Optimizing CNN-Based Transfer Learning through Fine-Tuning and Adaptive Augmentation for Chili Plant Disease Detection

Abstract

Chili peppers (Capsicum annuum L.) are a strategic horticultural commodity in Indonesia, but their productivity is often hampered by pathogen infections that cause leaf diseases such as anthracnose, leaf spot, and yellow virus. Early detection by farmers is still dominated by subjective visual observation and prone to misdiagnosis due to the similarity of symptoms between diseases. Although Deep Learning technology through Convolutional Neural Networks (CNN) offers an automated solution, implementation in real-world conditions still faces significant challenges such as lighting variations, complex backgrounds, and limited local datasets. This often leads to a drastic decrease in model performance compared to testing in a controlled environment. To address these issues, this study proposes an optimization of the transfer learning strategy on the MobileNetV2 architecture by integrating progressive layer-wise fine-tuning and adaptive data augmentation techniques. The fine-tuning method is carried out gradually on the pre-trained model layers, while adaptive augmentation dynamically manipulates images based on environmental characteristics to improve model robustness. The results of this study, which include multi-class classification on cross-location image data, are projected to be able to boost the accuracy and generalization ability of the model in heterogeneous field conditions. Practically, this research provides a framework for a more precise and robust disease detection system to accelerate the implementation of precision agriculture in the future.

Plant phenotyping relevance

植物葉の病徴を画像から分類するCNN手法の改良が研究の中心であり、病害状態の表現型推定に該当する。転移学習のファインチューニングと適応的画像拡張、異なる圃場条件での頑健性評価を扱っている。

Code and data availability

The paper states its chili leaf image dataset was supplemented with data from a supporting repository, cited as a public Mendeley Data deposit (reference [2]). This is a public plant-image dataset directly used for the paper's disease-classification phenotyping. No authors' analysis code or trained model checkpoint is,