Spatiotemporal Fusion of Soil Freeze/Thaw Datasets at Decision-Level Based on Convolutional Long Short-Term Memory Network

The transition of soil freeze/thaw (FT) state plays a critical role in ecosystem, hydrological, and biogeochemical processes. However, obtaining representative soil FT state datasets with a long time sequence and fine spatiotemporal resolution remains challenging. Here, a decision-level spatiotemporal data fusion algorithm based on Convolutional Long Short-Term Memory networks (ConvLSTM) for soil FT product is proposed to enhance the enhanced L3 freeze/thaw product of the Soil Moisture Active Passive Mission (SMAP_E_FT). In the algorithm, the feature of long-time sequence from the Freeze/Thaw Earth System Data Record product (ESDR_FT) and the land surface dataset produced from the fifth generation ECMWF atmospheric reanalysis (ERA5-land) are sucked in. The results of the validation based on eight dense in-situ networks in Genhe, Maqu, Naqu, Ngari, Pali, Saihanba, Shandian river basin (SDR), and Tianjun show that the correlation coefficient between the fusion product (ConvLSTM_FT) and SMAP_E_FT is 0.9498, and the deviation is 0.0140, which maintains the classification accuracy. This study provides a possible way of temporally extending the SMAP_E_FT.