Study On Soil Salt Inversion Based On Multisource Data And Machine Learning Algorithm In The Ebinur Lake Wetland National Nature Reserve

Soil salinization restricts the development of agriculture and has always been one of the most prominent ecological and environmental problems in the arid region of Northwest China.In the Ebinur Lake Wetland National Nature Reserve in northwestern China,due to the influence of topography,climate and human factors,the problem of soil salinization has become more serious,the ecological environment has deteriorated sharply,and the local agricultural economy has been threatened.Therefore,the management of saline soil is imminent.However,the acquisition of soil salinity and the spatial distribution in different seasons are very important for the management of saline soil,so this article takes the Ebinur Lake Wetland National Nature Reserve as the research object,selects the soil salt data of August 2018(dry season)and May 2019(wet season),and compares and analyzes the dynamic changes of soil salt and moisture,And spatial distribution,and construct traditional,two-dimensional and three-dimensional spectral indexes through multi-spectral remote sensing images Sentinel-2,Landsat 8 OLI and HJ1A-CCD2,select the soil salinity data of August 2018(dry season)and May 2019(wet season),contrast and analyze the dynamic changes and spatial distribution of soil salinity and moisture in the dry and wet seasons,through the multi-spectral remote sensing image Sentinel-2,Landsat 8 OLI and HJ1A-CCD2 to construct traditional,two-dimensional,three-dimensional spectral index.Finally,based on multiple linear regression(MLR),partial least squares regression(PLSR),random forest(RF)and BP neural network(BPNN)to construct soil salt inversion models,and through the coefficient of determination(R~2),root mean square error(RMSE)and relative analysis error(RPD)and other model accuracy evaluation to select the best soil salinity inversion model.And the accuracy of the model was verified by the soil salt data in July2017(dry season)and May 2018(wet season),and finally used in the dry and wet season soil salt inversion.The main conclusions are as follows:(1)According to the different layers of soil,the surface soil has more salt and less water,as the depth increases,the salt decreases and the water increases.According to the spatial distribution of soil water and salt,soil water and salt in the dry and wet seasons have obvious differences,the soil salt content in the southwest of the dry season is large and the wind force in the wet season is strong,resulting in uneven distribution of the surface soil salt,and the soil water content in the wet season is higher than that in the dry season.(2)According to the spectral index constructed after differential transformation of Sentinel-2 image,the correlation between the traditional spectral index and soil salinity is the lowest,and the three-dimensional spectral index is the highest,and the correlation between the three-dimensional spectral index and soil salinity in the first-order data is higher than that in the original and second-order data.Moreover,the correlation between the three-dimensional spectral index and soil salinity in the first-order data is higher than that in the original and second-order data.By comparing and analyzing the three-dimensional spectral index constructed by the Landsat 8 OLI and HJ1A-CCD image data,the spectral index constructed by the first-order data is highly correlated with soil salinity,which is consistent with the analysis results of the three-dimensional spectral index of Sentinel-2 data.Among the three-dimensional spectral indexes constructed by Sentinel-2and Landsat 8 OLI data,most of the three-dimensional spectral indexes include the far infrared band,and the correlation between the three-dimensional spectral index constructed by HJ1A-CCD2 data without far infrared band and soil salinity is much lower than that of Sentinel-2 and Landsat 8 OLI data,it shows that the spectral index constructed by shortwave infrared is more sensitive to soil salinity,which is consistent with previous research results.(3)The PLSR model constructed by the Sentinel-2 first-order three-dimensional spectral index is the best soil salinity inversion model.Its validation set R~2 is 0.754,RMSE is 6.216,RPD is 1.474,RPD is greater than 1.4,and it has predictive ability.The model is validated by soil salt data of July 2017(dry season)and may 2018(wet season),in the dry season,R~2 is 0.576,RMSE is 2.244,RPD is 2.351,the accuracy is high,the difference between the measured value and the predicted value is small and the fitting degree is high,in wet season,R~2 is 0.664,RMSE is 2.746,RPD is 1.460 respectively,most of the points are distributed around the 1:1 line,the predicted value is close to the measured value and the difference is small,so the model can be used for soil salinity inversion for dry and wet seasons.Therefore,PLSR model is applied to soil salt inversion for dry and wet seasons.It can be seen that soil salt decreases gradually from the lakeside to the periphery,which is consistent with the field survey results.