Dust source clay content and salinity estimation using VNIR spectrometry

The spectral behavior of soil will change through degradation, which makes it difficult to retrieve soil properties using previously developed models. This study aims to use linear [including partial least squares regression (PLSR) and ratio soil index (RSI)] and nonlinear [including partial least squares-backpropagation neural network (PLS-BPNN) and partial least squares-random forest (PLS-RF)] models to estimate soil electrical conductivity (EC) and clay content in dust sources. For this purpose, 142 soil samples were collected in Khuzestan province. After laboratory spectroscopic analysis, the area and depth of diagnostic absorption features (AFs) of continuum removed (CR) spectra were calculated. The results revealed that with increasing clay content, the depth of AFs at 1400, 1900, and 2200 nm will increase. Meanwhile, an increase in the soil salinity will increase the depth and area of AFs in 1450 and 1915 nm and decrease the depth and area of AF in 2200 nm. Spectral ranges of 2100-2300 and 1400-1600 nm were identified as the most important portions of the visible and near-infrared spectrum for analyzing clay content and EC, respectively. The RSI method performed poorly in soil salinity and clay content estimation. PLSR and PLS-RF methods overestimated clay content and salinity in low values. The PLS-BPNN model had the best performance for estimating clay content (RPIQ = 4.51) and EC (RPIQ = 4.76). Considering the expected non-linear relationship between soil properties and corresponding spectral reflectance, the results of this study were acceptable.