Research On The Generation Of High-resolution Soil Moisture Dataset Based On Machine Learning

High-quality soil moisture data have practical significance for the research and application of Earth system science.Since soil moisture is an important component in earth system research,it not only has a significant impact on climate change,biological ecology,and hydrological system,but also has a wide range of applications for agricultural production,disaster warning,ecological environmental protection,and soil management.However,the currently available reanalysis and remote sensing soil moisture data at home and abroad have certain limitations,such as low resolution and limited to topsoil,which is not conducive to further research and application.Therefore,this paper conducts in-depth research on generating high-resolution soil moisture datasets based on machine learning models.The main work is as follows:In this study,a novel soil moisture dataset,SMCI1.0(Soil Moisture of China by insitu data version 1.0),was generated by a robust machine learning approach at the national scale using 1,789 in-situ soil moisture,ERA5-Land,leaf area index,land cover type,topography,and soil texture as a benchmark.SMCI1.0 contains soil moisture with a high spatiotemporal resolution of 1km×1km from 2000 to 2020,at a daily scale,10 depth layers(10 to 100cm)spaced at 10 cm intervals.In the course of the study,the performance of various machine learning methods for generating soil moisture datasets was evaluated and analyzed,including Random Forest,RF),Cat Boost,e Xtreme Gradient Boosting(XGBoost),Muti-Layer Perception(MLP).By comparing the results of soil moisture data sets generated by various models and considering the wide application of the algorithm,the overall optimal random forest model was selected to generate SMCI1.0 data set.In terms of data acquisition and preprocessing,we strictly control the quality of the obtained field observed soil moisture data,and select predictors with physical significance for soil moisture,which provides strong data support for random forest modeling.Two sets of experiments were conducted to verify the accuracy of SMCI1.0.The first experiment is year-to-year,and the results show that ub RMSE is 0.043 ～ 0.053,R is0.867 ～ 0.908;The second experiment is station-to-station,and the results show that ub RMSE is 0.045 ～ 0.051,and R is 0.866 ～ 0.893.Through experimental analysis,compared with other popular soil moisture datasets(such as ERA5-Land,SMAP-L4,and So Mo.ml),SMCI1.0 has higher quality advantages such as high resolution,continuous spatial distribution,long coverage time span,more detailed description of deep soil moisture information,and low error.At the same time,it is found that the error of SMCI1.0 is mainly distributed in the North China monsoon region.In general,this study conducted two groups of experiments: year-to-year and station-to-station,and the highprecision experimental results estimation ensured the applicability of SMCI1.0 in the study of spatiotemporal patterns.The SMCI1.0 soil moisture dataset based on in-situ can be an effective complement to existing model-based and satellite datasets.This data set can be used for a variety of hydrological,meteorological,and ecological analyses and modeling,especially in applications that require high-quality,high-resolution soil moisture data.The dataset is published on the contribution platform of the National Tibetan Plateau Data Center.The DOI link for this dataset is http://dx.doi.org/10.11888/Terre.tpdc.272415.