A crop attribute-aware deep learning framework for farmland soil moisture retrieval under dense crop canopies

High-resolution soil moisture retrieval is critical for agricultural water management, but dense canopies severely degrade its accuracy. Vegetation attenuation and volume scattering weaken soil moisture-related signals, while dynamic crop structure and phenology further destabilize retrieval relationships. Addressing vegetation-induced interference therefore remains a major challenge for farmland soil moisture monitoring. This study proposes a Crop Attribute-Aware Multi-Branch Neural Network (CAA-MBNN) integrating Sentinel-1 SAR and Sentinel-2 optical data to improve soil moisture retrieval under varying canopy conditions. Crop type, plant height, and phenological stage serve as auxiliary supervision targets during training to learn semi-implicit vegetation representations, enabling 10 m regional soil moisture mapping without requiring these attributes during inference. Six field campaigns were conducted in the Hetao Irrigation District from August 2024 to August 2025, yielding more than 1000 in situ samples across five major crops and different growth stages, collected within ±1 day of Sentinel-1 acquisitions. Based on pooled independent test predictions from five-fold cross-validation, CAA-MBNN achieved an overall Pearson correlation coefficient (R) of 0.774 and a root mean square error (RMSE) of 0.052 m³ m - ³ , maintaining R = 0.724 and RMSE = 0.052 m³ m - ³ under dense vegetation (normalized difference vegetation index, NDVI≥0.6), and outperformed conventional multilayer perceptron (MLP) and random forest (RF) models and their enhanced versions with measured crop attribute inputs. Branch contribution analysis showed that the phenology, plant height, and crop type branches increased R by 16.9%, 16.1%, and 9.5%, respectively, while the complete CAA-MBNN achieved the largest improvement in R (20.9%) relative to the model without auxiliary branches. In cross-year validation, CAA-MBNN maintained robust performance, achieving the highest R (0.622) and lowest RMSE (0.037 m³ m - ³) among all models. These results demonstrate that structured incorporation of crop attribute information enhances soil moisture retrieval accuracy, robustness, and interpretability, especially under dense crop canopies, supporting large-scale, high-resolution soil moisture mapping in heterogeneous irrigated croplands.