Soil moisture retrieval and spatiotemporal variation analysis based on deep learning

Soil moisture is a key factor in soil-atmosphere energy and material exchanges, playing a crucial role in the hydrological cycle and agricultural management. Traditional monitoring methods are limited in large-scale and real-time applications, and the complex mechanisms of soil processes complicate modeling. However, deep learning provides a promising approach to capturing the complex nonlinear relationships between feature parameters and soil moisture content. Here, Sentinel-1 and Landsat data, along with in-situ measurements (0–10 cm) from the Wei-Ku Oasis, were used to extract 36 feature parameters. The Boruta algorithm and correlation analysis were applied to select key variables. Nine deep learning models, including three basic architectures (Convolutional Neural Networks (CNN), Long Short-Term Memory Networks (LSTM), Transformer) and six hybrid structures (CNN-LSTM, LSTM-CNN, CNN-with-LSTM, CNN-Transformer, GAN-LSTM, Transformer-LSTM), were systematically compared to evaluate the impact of neural network structure on model performance. The optimal model was then used to perform spatiotemporal mapping of soil moisture. Results indicated that both single-structure and hybrid models were effective for soil moisture retrieval, with CNN-based models (either standalone or hybrid) performing better. Among them, the CNN-LSTM hybrid model achieved the best performance with an R²of 0.72 on the test set. The soil moisture map produced by the optimal model reveals a spatial distribution pattern in the Weiku Oasis, characterized by higher moisture levels in the center and lower levels at the periphery. Temporally, from 2017–2024, soil moisture at the 0–10 cm depth exhibited an overall increasing trend. We demonstrate that the design of efficient neural network architectures is essential for soil moisture inversion, and provides valuable insights for deep learning applications in hydrological parameter estimation and other challenges in complex environmental contexts.