MACHINE LEARNING PARA DETECÇÃO DE Pratylenchus brachyurus NA CULTURA DA SOJA UTILIZANDO IMAGENS MSI/SENTINEL-2

Abstract

Soybean is the most important crop for agribusiness. Nematodes can threat soybean yield and worldwide economy. Remote sensing is a science with many technological tools used to prevent yield losses by nematodes. Monitoring and remote sensing classifications have potential for better decision-making. Classifiers based on machine learning are increasingly used as alternatives to traditional image classification techniques. This study aimed to evaluate the potential of machine learning algorithms for temporal classification of orbital images obtained by the multispectral imager (MSI) coupled to Sentinel-2 satellite. Vegetation indexes were used to predict clusters of plants infected by the nematode Pratylenchus brachyurus in soybean. For validating the occurrence of these phytoparasites in the field, an exploratory flight was conducted using a quadricopter drone to identify possible clusters of infected plants. Five zones showing visible symptoms of nematode-infested plants were identified. Soil and plant roots were collected for laboratory analysis. Orbital images were chosen throughout the crop cycle from October 2018 to January 2029. The data were divided into four classification models using the Random Forest classifier. The classification accuracy was determined by a 10-fold cross validation, global accuracy, and Kappa index. An importance metric was calculated using the gini index for the variables used on models. The best model for detection of P. brachyurus used all available predictor variables. The vegetation index GLI was the most sensitive to identify clusters of nematode-infested plants. Thus, the Random Forest classifier presented satisfactory results and can be used as an alternative to multitemporal analysis to predict the occurrence of nematodes in soybean crops.