A transformer-based attention network for high-precision soil moisture retrieval using CYGNSS data

In recent years, NASA’s Cyclone Global Navigation Satellite System (CYGNSS) has attracted significant attention for soil moisture (SM) retrieval through reflected signal analysis. The improvement of SM retrieval accuracy has long been a hot spot in this field. Most research restricts the utilization of the Delay-Doppler Map (DDM) to peak power, which cannot fully capture the characteristics of the DDM. A model named Transformer-Neural Network (TNN) is proposed in this paper to improve SM retrieval accuracy. The proposed model comprises a Vision Transformer (ViT) alongside a neural network. Feature extraction through ViT enables more effective acquisition of overall DDM information. The extracted DDM information is used to establish the connection with SM through a neural network, which improves SM retrieval accuracy. The results show that high-precision SM retrieval is achieved with our proposed model, in comparison with other machine learning (ML) models. The value of R reaches 0.88, and the value of root mean square error (RMSE) is as low as 0.04. The proposed model also exhibits good performance across various land cover types. Moreover, the SM values predicted by the model show better consistency with the Soil Moisture Active Passive (SMAP) observations. The proposed model effectively adopts the ViT image-processing strategy, providing a new method to improve the precision of GNSS-R SM retrieval.