Early Detection of Chili Leaf Diseases in AI-Based Agroforestry Systems Using Convolutional Neural Networks

Abstract

Early detection of chili plant diseases is essential for preventing yield loss, yet practical deployment remains challenging due to inconsistent illumination, leaf orientation variability, and the limited computational capacity of low-cost imaging hardware. This study proposes an integrated detection framework combining an ESP32-CAM acquisition pipeline, MQTT-based transmission, and a Python inference engine running a fine-tuned ResNet-18 model optimized for real-world noise conditions. The research aims to determine whether domain-aligned fine-tuning meaningfully improves generalization performance compared to older non-optimized models under field-like variability. Using a four-day observational design with two leaf subsets, the fine-tuned models consistently outperformed their non-fine-tuned counterparts in overall accuracy, per-class stability, and positional robustness. Real-time deployment using the Telegram Bot API successfully delivered classification results and images with low latency, demonstrating operational feasibility for remote plant health monitoring. These findings indicate that targeted fine-tuning is essential for transforming CNN-based classifiers from laboratory prototypes into stable, field-ready systems capable of supporting early disease detection in resource-constrained agricultural environments. Additionally, due to chili is widely cultivated in agroforestry systems in Indonesia, the proposed early disease detection framework offers substantial benefits for maintaining productivity in heterogeneous microclimatic conditions where manual diagnosis is more difficult.