Soybean rust detection and disease severity classification by remote sensing

• Autor(a) principal: Negrisoli, Matheus Mereb [UNESP]
• Data de Publicação: 2022
• Outros Autores: Negrisoli, RaphaelMereb [UNESP], Silva, FlavioNunes da [UNESP], Lopes, Lucasda Silva [UNESP], Souza Junior, Franciscode Sales de [UNESP], Velini, Edivaldo Domingues [UNESP], Carbonari, Caio Antonio [UNESP], Rodrigues, Sergio Augusto [UNESP], Raetano, Carlos Gilberto [UNESP]
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: Repositório Institucional da UNESP
• Texto Completo: http://dx.doi.org/10.1002/agj2.21152; http://hdl.handle.net/11449/237864

The detection and monitoring of soybean rust (SBR) through remote sensing is promising because of the importance of the crop and the aspects of the disease. We evaluated the effects of different levels of SBR severity on soybean [Glycine max (L.) Merr.] leaflets reflectance aiming for the construction of a disease classification model. Leaflet reflectance was evaluated on two cultivars (susceptible and partially resistant) at four disease severity levels: healthy, low, moderate, and high. Leaflets were collected in the field and taken to the laboratory for spectral evaluation through the spectrophotometer UV 2700 coupled with Integrating Sphere Attachment ISR-603, in the range of 270-1000 nm. The feasibility of using a collection of vegetation indices (VIs) and data dimensionality reduction through multiple factor analysis (MFA) was evaluated, and a classification model was constructed. Ten algorithms were assessed based on precision, sensibility, and accuracy parameters, using 80% of the dataset as training data and 20% as testing dataset. The visible range and red edge region contributed more significantly to the disease prediction and classification model. The MFA performed satisfactorily in the dimensionality reduction and unveiled the effect of specific wavelengths on the classification of each class. Most of the VIs studied had high correlation performance across the severity classes. Classification accuracy and precision were >70% for all models. Linear support vector machine with the collection of VIs achieved the best results. This study provides a practical path for developing a detection model to be integrated into SBR management programs.