons provides valuable insights into typical and atypical crop behavior, supporting assessments of climate vulnerability, resilience, and adaptive management strategies. The flexibility of the method allows seamless application across various S1 features, acquisition geometries, and crop types, demonstrating strong potential for upscaling to regional or national monitoring as well as for broader studies of phenological dynamics.

This work establishes a data-driven pathway toward advanced agricultural management by linking temporal S1 observations with crop performance indicators, thereby enhancing informed decision-making in a sector increasingly challenged by climate change.

How to cite: Löw, J., Conrad, C., Hill, S., Ullmann, T., and Otte, I.: Phenological Alignment and Divergence in Agricultural Systems Derived from Sentinel-1, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10857, https://doi.org/10.5194/egusphere-egu26-10857, 2026.

EGU26-10857

Download Close

How to cite: Lauber, T., Turkoglu, M. O., Senti, D., and Aasen, H.: National-Scale, Multi-Year Crop and Grassland Mapping from Satellite Image Time Series in Switzerland , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18032, https://doi.org/10.5194/egusphere-egu26-18032, 2026.

EGU26-18032

Download Close

Soil erosion by water is a widespread environmental problem with significant impacts on soil fertility, crop productivity, and ecosystem sustainability. In Switzerland, up to 10% of arable land is at a higher erosion risk [1], primarily due to unadapted farming methods, and could benefit from control measures. The combination of susceptible terrains with disturbances from reworking the soil or low soil coverage can exacerbate erosion risk. Reliable, spatially explicit information on soil cover dynamics is therefore essential for identifying erosion-prone areas and supporting sustainable land management.

A commonly used framework to assess erosion risk in agricultural systems is the Revised Universal Soil Loss Equation (RUSLE), in which the crop cover and management factor (C-factor) represents the protective effect of vegetation and farming practices against soil loss. The C-factor varies over time as a function of crop growth, harvest, residue management, and bare-soil periods [2], making its accurate estimation challenging at large spatial scales. For arable land in Switzerland, the annual average erosion indicator computed within the national agri-environmental monitoring programme [3] is based on generic crop calendars and assumed field management practices, leading to inaccuracies in the representation of crop cover and on-field management.

The advent of satellite data provides large-scale access to frequent and high-resolution observations (e.g. 5 days and up to 10 for Sentinel-2) that enable continuous monitoring of land surface conditions. Fractional cover can be retrieved at pixel level using spectral mixture analysis (SMA), which decomposes the mixed satellite signal into proportions of soil, photosynthetic vegetation, and non-photosynthetic vegetation [4].

In this research, we present an automated framework for producing high-resolution, temporally consistent fractional cover maps over Switzerland. We first establish SMA-based regression models by constructing a representative dataset of pure photosynthetic vegetation, non-photosynthetic vegetation, and soil spectra from Sentinel-2 imagery, capturing the diversity of crop types, management practices, and soil conditions across the country. Synthetic spectral mixtures with known proportions of each cover type are created and used as a training dataset for neural network models. The trained models are then applied to Sentinel-2 data to generate nationwide fractional cover time series. We further post-process the outputs to reduce cloud contamination, enforce temporal consistency, and aggregate predictions to regular timestamps and administrative units.

The resulting fractional cover product provides updated, spatially explicit inputs for C-factor estimation within the RUSLE framework, enabling up-to-date assessment of erosion risk at national scale. Beyond soil erosion modelling, the proposed approach offers a product for large-scale monitoring of vegetation and soil dynamics in agricultural landscapes.

[1] V. Prasuhn et al., “Der Agrarumweltindikator Erosionsrisiko,kulturspezifische C-Faktoren sowie eine Karte des aktuellen Erosionsrisikos der Schweiz,” tech.rep., Agroscope, 2023.

[2] P. I. A. Kinnell, “Event soil loss, runoff and the Universal Soil Loss Equation family of models:A review,” Journal of Hydrology, 2010.

[3] A. Gilgen et al., “New approach to calculateagri-environmental indicators using greenhouse gas emissions in Switzerland as an example”, Pre-print. 10.2139/ssrn.5640831, 2025.

[4] F. Lobert et al., “Unveiling year-round cropland cover by soil-specific spectral unmixing of Landsatand Sentinel-2 time series,” Remote Sensing of Environment, 2025.

How to cite: Ledain, S., Gilgen, A., and Aasen, H.: Large-Scale, High-Resolution Fractional Cover Mapping from Sentinel-2 for Agri-Environmental Monitoring, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-17433, https://doi.org/10.5194/egusphere-egu26-17433, 2026.

EGU26-17433

Download Close

Agroforestry, the integration of trees on productive agricultural land (Mosquera-Losada et al., 2018), can be identified through remote sensing methods by the combination of land cover and tree cover maps. Previous work has classified agroforestry as agricultural land with greater than 5% tree cover (Lawson et al., 2025; Zomer et al., 2016).

However, there exist several regional, European, and global land cover and tree cover products that could be suitable for agroforestry identification, but these products vary in resolution, data inputs, and methodology of production. Our work benchmarked the performance of four land cover maps(Büttne et al., 2021; Karvatte et al., 2021; Schultz et al., 2025; UKCEH, 2022) and nine tree cover maps(Brandt et al., 2024; Copernicus, 2023, 2025; Hunter et al., 2025; Lang et al., 2023; Tolan et al., 2024; Weinstein et al., 2020) that were capable of mapping trees outside of woodlands.  We evaluated the datasets’ ability to identify agroforestry on 25 agroforestry sites across the United Kingdom, including a mix of silvoarable and silvopastoral systems, as well as planting ages, densities, and species, as well as nearby agricultural (no trees) and woodland (no agriculture) control fields.

