A climate-adaptive transfer learning framework for improving soil moisture estimation in the Qinghai-Tibet Plateau

Soil moisture (SM) plays essential roles in revealing complex interaction mechanisms among air–soil-water-plant processes. In the Qinghai-Tibet Plateau (QTP), the in-situ SM data is sparse and limited, satellite-based SM data has short period, while reanalysis SM data has advantages on long-term and high spatiotemporal resolution but has relatively high error. In this study, to improve soil moisture estimation in the QTP, we aim to propose a Climate-Adaptive Transfer Learning (CATL) framework for data scarce region based on reanalysis data (ERA5-LAND dataset) and the in-situ data (International Soil Moisture Network (ISMN) data). Specifically, regarding the QTP as the target region, selecting the areas with similar climate types with QTP as the source region, we train the CNN-LSTM fusion model in the source region and then transfer it to the target region via fine-tuning strategy. Results indicate that the produced soil moisture data based on CATL framework achieves CC of 0.755 and ubRMSE of 0.042, which has better quality than SMAPL3 during 2015–2019. Additionally, the CATL framework also produced the historical SM data reconstruction during 1960–2010, with CC increased by 11.3 % and ubRMSE reduced by 1.5 % compared with the original ERA5-Land reanalysis data. Furthermore, compared to the direct fine-tuning strategy (without climate adaptive), the CATL framework showed an increase of CC with 2.6 %, and decreases in RMSE, MAE, and ubRMSE of 5.3 %, 4.2 %, and 7.5 %, respectively. Finally, an improved soil moisture dataset (daily, 0.05°) ranging from 1960 to 2019 is produced for the QTP. This study provides a new tool for soil moisture estimation improvement in data-scarce region which will also benefit basin hydrology and water resources management.