Image analysis and machine learning based methods for disease detection in soybeans

Plant phenotyping is important for genetic enhancements and plant biology research. There is a lot of work done to improve yield of crop plants, by selecting good genotypes to cross-breed in an effort to curb diseases or genetic deficiencies in these crops. In order to select these genotypes, one would have to perform phenotyping. Currently, plant phenotyping is based on visual assessment, where a breeder or researcher would have to visually inspect each plant and visually rate them. Visual rating is inefficient and can be inconsistent due to intra-rater repeatability or inter-rater reliability issues leading to incorrect visual scores. Not only that, it is also labor intensive and time consuming. Hence, there is a need to develop new tools amenable to high throughput phenotyping (HTP) for large scale plant genotype assessments. This requirement for high throughput phenotyping is applicable in a variety abiotic and biotic stresses. We developed a HTP framework which utilizes digital images in an effort for disease detection. This framework enabled us to accurately assign disease ratings to soybean plants that were affected by iron deficiency chlorosis (IDC). Utilizing image analysis techniques, we successfully extracted features pertaining to IDC and trained classification models on these features. A hierarchical classifier, based on linear discriminant analysis and support vector machine classifiers, produced the highest accuracy of 96%. Also, this framework was successfully implemented as a cellphone app. We envision to utilize hyperspectral imaging in the future for more accurate disease detection, prior to symptoms being visible.