干旱区生态环境知识资源中心

Soil salinization, which is one of the most important land degradation problems in arid and semiarid regions, has a significant impact on ecological equilibrium. Hyperspectral remote sensing, with a large number of measured wavelength bands and a high resolution, has gradually become a popular technology to investigate soil salinization. In this study, a model based on field-derived spectra was developed for soil salinity and the quantitative relationships between the soil spectrum and vegetation spectrum with the soil salt content (SSC) and soil electrical conductivity (EC). A field study was performed in Minqin County, China. A genetic algorithm (GA), partial least squares regression (PLS), and back-propagation neural network (BPNN) were used for modeling. The results showed that GA has a relatively strong ability for band selection. After the selection, the predictive ability of the GA-PLS model was better than the PLS model based on the full spectra. The BPNN model built by selected bands (GA-BP) was superior to the GA-PLS linear model. The models built using the soil spectrum after band selection have a high predictive ability. The R-2 and ratio of prediction to deviation (RPD) for SSC were 0.68 and 1.76 for GA-PLS and 0.72 and 1.89 for GA-BP, respectively. There were no significant correlational relationships between the normalized difference vegetation index and SSC or EC. The GA-BP model fitted using the vegetation spectrum was superior to a single vegetation index model, whereas the predictive ability of SSC (0.56 for R-2 and 1.47 for RPD) was not high due to influences such as plant species differences and vegetation coverage.