Soil moisture retrieval and trend prediction using multi-temporal remote sensing data: An interpretable deep regression approach

High-precision retrieval of soil moisture (SM) and prediction of its trends are crucial for research in agriculture, meteorology, and related fields. Multi-temporal multi-source remote sensing data can provide temporal variation information of various features. This study aims to achieve high-precision retrieval and prediction of SM by leveraging multi-temporal features. Compared to physics-based models, data-driven regression models exhibit significant advantages in handling multi-dimensional complex features. However, the lack of effective interpretation of their operational mechanisms remains a limitation in current data-driven model research. Given the superior performance of Transformer networks in processing multi-sequence features, this study constructs a deep regression model based on the Transformer architecture for SM extraction. To interpret the SM regression process of this model, the study quantifies the influence of input features on regression, analyzes the temporal variations of intermediate hidden features, and evaluates the output performance to elucidate the feature extraction and regression mechanisms. Experiments were conducted in the Pacific Northwest region. Analysis of feature derivatives and intermediate hidden features reveals that the Transformer intelligently allocates appropriate attention to data at different time points, resulting in stronger feature influence closer to the retrieval or prediction date. The experimental results indicate that multi temporal information is beneficial for SM retrieval and prediction, while assigning appropriate attention to features at different time points is more advantageous for predicting SM trends. This study provides a practical approach for deep regression-based SM retrieval and prediction and offers insights into interpreting the SM regression mechanisms of the Transformer.