Soil Salinity Inversion In Juyanze Based On Variable Optimization And Machine Learning

Soil salinization is a global problem that destroys land resources and seriously threatens the security and stability of ecosystem.It is of far-reaching significance for ecological environment protection and sustainable use of land resources to acquire soil salt information quickly and accurately in a wide range.Remote sensing technology is one of the important technical means of soil salinization monitoring,among which the remote sensing retrieval of soil salt with multi-source and multi-parameter combination has attracted much attention,but this kind of research lacks the comparison of the efficiency and advantages of various parameters in salt retrieval.In recent years,polarimetric synthetic aperture radar technology（Pol SAR）has developed rapidly.With rich polarization scattering information,fully polarized radar images can reflect the scattering mechanism between different targets and provide rich features and information of ground objects.Therefore,it has been widely used in the research fields of ground objects classification and surface parameter inversion,providing a new technical means for extracting soil salt information,but relevant research is scarce.There are many environmental factors affecting soil salinity,how to express the complex relationship between multiple environmental factors and soil salinity is the key to quantitatively describe soil salinization information.Compared with traditional methods,machine learning algorithms such as random forest（RF）,support vector machine（SVM）,back propagation artificial neural network（BP-ANN）and relevance vector machine（RVM）can effectively solve the nonlinear problem between soil salinity and environmental factors,and are not limited by the type and number of input environmental variables.However,its effect in salt inversion also needs to be widely verified.The problem of soil salinization is prominent in northwest China.The optimization of soil salt sensitive variables and the inversion of machine learning modeling have important theoretical significance and application value for soil salt monitoring and ecological environment protection in arid areas.In this paper,the Juyanze area in the west of Inner Mongolia Autonomous Region was taken as the research area.Sentinel-2 and Landsat-8 optical imagery,Radarsat-2 Pol SAR imagery and SRTM DEM data,82 environmental variables of 6 categories were extracted,grouping variables and their combinations,the variable optimization strategy was used to screen the dominant sensitive variables of soil salt monitoring in each group,and the environmental variables with strong sensitivity and universality were identified in combination with previous research results.Secondly,based on the optimal variable combination scheme and machine learning method,a variety of soil salinity machine learning inversion models were constructed,and the accuracy of each model was compared and analyzed,and the inversion efficiency and comparative advantages of 6 types of salt indicator environment variables were evaluated,and the optimal parameter combination scheme and the best model for soil salt inversion by machine learning were confirmed.Finally,based on the best model,soil salinity inversion and salinity classification were realized in Juyanze area.Combined with the results of saline land classification,the spatial distribution pattern of soil salinity in the study area was discussed.The main conclusions are as follows:（1）Near infrared band(B_8a)and red band（B₄）,differential vegetation index（DVI）and canopy salinity response vegetation index（CRSI）,normalized salinity index（NDSIre1,NDSIre2）,spirochaeta scattering(_kh)and average scattering angle（α）,land surface temperature（LST）,topographic wetness index（TWI）and total catchment area（CA）are the salinity indicator variables with forward importance in the six variable types.Short-wave infrared band(₁₁),canopy response salinity index（CRSI）,salinity index II red-edge3（S2re3）,single scattering(_Odd),land surface temperature（LST）and total catchment area（CA）have strong sensitivity and universal applicability to soil salinity monitoring.The optimized salt sensitive variables are not identical with the most important salt indicator variables,which indicates the necessity of the optimization of environmental variables.（2）The effects of 6 types of environmental variables to soil salinity inversion ranking in descending order was topographic factor,polarization radar parameter,surface temperature,salinity index,vegetation index and band reflectance.The multi-variable combination can effectively improve the accuracy and stability of the model,but more variables in the inversion is not always better.The R² of RF model is generally higher than that of SVM,BP-ANN and RVM model,and RMSE is also lower than the other three models,indicating that RF model is the best model for soil salinity inversion in this study area.After comprehensive consideration,the RF model constructed by Scheme 11,which contains 6 types of variables and is rich in information,is the optimal model for soil salt inversion in the study area,and Scheme 11 is the optimal parameter combination scheme for soil salt inversion.（3）After optimization,the RF model inversion results of reflectance,vegetation index,salinity index,polarization radar parameter,surface temperature,topographic factor and optimal parameter combination scheme showed that the overall pattern of soil salinity distribution was consistent,and there were overestimates or underestimates of salinity in the vicinity of Swan Lake and Juyanze ancient Lake basin in the east and west.The salted land classification indicates that RF has better classification effect than Wishart,which means that the features of decomposed polarized targets can reflect the scattering mechanism among different targets,and provide rich feature information of ground objects,and furthermore facilitate the discrimination of various ground objects.According to the overall accuracy,Kappa coefficient and classification results,the RF model possession the optimal parameter combination and integration both spectral characteristic information and polarization scattering information has better classification effect.（4）Combine the results of soil salinity inversion and saline land classification,it can be seen that the salinity in the northeast and near Swan Lake is low,while with a high value in the southwestern ancient lake basin.mountainous and Gobi are located in the northeastern region and there for the salt content is low.There are many tidal flats near Swan Lake,and xerophytic vegetation grows intensively,resulting in low salt content.The southwestern region is the convergence center of water and salt,and after the lake water evaporates and dries up,the surface salt precipitates out in large amounts,which leads to high salt content in this region.In addition,the presence of low salt areas herein may be due to the influence of vegetation cover,Yardang landform and soil type.