We found that a number of datasets used in previous studies underperformed when distinguishing agroforestry from control fields as well as the previously utilised pixel-based approaches being unsuitable to identify agroforestry fields as a whole. Datasets with coarse resolutions (>10m) often confused proximal small woodlands for trees within agricultural fields. Many datasets struggled to map trees in silvoarable systems, likely due to their linear arrangement differing from that of other trees outside of woodlands.

In addition, the majority of datasets were unable to identify agroforestry sites planted since 2000, suggesting a 20-year lag in identification. The only tree identification method capable of identifying young sites was the fine-tuning of the DeepForest tree detection model (Weinstein et al., 2020) with local data, suggesting a need for tree cover datasets that are capable of identifying seedlings and saplings.

We used the best-performing datasets (balanced accuracy 83-87%) to create a map of agroforestry, with quality flags to signify agreement between datasets. We conclude that there is 517,300 ha (3% of UAA) of area under agroforestry in the United Kingdom. Our map could have further use cases for calculating the uptake of agroforestry, as well as its benefits to people and nature, such as carbon storage, biodiversity impacts, farm income, and health of crops and livestock. We also conclude the need for model training on agroforestry trees, to both identify young trees and those with complex planting arrangements.

How to cite: Machine, A., Burns, M., and Balzter, H.: Evaluating land and tree cover datasets for the identification of agroforestry in temperate Europe, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12956, https://doi.org/10.5194/egusphere-egu26-12956, 2026.

EGU26-12956

Download Close

Global food security is facing increasing pressure from population growth and climate change. Rice, the staple food for over half of the world’s population, is essential to nutritional supply and social stability, especially in developing regions. Highly dependent on water resources, rice production is highly sensitive to climate change and extreme events, and its changes affect global carbon emissions backward. Therefore, timely, accurate, and high-resolution global rice distribution information is indispensable for agricultural management and hunger elimination to achieve Sustainable Development Goal 2 (Zero Hunger) of the United Nations.

Overcoming the uncertainty of optical remote sensing data acquisition with all-weather, all-time, and stable revisits, Synthetic aperture radar (SAR) provides a highly promising solution for the timely acquisition of global rice cultivation. Deep learning models provide strong interpretability of rice scattering patterns and superior generalization capabilities among different agricultural scenarios. Combining the most advanced computation technology with the big remote sensing data, we developed a Time-Series-to-Vision Rice Classification Model (T2VRCM). Instead of learning from the remote sensing image stacks, T2VRCM learns the intrinsic feature variations during rice growth with standardized 2D visual representations, so that problems such as irregular sampling and insufficient modalities can be avoided. A novel global rice dataset, GlobalRice20, was achieved, providing comprehensive and consistent global rice cultivation data in 2015 and 2024 at 20 m resolution for the first time. An overall accuracy of 92.33% was achieved with rigorous validation against over 160,000 reference samples, enabling promising spatiotemporal analysis over the first decade of the SDGs.

Our team has been dedicated to large-scale rice mapping using intelligent computation methods, advancing from national and regional to global scales. Starting with the classic U-Net model, we produced the first 20 m interannual rice maps for Southeast Asia (2019–2021) using time-series Sentinel-1 data. We then proposed an optical–SAR fusion strategy using stacked random forests to generate EARice10, a 10 m rice distribution product in 2023 with comprehensive coverage of four East Asia countries. To overcome global spatial heterogeneity, we further upgraded the framework to the Explainable Mamba U-Net (XM-UNet). With strong generalization, the model provides a physically explainable interpretation of multi-temporal Sentinel-1 SAR data and possesses robust generalization capabilities in countries with diverse cultivation patterns. In addition, we constructed the world's first plot-level rice dataset, Plot-Rice v1.0, with the SAM-2 model and Sentinel-1/2 features. Covering various climatic zones, the dataset supports multiple mainstream deep learning models and demonstrates strong transferability among cross-regional and cross-annual scenarios.

As a result of the achievements outlined above, we provided the 20 m global rice product in 2023 to support the assessments of the UN’s SDG2 indicators, as detailed in the Reports on Big Earth Data in Support of the Sustainable Development Goals in 2024 and 2025. This study reveals the up-to-date progress of our research to address advanced intelligent models in global rice mapping. Meanwhile, we are developing an in-season rice mapping methodology to enhance the timeliness of rice distribution information, in comparison to the current mainstream post-season rice mapping methods.

How to cite: Xu, L., Zhang, H., Guo, H., Song, M., Zuo, L., and Xie, Y.: Global Rice Mapping Driven by Intelligent Models and Big Earth Data Supporting Progress Assessment of SDG 2, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13294, https://doi.org/10.5194/egusphere-egu26-13294, 2026.

