💻 computer science

Hamiltonian Full Node Coverage Graph Attention Network with Fuzzy C-Means Superpixel Graph Learning for Banana Leaf Disease Classification

This paper proposes a novel framework combining Hamiltonian Full Node Coverage Graph Attention Networks with Fuzzy C-Means superpixel graph learning to achieve highly accurate and reliable classification of banana leaf diseases by effectively modeling both local and global spatial dependencies.

Original authors: S Pavithra, P. S. Eliahim Jeevaraj

Published 2026-08-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: S Pavithra, P. S. Eliahim Jeevaraj

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery, but instead of looking at a crime scene, you are looking at a banana leaf. In the world of agriculture, spotting a sick leaf early is a superpower; it can save a whole farm from disaster. For a long time, computers tried to do this detective work by looking at pictures like a human eye does: scanning a grid of pixels, left to right, top to bottom. This method, called a Convolutional Neural Network (or CNN), is like reading a book one word at a time in a straight line. It's good, but it sometimes misses the big picture. It might see a brown spot, but it struggles to understand how that spot connects to the rest of the leaf's "story."

To get smarter, scientists started using a different tool called Graph Attention Networks (GAT). Think of this as turning the leaf into a social network. Instead of a grid, the leaf is broken into little puzzle pieces (called superpixels), and each piece becomes a "person" in a party. These pieces talk to their neighbors to share gossip about what's happening. But here's the catch: in a normal party, people only talk to the folks standing right next to them. If a disease starts at the bottom of the leaf and spreads to the top, the pieces at the top might not hear about it until it's too late. They miss the long-distance connections. This paper asks: what if we could create a special route that visits every single piece of the leaf exactly once, ensuring no secret is left unshared along the way? That is the big idea behind this new research.

The Detective's New Map

The researchers, S. Pavithra and P.S. Eliahim Jeevaraj from Bishop Heber College in India, have built a new system they call the "Hamiltonian Full Node Coverage Graph Attention Network" (HFNC-GAT). It's a mouthful, so let's break it down into a fun story.

First, they take a picture of a banana leaf and slice it up into fuzzy, colorful puzzle pieces using something called "Fuzzy C-Means." Imagine the leaf is a painting, and instead of cutting it into perfect squares, they cut it into organic blobs that match the natural shapes of the disease spots. Because diseases often have fuzzy, blurry edges, these "fuzzy" pieces are better at capturing the truth than sharp, rigid cuts.

Next, they turn these puzzle pieces into a graph. Each piece is a node (a character), and they are connected by lines (edges). But here is where the magic happens. Usually, these nodes only chat with their immediate neighbors. The researchers, however, decided to add a special rule based on a "Hamiltonian path." In math, a Hamiltonian path is a route that visits every single spot in a city exactly once without getting lost. In their system, this means they create a special "super-highway" that forces the information to travel in a specific sequence, visiting every single puzzle piece along the way to ensure the whole leaf is covered. It's like a game of "telephone" where the message doesn't just whisper to the person next to you; it gets a special pass to run across the whole room in a set order to make sure everyone hears the news.

Finally, they use a "Graph Attention Network." This is the brain of the operation. Imagine the puzzle pieces are all shouting at once. The attention network is the smart moderator that listens and decides, "Hey, this specific piece of the leaf with the weird texture is really important! Let's listen to it more closely," while ignoring the healthy, boring parts. This helps the computer focus exactly on the disease symptoms, like the borders of a spot or the color changes.

What They Found

The team tested this new "super-connected" detective system against the old-school methods. They used two different sets of banana leaf photos, some from real farms and some from online databases, covering diseases like Cordana leaf spot, Panama disease, and Sigatoka.

The results were quite impressive. When they ran the test, their new HFNC-GAT model achieved a test accuracy of 91.86% on one dataset and 88.89% on the other. To put that in perspective, the standard "grid-reading" computers (CNNs) and the older "neighbor-only" graph computers (GCNs and standard GATs) were much less accurate. The standard GAT models, for instance, achieved test accuracies of 82.08% and 83.74%, while the CNN baselines struggled even more, often falling below 40%.

The researchers also checked how well the model agreed with the experts using a score called "Cohen's Kappa," where their model scored 0.8893 and 0.8242 on the respective datasets, and a measure called "Matthews Correlation Coefficient" (MCC) of 0.2441 and 0.8250. These numbers suggest the model isn't just guessing; it's genuinely understanding the patterns. They even used a technique to visualize what the model was looking at, showing that it correctly focused on the diseased areas rather than the healthy green parts.

Why It Matters

The paper suggests that by forcing the computer to look at the entire leaf as a connected web, rather than just isolated patches, we can spot diseases earlier and more reliably. The authors argue that the old way of just looking at neighbors isn't enough because diseases can spread in tricky, long-range ways that a simple grid misses. By using this "Hamiltonian full node coverage," they ensure that no part of the leaf is left out of the conversation.

While the results are strong, the researchers are careful to note that this is a specific solution for banana leaves. They suggest that in the future, this idea could be tweaked to help with other crops or even combined with new types of AI. But for now, they've shown that when you give a computer a map where every part of the leaf is connected, it becomes a much better detective for keeping our bananas healthy.

