Cotton Leaf Defect Classification System 🌿🦠

Intelligent diagnosis of cotton diseases for sustainable agriculture using Deep Learning.

📖 Overview

Cotton (Gossypium) is a crucial crop for rural economies in Ecuador. However, its production is threatened by foliar diseases such as Curl Virus, Leaf Reddening, and Leaf Spot Bacterial Blight, which reduce photosynthetic capacity and fiber quality.

This project implements and compares two Deep Learning architectures to detect these diseases automatically:

1. VGG16 (Baseline): A classic CNN using Transfer Learning.
2. SBTAYLOR-KAN modified (Proposed): A novel hybrid architecture combining Convolutional Neural Networks (CNN) with Kolmogorov-Arnold Networks (KAN) and Taylor Series expansion for learnable activation functions.

The proposed SBTAYLOR-KAN modified model achieved 94.80% accuracy, significantly outperforming the baseline while being computationally lighter (only ~241k parameters).

🎯 Objectives

• Early Detection: Provide a tool for farmers to identify phytosanitary problems early.
• Optimization: Reduce the use of chemical treatments by targeting only affected areas.
• Innovation: Validate the viability of KANs (Kolmogorov-Arnold Networks) as an alternative to traditional MLPs in computer vision.

🧠 Model Architecture: SBTAYLOR-KAN

Unlike traditional CNNs that use fixed activation functions and dense classification layers, our proposed model integrates:

1. CNN Block (Feature Extraction): 4 sub-blocks of Conv2D $\rightarrow$ BatchNorm $\rightarrow$ ReLU $\rightarrow$ MaxPool (Filters: $32 \rightarrow 64 \rightarrow 128$).
2. Adaptive Pooling: Reduces feature maps to a compact representation ($128 \times 1 \times 1$).
3. Taylor Series Function: Enhances feature representation using 5 terms (odd functions/sine approximation).
4. KAN Classifier: Replaces dense layers with KANLinear layers, which use learnable B-splines on edges.
   • Structure: $128 \rightarrow 256 \rightarrow 128 \rightarrow 4$ (Classes).

📊 Dataset

The model was trained on a composite dataset merging SAR-CLD-2024 and COT-AD, containing images of cotton leaves classified into 4 categories:

• Fresh Leaf (Healthy)
• Curl Virus
• Leaf Reddening
• Leaf Spot Bacterial Blight

Preprocessing: Images were resized to $224 \times 224$ pixels, normalized, and augmented to ensure robustness.

🚀 Key Results

Metric	VGG16 (Transfer Learning)	SBTAYLOR-KAN (Proposed)
Accuracy	88.73%	94.80%
Inference Time	Slower	~300ms (Real-time ready)
Convergence	Slow	Fast

Conclusion: The SBTAYLOR-KAN model offers superior accuracy and clearer decision boundaries (as seen in the confusion matrix) with a fraction of the parameters.

🛠️ Technologies Used

• Language: Python 3.x
• Deep Learning: PyTorch (SBTAYLOR-KAN), TensorFlow/Keras (VGG16)
• Data Processing: NumPy, Pandas, torchvision
• Visualization: Matplotlib, Seaborn, Grad-CAM (for interpretability)
• Environment: Jupyter Notebooks / Google Colab

💻 Installation & Usage

1. Clone the repository:

   git clone [https://github.com/BryanEstrada003/cotton-leaf-defect-classification.git](https://github.com/BryanEstrada003/cotton-leaf-defect-classification.git)
   cd cotton-leaf-defect-classification

2. Install dependencies:

   pip install -r requirements.txt

3. Run Inference:

   from models import SBTAYLOR_KAN
   import torch
   
   # Load model
   model = SBTAYLOR_KAN(num_classes=4)
   model.load_state_dict(torch.load('weights/best_model.pth'))
   
   # Predict
   prediction = model.predict('path/to/leaf_image.jpg')
   print(f"Diagnosis: {prediction}")

👥 Authors

Group 7 - Artificial Intelligence (PAO II 2025)

• Michael Estrada Santana - GitHub Profile
• Melissa Ayllon Gutiérrez - GitHub Profile
• Juan Pablo Plúas Muñoz - GitHub Profile

Advisor: Enrique Pelaez
Institution: Escuela Superior Politécnica del Litoral (ESPOL) 🇪🇨

---

References:

[1] P. Bishshash, M. A. S. Nirob, M. H. Shikder, y A. Sarower, «SAR-CLD-2024: A comprehensive dataset for Cotton Leaf Disease Detection». Mendeley Data, 2024.

[2] A. Ali et al., «COT-AD: Cotton Analysis Dataset», arXiv preprint arXiv:2507.18532, 2025.

[3] Anonymous, «Kolmogorov–Arnold Networks: A Critical Assessment of Claims and Empirical Evidence», arXiv preprint arXiv:2407.11075, 2024, [En línea]. Disponible en: https://arxiv.org/abs/2407.11075

[4] K. Fatema, E. A. Mohammed, y S. S. Sehra, «Taylor-Series Expanded Kolmogorov–Arnold Network for Medical Imaging Classification», arXiv preprint arXiv:2509.13687, 2025, [En línea]. Disponible en: https://arxiv.org/abs/2509.13687