Computational Intelligence Applied to Soil Quality Index Using GIS and Geostatistical Approaches in Semiarid Ecosystem

Abstract

The importance of soil quality is increasing every passing day for sustainable agriculture. In recent years, the investigation of the classification of soil quality with some classification methods known as machine learning algorithms draws attention. The study carried out for this purpose was hold on the farmland of Isparta University of Applied Sciences. Soil quality index was determined with a linear combination technique approach and analytical hierarchical process (observed values) and estimated by decision trees (predicted values). Total and minimum data sets (27 and 15 indicators, respectively) were evaluated by both methods, and all four outputs were compared. Deterministic (Inverse Distance Weighted-1, 2, 3 powers and radial based functions-completely regularized spline, spline with tension, multiquadric) and scholastic (spherical, exponential, Gaussian belonging to ordinary kriging, simple kriging and universal kriging) models were used in the creation of the distribution maps of observed and predicted values. No statistically significant differences were found in the comparison of soil quality index obtained using both data sets (P>0.05). In the decision tree where organic matter was determined as the root node, quality classes can be predicted at 91.1% by separating sand, wilting point, and EC properties into branches as an internal node. Area under the curve value in evaluating the estimation accuracy was found as 0.991, 0.960, and 0.943 for I, II, and III classes, respectively (P=0.00). It was determined that estimation can be done with 91.7% sensitivity and 90.9% specificity at 0.38 cut-off value for class III soils. Consequently, the highest accuracy in distribution maps of predicted and observed soil quality index values were found with the Gaussian semivariogram model of the ordinary and simple kriging for both data sets.