A Modified Hierarchical Vision Transformer for Soil Moisture Retrieval From CYGNSS Data

Abstract This research introduces a new deep learning (DL) framework, multi‐head self‐attention‐aided vision Transformer (MSA‐ViT), for soil moisture (SM) retrieval using Cyclone Global Navigation Satellite System (CYGNSS) data. We first assess the sensitivity of CYGNSS reflectivity to SM, demonstrating a strong physical linkage through coherent scattering theory. The proposed MSA‐ViT model integrates this physical understanding with DL to capture nonlinear interactions between SM, surface roughness, and vegetation attenuation. Using data from January 2020 to December 2024, we aggregated observations over multiple temporal scales (3–60 days) to capture diverse hydrological patterns. The MSA‐ViT model was initially trained using 10‐day averaged data and subsequently tested across varying temporal scales to confirm its ability to reflect SM dynamics. Comparative experiments with conventional techniques such as linear regression and shallow neural networks, alongside other established DL models, demonstrated the outperformance of the proposed MSA‐ViT‐based approach. Following the initial validation, the training data set was expanded with a broader range of temporal patterns to enhance the model's generalization capabilities. Further evaluation was conducted through time series analysis, comparing the model's 3‐day retrievals with the Soil Moisture Active Passive data, the CYGNSS L3 SM V3.2 product, the in situ International Soil Moisture Network measurements and the Global Precipitation Measurement records, which showed consistent alignment with SMAP SMs and clear seasonal variability. Results also demonstrate improvement over the current CYGNSS L3 product on both precision and coverage. This comprehensive validation across large watersheds and diverse spatiotemporal scales attests to the model's robustness and its applicability for different ecosystem types.