基于XGBoost与地理加权回归的吉林省西部土壤盐渍化反演OACSTPCD

In order to timely and accurately applying multi-source remote sensing data to extract soil salinization inversion characteristics in arid and semi-arid areas and obtain spatial distribution data of soil salinization degree.Taking Da'an City in western Jilin Province as the research area,Sentinel-1 SAR,Sentinel-2 MSI multi-source remote sensing data and DEM data were used to construct a soil salt content(SSC)inversion feature set.Combined with the BorutaShap algorithm to optimize features,the coupling Geographically Weighted Regression(GWR)with XGBoost regression was used to construct a soil salinization inversion model.The results were compared with XGBoost regression and GWR inversion results.The results show that:The SSC inversion features of this study are concentrated.Salinity index and vegetation index achieve high importance ranking in the BorutaShap algorithm and are important features of SSC inversion in Da'an City.In the soil salinization inversion model constructed,the R2 and RMSE of the GWR model are 0.48 and 4.83 g/kg,respectively,and the R2 and RMSE of the XGBoost regression model are 0.54 and 4.35 g/kg,respectively.The prediction accuracy of the soil salinization inversion model constructed by coupling GWR and XGBoost regression is significantly improved,with R2 and RMSE reaching 0.63 and 3.71 g/kg,respectively.According to the inversion results of this model,there is strong spatial heterogeneity in the distribution of various types of saline soil in Da'an City.The SSC shows a gradually decreasing trend from southeast to northwest,which is basically consistent with the field survey.In summary,the coupled GWR and XGBoost regression model constructed in this study fully considers the spatial heterogeneity and nonlinear relationship of the inversion characteristics,which effectively improves the SSC inversion accuracy.A more realistic SSC spatial distribution can be obtained.This model can be used for the inversion of soil salinity content in arid and semi-arid areas.