Abstract:
Eggplant is a popular crop in developing countries and is severely affected by leaf
diseases that influence the harvest and the quality. Proper and timely diagnosis is a
key to the sustainability of crop management. This paper explores how contemporary
deep learning models can be used to automatically label four typical eggplant leaf
states including Leaf Spot, Mosaic Virus, Insect Pest and Healthy. A comprehensive
dataset was created by merging 1,008 field-collected images with publicly available
images. The manually curated dataset contained 2,032 images, where duplicates, poor
images, and incorrect labels were removed and stratified by splitting (70% training,
20% validation, 10% testing) was presented. We compared three architectures: Vision
Transformer (ViT), YOLOv8n, and YOLOv11n using the same training protocols in
the local and the cloud. They are Precision, Recall, F1-score and mean Average
Precision (mAP). YOLOv8n was the most balanced and robust model to deploy in real
time, with an overall precision of 0.79, recall of 0.80, F1-score of 0.81, and mAP50 of
0.87. Compared to these, YOLOv11n had high recall rates (0.87) and mAP50 (0.86)
with marginally low precision (0.72) whereas ViT had high class-wise F1-scores but
poor object-level localization and low run-time performance. YOLOv8n was the most
optimal choice between detection and inference speed, suitable in real field practice.
This article shows that lightweight detection frameworks, such as YOLOv8n, can play
a crucial role in disease surveillance and diagnosis, which can facilitate early
intervention and support sustainable agriculture by using AI-based mobile and edge
deployment
