Remote Sensing Inversion Of Soil Salinity In Cotton Field Of Typical Oasis Irrigation Area In Southern Xinjiang

In order to explore the effective method of extracting soil salinity quickly in oasis irrigation area,and provide reference for soil salinization control.Based on Landsat 8 OLI multispectral images and measured soil salinity content in spring,summer and autumn of 2019 and 2021,the remote sensing inversion model of soil salinity content was constructed by Multiple stepwise regression,Partial least squares regression,Extreme learning machine,Support vector machine and Back propagation neural network.The effects of input variables and modeling methods on the model accuracy were explored.The best inversion model of soil salinity in each season was determined by comparison,and the corresponding seasonal surface soil salinity content was inverted quantitatively.The spatial distribution,seasonal variation and inter-annual variation characteristics of soil salinity content in the study area were analyzed according to the inversion results.The main conclusions are as follows:(1)Descriptive statistics were made on soil salinity at sampling sites.It was found that cotton soil in2019 and 2021 was mainly non-salinized soil and mildly salinized soil.The variation coefficients of soil salinity in spring,summer and autumn were 0.67,0.56 and 0.67,showing moderate variability.The spectral reflectance of moderately salinized soil is higher than that of lightly salinized soil and non-salinized soil.(2)Through the correlation analysis of soil salinity with the bands and spectral indices of multi-spectral images,it was found that the salinity indices of coastal band(b1),blue wave band(b2),green band(b3),red band(b4)and SI1,SI2,SI3,SI4,S3,S4,S5 passed the significance test P<0.01 in spring;In summer,except for SI4 salinity index,the other 32 spectral parameters were significantly correlated with soil salinity(P<0.01),vegetation indices was negatively correlated with soil salinity content.Soil salinity indices of b1,b2,b3,b4 and NDSI,SI-T,SI1,SI3,S1,S2,S3,S4 and S5 in autumn were significantly correlated with soil salinity(P<0.01).(3)After comprehensive analysis of modeling and verification effects,it was found that the BP neural network models based on the full variable group were the optimal soil salinity inversion models in all seasons.In spring,the modeling set coefficient of determination(R_c~2)and the root mean square error(RMSE_c)is0.705 and 1.133 respectively.The verification set coefficient of determination(R_v~2)and the root mean square error(RMSE_v)of validation set is 0.556 and 1.409 respectively.In summer,R_c~2 is 0.830,RMSE_c is 1.128,R_v~2 is 0.767,RMSE_vis 1.552.In autumn,R_c~2 is 0.634,RMSE_c is 1.379,R_v~2 is 0.527,RMSE_v is 1.742.(4)According to the inversion results,the soil of cotton field in spring,summer and autumn of 2019 and2021 is mainly non-salinized soil and mildly salinized soil.The seasonal variation process of soil salinity content was similar in the two years.The soil salinity content decreased from spring to summer,and increased from summer to autumn,and the soil salinity content was autumn>spring>summer.The interannual variation process of soil salinity content in each season was similar,and soil salinity content in all seasons were reduced from 2019 to 2021.