EGU26-13294

Download Close

Accurate mapping of weed infestations in fallow farmlands is critical for supporting sustainable weed management and reducing unnecessary chemical inputs. Previous work has demonstrated the effectiveness of convolutional encoder–decoder architectures, such as U-Net, for weed segmentation from satellite imagery; however, these approaches are typically constrained by sensor-specific training, limited cross-site generalisation, and sensitivity to variations in spectral and spatial resolution.  In this study, we investigate the application of the ANYSAT foundation model [1] for sensor-agnostic weed segmentation across heterogeneous fallow agricultural farms across Australia. Building on top of an established U-Net-based workflow, we evaluate whether a foundation model pretrained on diverse Earth observation data can improve robustness and transferability across multiple satellite sensors without explicit sensor-dependent retraining. Multi-spectral satellite imagery from different platforms is used to fine-tune ANYSAT for semantic segmentation of weed presence in fallow paddocks, with human-curated and U-Net-refined weed masks serving as supervisory labels.  We design a systematic evaluation strategy based on leave-one-farm and leave-one-region validation to test model robustness under spatial and spectral variability. Rather than focusing on achieved performance, this work emphasises assessing feasibility, identifying the strengths and limitations of foundation-model-based segmentation for this task, and outlining key considerations for operational deployment in data-sparse agricultural settings. By framing weed detection as a sensor-agnostic problem, this study provides a structured pathway for testing foundation models in agroecosystem monitoring. It contributes to understanding how emerging Earth observation foundation models can be adapted for practical agricultural applications. 

[1] Astruc, Guillaume, et al. "AnySat: One Earth Observation Model for Many Resolutions, Scales, and Modalities." Proceedings of the Computer Vision and Pattern Recognition Conference. 2025. 

How to cite: Deo, R., Filippi, P., and Bishop, T.: Weed Segmentation in Fallow Farmlands Using the ANYSAT Foundation Model , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15427, https://doi.org/10.5194/egusphere-egu26-15427, 2026.

EGU26-15427

Download Close

Recent advances in Earth Observation (EO) data and multimodal Geo-Foundation Models have sharply improved the ability to generate accurate crop-type maps by leveraging rich spatio-temporal representations. These models are inherently scalable across diverse and heterogeneous agricultural landscapes, thus exhibiting strong generalisation. However, timely and high-quality reference data remain a major bottleneck for reliable agricultural mapping and monitoring. Agricultural landscapes are highly dynamic, with frequent crop rotations that require seasonal or annual updates. In addition, European agriculture is increasingly affected by weather extremes (e.g., droughts, hail, and storms), which are expected to intensify in both magnitude and frequency.

Traditional approaches rely on time-consuming and costly manual annotations or field surveys, which are difficult to sustain on a continuous basis and at large spatial scales (e.g., continental monitoring). In this context, geo- and time-tagged field photos represent a promising complementary data source. Each field photo can be linked to satellite image time series acquired over the same location up to the acquisition date. Compared to conventional in-situ surveys based on manual annotations, the combined use of satellite image time series and field photos provides a richer semantic representation of agricultural areas. While satellite image time series capture the temporal dynamics of crop development, field photos offer ground-level information at high resolution on crop condition, phenological stage, and management practices.

Despite their potential, the operational use of field photos in agricultural monitoring remains limited, in part due to challenges in translating heterogeneous images into structured information. Recent advances in Vision–Language Models (VLMs) have unlocked substantial progress in the automatic interpretation and semantic extraction of information from raw field photos. By aligning visual features with semantic concepts expressed in natural language, VLMs provide a powerful mechanism for mapping unstructured field photos to standardised crop-type labels.

This study investigates the potential of combining satellite image time series and geo-tagged field photos to expand, update, and complement existing reference datasets to support continuous large-scale agricultural monitoring. Preliminary results of mapping seven crop types (i.e., maize, wheat, rape, sugarbeet, oat, barley, and sunflower) in Europe indicate that, even in a zero-shot setting and when using simple prompts, the CLIP VLM can correctly identify crop types from field photos when a distinct phenological stage is visible. Incorporating phenological information derived from the temporal patterns of satellite image time series is therefore crucial, as it allows for the filtering of irrelevant images (e.g., post-harvest fields) and the selection of samples for which reliable classification is feasible. Furthermore, when consistency of label predictions obtained independently from field photos (using CLIP) and from Sentinel-1 and Sentinel-2 time series (using a simple Random Forest classifier) is used as a data reliability strategy, highly accurate classification performance across all considered crop types can be obtained. Overall, these findings highlight the strong potential of jointly exploiting satellite image time series and geo-tagged field photos for the efficient and reliable preparation of crop-type reference datasets.

How to cite: Paris, C., Celik, M. F., Maurogiovanni, S., Sedona, R., Cavallaro, G., Cartuyvels, R., and Marsocci, V.: From Satellite Data and Geo-tagged Field Photos to Reliable  Agricultural Reference Data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16567, https://doi.org/10.5194/egusphere-egu26-16567, 2026.

