Performance of random forest and buffer analysis of Sentinel-2 data for modelling soil salinity in the Lower-Cheliff plain (Algeria)

Modelling approaches are becoming an efficient tool in the forecasting of the salinization spread impact on the local and global scales. In the Lower-Cheliff plain, updating the information on soil salinity expansion is more than required as it continues to damage the agricultural environment in there. Through this study, we adopted an artificial methodology that consists of a combination of buffer analysis using Sentinel-2 data and machine learning modelling to assess their aptitude in the prediction and mapping of soil salinity. The adopted random forest (RF) algorithm included the reflectance information from of the bands from Blue (rho(Blue)), Green (rho(Green)), Red (rho(Red)), Near infrared (NIR), Vegetation red-edge (VRE) and Shortwave Infrared (SWIR), optimized with the geospatial buffering based on the 91 soil random samples collected during the summer of 2019 and measured for the Electrical Conductivity (EC) in the laboratory. The outputs from the geospatial buffering refined the goodness of the correlation between field data and the variables set from bands reflectance and selected salinity indices. The obtained coefficient of determination (R-2) with Multiple Linear Regression (MLR) and Partial Least Square regression (PLS) models proved an improvement in the multivariable prediction of soil salinity with the optimized Sentinel-2 data (R-2 = 0.61 and 0.68 + RMSE of 3.36 and 2.87 dS m(-1) respectively), compared to the results from the 2015 multiple regression model, Given the high values of R (2) = 0.77 and 0.95 and the low values of RMSE = 2.29 and 1.18 dS m(-1) respectively; the RF regressor was very efficient in predicting EC either with all variables or with selected variables from the variable importance measure based on the mean decrease accuracy (MDA). The random forest classifier performed top of the range classification with an overall Accuracy (OA) about 99%, estimating that over 75% of the study area surface is suffering from the salinization extent.

Automated summary

Modeling approaches, utilizing a combination of buffer analysis and machine learning with Sentinel-2 data, prove to be an efficient tool in forecasting and mapping soil salinity expansion. The Random Forest algorithm shows significant improvement in predicting soil salinity, with high accuracy in classification and estimation of the extent of salinization impact.