Comparison of Decision Tree and Neural Network Methods in Predicting Soil Salinity in the West of Lake Urmia

Soil salinization is one of the most important soil degradation phenomena in arid and semi-arid regions. In recent years, indirect methods have been used to estimate soil salinity. For this purpose, 100 samples were taken from a depth of 0-30 cm around Lake Urmia and tested, and soil salinity was estimated using Landsat 8 satellite image indicators and digital elevation model. In order to model soil salinity, decision tree models and artificial neural network were used. Accordingly, the data were divided into educational duality (80%) and evaluation (20%). The results of evaluating the models based on the square root indices of error, mean error and coefficient of explanation showed that the decision tree model has the highest accuracy in predicting soil properties. The results of kappa coefficient and overall accuracy obtained from the two models showed that the decision tree model with having kappa coefficient percentage (56.56) and overall accuracy (73.46) had a greater agreement with the soil salinity of the region. In general, based on the obtained results, it was shown that CRSI and NDSI indices are the most important parameters for predicting soil salinity class and have the highest correlation with terrestrial data. Therefore, in the future studies, it is suggested to use tree models and CRSI and NDSI indices to prepare a digital soil salinity map.