🤖 AI

Healthy Harvests: A Comparative Look at Guava Disease Classification Using InceptionV3

This study compares the performance of InceptionV3 and ResNet50 deep learning models in classifying guava diseases (Anthracnose, Fruit flies, and Healthy), demonstrating that the InceptionV3 model, enhanced by data augmentation and mixing techniques, achieved a superior accuracy of 98.15%.

Original authors: Samanta Ghosh, Shaila Afroz Anika, Umma Habiba Ahmed, B. M. Shahria Alam, Mohammad Tahmid Noor, Nishat Tasnim Niloy

Published 2026-02-12

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Samanta Ghosh, Shaila Afroz Anika, Umma Habiba Ahmed, B. M. Shahria Alam, Mohammad Tahmid Noor, Nishat Tasnim Niloy

Original paper licensed under CC BY 4.0 (http://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

The Story of the "Smart Fruit Doctor"

Imagine you are a farmer with a massive orchard of guava trees. You work incredibly hard, but there’s a problem: tiny, invisible enemies are attacking your harvest. Some are fungi (Anthracnose) that rot the fruit from the outside, and others are fruit flies that sneak inside to eat the fruit from within.

If you don't catch these "intruders" early, you lose your entire income. Checking every single fruit by hand is like trying to read every single grain of sand on a beach—it’s exhausting, slow, and you’re bound to miss something.

This research paper describes the creation of a "Digital Fruit Doctor"—an AI system that can look at a photo of a guava and instantly tell you if it’s healthy or sick.

1. The Training: Teaching the Brain

To make this "doctor" smart, the researchers didn't just give it a textbook; they gave it a massive photo album. They took 473 original photos of guavas and used a technique called "Data Augmentation."

The Analogy: Imagine teaching a child to recognize a dog. If you only show them one picture of a Golden Retriever sitting down, they might not recognize a Husky running. So, you show them the dog from different angles, in different lighting, and even "mix" pictures together (using techniques called CutMix and MixUp). This is like taking two different photos and blending them together like a smoothie, forcing the AI to learn the essence of the disease rather than just memorizing one specific picture.

2. The Two Specialists: InceptionV3 vs. ResNet50

The researchers hired two different "AI Specialists" to see who was better at the job:

The InceptionV3 Specialist (The Multi-Lens Photographer): This specialist uses many different types of magnifying glasses at the same time. One lens looks for big shapes, another for tiny textures, and another for colors. Because it looks at everything at once, it’s incredibly thorough.
The ResNet50 Specialist (The Deep Thinker): This specialist is like a person running a very long marathon. In deep learning, sometimes "thinking" too long makes the brain get confused (the "vanishing gradient" problem). ResNet50 uses "shortcuts" (residual learning) to make sure the important information doesn't get lost as it travels through its 50 layers of logic.

The Result: The InceptionV3 specialist won the gold medal, achieving a staggering 98.15% accuracy!

3. The "Why" Factor: Explainable AI (XAI)

One big problem with AI is that it’s often a "Black Box"—it gives you an answer, but it won't tell you why. If a doctor tells you that you're sick but can't show you where the problem is, you wouldn't trust them.

To fix this, the researchers used tools called SHAP and LIME.

The Analogy: Think of this like a Highlighter Pen. When the AI looks at a guava and says, "This is diseased," the XAI tool comes in and highlights the exact spots on the fruit that triggered that decision. It might highlight a dark brown rot spot or a tiny puncture from a fly. This proves to the farmer that the AI isn't just guessing—it’s actually looking at the symptoms.

The Bottom Line

This paper isn't just about math and code; it's about food security. By creating a highly accurate, "explainable" digital doctor, we can help farmers protect their crops, reduce waste, and ensure that when you buy a guava at the market, it’s healthy, delicious, and ready to eat.

Technical Summary: Healthy Harvests: A Comparative Look at Guava Disease Classification Using InceptionV3

1. Problem Statement

Guava is a vital economic crop in South Asia, providing essential nutrition and income for smallholder farmers. However, its production is significantly threatened by diseases—most notably Anthracnose (a fungal infection) and Fruit Fly infestations (larval damage). Manual inspection of these diseases is time-consuming, prone to human error, and difficult to scale. There is a critical need for an automated, reliable, and interpretable system to identify these diseases early to ensure food security and crop yield sustainability.

2. Methodology

The researchers proposed a deep learning-based framework to classify guava into three categories: Anthracnose, Fruit Flies, and Healthy.

Dataset & Preprocessing: The study utilized a dataset of 473 original images from Mendeley Data. To overcome the limitation of a small sample size, the authors applied advanced data augmentation, expanding the dataset to 3,784 images. All images were resized to $256 \times 256$ pixels in RGB mode.
Augmented Regularization: To enhance model robustness and generalization, two specific data mixing techniques were implemented:
- CutMix: Replacing a patch of one image with a patch from another.
- MixUp: Linearly blending two images and their corresponding labels.
Model Architectures: Two state-of-the-art Convolutional Neural Network (CNN) architectures were compared:
- InceptionV3: Utilizes multi-scale convolutional filters in parallel to capture diverse features efficiently.
- ResNet50: Employs residual learning (skip connections) to mitigate the vanishing gradient problem in deep networks.
Explainable AI (XAI): To move beyond "black-box" predictions, the study integrated SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations). These tools generate heatmaps to visualize which specific regions of the fruit (e.g., rot spots or puncture marks) influenced the model's decision.

3. Key Contributions

Comparative Analysis: A rigorous performance comparison between InceptionV3 and ResNet50 specifically for guava pathology.
Enhanced Robustness: The application of CutMix and MixUp augmentation to improve the model's ability to handle real-world agricultural variations.
Interpretability: The integration of XAI (SHAP/LIME) to provide transparency, ensuring that the model is making decisions based on actual disease symptoms rather than background noise.
Benchmark Improvement: Demonstrating superior accuracy compared to existing literature in the field of guava disease detection.

4. Results

The experimental results indicate that InceptionV3 outperformed ResNet50 across nearly all metrics:

Metric	InceptionV3	ResNet50
Overall Accuracy	98.15%	94.46%
Validation Accuracy	~88%	~68%
Performance Trend	High stability/low overfitting	Higher fluctuation/overfitting tendencies

InceptionV3 showed exceptional precision and recall, particularly in identifying Anthracnose (1.00) and Healthy fruits (1.00).
Confusion Matrix Analysis: InceptionV3 exhibited minimal confusion, primarily misclassifying a very small number of healthy samples as fruit flies.
XAI Validation: SHAP and LIME visualizations confirmed that the models' "attention" aligned with the actual physical damage on the fruit, validating the biological relevance of the model's features.

5. Significance

This research provides a highly accurate and interpretable technological solution for precision agriculture. By achieving over 98% accuracy, the proposed InceptionV3 model offers a reliable tool that could be integrated into mobile or edge devices for real-time field use. This enables farmers to detect diseases early, reduce chemical usage through targeted treatment, and ultimately protect their livelihoods and the global food supply.