Agricultural Plant Health Classification (Bean Crop)

Project Overview

This project is a proof-of-concept machine learning system for classifying agricultural plant images into healthy and unhealthy categories. It demonstrates an end-to-end ML workflow — from data loading and preprocessing to model training, evaluation, and reflection — using a real-world image dataset.

The project intentionally progresses from classical machine learning baselines to an exploratory deep learning approach, highlighting trade-offs, limitations, and future directions relevant to AI-powered products.

Problem Statement

Early detection of plant stress and disease is critical for improving crop yield and reducing agricultural losses. Manual inspection is time-consuming, subjective, and difficult to scale.

The goal of this project is to automatically classify Bean plant images into:

• Healthy (label 0)
• Unhealthy (label 1)

Dataset

Source: Mendeley Data – Bangladeshi Vegetables Dataset
Selected Crop: Bean

Only the Bean crop was used in this project to:

• Reduce inter-crop variability
• Focus on binary classification
• Clearly demonstrate the ML workflow

Class Distribution

• Healthy: 632 images
• Unhealthy: 13 images

The dataset is severely imbalanced, which significantly affects model performance and evaluation.

Approach

1. Data Preprocessing

• Images loaded from directory structure (Healthy, Unhealthy)
• Converted to RGB format
• Resized to 64 × 64
• Pixel values normalized to [0, 1]

Two representations were used:

• Flattened images for classical ML models
• 3D image tensors for the CNN

2. Feature Representation

• Raw pixel intensities
• Flattened to 1D vectors for classical models
• Preserved spatial structure for CNN experiments

This design allows direct comparison between traditional ML and deep learning approaches.

3. Models Implemented

Classical Machine Learning (Baseline)

• Logistic Regression
• Support Vector Machine (SVM)
• Random Forest Classifier

These models were selected to:

• Build interpretable baselines
• Apply classification concepts learned from regression
• Highlight the limitations of classical models on image data

Deep Learning (Exploratory)

• Simple Convolutional Neural Network (CNN)
  • Convolution + pooling layers
  • Fully connected classification head
  • Binary output using sigmoid activation

The CNN was implemented as a proof-of-concept to assess suitability for computer vision tasks.

---

4. Evaluation Metrics

• Accuracy
• Precision, Recall, F1-score
• Confusion Matrix

Special attention was paid to minority class performance due to dataset imbalance.

Results

Classical Models

• Accuracy ranged between 94% – 98%
• Models performed very well on the Healthy class
• All classical models failed to correctly identify Unhealthy samples

This demonstrates that: High accuracy does not imply effective classification when classes are imbalanced.

CNN Experiment

• The CNN learned spatial features directly from images
• Training accuracy fluctuated across epochs
• Performance was limited by:
  • Small dataset size
  • Severe class imbalance
  • Lack of data augmentation

Despite limitations, the CNN confirms the appropriateness of deep learning for this problem domain.

---

Challenges Faced

• Severe class imbalance (632 vs 13 samples)
• Learning and applying image-based classification within a short timeframe
• Implementing image loading and preprocessing pipelines
• Understanding and interpreting classification metrics beyond accuracy
• Managing deep learning instability with limited data

Limitations

• Flattened pixel features discard spatial information
• Classical ML models are not well-suited for complex visual patterns
• Minority class underrepresentation limits generalization
• CNN experiment was constrained by time and data availability

Key Takeaway

This project establishes a strong baseline and demonstrates an understanding of:

• ML experimentation workflows
• Model limitations
• The transition from classical ML to modern computer vision systems