Inversion Model Of Soil Salt Content In Hetao Irrigation Area Based On Sentinel Series Satellite Remote Sensing Data

Soil salinization is one of the major factors contributing to global land degradation and seriously affecting sustainable agricultural development.Accurate and stable monitoring of soil salinity can contribute to the management and restoration of salinized soils.In order to evaluate the potential of multispectral satellite remote sensing data and radar satellite remote sensing data to synergistically estimate soil salinity and create more knowledge about the ability of remote sensing to monitor soil salinity,this paper took the irrigation area of Sha Trench Canal in the river-loop irrigation area as the study area and collected samples of surface soil salinity from April to August,respectively,and simultaneously acquired Sentinel-1 radar remote sensing images and Sentinel-2 multispectral remote sensing images that matched the time of soil sample collection.The image texture features of bare soil and vegetation periods were extracted using different window sizes,and the optimal window sizes of image texture windows in different periods were analyzed,and a monitoring model of soil salinity was constructed with texture features of optimal texture window sizes to evaluate the effectiveness of image texture features in monitoring soil salinity.Based on this,the image texture features and spectral/radar variables with the best window size sensitive to soil salinity were obtained by using the variable selection method,and the soil salinity monitoring models with different machine learning algorithms for bare soil and vegetation periods were constructed,and finally the model accuracy was evaluated using the model evaluation index to obtain the best inversion model and spatial distribution status of soil salinity.The main findings obtained in this paper are:(1)The influence of image texture on soil salinity monitoring in different periods was explored,and the optimal window size for image texture features was extracted.Among them,the RF importance metric shows that HOM,ENT,COR and CON are important for S-1 in the bare soil period;while VAR,MEAN,ENT and COR are important features for monitoring soil salinity in S-2 in the bare soil period;CON,HOM,DIS and ENT are important features for monitoring soil salinity in S-1 in the vegetation period;VAR,CON,HOM and ENT are more important for monitoring soil salinity in S-2 of vegetation period.Among the eight textures in different periods of different data sources,ENT always ranked the top four in importance,so ENT is a stable texture feature under different periods of different data sources.The RF model constructed from the texture features of vegetation period S-2 was the most effective in monitoring soil salinity with R~2 of 0.399 and RMSE of0.136.(2)A model based on Sentinel-1 image texture and Sentinel-2 salinity spectral index was constructed to estimate the salinity of bare soil in collaboration,and the spatial distribution of soil salinity in the study area was mapped.The prediction performance of multi-source sensors were all better than that of single-source sensors(e.g.,R~2 of strategy III in the RF model was 0.05 and 0.24 higher than that of strategies I and II,respectively;RMSE was 0.015%and 0.027%lower;MAE was 0.01%and 0.045%lower,respectively).Texture features were the main explanatory variables for predicting SSC,followed by salinity spectral index and remote sensing data.CON was the most important predictor variable affecting the spatial distribution of soil salinity.The RF model outperformed the SVM and ELM models in predicting SSC.More importantly,these three machine learning models obtained higher accuracy under low salinity conditions in the present study area.These three models can perform better in the low salinity region.The RF model based on Strategy III obtained the SSC inversion map of the study area in April,and the soil salinity was relatively low in the southern part of the study area.Salinized soils were mainly concentrated in the northwest to central part of the study area.(3)A model for soil salinity monitoring under vegetation cover conditions was constructed in collaboration with Sentinel-1 improved polarization combination index and Sentinel-2 multispectral image texture.The correlation between the radar backscatter coefficients and their polarization combination indices with soil salinity after the removal of vegetation effects by the water cloud model was improved to some extent.Comparing the performance of different variable selection methods with different machine learning methods coupled models for inversion of soil salinity,the OOB variable screening method with three machine learning methods,RF,ELM and Cubist,had the best accuracy of coupled models,with R~2 above 0.75 for both modeling and validation sets,and the smallest RMSE and MAE for the validation set.Comparing the performance of the coupled model of OOB variable screening method with three machine learning methods,RF,ELM and Cubist,the OOB-Cubist coupled model has the highest accuracy and the R_v~2/R_c~2 is 0.955,and the coupled model has good robustness.