Potential of multispectral and hyperspectral data to detect saline-exposed soils in Brazil

Abstract

Irrigation-induced soil salinization is an important land degradation process in northeastern Brazil. We used multispectral and hyperspectral sensors to detect saline-exposed soils in an area cultivated with irrigated rice. Spectral mixture analysis (SMA) was applied to Operational Land Imager (OLI)/Landsat-8 data to identify exposed soils. By measuring the electrical conductivity (EC) of soil samples from 53 sites, we classified them into saline and non-saline. The surface reflectance Thematic Mapper /Landsat-5 product was used to inspect the normalized difference vegetation index (NDVI) variations over time (1984–2011) at the sites. Using OLI/Landsat-8 and Hyperion/Earth Observing One, we obtained five salinity indices and scores from principal component analysis applied to exposed soil pixels. These indices along with the first principal component (PC1) were regressed against EC to estimate soil salinization. Different metrics and support vector machine (SVM) were tested to discriminate saline and non-saline soils. The results showed that exposed soils detected by SMA had NDVI with a lower mean and standard deviation over time in the saline areas due to vegetation growth limitation. NaCl absorption bands were not observed in the Hyperion spectra due to atmospheric water vapor. Therefore, soil salinity detected by OLI or Hyperion was due to soil brightness rather than absorption bands. Because most salinity indices and scores expressed brightness to some extent, they were correlated with EC, especially the Salinity Index and PC1. However, compared with OLI, the narrow-band salinity indices of Hyperion produced a lower root mean square error for EC estimates, better discrimination between saline and non-saline soils using the Euclidean distance and spectral angle metrics, and higher SVM classification accuracy.