Method for predicting soil salinity concentrations in croplands based on machine learning and remote sensing techniques

Abstract

To assess the potential of machine learning methods for predicting and mapping soil salinity in highly vegetated and slightly salt-affected croplands, a random forest (RF) regression model was constructed based on the measured soil salinity and spectral indices of remote sensing data, Landsat 8 Operational Land Imager (OLI), and Moderate Resolution Imaging Spectroradiometer (MODIS). The model was evaluated under different soil salinity conditions and with different resolutions of imagery data (Landsat 8 OLI and MODIS) to compare it with the multivariate adaptive regression splines (MARS) model. The results indicated that a higher accuracy estimation can be achieved from the RF model under conditions of poor correlations between the soil salinity and the spectral indices in the highly vegetated croplands, with a cross-validation determination coefficient ( R 2 ) = 0.86, root-mean-square error ( RMSE ) = 1.83, and ratio of performance to deviation ( RPD ) = 2.7; these were better than the MARS model ( R 2 = 0.81, RMSE = 4.8, and RPD = 1.96). The performance of the MARS model shows sensitivity to the variety of salinity levels. Estimated total salinity maps of Jiashi County were produced using the RF model based on OLI and Moderate Resolution Imaging Spectroradiometer (MODIS) data for September 2015 and April 2016, respectively. Acceptable mapping accuracy was achieved from the cross validation of the salinity mapping results for September 2015 ( R 2 = 0.84 and 0.81 for OLI and MODIS data, respectively) and April 2016 ( R 2 = 0.85 and 0.82 for OLI and MODIS data, respectively). The results indicated that the method developed is reliable and accurate for digital soil salinity mapping of slightly and moderately salt-affected croplands.