A new proposal for automatic identification of multiple soybean diseases

This work proposes a combination of digital image processing techniques consisting of color moments, local binary patterns (LBP) and bag of visual words (BoVW) model for automatic detection of soybean diseases based on the analysis of color, texture and local characteristics of spots on affected leaves. The characteristics extracted after data collection and the application of the cited techniques are grouped and used as input to the support vector machine (SVM) for the purpose of disease classification. The images used in the development and validation of the proposed identification system were obtained from Digipathos, a database provided by the Brazilian Agricultural Research Agency (Embrapa), consisting of 354 images related to 8 typical soybean diseases, namely, bacterial blight, soybean rust, copper phytotoxicity, soybean mosaic, target spot, downy mildew, powdery mildew and Septoria brown spot. Performance analyses are presented separately for the techniques used in several stages of data processing and for the overall system consisting of the combination of these techniques and the SVM classifier. The proposed system achieves an success rate of 75.8%, representing an improvement of approximately 17% over results of methods in the existing literature.