Tomato Disease Detection Using Multispectral Imaging with Deep Learning Models

Abstract

Identifying plant diseases plays a pivotal role in managing plant health. It poses significant challenges due to the diverse manifestations of diseases. This study investigates the impact of environmental conditions and imaging factors on analysing multispectral images captured through six filters spanning various portions of the Near Infrared spectrum for tomato disease identification. Two distinct datasets were compiled for analysis. Dataset 1 comprised uniform images across all filters. Dataset 2 incorporated variations in image capture to explore the influence of environmental and imaging factors on filter performance. Classification between healthy and diseased states was conducted on both datasets utilising the popular Convolutional Neural Network, Vision Transformer, Hybrid Vision Transformer, and Swin Transformer models. Among all the filters tested, K590 demonstrated the highest average accuracy, reaching 88.69 % for Dataset 1 and 93.31 % for Dataset 2. Furthermore, this filter consistently outperformed others in terms of precision and recall across both datasets. ViT-BI6 emerged as the most effective model across all evaluation metrics, with an average accuracy of 89.92%. Furthermore, comparisons were drawn with prior literature, encompassing balanced and unbalanced datasets for tomato disease classification tasks. The findings indicated that environmental and imaging factors do not significantly influence disease classification using multispectral imaging.