Accurate soil moisture (SM) information at sub-kilometer resolution is essential for hydrologic modeling, weather forecasting, precision irrigation, and drought monitoring. Satellite missions such as Soil Moisture Active Passive (SMAP) have transformed global SM monitoring, but their coarse resolution, shallow sensing depth, and multi-day revisit cycle limit local-scale applications. Retrieval accuracy further declines over dense vegetation and complex terrain. This study presents a deep learning downscaling framework that integrates convolutional neural networks and long short-term memory networks (CNN-LSTM) to produce daily SM maps at 100 m resolution for surface (5 cm) and root-zone (20 cm) depths. The model was trained with seven years of SM data from similar to 650 stations across the U.S., using dynamic predictors from SMAP and MODIS (i.e., brightness temperature, roughness coefficient, surface reflectance in red, near-infrared, and shortwave infrared) and static features predictors from elevation, land cover, and Soil Landscapes of the U.S. (SOLUS100) databases. Compared to SMAP level-3 SM, the CNN-LSTM improved surface SM accuracy, with median correlation coefficient (R) increasing from 0.61 to 0.82, unbiased root mean square error (ubRMSE) decreasing from 0.06 to 0.05 cm(3) cm(-3), and Kling-Gupta Efficiency (KGE) increasing from 0.30 to 0.61. Root-zone SM achieved R similar to 0.72, ubRMSE similar to 0.05 cm(3) cm(-3) and KGE similar to 0.53. The model also outperformed SMAP in forested and mountainous areas, capturing SMAP sub-pixel SM variability across the three well-instrumented watersheds and Florida. In Florida, where no training data was used, performance decreased but was still better than SMAP. These findings indicate that the CNN-LSTM framework bridges the spatial resolution gap in SM products and support small scale agricultural and hydrological applications.
更多
原文链接
分享
引用
