Digital mapping of soil texture components in the Sirjan region using machine learning models

This study aims to evaluate and compare the efficiency of Artificial Neural Network (ANN), Regression Tree (RT) and Neuro-Fuzzy (ANFIS) models using a digital soil mapping framework to predict soil texture in a part of Sirjan province. Sampling was carried out at 84 observation points with a regular grid of 2x2 km, and soil texture components were determined from the soil surface depth of 0 to 30 cm. Auxiliary variables included primary and secondary derivatives of the digital elevation model (DEM), a geomorphological map and remote sensing (RS) spectral indices. The appropriate variables selected using the Principal Component Analysis (PCA) feature selection method. Based on PCA, eight topographic variables and six vegetation indices and spectra from RS selected to predict soil texture components (sand, silt and clay). The efficiency of the models was evaluated using coefficient of determination (R2), mean error (ME), root mean square error (RMSE) and normalised root mean square error (nRMSE). The RMSE values in the neuro-fuzzy model compared with the regression tree model. The results of the neuro-fuzzy model were 1.43% for clay, 1.98% for sand and 2.1% for silt, which were 4.32%, 5% and 4.54% lower respectively compared to the regression tree model. The results of this study showed that the ANFIS model was more accurate in predicting clay, silt and sand compared to ANN and RT. Also, the geomorphology map, topographic wetness index, multi-resolution valley bottumn flatness index and Landsat 8 bands 5 and 6 had the highest relative importance in predicting soil texture components.