Point-surface Fusion Of Ground Measurements And Satellite Observations For Quality Improvement And Retrieval Of Soil Moisture Using Machine Learning

Soil moisture（SM）is a key component of the global water cycle.A variety of techniques are available for long-term SM monitoring.The ground measurement technique allows for direct access to high-precision SM data,but is limited by factors such as small spatial support and high maintenance costs,making it difficult to monitor continuously on a large scale.On the other hand,microwave remote sensing is an effective tool for the inversion of long-term and large-scale surface SM,which can compensate for the inability of ground-based measurement in terms of observing scales.However,since quantitative remote sensing inversion is typically an ill-posed problem,traditional SM retrieval algorithms based on microwave radiative transfer models have many drawbacks in terms of accuracy,computational costs,and product release delay.In view of the fact that machine learning（ML）algorithms can circumvent the model parameterization process,flexibly combine heterogeneous multi-source data,and have the advantages of strong approximation ability and high computational efficiency,and taking account of the complementarity between microwave satellite observations and ground-based measurements,this paper proposes an ML-based point-surface fusion framework for satellite SM correction and retrieval,with the goal of producing high-precision SM products.（1）ML-based point-surface SM correctionIn order to solve the ill-posed inversion problem of traditional satellite SM retrieval algorithms and improve the quality of existing satellite SM products,an ML-based point-surface SM correction method is presented.By using satellite SM products as the input of ML algorithms,the in-situ SM as the training target,the ML algorithms are able to establish the relationship between the point and the surface SM,whereby the quality of the original satellite SM product can be improved based on the established relationship.Meanwhile,the effectiveness of the Global Navigation Satellite System-Interferometric Reflectometry（GNSS-IR）-based SM estimates in mitigating the point-surface scale mismatch issue is investigated by incorporating them into the ML training process.By conducting experiments in the western part of the continental United States（CONUS）,it was found that among the three ML algorithms used in this study（i.e.,the multiple linear regression（MLR）,the back-propagation neural network（BPNN）and the generalized regression neural network（GRNN））,the GRNN model performed the best overall,with cross-validated correlation coefficient（R）and unbiased root-mean-square error（ub RMSE）of 0.81 and 0.059 cm³ cm^-3,respectively.In addition,the GRNN-corrected SM product was superior to the original Soil Moisture Active Passive（SMAP）satellite’s passive SM product in terms of spatial patterns and temporal dynamics.（2）ML-based point-surface SM retrievalIn the previous method,the official satellite SM products are used as the input data,but they are derived from satellite observations through the inversion of microwave radiative transfer models,which went through a process of error propagation and accumulation;therefore,in order to circumvent this process,an ML-based point-surface SM retrieval method is suggested by replacing the input data with microwave brightness temperature observations.The potential of this method for high-precision SM retrieval is further explored by employing the extended triple collocation technique to screen reliable in-situ sites,whereby the spatially representative error caused by the scale mismatch issue can be weakened.The results of experiments in the entire CONUS showed that the GRNN trained on reliable in-situ SM data substantially outperformed MLR and BPNN in terms of overall performance and average evaluation metrics over individual SM networks from the International Soil Moisture Network（ISMN）,with cross-validated R and ub RMSE of 0.88 and 0.050 cm³ cm^-3,respectively.In addition,compared to the official SMAP SM product,ERA-Interim SM reanalysis data and in-situ SM measurements from ISMN,the GRNN-based point-surface SM retrievals had the best spatial performance and temporal consistency with ground SM reference data.