EGU26-16567

Download Close

As a globally vital oilseed, rapeseed necessitates precise in-season mapping to support field management. Furthermore, the accurate retrieval of peak flowering dates is critical for yield estimation, as this phenological stage directly correlates with crop productivity. While state-of-the-art methods have advanced both crop mapping and phenology retrieval, existing approaches predominantly address these tasks in isolation, thereby neglecting their inherent phenological interdependence. Specifically, in-season mapping is often confounded by early-season phenological heterogeneity across fields and regions, whereas flowering retrieval typically relies on the prerequisite of an accurate a priori crop map. To address these limitations, this study introduces a multi-task Transformer-based framework that simultaneously maps rapeseed and retrieves peak flowering dates using Sentinel-1 and Sentinel-2 time series. Reliable training samples were automatically generated via phenology-based rules applied to cloud-free time series. To enhance the robustness against cloud contamination, a data augmentation strategy was introduced that masks Sentinel-2 observations using real-cloud temporal masks to simulate realistic data unavailability. The proposed architecture integrates a dual-task framework with adaptive loss weighting to dynamically balance learning gradients between tasks. Extensive validation across 13 European countries, covering a flowering gradient of up to two months, demonstrates that the proposed method achieves an F1-score of 0.89 for rapeseed mapping four months prior to harvest, and a Mean Absolute Error (MAE) of 6 days for peak flowering retrieval. These results substantially outperform both conventional sequential single-task baselines and specialized state-of-the-art methods. Furthermore, independent validation against phenological records from the German Weather Service (DWD) further confirm the robustness of the proposed method in flowering retrieval. To provide interpretable insights into the model's effectiveness, we analyzed Transformer attention maps and band importance. These visualizations substantiate that the multi-task model effectively extracts task-shared spectral-temporal features, offering a clear and interpretable basis for its enhanced generalization. Overall, this study presents a practical, scalable solution for integrated, large-scale rapeseed monitoring, demonstrating a robust framework that is adaptable to the integrated monitoring of other crops.

How to cite: Zang, Y., Chen, X., Shen, M., Yang, W., Vrieling, A., Paris, C., Qiu, B., Xia, L., Wu, S., and Chen, J.: Integrated In-season Rapeseed Mapping and Flowering Retrieval: A Multi-task Transformer Framework Using Sentinel-1 and Sentinel-2, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-14192, https://doi.org/10.5194/egusphere-egu26-14192, 2026.

EGU26-14192

Download Close

Indonesia is currently the third major producers of coffee in the world, with approximately 20% of its exports destined for European Union (EU) markets. Recent policy developments, such as the EU Deforestation Regulation (EU-DR), impose stringent traceability requirements on coffee imports to EU, requiring spatially explicit linkage between coffee products and their production area. Consequently, there is an urgent need for accurate coffee mapping to support compliance, monitoring, and benchmarking. In this study, we map coffee distribution across Indonesia using multi-temporal imageries, by integrating optical imageries from Harmonized Landsat Sentinel-2 (HLS) dataset, radar imageries from Sentinel-1, and auxiliary environmental data (i.e., topography and distance to human settlement) using Random Forest classification in Google Earth Engine. Specifically, we aim to (1) produce annual maps of monoculture and coffee agroforestry distribution from 2016 to 2024 and (2) assess coffee-related land use changes over this period. The resulting maps will provide critical information on regional coffee distribution to support sustainable land management and future carbon modelling.

How to cite: Satriawan, T. and Luo, X.: Mapping the distribution of coffee agroecosystems in Indonesia from 2016 to 2024, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15565, https://doi.org/10.5194/egusphere-egu26-15565, 2026.

EGU26-15565

Download Close

Detection of biotic and abiotic stress at the field level is an important crop monitoring task with varied applications, including the delineation of management zones and targeted management interventions. Satellite remote sensing provides extensive spatial and temporal coverage for this purpose; however, automated stress detection is constrained by a lack of field-level ground-truth data required to train supervised models. This study develops and evaluates an unsupervised anomaly-detection workflow for identifying biotic and abiotic crop stress using openly available Sentinel-2 satellite imagery, without relying on ground-truth labels. The study develops an Isolation Forest–based method incorporating within-season time-series data that include spectral bands and vegetation indices. Unlike traditional statistical anomaly-detection methods, this model-based technique accommodates multivariate inputs and does not require the assumption of a normal data distribution. Multiple feature configurations were assessed, including visible, red-edge, near-infrared, and shortwave infrared bands, their combinations, and selected vegetation indices. Anomaly scores were computed across multiple image acquisition dates, and only regions consistently identified as anomalous over time were retained as persistent stress signals. The framework was evaluated across three different stress scenarios; frost damage, Septoria disease incidence, and nitrogen deficiency. Results show that the proposed approach successfully detected stress patterns across all sites, achieving accuracies of up to 83%. In addition, the experiments identified key spectral features that were particularly informative for detecting each specific type of stress. This workflow offers a scalable and operationally feasible option for crop stress detection in agricultural systems where ground-truth data are limited. 

How to cite: Joshi, A., Filippi, P., and Bishop, T.: Unsupervised detection of biotic and abiotic crop stress using Sentinel-2 time series and Isolation Forest , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6052, https://doi.org/10.5194/egusphere-egu26-6052, 2026.

EGU26-6052

Download Close

Reliable quantification of agricultural water use and greenhouse gas (GHG) emissions is essential for understanding and mitigating the environmental footprint of food production. However, it remains challenging due to the limited spatial representativeness of in-situ measurements and the strong influence of vegetation dynamics, management practices, and weather variability. Eddy-covariance (EC) observations provide direct and high-frequency measurements of evapotranspiration (ET) and GHG fluxes, but their footprint is inherently local, constraining their applicability for regional and national assessments. Satellite remote sensing (RS) offers spatially continuous information on vegetation status and land cover, yet its effective integration with flux observations for process-relevant upscaling remains limited.

