In recent years, indirect methods have been used to estimate soil salinity in agricultural lands. In this research, the electrical conductivity of 93 soil samples from 0 to 30 cm and 0 to 100 cm was measured using the hypercube technique at Sharifabad-Saveh Plain, Iran. Land area parameters such as TWI, TCI, STP, DEM, and LS were used as topographic variables and spatial indices of salinity and vegetation were derived from Landsat 8 images. Soil salinity off crops and gardens was determined at 0-30 cm and 0-100 cm. The data were divided into two series: the training set (70%) and the test set (30%). In order to model and predict salinity, models such as an artificial neural network (ANN), integration of neural network and genetic algorithm (ANN-GA), PLSR, and decision tree (DT) were used. The results of the models' evaluation based on MSE and R2 indices showed that the ANN-GA model has the highest accuracy in predicting soil properties. This model improved the accuracy of soil salinity prediction by 28%, 42%, and 23% in 0-30 cm and by 20%, 28%, and 25% at 100 cm than ANN, PLSR, and DT. The result showed the 2 dS/m EC at alfalfa and cucurbits farmlands while pistachio orchards have low salinity and bare lands have moderate and high salinity.
Keywords: Data mining; Machine learning; Pedometrics; Satellite image; Soil classification map.