Estimation Of Soil Salinity In Arid Area Based On Multi-Source Remote Sensing

Due to the influence of natural conditions and human activities,secondary salinization has been intensified over the years,which brings a great challenge to the sustainable utilization of soils.In the arid area located at northwestern part of China,the problem of soil salinization is particularly severe,approximately 30%of the cultivated land was salinized.This study selected three fields with different land cover(bare soil,sparse vegetation,and dense vegetation)in Aksu,Xinjiang as our study area.Electromagnetic induction measurements and spatial interpolation of soil salinity were conducted to assess the degree of soil salinization and to explore the spatial distribution of surface soil salinity in the study area.In addition,UAV-borme hyperspectral and satellite-borne multispectral images were used to analyze the spectral characteristics of soil salinity.Plenty of salt-sensitive spectral indices have been extracted from narrow-band and broad?band spectra,their correlation relationship with soil salinity were presented via Pearson's correlation coefficients.Finally the quantitative prediction models of soil salinity were developed from remote sensed data,and the prediction maps of soil salinity were plotted afterwards.The results of this research could make a great contribution to the monitoring and control of soil salinization.The main research contents and goals of this research are as followed:(1)Spatial distribution and variation of soil salinity based on semi-variogram fittingThe results of this study indicate that the soil salinization problem in the study area is serious,and the average ECi:s detection values of the three sample areas are above 10 dS m-1.Among them,the salinization of the bare soil area is the most severe among all the three fields,and the ECi:s content of the soil samples collected in the field is between 20.25?54.90 dS m-1.Highly salinized land cannot be cultivated or utilized,causing damage to local agricultural activities and environmental health.Thus the treatment of salinized areas could no longer be delayed.Through the semi-variogram fitting,the study found that there is a strong spatial autocorrelation of soil salinity existed in the study area.With the increase of vegetation coverage in all the three fields,the autocorrelation range of soil salinity continues to expand.With the decrease of soil salinity in the sample area,the spatial variability of salt in the near distance becomes more intense,and the spatial correlation of soil salinity is continuously weakened.Using the fitting model of the semivariogram,the Ordinary Kriging(OK)interpolation and mapping of the soil salinity in the study area showed that the high salt value in the bare soil field was mainly concentrated at the northwest and southeast regions.While the high salt value in the dense vegetation field is mainly concentrated at the southwest and northeast regions.(2)The spectral characteristics of soil salinity based on UAV hyperspectral and satellite multispectral dataThe hyperspectral curves of different fields are obviously different from each other.The reflectivity of bare soil is relatively low,whereas the reflectivity of salt crust is relatively high and tends to increase with increasing wavelength.The red edge effect is obvious in the spectrum of halophyte.By extracting and comparing a series of different broad-band spectral indices and narrow-band spectral characteristics,it is found that the salt indices tend to have a higher correlation with soil salinity than the vegetation indices.For many narrow bands provided by hyperspectral data,the red and near-infrared bands are more correlated with soil salinity,and the spectral first derivative is more correlated than the original spectrum and soil salinity.In addition,normalized difference indices were calculated using all the possible 2-band combinations,and the correlations with soil salinity were analyzed.Finally three optimal narrow-band spectral indices were established to later predict the soil salinity of the study area.(3)Quantitative prediction models of soil salinity based on remote sensed dataUsing hyperspectral first derivative,narrow-band and broad-band spectral indices as independent variables,an attempt was made to establish a quantitative prediction model between remote sensing data and soil salinity using partial least squares regression(PLSR)and random forest regression(RFR).Lin's concordance correlation coefficient(CC),ratio of prediction to deviation(RPD)and root mean square error(RMSE)were used as evaluation indicators to measure the prediction accuracy of different models.The results show that compared with the partial least squares regression model,the random forest regression model can provide better prediction of soil salinity better use the spectral data to.For the regression model of the same area,the prediction accuracy of the model in the bare soil area is the highest,and the verification set CC of the random forest regression model is as high as 0.95 and the RPD is as high as 3.43.Based on the random forest regression model,the soil salinity in different areas was mapped.The results showed that the high salt value in the bare soil sample area was mainly concentrated in the north-western region,and the salt content in the western and north-eastern regions of the sparse vegetation sample area was higher,while in the dense vegetation field,the spectral inversion model of the region is characterized by the spectral interference of the halophyte,the obtained soil salt map is relatively fragmented,and does not reveal the distribution pattern of soil salinity well.