Estimating crop biophysical parameters from satellite-based SAR and optical observations using self-supervised learning with geospatial foundation models

Accurate knowledge of vegetation water content (VWC) and crop height is crucial for agricultural management, environmental monitoring, and for satellite-based retrieval algorithms for geophysical variables. Traditional methods to estimate VWC, primarily rely on optical indices, which has limitations of biomass saturation, and sensitivity to atmospheric conditions. This study introduces a novel application of geospatial foundation models (GFMs), leveraging extensive, unlabeled datasets through self-supervised learning to enhance the skill of VWC and crop height estimation. We developed a comprehensive model integrating Sentinel-1 A C-band SAR and Sentinel-2 A/B indices with weather parameters to estimate soybean and corn VWC and crop height.

Our research study area spans a variety of climatic zones and management practices, from the humid continental climate of Iowa and Michigan to the subtropical environment of Florida, encompassing both irrigated and non-irrigated fields as well as diverse tillage practices. We compared the performance of Single-Task Learning GFM (STL-GFM), Multi-Task Learning GFM (MTL-GFM), and machine learning techniques including Random Forest (RF), and XGBoost (XGB) to evaluate their effectiveness in estimating VWC and crop height.

Results demonstrated that STL-GFM outperforms other methods in accuracy and generalizability. For VWC estimation, STL-GFM achieved R² values of 0.90 and 0.89 for soybean and corn, respectively. For crop height, R² values reached 0.95 for soybean and 0.98 for corn. The integration of SAR, optical, and climate data provided more reliable estimations than using individual data sources.

Feature importance analysis identified NDVI, NDWI, VH backscatter, and precipitation as key drivers for accurate VWC and height estimations. The red-edge band emerged as significant for VWC estimation but showed limited importance for height prediction. Notably, surface roughness demonstrated a noticeable impact on corn VWC and height estimations, while soil moisture exhibited less influence than initially anticipated. Notably, without directly incorporating soil moisture and surface roughness data, but by including diverse field conditions in training and validation, the STL-GFM model demonstrated strong generalization capabilities. This study highlights the potential of GFMs in advancing crop monitoring techniques, offering more reliable data for precision agriculture, and supporting sustainable farming practices across diverse agricultural landscapes.