Abstract
Florida is the largest fresh tomato producer in the USA. Due to the subtropical climate, especially the hot and humid weather, the devastating tomato bacterial spot disease (TBS), caused by the bacteria Xanthomonas perforans, has been a threat to tomato crops for decades. Prevention through early disease detection is a crucial management strategy to minimise losses. Diagnostics for this disease that are commonly used are a visual diagnosis of disease symptoms or laboratory assays. These procedures are time-consuming, require specialised personnel, and are inefficient in detecting the disease in its early stages. Hyperspectral imaging provides a non-invasive tool for early disease detection in various crops. In this study, five optimised machine learning algorithms (MLAs), such as Linear Discriminant Analysis (LDA), Partial Least Squares Discriminant Analysis (PLSDA), Weighted K-Nearest Neighbours (W–KNN), Support Vector Machine (SVM), and Ensemble Boosting Tree (EBT), were utilised for early detection of TBS in transplant houses. Two normalisation procedures, a standardisation method, and unit vectorisation, were applied to preprocess the data in a dynamic range expansion process. The input variables to the models included time-lined events of disease progression and reduced datasets. To create the reduced datasets, a novel data reduction methodology was developed, which was able to identify 12 key (significant) wavelengths (from the 300 initial scans of the hyperspectral camera) that carry the highest weighted spectral components for accurately detecting TBS-affected plants. The PLSDA model presented the highest F1 score (90%) on early detection of TBS.
•Early detection of bacterial spot disease was achieved in tomato transplant houses.•Optimized machine learning algorithms and hyperspectral imagery were utilized.•A novel data reduction technique identified 12 significant wavelengths.•A low-cost optical device for monitoring transplant houses can be developed.
