Spatiotemporal Seamless Estimation of Global Surface Soil Moisture Using Triple Collocation, Machine Learning, and Data Assimilation

Accurate and spatiotemporal seamless soil moisture (SM) products are important for hydrological drought monitoring and agricultural water management. Currently, physically-based process models with data assimilation are widely used for global seamless SM generation, such as Soil Moisture Active Passive level 4 (SMAP L4), the land component of the fifth generation of European Reanalysis (ERA5-land) and Global Land Data Assimilation System Noah (GLDAS-Noah). These datasets are usually produced using high-performance computation platforms and may subject to potential uncertainties from model structure and parameters, limiting their practical application capacity in a flexible way in local or global areas. Here, we proposed a data-driven artificial intelligence (AI)-based method to generate spatiotemporal seamless daily soil moisture data using triple collocation, machine learning and data assimilation. Specifically, the triple collocation correlation coefficients (TCR) method is employed to combine different SM datasets in order to obtain high-accuracy label data for model training first. A LightGBM machine learning (ML) model is constructed to simulate global daily soil moisture at 0.25◦ in an autoregressive way, using ERA5 meteorological forcings and MSWEP precipitation data as inputs. In addition, the satellite-based soil moisture SMAP level 3 (SMAP L3) is assimilated into the developed machine learning model using the simple Newtonian nudging technique to update the soil moisture simulation states. The incorporation of data assimilation into machine learning mimics the idea of physical models and brings much room for adaptable soil moisture simulations. The developed data-driven model is examined over global land areas from March 31, 2015 to May 31, 2023 with a ten-fold cross validation scheme, evaluated using 1094 in-situ soil moisture stations from International Soil Moisture Network (ISMN). The results indicate that the ML-based assimilated soil moisture dataset (MLDA) demonstrates a median correlation (R) of 0.741 and an unbiased root mean square error (ubRMSE) of 0.0437 m<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup>/m<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup>, better than SMAP L4 (R=0.717, ubRMSE=0.0452 m<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup>/m<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup>, ERA5-land (R=0.706, ubRMSE=0.0452 m<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup>/m<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup>) and GLDAS (R=0.633, ubRMSE=0.0501 m<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup>/m<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup>). Compared to the three model-based soil moisture products, the ML-DA dataset exhibits superior performance in time and space and also in dry-wet zones. Therefore, the developed ML-DA framework offers significant potential for accurate, spatiotemporal soil moisture simulations globally.