An Explainable Hybrid CNN-ViT Framework for Ethiopian Coffee Leaf Disease Classification

In the highlands of Ethiopia, coffee is more than a crop; it is the heartbeat of the economy and the livelihood of millions. For centuries, farmers have relied on their eyes and generations of local knowledge to spot the signs of disease on their plants. They look for the yellowing of a leaf or the appearance of a strange spot, using this visual intuition to decide how to protect their harvest. Yet, this traditional method has limits. Early signs of infection can be subtle and easily missed, and by the time a disease becomes obvious to the human eye, it may have already spread too far to stop. In recent years, scientists have turned to computers to help, teaching them to recognize patterns in images that humans might overlook. These computer systems, known as deep learning models, act like digital eyes that can scan thousands of leaves in seconds, identifying specific illnesses with remarkable speed. However, a major hurdle remains: these powerful computers often work like black boxes. They give an answer, but they do not explain why they chose it, leaving farmers and experts unsure if they can trust the diagnosis.

A team of researchers from the University of Gondar and Mekdela Amba University has developed a new approach to bridge this gap between raw power and human understanding. They created a system designed specifically to diagnose four common coffee leaf diseases in Ethiopia: Cercospora, Leaf Rust, Phoma, and the healthy state of the plant. Instead of relying on a single type of computer vision, they built a hybrid framework that combines two different strengths. One part of the system acts like a microscope, zooming in to see tiny textures and local details on the leaf surface. The other part acts like a wide-angle lens, stepping back to understand the overall shape and context of the entire leaf. By merging these two perspectives, the system learns to see the disease more clearly than either method could alone. Crucially, the researchers also equipped this system with a tool that highlights exactly which parts of the leaf the computer is looking at when it makes a decision. This transparency allows a farmer to see the computer's reasoning, confirming that the machine is focusing on the actual disease spots rather than random noise or background shadows.

The researchers tested their new framework on a large collection of 12,000 high-quality images of coffee leaves gathered from major research centers and farms across Ethiopia. They split this collection into groups, using most of the images to teach the computer and the rest to test its knowledge. The system was trained to distinguish between the four categories, learning to tell the difference between a healthy leaf and one suffering from fungal infections. When the testing phase began, the hybrid system proved to be exceptionally accurate. It correctly identified the condition of the leaves in 98.67 percent of the test cases. This performance was slightly better than other popular computer models the researchers compared it against, which achieved scores ranging from 98.06 percent to 98.56 percent. The new system was particularly good at telling the difference between healthy leaves and diseased ones. However, distinguishing between the two visually similar diseases, Leaf Rust and Phoma, remained the most challenging task for all models, with the majority of misclassifications occurring when the system confused Leaf Rust with Phoma. Despite this difficulty, the hybrid framework still managed to outperform the older models in classifying Leaf Rust, demonstrating the advantage of combining local texture analysis with global context even for the most difficult cases.

Beyond the numbers, the most significant finding lies in how the system explains itself. The researchers used a visualization technique to generate heatmaps, which are colored overlays that show where the computer's attention is focused on the image. When the system identified a disease, these heatmaps consistently highlighted the specific lesions, spots, or rust pustules that plant pathologists know to be the signs of illness. For example, when diagnosing a leaf with Phoma, the computer focused precisely on the necrotic spots and their edges. When looking at a healthy leaf, the computer showed no such focus, indicating it saw no disease patterns. This alignment between the computer's focus and human expert knowledge is vital. It transforms the system from a mysterious oracle into a trustworthy partner. A farmer or an agricultural extension worker can look at the image and the heatmap together, seeing exactly why the computer made its call. This visual proof builds the confidence necessary for these tools to be adopted in real-world farming, where trust is as important as accuracy.

The study acknowledges that while the results are strong, the work is not yet finished. The dataset used for training came from specific regions in Ethiopia, and the system may need further testing in different environments to ensure it works everywhere. Additionally, the hybrid model requires more computing power than simpler systems, which could make it harder to run on basic smartphones without further optimization. However, the researchers have laid a clear path forward. They suggest that future versions of this technology could be compressed to run on mobile devices, integrated with weather data for even better predictions, and expanded to cover more types of plant problems. By combining the ability to see fine details with the ability to understand the whole picture, and by making that process visible to the user, this research offers a promising step toward a future where technology supports, rather than replaces, the wisdom of Ethiopian coffee farmers. The goal is not just to detect disease, but to do so in a way that people can understand, trust, and use to protect their crops and their livelihoods.