A novel AI framework for soil moisture estimation using sentinel-2 multispectral imagery and temporal convolutional networks

Abstract

Accurately assessing soil moisture is indispensable for agriculture, water resources management, and hydrology studies. The priority of these forecasts is even higher, especially in areas with different climatic conditions and terrain characteristics. Most satellite-based approaches currently in use are not able to effectively account for spectral information, spatial characteristics, and seasonal changes all at once. With this problem in mind, we have proposed a new deep learning approach that uses a combination of Sentinel-2 multi-spectral satellite images and ground-based soil moisture measurements. We have used the Spectral Fusion Enhancement approach to prioritize important spectral bands, Convolutional Neural Networks to identify terrain features, and Dynamic Dilated Temporal Convolutional Networks to analyze seasonal changes. The study used 3,801 Sentinel-2 satellite image patches collected from 26 districts of Andhra Pradesh state by Google Earth Engine and the actual soil moisture values obtained from APWRIMS. According to the results of the experiment, the proposed model achieved values of 4.52 RMSE, 1.67 MAE, and 0.94 R2, showing better performance than other ordinary models. The ablation study also demonstrated that the predictive accuracy is significantly improved when each component is worked together. Therefore, this approach is useful for reliably estimating soil moisture at the regional level and is crucial in precision agriculture and water management.