Spectral Characteristics Of Salt-affected Farmland Soil And Its Remote Sensing Inversion And Classification

Soil salinization and irrigation-induced soil secondary salinization are the main obstacle and restricting factorsfor agricultural development in arid and semi-arid. These problems also affect the stability of oasis agro-ecologicalenvironment. Xinjiang is one of the concentrated distribution areas of salt-affected soils in China. In this study, theresearch material was farmland soil in typical saline area in Changji Hui Autonomous Prefecture of Xinjiang. Aftersampling and determining of soil spectra in field, we analyzed the relationship between soil salinity and soil spectralcurve characteristics. Using continuum-removed and first derivate methods, the sensitive bands which reflected soilsainlization level were selected. Hyperspectral quantitative prediction model was constructed. Based on TM images,and remote sensing quantitative retrieval and decision tree classification methods, classification of soil salinization inthe area was studied. The distribution characteristics of soil salinization were explored in Changji Hui AutonomousPrefecture. We also analyzed the causes of soil salinezation and proposed an amelioration and utilization model. Mainconclusions are listed below,(1)There were four obvious water vapor absorption bands of soil salinization in wavelength range, i.e.,1350to1730nm,1800to2100nm,2130to2270nm and2300to2500nm.(2)Generally the saline soil spectral curves changed gently, spectral characteristics were similar in morphologyand were parallel substantially. There was a certain negative correlation between spectral characteristics and soilsalinity in the visible light band350~460nm. However, a positive correlation could be found in the near-infraredbands1800to around1901nm, which exhibited a trend that the higher the salt content was, the higher the reflectivity.(3) Through correlation analysis among normalized spectrum, the first derivative spectra and soil salinity, wefound that the sensitive bands which reflected soil salinization level involved347-377nm and1800-1911nm, whenusing continuum-removed and first derivate methods, The spectrum prediction models of normalized and the firstderivative were established, i.e., y=316.485x1865-279.946and y=11.503-1239.249x360+726.153x375. Modelexamination showed that the first derivative model, with an average relative error of11.9%, was better than thecontinuum-removed modle.(4) The correlation coefficients of each band about TM images in the study area were statistical analyzed.Significant correlation relationships were found among the bands1,2, and3. Based on the OIF method and thecorrelation coefficients between each bands, we found that the combination of band1and band3of TM images hadthe highest correlation coefficient with soil salinity.(5)Using TM images and combined with a large number of field-measured data, the classification accuracies ofremote sensing quantitative retrieval and decision tree were compared. Soil salinization of farmland in the study areawas classified. There were68%non-salinization cultivated land in the study area, middle level of salinization ofcultivated land accounted for21.36%, and the moderate and severe level of salinization of cultivated land were8.17%and5.79%of the total arable land, respectively. In this study, due to the impact of the accuracy of remote sensingimages and the number of band, inversion of soil salinity classification accuracy was81.5%, Kappa coefficient was78.8%.