Water resource management in agricultural irrigation districts is increasingly dependent on high-quality, multilayer soil moisture data. However, the currently available remote sensing soil moisture products typically fail to offer high spatial resolution suitable for field-scale monitoring, especially for studying dynamic changes in soil moisture (SM) within the crop root zone. Leveraging high-spatial-resolution radar remote sensing technologies and deep learning algorithms, we propose a multi-layer soil moisture inversion framework based on feature-level fusion models. The framework aims to estimate soil moisture at depths of 0-20 cm, 20-40 cm, and 40-60 cm with a spatial resolution of 10 m & times; 10 m for arid agricultural area with shallow groundwater. This method first uses Sentinel-1A VV/VH dual-polarization radar data, land surface temperature (LST), actual evapotranspiration (ETc act), saturated hydraulic conductivity (Ksat), and field monitoring layered soil moisture and groundwater depth (GWD) data from 59 regional sites over a complete growing season as input variables for feature extraction by the LSTM deep learning model. The extracted 64-dimensional features are then fed into the XGBoost machine learning model for modeling, ultimately generating high spatial resolution layered soil moisture data. The framework was evaluated using continuous monitoring data from 11 validation sites. The results show that, compared to using LSTM and XGBoost models individually, the feature-level fusion model-LSTMXGBoost-performs better in simulating soil moisture at three different depths, with the most notable performance improvement at 40-60 cm depth (R2=0.726, RMSE=0.044 cm3/cm3, MAE=0.027 cm3/cm3). Moreover, the LSTM-XGBoost model demonstrates stronger temporal dynamic capturing ability for soil moisture simulations, with R2 greater than 0.56 at all three depths. Furthermore, we found that ETc act and GWD play a dominant role in simulating surface soil moisture (0-20 cm) and deep soil moisture (40-60 cm), respectively in the area featured by arid climate condition and shallow groundwater. The proposed framework effectively enhances the high-accuracy prediction of layered soil moisture at field scales, providing strong support for agricultural irrigation management and soil moisture monitoring.