PREDICTION AND CLASSIFICATION MODELS OF IRON DEFICIENCIES IN COMMON BEAN (Phaseolus vulgaris L.) LEAVES USING BAYESIAN REGULARIZED NEURAL NETWORKS AND CLASSIFICATION TREES

Abstract

The detection methods for iron (Fe) deficiencies in crops, such as common bean (Phaseolus vulgaris L.), are valuable decision-making tools that predict the nutritional status of plants in early stages. For this research, Bayesian regularized neural networks (BRNN) and classification trees were applied to predict iron deficiencies based on SPAD-502 readings that estimated the greenness index in common bean leaves. An experiment was carried out with eight treatments at different levels of Fe concentration in the nutrient solution (0, 20, 40, 60, 80, 100, 150 and 200 %). For seven weeks, the average green index measurements of the three leaflets of five replicates corresponding to the eight treatments were taken, and the collected data were used to adjust the statistical models mentioned above. With BRNN, the correlation between observed and predicted values was 0.77 for the training data set and 0.54 to 0.71 for the test data. In the case of classification trees, in the training stage, the percentage of correct classifications was 56.25 %, and when the validation procedure was carried out, it decreased almost 30 %. Thus for this type of research, the use of BRNN constitutes a valuable tool for the prediction of early deficiencies of Fe in common bean crop.