Prediction of the Calcium and Magnesium Content in Soils through a Generalized Regression Neural Networks and Genetic Algorithms

Abstract

Users need to discern how the soil characteristics at locations of their interest are, but soil properties can be determined only in a small number of sampling points. Therefore, it is necessary to predict how the soil is at points that have not been sampled. This study proposes a system for predicting soil property values, based on Generalized Regression Neural Networks and Genetic Algorithms. The Generalized Regression Neural Network is particularly useful when the amount of data is small, as is common in soil inventories. The proposed system calculates the mean square error, mean absolute error and the coefficient of determination as indicators of the prediction error. It also calculates the proportion of points which are generated unpredictably in a resulting map. This information helps the user to select the best combination of input variables and system parameters, according to their needs. The system allowed generating maps of calcium and magnesium concentrations in the soil, from a digital elevation model, satellite image and the values measured in a limited number of sampling points in a cross section of the Caramacate river basin (Aragua state, Venezuela). The selection of input variables to the network and the value of the smoothing parameter which is generated using a Genetic Algorithm, allowed to minimize the prediction error and the percentage of points rated. The results revealed that the selection of input variables to the network is crucial for the success of the prediction.