AI-based SAR-to-optical GAI regression for crop monitoring

Bouchat, Jean [UCL] Deffense, Quentin [UCL] Defourny, Pierre [UCL]

The green area index (GAI) is a key biophysical variable for crop monitoring. The most accurate methods for its large-scale estimation rely on optical remote sensing data. However, these can be hampered by frequent cloud cover. In this context, synthetic aperture radar (SAR) offers the advantage of being able to provide dense time series that can be used to complement the sparse GAI series derived from optical data. In this study, SAR-to-optical GAI regression is performed using a transformer neural network with past and current values of SAR backscatter and interferometric coherence, as well as past values of GAI when available. Sentine-1 and -2 data acquired from 2018 to 2021 over the Hesbaye region of Belgium are used for cross-validation. The model is trained on three growing seasons and tested on the fourth for each fold. The results show that the model can successfully predict Sentinel-2-derived GAI with an average R2=0.88 and RMSE=0.74, outperforming methods relying on radiative transfer model (e.g., Water Cloud model) inversion. The method is also validated with data collected in situ in eight maize fields in Belgium (R2=0.87 and RMSE=0.75). These promising results pave the way for the generation of accurate, dense GAI time series throughout the growing season, allowing for timely crop monitoring in cloud-prone regions.