In this contribution, we provide first insights from a synthesis of recent field-scale literature, comprising over 300 ET studies and more than 400 GHG-focused studies, to assess how remote sensing information has been incorporated into ET and GHG flux modelling. Our review indicates a clear divergence in modelling development trajectories across flux types. Earlier ET studies were largely dominated by physically based formulations, such as Penman–Monteith and surface energy balance models. Over the past five years, ET modelling has shifted toward data-driven and machine-learning approaches, enabling the integration of a broader range of satellite-derived predictors, including vegetation indices and shortwave infrared (SWIR)-based indicators related to soil moisture conditions. Net ecosystem exchange (NEE) exhibits a similar transition from process-based to data-driven modelling frameworks, reflecting improved data availability and methodological flexibility.

In contrast, modelling of other GHG fluxes, particularly CH₄ and N₂O, remains largely confined to process-based approaches, with DNDC and DayCent being the most widely applied models. This persistence primarily reflects the limited availability of long-term, high-quality ground-based GHG flux measurements. Moreover, RS-based information on soil moisture and temperature, vegetation status, and land-use or management practices offers potential to better inform and constrain GHG flux estimates in agricultural systems. These findings highlight a persistent gap between the availability of spatially explicit satellite information and its current use in GHG flux modelling, pointing to substantial opportunities for improved integration of remote sensing and in-situ flux observations in future upscaling efforts.

How to cite: Jia, A., Aasen, H., and Buchmann, N.: Integrating Eddy-Covariance and Satellite Data to Upscale ET and GHG Fluxes across Swiss Agricultural Systems, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13185, https://doi.org/10.5194/egusphere-egu26-13185, 2026.

EGU26-13185

Download Close

Machine learning methods provide a powerful basis for developing flexible, non-parametric models of complex phenomena and have demonstrated strong predictive capabilities across many areas of the physical sciences generally and the earth sciences specifically. While machine learning methods have been demonstrated to be flexible predictive tools capable of integrating diverse data streams, they present significant challenges in terms of interpretability and generalizability. This is especially true in the context of ecohydrological or biophysical systems, where the objective is often to develop a better understanding of the underlying system rather than exclusively improve predictive performance. There is a growing recognition that interpretability, physical consistency, and data complexity are key challenges in the successful adoption of machine learning methodologies. Here we evaluate the application of machine learning methods to produce models for land-atmosphere water vapor exchange across a set of diverse agricultural systems. Specific focus is placed on the use of environmental and proximal sensing information to develop simple and effective models of evapotranspiration using both machine learning and hybrid modeling approaches that leverage the advantages of machine learning and biophysical simulation. Emphasis is placed on parsimonious model development and interpretability of model performance.

How to cite: Drewry, D., Cross, J., Shirazi, S., Gaur, S., Mallick, K., Aslan-Sungur, G., and Vanloocke, A.: Machine Learning and Proximal Sensing for Predicting Evapotranspiration of Agricultural Systems, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2168, https://doi.org/10.5194/egusphere-egu26-2168, 2026.

EGU26-2168

Download Close

The Mekong Delta contributes approximately 7–10% of the global rice trade, produced by about 1.5 million small-scale farmers. Understanding how field size affects rice yield is critical for advancing the sustainability and resilience of smallholder farming systems in the Mekong Delta. However, rice yield variability across field sizes remains poorly understood, due to the complexity of rice cropping systems and the lack of accurate field boundaries for smallholder farms in this region. In this study, we delineated field boundaries across the Mekong Delta using 3-m PlanetScope imagery and analyzed rice yield patterns across field sizes using near-infrared reflectance of vegetation (NIRv) as a yield proxy, derived from 10-m Sentinel-2 observations spanning 2019–2025. To delineate smallholder farms, we fine-tuned the Segment Anything Model (SAM), which generated field boundaries with an accuracy of 78%, an F1-score of 0.54, and a Matthew’s correlation coefficient (MCC) of 0.40. Using these boundaries, we assessed rice yield variability across field sizes and found that yield increased with field size (r = 0.30, p < 0.001). This relationship remained stable across years, indicating that smaller farms consistently experienced lower yields. This study contributes to understanding rice yield patterns within smallholder farming systems for better management practices in the Mekong Delta.

How to cite: Zhou, Q., Lin, Z., Guan, K., Wang, S., and Luo, X.: Assessing the dependence of paddy rice yield on field size across the Mekong Delta, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6102, https://doi.org/10.5194/egusphere-egu26-6102, 2026.

EGU26-6102

Download Close

The stabilization of soil organic carbon (SOC) in forest ecosystems is crucial for mitigating climate change. However, the interaction of mycorrhizal associations with environmental factors to influence SOC fractions globally remains poorly understood. Here, we synthesize 2,784 observations from 234 peer-reviewed studies to examine global patterns of particulate (POC) and mineral-associated organic carbon (MAOC) in forests dominated by arbuscular (AM) versus ectomycorrhizal (ECM) forests. Our results reveal that ECM forests possess 24% higher POC content and exhibit greater sensitivity to climate warming than AM forests. In contrast, AM forests sustain higher MAOC content, which shows less variability across climate gradients. Linear mixed-effects models indicate distinct responses of POC and MAOC to the interactive effects of mycorrhizal type and environmental drivers. Notably, POC content in ECM forests increases with stand age. While young AM forests contain higher levels of both POC and MAOC, middle-aged and mature ECM forests surpass AM forests in POC, with no significant difference in MAOC. Using existing data, we project global changes in these SOC fractions and propose a mycorrhiza-informed framework for forest carbon sequestration. Our findings underscore the pivotal role of mycorrhiza-environment interactions in SOC partitioning and stabilization, offering critical insights for refining global carbon models and guiding climate-smart forest management.

