Early Detection of Coffee Leaf Rust Through Convolutional Neural Networks Trained on Low-Resolution Images

TL;DR

This work addresses the urgent need for early detection of coffee leaf rust in Central America under limited data and computational resources by applying a convolutional neural network to low-resolution, grayscale images processed with a high-pass edge-enhancement filter. The approach demonstrates that edge-focused preprocessing enables strong performance, achieving over 90% across precision, recall, F1-score, and Dice, with 128x128 resolution offering the best trade-off between accuracy and efficiency. The findings suggest practical deployment potential on low-power devices for field-scale surveillance, reducing latency in disease management. The study contributes a practical pipeline for resource-constrained environments and outlines future directions involving unsupervised labeling and GAN-based data augmentation to bolster robustness against evolving rust strains.

Abstract

Coffee leaf rust, a foliar disease caused by the fungus Hemileia vastatrix, poses a major threat to coffee production, especially in Central America. Climate change further aggravates this issue, as it shortens the latency period between initial infection and the emergence of visible symptoms in diseases like leaf rust. Shortened latency periods can lead to more severe plant epidemics and faster spread of diseases. There is, hence, an urgent need for effective disease management strategies. To address these challenges, we explore the potential of deep learning models for enhancing early disease detection. However, deep learning models require extensive processing power and large amounts of data for model training, resources that are typically scarce. To overcome these barriers, we propose a preprocessing technique that involves convolving training images with a high-pass filter to enhance lesion-leaf contrast, significantly improving model efficacy in resource-limited environments. This method and our model demonstrated a strong performance, achieving over 90% across all evaluation metrics--including precision, recall, F1-score, and the Dice coefficient. Our experiments show that this approach outperforms other methods, including two different image preprocessing techniques and using unaltered, full-color images.

Figures (6)

• Figure 1: Stages of leaf rust beginning from the lesion emergence (Source: [4]).
• Figure 2: Image pre-convolution (left) and post-convolution (right).
• Figure 3: Test images post-convolution and downscaling with model predictions.
• Figure 4: Convolution process on a sample image.
• Figure 5: Histogram equalization on convolved (top) and grayscale images.