Evaluation of soil quality of cultivated lands with classification and regression-based machine learning algorithms optimization under humid environmental condition

Abstract

In soil science, machine learning algorithms are preferred for pedotransfer functions due to their rapid data acquisition and high prediction accuracy. The current study aims to evaluate the prediction of soil quality in agricultural lands dominated by the humid Black Sea climate using various algorithms. Both classification and regression-based algorithms (Random Forest-RF, Light Gradient Boosting-LGB, Extreme Gradient Boosting-XGBoost, k-nearest neighbors-kNN, Logistic Regression, multilayer perceptron-MLP, Linear Regression-LR and Bayesian Ridge- BR) were used in the method. The comparison of soil maps is also included. Furthermore, the present study evaluates the Grid Search optimization method with K-Fold Cross Validation (K = 5) for both classification and regression-based algorithms. The prediction of soil quality was performed using class-based and regression-based algorithms. As a result of the study, the RF and XGBoost algorithms achieved an approximate accuracy rate of 92 % in the class-based prediction. In regression-based predictions, the most successful algorithms were BR and LR, with an R² Score of 0.84. The Grid Search optimization method was used to improve the R² Score, resulting in an increase to 0.90 and 0.88 for BR and LR, respectively. The optimized hyperparameters showed improved performance in predicting the soil quality index. The present study found that Gaussian and Spherical models had the lowest prediction errors in spatial distribution maps. Tree-based algorithms were found to be suitable for class-based prediction of soil quality, while the linear regression method was appropriate for regression predictions. This study is characterized by a rainy climate resulting in acidic soils with high organic matter content. Planning of new studies in different climates and soil properties is recommended.