Potato Tuber Disease Classification using Deep Learning and Grad-CAM

Introduction

Potato is one of the most widely cultivated and consumed vegetables globally. It is a vital source of food, income, and livelihood for millions of farmers, especially in developing countries. Among the different types of potato production methods, potted potato plants—grown in controlled environments like greenhouses or small-scale home setups—have become increasingly popular. These systems help farmers manage soil conditions, monitor growth, and reduce pest exposure.

However, even in such controlled environments, tuber diseases such as blackspot bruising, brown rot, dry rot, and soft rot can affect crop yield and quality significantly. Manual inspection is time-consuming and prone to human error. Here, deep learning-based image classification provides an efficient, scalable solution for disease identification.

This project leverages CNNs and explainable AI to classify tuber diseases with high accuracy and interpretability.

Dataset and Classes

The dataset is divided into:

• Training: 70%
• Validation: 15%
• Testing: 15%

Classes:

• Blackspot Bruising Disease
• Healthy Potato
• Potato Brown Rot Disease
• Potato Dry Rot Disease
• Potato Soft Rot Disease

Models Used

1. EfficientNetV2B0

• How it works: Scales network depth, width, and resolution using a compound coefficient. It includes fused MBConv blocks for faster and more efficient learning.
• Why chosen: Lightweight yet powerful; delivers state-of-the-art results with fewer parameters.
• Efficiency: Achieved 98.48% test accuracy with an F1-score of 0.98.

2. DenseNet121

• Working principle: Each layer receives input from all previous layers (dense connections), which improves feature propagation.
• Reason for selection: Effective in learning subtle features of diseases with fewer training epochs.
• Performance: Test accuracy of 97.72%, F1-score 0.98.

3. Xception

• How it works: Utilizes depthwise separable convolutions, allowing better model efficiency and deeper architecture.
• Why chosen: Excellent at learning abstract patterns from complex tuber textures.
• Efficiency: Reached 97.53% accuracy with strong generalization.

4. InceptionResNetV2

• Architecture: Combines the Inception module with residual connections, enabling very deep networks to train efficiently.
• Why chosen: Captures multi-scale features and mitigates vanishing gradients.
• Performance: Delivered 93.55% accuracy.

5. NASNetMobile

• How it works: Designed through neural architecture search (NAS), this model balances efficiency and accuracy for mobile deployments.
• Use case: Ideal for real-time classification on edge devices.
• Result: 87.10% accuracy, slightly lower but efficient on lightweight setups.

Explainable AI with Grad-CAM

To understand and visualize the model’s decision-making process, Grad-CAM (Gradient-weighted Class Activation Mapping) is applied. It highlights the regions of the image that were important for the classification, offering transparency and trust—critical for agricultural applications.

Result Summary

Model	Accuracy	Precision	Recall	F1-Score
EfficientNetV2B0	98.48%	0.98	0.98	0.98
DenseNet121	97.72%	0.98	0.98	0.98
Xception	97.53%	0.98	0.98	0.98
InceptionResNetV2	93.55%	0.94	0.94	0.94
NASNetMobile	87.10%	0.91	0.87	0.88

Use Cases and Future Scope

• Integration into grading machines to sort diseased tubers.
• Real-time camera-based disease detection in warehouses.
• Training models for early-stage (immature tubers) disease identification.
• Extending to nutrient or pesticide deficiency detection.