Computational intelligence applied to the least limiting water range to estimate soil water content using GIS and geostatistical approaches in alluvial lands*

The use of machine learning methods in pedotransfer functions has attracted considerable attention in recent years. These methods are fast and effective in solving complex events. The least limiting water range (LLWR) feature is very important in terms of water uptake by the plant and root development in agricultural production. In this study, the predictability of the LLWR feature was investigated with artificial neural networks, deep learning (DL) and the k-nearest neighbour (k-NN) algorithm from machine learning methods. Estimated values obtained from the model with the best estimation accuracy and observed values were evaluated through a geostatistical method from which their spatial distribution maps were created. In the present study, which was carried out on alluvial lands with different soil properties, the LLWR values of soils vary between 5.5% and 25.9%. Field capacity, bulk density, clay, organic matter, and lime content properties, which have a high correlation with the LLWR, were taken into consideration in the estimation methods. DL was determined as the best estimation method (mean absolute error [MAE]: 0.94%; root mean square error [RMSE]: 1.45%; coefficient of determination [R-2]: 0.93), and the worst was k-NN (MAE: 2.00%; RMSE: 2.55%; R-2: 0.77) for the LLWR. In addition, the LLWR can be estimated with high accuracy by using ReLU and softmax functions in the DL method. The study shows that distribution maps created with LLWR values obtained by observed data and the DL method have a very similar pattern.