Detecting bacterial pustules on soybean plants by hyperspectral imaging

Bacterial pustules are a major threat to soybean cultivation but are difficult to detect early because they manifest on the bottoms of leaves. In this study, hyperspectral imaging was applied to detect bacterial pustules on soybean plants. Images were preprocessed, and representative central wavelengths (i.e., bands) were identified through a two-sample t-test to calculate vegetation indices (VIs) for non-inoculated (i.e., control) and inoculated (i.e., treated) groups. Three machine-learning models were applied to classify infected soybean plants based on the VIs: partial least-squares discriminant analysis, support vector machine, and random forest (RF). The best classification performance was achieved by the RF model using five VIs with an overall accuracy (OA) of 0.89 and kappa coefficient (KC) of 0.77. The RF model also achieved an OA of 0.77 and KC of 0.55 when tested on a dataset before the expression of symptoms. The results of this study can potentially be applied to developing a multispectral image sensor that can be mounted on various platforms for the early detection of bacterial pustules on soybean crops.