Método automático para detecção de nematóides em lavoura cafeeira usando imagens aéreas

Abstract

Precision agriculture presents several challenges, amongst them the detection of diseases and pests in agricultural environments. This work describes an automatic method for the detection of nematodes pest in coffee crops using aerial images. Currently, aerial imagery collection combined with the use of machine learning (ML) based computational techniques have great potential in tasks involving pest detection in several crops, such as eucalyptus, soybean, among others. However, there is a scarcity of studies for the coffee culture. The proposed method uses two distinct strategies for the feature extraction and identification of the regions with the presence of nematodes. The first one based on blocks with the KNN, RF and SVM algorithms to classify the regions in pest and non-pest. The second one based on Convolutional Neural Networks (CNN) with state-of-the-art architectures U-Net and PSPNet to classify areas into healthy, pest and soil. The influence on height variation was also evaluated using the U-Net architecture. Results demonstrate the viability of the proposed method, with an average F-measure of 0.64 for the RF using 10-fold cross validation and an average F-measure of 0.69 for the U-Net architecture for the test set.