Modeling Soil Salinity in Khuzestan Lands Susceptible for Dust Production Using Visible-Near Infrared Spectroscopic Method

A broad area of saline and semi-saline lands of Khuzestan province have changed into centers susceptible to dust production due to eroded wind and lack of surface coating and low soil resistance. The objective of this study was to model the soil salinity of sensitive areas to dust production in Khuzestan Provenience usin spectrometry method of visible and near-infrared wavelengths (2500-350 nm). The least square multivariate regression model, artificial neural network and random forest model were used to estimate soil salinity. The main soil spectrum was determined using the FieldSpect machine. Also, preprocessing methods including Savitzky-Golay filter, the first derivative with the Savitzky-Golay filter (FD-SG), the second derivative with the Savitzky-Golay filter (SD-SG), the standard normalization method (SNV), and the continuum remove method (CR) were used to eliminate the noise and to increase the accuracy of the multivariate model. The results showed that the combined model partial least squares-artificial neural network model with assessment criteria (RPDcal = 3.40-2.65) has high accuracy for salinity estimation. In contrast, the combined model of least squares - random forest showed the lowest accuracy (RPDcal = 0.85-1.98). Preprocess of the main spectrum in two models (neural network and partial least squares regression) increased the relative accuracy of the model; while in the random forest model, preprocess reduced the accuracy of the model compared to the main spectrum. The ranges of 1800, 1900, 2000, 2300 and 1500 nm were recognized as "the key wavelengths" impressed by soil salinity. The key wavelengths can be used in remote sensing studies and mapping of soil salinity in areas sensitive to dust production in Khuzestan province.