A GNSS-R Soil Moisture Inversion Method Based on Multimodal Feature Fusion for the Yellow River Delta

As an emerging remote sensing technology, global navigation satellite system reflectometry (GNSS-R) provides moderate penetration and all-weather observation capabilities, making it attractive for soil moisture retrieval. However, GNSS-R-based inversion is challenged by surface heterogeneity, complex feature associations, and the difficulty of unified modeling among delay-Doppler map (DDM) imagery, time-delay waveforms, and auxiliary features. To address these issues, this study proposes a multimodal cross-branch fusion network (MCBF-Net) that combines multimodal fusion with deep learning. MCBF-Net consists of three parallel branches: a 1-D convolutional neural network for extracting features from time-delay waveforms, a multiscale fusion transformer for capturing global features of DDM imagery, and a backpropagation neural network for learning nonlinear patterns in auxiliary data. A multimodal feature fusion module and a fusion loss function are embedded to enhance cross-modal integration and branch synergy. This design enables comprehensive feature representation and significantly improves retrieval performance. Experiments were conducted in the Yellow River Delta, a region with urgent needs in ecological protection and agricultural sustainability. Soil Moisture ActivePassive (SMAP) satellite observations were used as reference data. The proposed method achieved an RMSE of 0.0284 cm³/cm³ and a correlation coefficient R of 0.9232, demonstrating high accuracy and robustness. Furthermore, the spatiotemporal distribution of soil moisture was analyzed, showing the models effectiveness for regional-scale soil moisture inversion. Overall, the study provides a feasible and efficient solution for refined soil moisture monitoring and supports sustainable agricultural and environmental management.