A multiscale deep learning model integrating satellite-based and in-situ data for high-resolution soil moisture predictions

Detailed and accurate soil moisture is critical for many applications, such as forecasting agricultural drought and pests and mapping landslides. Deep learning can perform extraordinarily well in soil moisture, streamflow, and model uncertainty estimation. However, these models may inherit disadvantages of training data, such as limited coverage of in situ data or low resolution/accuracy of satellite data. Here, we propose a novel multiscale DL scheme that learns from satellite and in situ data to predict daily soil moisture at 9 km. The model outperforms land surface models, the SMAP satellite product, and a candidate machine learning model. Based on spatial cross-validation, it achieved a median correlation of 0.901 and a root-mean-square error of 0.034 m3/m3 over sites in the conterminous United States. Our scheme generally applies to topics in the geosciences with multiscale data, breaking the limitations of a single dataset.