How to cite: Chen, J., Liu, S., and Sun, S.: Mycorrhiza-mediated distribution of particulate and mineral-associated organic carbon across global forests, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12127, https://doi.org/10.5194/egusphere-egu26-12127, 2026.

EGU26-12127

Download Close

Information on the rice agricultural system, including not only planting area but also phenology and cropping intensity, is critical for advancing our understanding of food and water security, methane emissions, carbon and nitrogen cycles, and avian influenza transmission. However, existing efforts primarily focus on mapping planting area and lack a comprehensive picture of the rice agricultural system. To address this gap, we propose a remote sensing-based comprehensive framework for mapping the rice agricultural system in China: First, we identified valid growth cycles of crop by using 30-m Sentinel-2 and Landsat fused data; Second, we applied a well-established phenology-based algorithm to map rice planting areas, by extracting the flooding and rapid growth signals in the transplanting and rapid growth temporal windows; Third, the rice-specific phenology phases (i.e., transplanting, tillering, heading, and mature) were identified using a phenology extraction method tailored to the physiological characteristics of rice; Finally, rice cropping intensity was determined based on detailed phenological phases and planting area data. Due to the accurate identification of crop cycles and pixel-level temporal windows at the national scale, the generated rice planting area maps exhibit a high accuracy across China, with both overall accuracy and F1 scores exceeding 0.8, based on validation with over 13,000 field samples. Improvements in the extraction method have addressed the lag in phenology detection caused by rice's flooded environment, leading to more accurate phenological information. As a result, the phenological data shows reliable accuracy (R² of 0.6–0.8 and RMSE of 8–15 days), facilitated by the mutual enhancement of rice planting area and phenology mapping. The resultant rice phenology and cropping intensity maps are the first of their kind with 30 m resolution. Together, the resultant rice planting area, rice phenology, and cropping intensity maps provide, for the first time, a comprehensive picture of China's rice agricultural system, better supporting multiple targets related to Sustainable Development Goals.

How to cite: Zhao, Z., Dong, J., Yang, J., Liu, L., You, N., Xiao, X., and Zhang, G.: From rice planting area mapping to rice agricultural system mapping: A holistic remote sensing framework for understanding China's complex rice systems, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11089, https://doi.org/10.5194/egusphere-egu26-11089, 2026.

EGU26-11089

Download Close

Over the last decade, innovations in satellite remote sensing (RS) and data science have widened the scope and relevance of agricultural monitoring and management applications at farm and territorial scales. Recently launched and upcoming hyperspectral satellite missions (e.g., ASI-PRISMA, DLR-ENMAP, Planet-Tanager, ESA-CHIME, Kuva-Hyperfield) provide high spectral resolution (< 10 nm) across the 400-2500 nm range, opening new frontiers for assessing biophysical and biochemical functional traits of agroecosystems, while advancements in machine learning (ML) and artificial intelligence (AI) allow the efficient exploitation of the information content of high-dimensionality spectral datasets.   
Here we summarize the results and lessons learned from three experiments in different European agricultural systems (croplands and grasslands), analysing how the synergy between hyperspectral imaging spectroscopy (field and satellite), ML, and foundation models could support adaptive agroecosystem management through the retrieval of vegetation and soil properties related to the nutrient cycle. The three case studies are:(1) Assessment of biomass and nutritional quality of Alpine pastures: Gaussian Process Regression (GPR) models were calibrated on 250 vegetation samples and field spectra collected in 2024 and 2025 from semi-natural pastures in Valtournanche (Aosta, Italy) and Val Camonica (Brescia, Italy). PRISMA-derived maps for biomass and leaf-level protein, and fiber content showed good accuracy against in-situ data (LAI [-] R² = 0.71, RMSE = 0.89; Biomass [g · m^-2] R² = 0.67, RMSE = 178.71; DM [%] R² = 0.65, RMSE = 2.70; CP [%] R² = 0.58, RMSE = 0.52; ADF [%] R² = 0.45, RMSE = 2.42; NDF [%] R² = 0.50, RMSE = 0.61), allowing mapping of pasture metabolizable energy in support of grazing management.

(2) Monitoring of nutritional status of paddy fields: GPR models were developed on 200 vegetation samples and field spectra collected in 2024 and 2025 in several fields located in the Ferrara region (Italy). These models, demonstrated on PRISMA and EnMAP time-series, effectively monitored crop development across a temporally and spatially independent test set (LAI [-] R² = 0.83, RMSE = 0.30; Fresh Biomass [g · m^-2] R² = 0.72, RMSE = 627.67; LCC [μg · cm^-2] R² = 0.58, RMSE = 3.40; LNC [μg · cm^-2] R² = 0.34, RMSE = 22.51; CNC [g · m^-2] R² = 0.56, RMSE = 0.77).

(3) Retrieval of Soil Organic Carbon (SOC) and soil Nitrogen (N) on arable lands: A transformer-based, sensor-agnostic deep learning architecture was fine-tuned on open global spectral libraries. When applied to EMIT and Tanager-1 imagery over the Po Plain (Italy) and Northern Netherlands regions, the model yielded high accuracy (SOC [%] R² = 0.61, MAE = 0.37; N [%] R² = 0.68, MAE = 0.12) against 289 independent field observations.

