High-resolution soil moisture estimation via multi-source data merging and machine learning-based downscaling over Taiwan

Soil moisture (SM) is a pivotal variable that regulates hydroclimate at regional to global scales. Various techniques for acquiring SM data have demonstrated their strengths and limitations in accuracy, spatiotemporal coverage, and resolution. This study employs the triple collocation method to merge SM data from the Climate Change Initiative (CCI), Soil Moisture Active Passive (SMAP), and the High-Resolution Land Data Assimilation System (HRLDAS) products over Taiwan from 2018 to 2021. Afterward, the merged SM product at 0.25° is downscaled to 3 km using a machine learning model, namely convolutional long short-term memory networks. The original (whether merged or downscaled) and final SM products (merged and downscaled) are validated against in-situ measurements. The merged and downscaled SM, which demonstrates the best overall performance (with bias, RMSE, and ubRMSE 13–68% lower than those of other products) and reasonable fine-scale responses to heavy rainfall events, is expected to support various hydrometeorological applications.