A Hybrid Compact Convolutional Transformer with Bilateral Filtering for Coffee Berry Disease Classification

Area of Science:

• Agricultural Science
• Plant Pathology
• Computer Vision
• Deep Learning

Background:

• Coffee berry disease (CBD), caused by Colletotrichum kahawae, poses a significant threat to global Coffee arabica production, resulting in substantial yield losses.
• Current methods for CBD detection are often subjective and inefficient, especially in resource-limited agricultural settings.
• Existing deep learning research primarily focuses on leaf diseases, with limited attention to berry-specific infections like CBD.

Purpose of the Study:

• To develop a lightweight and accurate deep learning model for the classification of healthy and CBD-affected coffee berries.
• To address the gap in research concerning berry-specific plant disease detection using advanced AI techniques.
• To provide a potential solution for real-time, low-resource deployment in sustainable coffee farming.

Main Methods:

• A dataset of 1737 coffee berry images was preprocessed using bilateral filtering and color segmentation.
• A Compact Convolutional Transformer (CCT) model, combining convolutional branches and a transformer encoder, was employed for feature extraction and classification.
• The CCT model was integrated with a Multilayer Perceptron (MLP) classifier and optimized using early stopping and regularization techniques.

Main Results:

• The proposed CCT model achieved a validation accuracy of 97.70% and 100% sensitivity for detecting CBD.
• CCT-extracted features demonstrated strong performance with traditional classifiers like SVM (82.47% accuracy, AUC 0.91) and Decision Tree (82.76% accuracy, AUC 0.86).
• The CCT system exhibited superior accuracy (97.5%) with significantly fewer parameters (0.408 million) and faster training times (2.3 s/epoch) compared to pretrained models.

Conclusions:

• The developed lightweight CCT model offers a highly accurate and sensitive solution for identifying coffee berry disease.
• The model's efficiency and low resource requirements make it suitable for real-time applications in sustainable coffee production.
• This study highlights the potential of advanced deep learning architectures for addressing critical challenges in agricultural plant pathology.