The Global Navigation Satellite System Reflectometry (GNSS-R) technology has pioneered an innovative technical pathway for surface parameter inversion and has been widely applied in soil moisture retrieval. Addressing the limitations of conventional soil moisture inversion methods in fully exploiting GNSS-R features-particularly their inability to effectively integrate heterogeneous information including imagery, waveform, and auxiliary parameters from GNSS-R data, this paper proposes a Hybrid Convolutional Neural Network (HCNN) model that combines 2D and 1D architectures. The model establishes an end-to-end inversion framework by extracting spatial distribution characteristics from Delay-Doppler Maps (DDM) through 2D CNN, capturing dynamic sequential features from time-delay waveforms using 1D CNN, while incorporating auxiliary parameters. Experimental validation was conducted in the Yellow River Delta High-Efficiency Eco-Economic Zone, with SMAP data serving as the soil moisture reference. Results demonstrate that the HCNN model achieves significantly superior inversion accuracy compared to conventional methods, attaining a correlation coefficient of 0. 9074 and root mean square error of 0. 0311cm3/ cm3. Comparative analyses with traditional multiple linear regression, Backpropagation (BP) neural network models, and random forest models further validate its advantages in soil moisture inversion tasks. Ablation experiments specifically investigate the individual contributions of DDM and time-delay waveform. features, revealing that their synergistic integration is critical for enhancing inversion accuracy. This study provides a more precise and reliable solution for GNSS-R-based soil moisture inversion.
