How Do Secondary Plant Health Symptoms Degrade Deep Learning Performance in Black Pepper Leaf Analysis?

Abstract

Black pepper is an important agricultural commodity in Sarawak and contributes to the livelihoods of local farming communities and the national economy. However, black pepper plants are susceptible to diseases and nutrient deficiencies that can substantially reduce plant growth and crop yield. Timely intervention is constrained by limited personnel and the logistical challenges associated with assessing remote farms. Machine-learning-based diagnostic systems offer a feasible means of addressing these constraints because of their effectiveness in identifying crop health conditions. Nevertheless, the presence of secondary disease or nutrient-deficiency symptoms in leaf images may complicate classification. This study established a diagnostic benchmark by evaluating six Convolutional Neural Network (CNN) models on datasets containing single-symptom and hybrid-symptom images. EfficientNet-B0 achieved the best overall classification performance, whereas ResNet-V2-50 was the most resilient to the introduction of hybrid images. The custom CNN also achieved performance comparable to that of the established architectures. Across the evaluated models, the inclusion of secondary symptoms generally reduced classification performance. These findings indicate a need for methods that mitigate the effects of secondary symptoms in leaf images to improve the reliability of machine-learning-based plant health diagnosis.