Technical Summary: Hamiltonian Full Node Coverage Graph Attention Network with Fuzzy C-Means Superpixel Graph Learning for Banana Leaf Disease Classification

Problem Statement
The classification of banana leaf diseases (including Cordana leaf spot, Fusarium wilt, and Sigatoka) is critical for agricultural output but faces significant challenges with traditional deep learning approaches. Standard Convolutional Neural Networks (CNNs) treat images as regular Euclidean grids, failing to capture the complex spatial arrangements and connectivity between distributed lesion regions on a leaf. Furthermore, CNNs often lack interpretability, acting as "black boxes" that hinder clinical trust and validation by agronomists. Existing graph-based methods often rely on local aggregation, neglecting long-range structural dependencies, and frequently utilize hard clustering for segmentation, which does not adequately account for the uncertainty and irregular boundaries of diseased regions.

Methodology
The authors propose a novel framework called Hamiltonian Full Node Coverage Graph Attention Network (HFNC-GAT), integrated with Fuzzy C-Means (FCM) superpixel graph learning. The methodology proceeds through the following stages:

Preprocessing and Segmentation:
- Banana leaf images are resized to 128×128 pixels and normalized.
- Fuzzy C-Means (FCM) clustering is applied to segment the image into superpixels. Unlike hard clustering, FCM assigns a degree of membership to each pixel, allowing for the representation of uncertain and irregular disease boundaries.
- Each superpixel is characterized by mean RGB values, texture attributes, and fuzzy membership scores.
Graph Construction:
- Each superpixel is treated as a node in a graph $G=(V, E)$ .
- Edges are established in two ways:
  - Spatial Adjacency: Connecting neighboring superpixels to preserve local structure.
  - Hamiltonian Path: An approximate Hamiltonian path is constructed to visit every node in the graph twice. This creates long-range connections, ensuring full node coverage and preserving global structural dependencies that standard local graph methods miss.
- Node features include mean RGB, texture statistics, and FCM membership values.
Graph Attention Network (GAT) Architecture:
- The graph is processed through a two-layer GAT architecture.
- The first layer utilizes 4 attention heads with a hidden dimension of 64 to learn diverse feature representations.
- The second layer uses 1 attention head to refine node embeddings.
- Attention Mechanism: Adaptive weights are assigned to neighboring nodes, allowing the model to focus on discriminative features such as lesion borders and textures.
- Global Mean Pooling: Aggregates node representations into a single graph-level embedding.
- Classification: A fully connected layer with Softmax outputs the disease class (Healthy, Cordana, Panama, Yellow/Black Sigatoka).
Training Strategy:
- The model employs a hybrid loss function: $L_{total} = L_{CE} + \lambda L_{coverage}$ .
- $L_{CE}$ is the standard cross-entropy loss.
- $L_{coverage}$ is a regularization term that penalizes the model if nodes are not utilized during the forward pass, ensuring the Hamiltonian full node coverage constraint is met and preventing information loss.
- Training is conducted using the Adam optimizer (learning rate 0.001) for 100 epochs with early stopping based on validation accuracy.

Key Contributions
The paper identifies three primary innovations:

FCM Superpixel Graph Learning: The use of Fuzzy C-Means to generate superpixels incorporates soft membership vectors into node embeddings, effectively handling the uncertainty and irregularity of disease boundaries.
Hamiltonian Full Node Coverage: The integration of a Hamiltonian path ensures that every node in the graph is traversed and connected. This guarantees a global structural representation and facilitates long-range dependency learning, addressing the limitation of local-only aggregation in standard Graph Convolutional Networks (GCNs).
Coverage Loss Regularization: A specific loss term is introduced to penalize unused nodes, forcing the model to utilize the entire leaf structure for feature propagation, thereby reducing overfitting and improving contextual learning.

Experimental Results
The framework was evaluated on two datasets: BananaLSD (Dataset I) and Banana Leaf Disease Dataset v4 (Dataset II). The proposed HFNC-GAT was compared against baseline CNNs (ResNet50, DenseNet, ShuffleNet) and graph-based models (GCN, standard GAT).

Accuracy: On Dataset I, the proposed model achieved 96.11% training accuracy and 91.86% test accuracy, significantly outperforming the standard GAT (82.08%) and GCN (47.76%). On Dataset II, it achieved 94.19% training and 88.89% test accuracy, compared to 83.74% for GAT.
Robustness Metrics: The model demonstrated high performance across multiple metrics, including Cohen's Kappa (0.8893 on Dataset I), Matthews Correlation Coefficient (MCC), F2-score, and ROC-AUC (0.9917 on Dataset I).
Ablation Studies: Experiments confirmed that the combination of FCM segmentation, Hamiltonian connectivity, and attention mechanisms collectively drives the performance gains. Removing any component resulted in lower accuracy and reduced interpretability.
Visualization: The study utilized radar graphs and Grad-CAM to visualize the model's focus on disease-specific regions, enhancing the explainability of the classification decisions.

Significance and Claims
The authors claim that the HFNC-GAT framework provides a robust, accurate, and interpretable solution for banana leaf disease classification. By explicitly modeling the spatial and structural relationships of diseased regions through a graph, the method overcomes the limitations of Euclidean-based CNNs. The integration of Hamiltonian paths ensures that global context is not lost, while the FCM segmentation handles the inherent uncertainty of disease boundaries. The study concludes that this graph-based approach yields superior predictive performance compared to existing CNN and standard graph learning baselines, offering a viable tool for precision agriculture and early disease management. The authors suggest that this framework lays the groundwork for extending graph-based diagnostics to other plant pathologies with complex spatial symptom distributions.

Technical Summary: Hamiltonian Full Node Coverage Graph Attention Network with Fuzzy C-Means Superpixel Graph Learning for Banana Leaf Disease Classification

More like this