Interpolation and artificial neural network to estimate soil spatial variability affected by land use and altitude

Abstract

Finding the most suitable methods, which may predict soil spatial variability is essential for proper handling of agricultural lands affected by land use types and altitude. There is not much data on the use of artificial neural network (ANN) outperforming the traditional methods such as the interpolation methods for predicting soil spatial variability. Accordingly, the interpolation methods of inverse distance weighting (IDW), kriging, and co-kriging, as well as ANN were tested to predict soil spatial variability of pH, salinity (EC), and cation exchange capacity (CEC) affected by land use type (cultivated and uncultivated lands, orchard, forestry and rangeland) and altitude (-20-0 (A1), 0–100 (A2), 100–500 (A3), and >500 m (A4)) in a 9545 km² research area. The chemical properties of the 249 soil samples (0–15 cm) were determined. Land use type indicated pH of 6.56 (forestry) to 7.32 (cultivated land), EC of 1.10 (forestry) to 2.87 dS/m (rangeland), and CEC of 17.71 (uncultivated land) to 37.01 meq/100 g soil (forestry). Altitude resulted in pH of 6.72 (A4) to 7.35 dS/m (A2), EC of 1.31 (A4) to 1.90 (A2) dS/m, and CEC of 20.07 (A1) to 34.45 meq/100 g soil. Although cross-validation method (using mean error (ME) and root means square error (RMSE)) indicated the accuracy of interpolation methods to predict soil spatial variability, ANN was the most suitable one. The proper training of ANN may precisely predict the spatial heterogeneity of soil chemical properties affected by land use type and altitude, useful for the appropriate handling of agricultural lands.