APRENDIZAGEM DE MÁQUINA PARA IDENTIFICAÇÃO DE PLANTAS DE SOJA SOB ATAQUE DE INSETOS USANDO DADOS HIPERESPECTRAIS

Abstract

The integration between the areas of remote sensing and machine learning has allowed an advance in the way of mapping agricultural fields and monitoring crops. This work investigates the ability of machine learning algorithms to classify soybean plants under insect attack, using reflectance spectroscopy measurements collected at the leaf level. To this end, tests were developed with different algorithms using a set of 991 spectral curves referring to healthy soybean plants under attack by pests, collected in eight consecutive days. These curves were measured by the EMBRAPA team, using a portable spectroradiometer, which records in the range of 350 to 2500 nm. Such curves were, initially, pre-processed to remove the regions of atmospheric absorption by water vapor, and then subdivided into a set of training, validation and testing of the machine learning algorithms. The Google Collabs interpreter was used and the algorithms were written in Python language, using libraries such as Skit Sklearn. Among the algorithms used, there are Random Forest, Decision Tree, Support Vector Machine, Logistic Regression and Extra-Tree. The Extra-tree has better performance (F1-score = 80.40%; precision = 81%; recall = 80%) in the proposed task. It is concluded that it is possible to process reflectance spectroscopy measurements with machine learning algorithms to monitor insect attack on soybean plants. It is recommended that the applied approach be tested in other cultures.