Método precoce para diagnóstico da doença “couro de sapo” da mandioca por infravermelho próximo

Abstract

The frogskin disease causes significant yield losses in cassava roots (Manihot esculenta Crantz) and in general, the diagnosis is carried out visually or using molecular techniques, usually diagnosed in old plants. The objective of this work was to analyze the feasibility of near infrared spectrometry (NIR) for early detection of the disease by using different calibration models. A NIR analysis was performed on 238 cassava accessions (120 healthy and 118 frogskin-infected). Six classification models were used: bayesian generalized linear model (BGLM), extreme learning machine (ELM), high dimensional discriminant analysis (HDDA), partial least squares (PLS), parallel random forest (PRANDF) and support vector machines with linear kernel (SVM). The predictive ability of the models was evaluated by the accuracy and Kappa index obtained by cross validation. The models are tested using all wavelength and wavelength selection by the recursive feature elimination (RFE) method based on random forest algorithm. The predictive models demonstrated high efficiency to distinguish healthy from frogskin-infected accessions (accuracy > 80%). Four models (SVM, BGLM, PRANDF, PLS) yielded almost perfect classification based on the Kappa index. The SVM and BGLM models presented high accuracy (99.07 and 98.92%, respectively) and reproducibility (0.98) to classify the cassava accession based on the frogskin symptoms. The selection of NIR wavelengths allowed the identification of high informative spectrum for building the classification models, but it reduced the classification performance of all models, except the ELM. The use of NIR is a viable alternative for the frogskin detection with inherent advantages of early and accurate detection, with high accuracy and low time-consuming compared to traditional methods of diagnosis.