Classificação de níveis de severidade da mancha de Cercospora em milho por meio de dados hiperespectrais e aprendizagem de máquinas

Abstract

Visual diagnosis of leaf diseases can be limited by their etiology and interaction with other environmental factors, which make it difficult to detect their early occurrence in crops. The objective of this study was to identify and classify different levels of severity of cercospora leaf spot using machine learning models, searching for the best algorithm and inputs that guarantee better performance. Healthy leaves and three levels of severity of cercospora leaf spot on corn leaves were evaluated by means of a hyperspectral sensor using the FieldSpec 4 Jr spectroradiometer from Analytical Spectral Devices (Boulder, USA). From the spectral data, 28 spectral bands were extracted and calculated with spectral values and vegetation indices. These configurations were used as input for the machine learning models. After obtaining the spectral data and separating them into bands, the data were subjected to machine learning analysis using the algorithms Artificial Neural Networks (ANN), REPTree decision tree (DT) and J48, random forest (RF), support vector machine (SVM) and used as the traditional classification method logistic regression (RL). When comparing the effectiveness of the three inputs between the different models, it was observed that for ANN and RF, the spectral bands were the most effective input. For the DT, J48 and RL models, the best results were achieved with SB or WL. In the case of SVM, the model proved to be more efficient when using the full spectrum (WL). The application of RL and SVM, depending on the available spectral resources, is a practical and accurate approach in monitoring cercospora in corn.