Our findings demonstrate that satellite hyperspectral spectroscopy can complement operational multi-spectral missions, adding key information about agroecosystems nutritional status.  Furthermore, we show that the use of pre-trained ML and AI models on global spectral libraries and field reflectance data allows accurate retrieval even in the absence of ground truth acquisition synchronous to the satellite overpass, offering a potential scalable solution for agroecosystems management and monitoring at landscape scale.

How to cite: Ceriani, R., Pepe, M., Boschetti, M., and Fava, F.: Advancing spaceborne remote sensing for agroecosystem adaptive management, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19714, https://doi.org/10.5194/egusphere-egu26-19714, 2026.

EGU26-19714

Download Close

Soil organic carbon (SOC) plays a key role in climate regulation and soil functioning. Although SOC stock and stability result from complex interactions between environmental, biological, and anthropogenic factors, the hierarchy of these drivers is strongly scale-dependent. At large spatial scales, climatic forcing often dominates SOC patterns, potentially masking the effects of land management. At local scales, where climatic variability is reduced, the relative influence of agricultural practices compared to landscape heterogeneity (e.g., topography and soil properties) remains poorly quantified, notably due to the scarcity of datasets combining soil properties and high-resolution management data. Clarifying this balance is essential for designing effective climate mitigation strategies in agricultural systems.

We investigated the drivers of SOC stock and stability within a 3,000 km² heterogeneous agricultural territory in Burgundy (France). The territory spans a diverse landscape, transitioning from western limestone plateaus to agricultural plains in the east. SOC measurements (0–20 cm) from 147 cropland sites were combined with 18 explanatory variables derived from in situ measurements, field survey, or satellite data and describing topography, climate, soil physico-chemical properties, vegetation dynamics, and contrasting agricultural management (diverse crop rotations, residue management, the use of cover crops, and organic amendments). SOC stable and active fractions were quantified using Rock-Eval® thermal analysis coupled with the PARTYsoc learning model. The SOC stocks averaged 41.7 ± 13.9 tC.ha^-1, with the active (C_a​) and stable (C_s) stocks representing 19.9 ± 8.3 tC.ha^-1 and 21.8 ± 6.1 tC.ha^-1, respectively. 

Random Forest models were used to capture non-linear relationships between SOC variables and their drivers, and SHAP (SHapley Additive exPlanations) values were applied to quantify the relative importance and direction of individual drivers. Model performance reached a coefficient of determination (R²) of 0.41 for SOC stocks, and 0.50 and 0.26 for C_a​ and C_s stocks respectively. The lower R² for C_s​ likely reflects missing explanatory variables related to historical land use or specific soil mineralogy.

SHAP analysis revealed that even at local scales (a few km), soil properties and climate remain the dominant drivers of SOC stock and stability in this study. Nevertheless, management-related factors, such as crop residue management and number of vegetation days during the intercrop periods, exert a stronger influence on SOC stock than topographic variables. Patterns differ among pools: active carbon is mainly influenced by CaCO₃, temperature, and precipitation, whereas clay content dominates the stable carbon fraction.

Our results demonstrate that while soil and climate largely control SOC stocks at local scales in the context of a highly heterogeneous terrain, agricultural management can meaningfully influence SOC dynamics and stability, highlighting opportunities for targeted strategies to enhance soil carbon sequestration.

How to cite: Girod, C., Barré, P., Fieuzal, R., Ceschia, E., Chemidlin Prevost-Bouré, N., Marron, P.-A., Ranjard, L., and Hermand, A.: Understanding Local Drivers of Soil Organic Carbon Stocks and Stability Using SHAP Analysis in an Agricultural Territory of Eastern France, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19004, https://doi.org/10.5194/egusphere-egu26-19004, 2026.

EGU26-19004

Download Close

In the Galapagos Archipelago, agricultural abandonment and biological invasions act as synergistic forces, creating "novel ecosystems" that threaten both endemic biodiversity and local food security. While historical land cover changes are well-documented, the mechanisms determining when and where productive land is lost to invasion remain obscured by complex interactions between climatic legacies and anthropogenic pressure. This study presents a unified spatiotemporal framework to assess the susceptibility of island agroecosystems to three critical transitions: agricultural abandonment, invasive species expansion, and invasive conversion (die-back).

We integrated dense Sentinel-1 SAR time-series (2018–2024) with high-resolution climatic variables (CHIRPS/TerraClimate) across the agricultural highlands (≈25,000 ha). Our hybrid workflow fuses satellite event-dating (Vertex AI + PELT) with epidemiological Case-Crossover designs to pinpoint specific climatic triggers, followed by Bayesian Spatial Modeling (R-INLA) and Random Forest classifiers to map landscape susceptibility.

Our results reveal distinct spatiotemporal fingerprints with direct implications for farm management. Temporally, agricultural abandonment is triggered by persistent drought stress (longer dry spells); spatially, risk is critically clustered in Silvopasture and Mixed Forest zones, identifying these productive assets as "stepping stones" to total land abandonment. Conversely, invasive expansion exhibits a "Rainfall Paradox": it is primed by short-term wetting pulses, while spatially, the models detect a process of "densification" within existing patches rather than purely frontier expansion. Finally, invasive retreat (die-back) is linked to extreme wet spikes and heat interaction, and is spatially confined to high-elevation climatic niches, supporting the "Environmental Filtering" hypothesis where native resilience limits invasive establishment.

