CLASSIFICATION OF ALFALFA (Medicago sativa L.) PHENOLOGY USING MACHINE LEARNING METHODS

Abstract

Alfalfa (Medicago sativa L.) is an important crop for food security and livestock sustainability. The accurate identification of its phenological stages, based on subjective observations, can be improved using machine learning methods that enable objective and efficient classification based on field data and remote sensors. The purpose of this study was to classify four phenological stages of alfalfa using Sentinel-2 images and machine learning models such as Support Vector Machine (SVM) and Multilayer Perceptron (MLP) neural networks. To this end, a dimensionality reduction process based on correlation analysis and Sequential Forward Selection (SFS) of features was integrated to optimize accuracy and computational efficiency. In this study, 41 Sentinel-2 images corresponding to 72 alfalfa plots during one agricultural cycle were analyzed. From the images, 86 texture, color, and vegetation index characteristics were extracted; subsequently, a correlation analysis was applied to eliminate redundant variables, reducing the set to 50 independent characteristics. On this subset, the SFS method was implemented with a gradient stopping criterion, which allowed the identification of the 29 variables with the highest discriminating power. As a result, the SVM model improved its accuracy from 70.1 to 82.2 % after the reduction of characteristics, while the MLP network achieved the highest overall accuracy (85 %, k = 0.77) with a configuration of 50-50 neurons in two hidden layers. The combination of correlation analysis and feature selection reduced dimensionality by 58 % without loss of accuracy. The MLP network outperformed the SVM in its generalization ability. This approach constitutes a low-cost operational alternative for the phenological monitoring of perennial crops using freely available satellite imagery.