Estimation Of Soil Salinity In Werigan-kuqa Oasis Coupling Multi-source Sensors And Machine Learning Algorithms

Soil salinization is one of the main types of soil degradation in arid areas,which seriously affects the rational allocation of soil and water resources and the healthy development of ecosystem in arid areas.Therefore,it is particularly important to realize the dynamic,rapid and accurate monitoring of soil salinization.Remote sensing technology had shown great potential in the research of continuous observation of soil salinization in a large area,among which optical remote sensing is the most widely used.However,the observation of optical image is easily affected by bad weather,and the advantage of microwave remote sensing which can be observed all day long makes up for this deficiency.The newly launched Sentinel satellites provided a large number of freely accessible remote sensing image data with high spatial and temporal resolution,which provided a new data source for quantitative evaluation of soil salinity.Based on this,this studied Sentinel-1 synthetic aperture radar(SAR)images of different polarization combinations to build radar index,the Sentinel-2 multi-spectral instrument(MSI)image build salinity index,vegetation index,supplemented by the terrain derived variable coupling machine learning model and soil water content,on the basis of the optimization variables were constructed based on radar sensor,optical sensor and the prediction model of soil salt multi-source data fusion,and map the spatial distribution of soil salt figure in the study area,able to jump to conclusions:(1)The soil salinity in the Werigan-Kuqa oasis decreases with the increase of depth and is alkaline.With the increase of depth,it changes from strong variation to moderate variation.The soil water content also increases with the increase of depth.In addition,the semivariance classical geological statistical analysis shows that the variation of soil salt at different depths is mainly caused by structural variation,and the random errors caused by short-distance errors and experimental errors are small.Among them,the surface soil salinity fits best with the corresponding theoretical semivariance model(R²=0.91).(2)Using the optimal salinity index,vegetation index and topographic index of Pearson correlation analysis,supplemented by soil water content as the predictor of EC,four types of soil salinity prediction models based on optical sensors were constructed,among which the prediction of XGBoost model.The accuracy is the best(validation set R²=0.71,RMSE=9.38 d S m^-1,MAE=7.86 d S m^-1,RPD=1.85),followed by RF(validation set R²=0.65,RMSE=10.21 d S m^-1,MAE=7.90 d S m^-1,RPD=1.70).In addition,although the importance of different indexes in the model is different,they still have the same order,that is,the salinity index is the most important,accounting for more than 45%,followed by It is the topographic index,vegetation index and soil moisture content.(3)In the radar sensor-based soil salt prediction model constructed with radar index,topographic index and soil moisture content,the RF estimation accuracy is the best(verification set R²=0.61,RMSE=10.77 d S m^-1,MAE=7.37 d S m^-1,RPD=1.61),where the topographic index is designated as the most important explanatory variable by the model,and its importance exceeds 50%,followed by radar index and soil moisture content.(4)Compared with a single sensor,the fusion of multiple sensors greatly improves the estimation accuracy and stability of the model.Both the RF and XGBoost models have excellent prediction accuracy(RPD>2.0).Relatively speaking,the RF prediction accuracy Better(validation set R²=0.81,RMSE=7.52 d S m^-1,RPD=2.30),in which the salinity index is designated as the most important explanatory variable,and its importance exceeds 40%,while other types of indexes are different The interpretability in the model is different.(5)The soil salinity predicted by different data sources and models has similar spatial distribution characteristics,that is,the degree of soil salinization continues to increase from the inside of the oasis to the edge.According to the prediction results of the best prediction model(RF)constructed with multi-source data as predictors,the area of non-saline soil in the study area accounts for only 16.26%,while the area of heavy soil accounts for more than 60%.The staining problem is serious.This study separately evaluated the prediction accuracy of different types of sensors and their combinations in soil salinization monitoring,and carried out mapping of the spatial distribution characteristics of soil salinization,laying a foundation for the dynamic monitoring of soil salinization at the regional scale.It also provides relevant theoretical support for improving the allocation of water and soil resources and preventing the continuous deterioration of the ecological environment.