Comparison of partial least square regression, support vector machine, and deep-learning techniques for estimating soil salinity from hyperspectral data

This study explored three techniques for estimating the soil salt content from Landsat data. First, the 127 items of in situ measured hyperspectral reflectance data were collected and resampled to the spectral resolution of the reflectance bands of Landsat 5 and Landsat 8, respectively. Second, 12 soil salt indices (SSI) summarized from previous literature were determined based on the simulated Landsat bands. Third, 127 measurement groups with Landsat bands and SSI were randomly divided into training (102) and testing subgroups (25). Three techniques including partial least square regression (PLSR), support vector machine (SVM), and deep learning (DL) were used to establish a soil salinity model using SSI and the simulated Landsat bands as independent variables (IV), respectively. Results indicated that PLSR with SSI performed best for both simulated Landsat 5 and Landsat 8 data. Compared with PLSR, SVM underestimated soil salt content, whereas DL obtained centralized simulations and failed to capture the lower and upper observations. We recommend the PLSR model with SSI as IV to estimate soil salt content because it can identify >66% moderate-to-high-saline soils, which indicates its great potential for soil salt monitoring in arid or semiarid regions. (c) 2018 Society of Photo-Optical Instrumentation Engineers (SPIE)