A multi-task-oriented spatial and temporal convolutional neural network for high-resolution soil moisture estimation

Soil moisture is an essential state variable that controls the hydrological processes of infiltration, runoff, evaporation, and precipitation. It affects various studies and applications including projection of drought or waterlogging, management of water resources, estimation of agricultural production, forecasting of weather conditions, and analysis of climate change. Currently, passive microwave remote sensing is one of the most effective technologies for large-scale monitoring of soil moisture dynamics. However, due to the insufficient continuity and resolution of passive microwave soil moisture data, it is difficult to meet the requirements of regional fine-grained hydrological research. There is an urgent need for the comprehensive improvement of passive microwave soil moisture products, to enhance their capabilities for various practical applications. Therefore, in this study, we propose a Multitask-Oriented Spatio-Temporal Convolutional Network (MOSTCN) to estimate daily soil moisture data with high accuracy, good continuity, and fine resolution. The multitask-oriented learning framework consists of gap-filling, multi-source fusion, and spatial downscaling, which can fully utilize multivariate spatio-temporal information. Firstly, the gap-filling process is employed to fill the gaps in the target dataset by considering the spatiotemporal correlation of soil moisture, which includes the partial convolution module, long short-term memory module, and cross-attention module. Subsequently, the multi-source fusion process is designed to combine the assimilated soil moisture data with the reconstructed satellite soil moisture data. The residual dense module and Laplacian attention module are introduced to extract multi-scale and multi-level features, which preliminarily enhance the spatial resolution of the soil moisture data. Lastly, the advanced residual dense convolutional module with embedded attention mechanisms is used to explore the complex nonlinear correlations between soil moisture and various ancillary factors, and then estimate the soil moisture data with high continuity and fine resolution. The experimental results demonstrate that the MOSTCN method has a robust effect on enhancing the continuity and resolution of passive microwave soil moisture data in China. Compared with in-situ measurements, the proposed method increases the correlation coefficient from 0.57 to 0.72 and decreases the root-mean-square error by 0.015 m3/m3. Its generalization is superior to that of the two baseline methods. This provides an efficient way to generate seamless and high-resolution soil moisture products.