By coupling AI-driven event detection with physics-aware spatial statistics, we demonstrate that invasive dynamics are pulsed by climate "windows of opportunity." The resulting risk maps provide a dual-purpose baseline for the Galapagos National Park and the Ministry of Agriculture, facilitating targeted interventions to protect native ecosystems while reinforcing the resilience of farming systems against climatic shocks.

How to cite: Benitez, F., Mena, C., and Gobin, A.: AI-assisted event-dating of invasive transitions in Galápagos agroecosystems: Disentangling climate triggers and landscape susceptibility using Satellite Imagery and Bayesian–ML, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16104, https://doi.org/10.5194/egusphere-egu26-16104, 2026.

EGU26-16104

Download Close

Climate stressors pose increasing risks to major oilseed cropping systems of groundnut, mustard, and soybean across South Asia, a region where these crops are critical for food security, livelihoods, and edible oil supply. Existing assessments often rely on aggregated climate indicators or generalized crop responses. This limits their usefulness for identifying suitable crop-specific adaptation options. This study advances current understanding of climate–oilseed interactions by adopting a physiology-based, adaptation-oriented framework that explicitly links biologically relevant climate stressors to the suitability of adaptation interventions under current and future climates.

We quantify multiple heat and rainfall-related stressors using crop-specific physiological thresholds and analyse their intensity and frequency under historical conditions and CMIP6-based future scenarios for the 2050s and 2080s. The analysis distinguishes between stress exposure over the full crop (cardinal) cycle and stress occurring during sensitive phenological windows, particularly the reproductive and pollination phases. Stressor projections are then linked to adaptation options using a logical, expert-reviewed heuristic framework that evaluates the feasibility and expected effectiveness of genetic, management, structural, irrigation, and financial interventions under increasing climate stress.

Our results show that the intensity of all heat-related stressors and the crop water deficit index is projected to increase substantially across oilseed-growing regions in South Asia. Rainfall-related stressors display mixed and spatially heterogeneous responses, reflecting uncertainty and regional differences in future precipitation patterns. Importantly, heat stress during the full crop cycle and during critical reproductive phases exhibits contrasting behaviour. Critical-phase heat stress is projected to increase mainly in frequency, implying more frequent exposure to damaging conditions during short, sensitive windows, whereas full-cycle heat stress is projected to majorly intensify in the future.

These changes have direct implications for adaptation planning. Genetic interventions and financial risk-transfer mechanisms emerge as the most consistently robust options across crops, regions, and emission pathways. In contrast, structural measures, nutrient management, and irrigation-based interventions progressively lose effectiveness as future heat and moisture stresses exceed the thresholds these measures can realistically buffer, with outcomes strongly dependent on emission trajectories.

By mapping transitions in stressor regimes and adaptation suitability, this study provides a first-order, spatially explicit basis for climate-smart adaptation planning in South Asian oilseed systems. The findings highlight the need for innovation focused on protecting critical phenological processes under future climate change.

How to cite: Barik, A., Shirsath, P., and Aggarwal, P.: Climate Stresses and Adaptation Pathways under Changing Climate for South Asian Oilseed Systems, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21855, https://doi.org/10.5194/egusphere-egu26-21855, 2026.

EGU26-21855

Download Close

Near-real-time daily high-resolution estimates of crop gross primary productivity (GPP) are crucial for accurate biomass and yield estimation. The multiplicative combination of near-infrared reflectance of vegetation and photosynthetically active radiation (NIRvP) serves as a biophysically grounded proxy that enhances the responsiveness of GPP estimation. However, a mechanistic model for accurately estimating GPP using NIRvP remains lacking, which limits the potential for enhanced crop productivity assessments. In this study, we developed a model based on the NIRvP and eco-evolutionary optimality (EEO) theory (NIRvP-EEO) with Sentinel-2 imagery and meteorological data on Google Earth Engine for crop GPP estimation without the need for calibration. Specifically, we integrated the Ball-Berry stomatal conductance model into NIRvP-EEO to balance carbon and water vapor fluxes. To enable near-real-time daily monitoring of crop GPP, we employed temporal-weighted interpolation and Whittaker-smoothing filtering methods to fill data gaps. Compared to benchmark models such as enhanced SatelLite Only Photosynthesis Estimation (ESLOPE), crop SLOPE (CSLOPE), GPP network (GPP-net) and P-model, the NIRvP-EEO model demonstrated improved daily GPP estimation for four major crops including corn, soybean, wheat and rice. We found that NIRvP-EEO could reliably GPP estimation not only in drought and heatwave years but also in flood years. Additionally, the model effectively captures the fine spatial details and interannual variations in GPP for these crops. By leveraging the Google Earth Engine platform, our model enables conduct near-real-time daily continuous monitoring of crop GPP at a high spatial resolution anywhere in the world.

How to cite: Yu, W., Ryu, Y., Zhang, H., Wang, S., and Feng, H.: NIRvP-EEO: An NIRvP-based eco-evolutionary optimality model for near-real-time daily crop gross primary productivity estimation at field scale, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-8451, https://doi.org/10.5194/egusphere-egu26-8451, 2026.

EGU26-8451

Download Close