Application of the GMDH Neural Network in Monitoring Soil Salinity of Saveh Plain using Spectral and Topographic Indicators

In recent years, indirect methods such as remote sensing and data mining have been used to estimate soil salinity. In this research, the electrical conductivity of 94 soil samples from 0 to 100 cm was measured using the Hypercube technique in the Saveh plain. 23 types of input data were used in the form of topographic and spectral categories. Land area parameters such as the Topographic Wetness Index (TWI), Terrain Classification Index (TCI), Stream Power Index (STP), Digital Elevation Model (DEM), and Length of Slope (LS) were considered as topographic inputs using Arc-GIS and SAGA software. Also, salinity spatial and vegetation indices were extracted from Landsat 8 images and were considered spectral inputs. The GMDH neural network was used to model salinity with a ratio of 70% for training and 30% for validation. The results showed that the soil salinity values were between 0.1 and 18 with mean and standard deviation of 5 and 4.7 dS/m, respectively. Also, the results of modeling indicated that the statistical parameters R2, MBE, and NRMSE in the training step were 0.80, 0.06, and 42.1%, respectively. The same values in the validation step were 0.79, 0.13, and 48.7%, respectively. Therefore, the application of spectral, topographic, and GMDH neural network indices for modeling soil salinity is effective.