Synergizing multi-collocation fusion and spatiotemporal deep learning to enhance global soil moisture memory characterization

Accurate identification of Soil Moisture (SM) memory is essential for monitoring climate and meteorological extreme events. Due to simulation or retrieval uncertainties and limited sampling frequency, SM memory metrics derived from various Land Surface Models (LSM) and satellite observations exhibited substantial discrepancies and biases. To address this, a fusion–reconstruction dual strategy was proposed. First, a Multiple Collocation Analysis (MCA)-based fusion framework was introduced to reducing SM estimation uncertainties by integrating multi-source microwave remote sensing products from the Soil Moisture Active Passive (SMAP), Soil Moisture and Ocean Salinity (SMOS), Advanced Microwave Scanning Radiometer 2 (AMSR2), and Advanced Scatterometer (ASCAT) with LSM products from the European Centre for Medium-Range Weather Forecasts Reanalysis 5 (ERA5)-Land, Modern-Era Retrospective analysis for Research and Applications Version 2 (MERRA2), and Global Land Data Assimilation System (GLDAS). Then, a spatiotemporal reconstruction scheme was performed using the Convolutional Long Short Term Memory (ConvLSTM) model, with residual interpolation conducted via Spatiotemporal Inverse Distance Weighting (ST-IDW), to improve sampling frequency of the fusion product. In-situ validation demonstrated that the fusion product achieved statistically significantly superior overall accuracy compared to all parent products, and the ConvLSTM effectively reconstructed missing information and enhanced the spatiotemporal data coverage without significantly degrading fusion accuracy. The SM memory timescale is primarily controlled by long-term evapotranspiration and a short-term, precipitation-driven memory metric, Formula: see text, was derived based on the harmonized, high-accuracy, spatiotemporally continuous global daily surface SM dataset from 31 March 2015 to 31 December 2020. Here, Formula: see text is defined as the fraction of precipitation that remains stored in the surface soil layer one day after reaching the land surface. The memory timescale quantifies the persistence of anomalously dry/wet SM states under atmospheric or climatic forcing, with a global mean of ~13 days. The global mean Formula: see text is 0.327, indicating that, on average, ~32.7% of precipitation is retained in the surface soil layer under the assumed 0–7 cm depth after one day. Memory timescale decreased with soil sand content but increased with clay content, while Formula: see text showed the opposite trend. Low-frequency sampling led to an underestimation of Formula: see text and the memory timescale. Additionally, Formula: see text was underestimated by 0.10–0.22 across L-band microwave satellite and LSM products, and memory timescale was overestimated by LSM products from GLDAS and MERRA2 (12.85 and 28.23 days) and underestimated by microwave products (1.3–8.4 days). The proposed dual strategy mitigated biases in both short- and long-term SM memory estimation caused by sparse sampling and inaccurate SM retrievals/simulations, extended satellite-based memory beyond the topsoil layer, and enhanced LSM-based memory through integrating satellite observations. This reliable and consistent SM memory characterization supports facilitated quantification of global terrestrial water resources and water cycle rates.