ASCAT soil moisture retrieval using deep learning: a focus on localization strategy

This study investigates the intercomparison of daily soil moisture (SM) retrieval from ASCAT (Advanced SCATterometer) observations using machine learning. The exploitation of spatial structure through convolutional neural networks (CNNs) is shown to significantly enhance retrieval performance compared to a standard multilayer perceptron (MLP), with spatial correlation with the target ERA5 SM increasing from 0.55 to 0.91 and temporal correlation from 0.61 to 0.73. Incorporating “localization” (i.e., a strategy to adjust the neural network (NN) behavior to local conditions) into the model is a key factor for improving retrieval quality, resulting in more accurate SM estimates, reduced regional biases, improved temporal dynamics, and more realistic representations of extreme SM events. Our NN-based retrievals show strong agreement with in situ SM measurements, achieving temporal correlations of 0.60 and 0.68 for the MLP and CNN models, respectively, in the contiguous United States (CONUS) during 2019. These findings underscore the critical role of spatial learning and localization in SM retrieval from remote sensing data such as ASCAT.