PlantDoc: Plant Disease Classification

Introduction

PlantDoc is a complete implementation of a plant disease classification system using a CBAM (Convolutional Block Attention Module) augmented ResNet18 architecture. The system is designed to accurately identify various plant diseases from images, leveraging attention mechanisms to focus on the most relevant features for diagnosis.

Plant diseases cause significant crop losses worldwide, with estimates suggesting 20-40% of global crop production is lost to pests and diseases annually. Early and accurate detection is crucial for effective management and sustainable agriculture. Traditional disease diagnosis relies on manual inspection by experts, which is time-consuming and often subjective.

This project implements a state-of-the-art deep learning approach that combines ResNet18 with attention mechanisms to improve classification accuracy for plant disease diagnosis, enabling faster and more reliable identification of diseases across 39 different plant disease classes.

Dataset Overview

The PlantDoc project utilizes a comprehensive plant disease dataset with the following characteristics:

Class Distribution

The dataset contains 39 classes representing different plant diseases and healthy plants, with some class imbalance:

The largest classes include Orange_Haunglongbing_Citrus_greening (8.96%), Tomato_Tomato_Yellow_Leaf_Curl_Virus (8.71%), and Soybean_healthy (8.28%), while many classes have approximately 1.63% of the total dataset (about 1,000 images each).

Image Properties

• Image Dimensions: Width and height range from 192-350 pixels, with median 256×256
• Aspect Ratio: Nearly all images (97.84%) are square with a 1:1 aspect ratio
• File Sizes: Range from 4.11 KB to 28.60 KB, with a median of 14.96 KB
• Color Information: Average RGB values of R=118.45, G=124.62, B=104.63

Data Preprocessing and Augmentation

To address class imbalance and improve model generalization, the project implements a comprehensive data pipeline including:

• Advanced preprocessing with the Albumentations library
• State-of-the-art augmentation strategies (RandAugment, CutMix, MixUp)
• Automatic data validation and integrity checking
• Class balancing techniques to handle uneven distributions

Methodology

The PlantDoc system leverages a dual-attention enhanced architecture to significantly improve classification performance. The key components of our methodology include:

CBAM: Dual Attention Mechanism

CBAM sequentially infers attention maps along two separate dimensions:

1. Channel Attention Module: Focuses on "what" is meaningful in the input image channels by applying attention weights to emphasize important feature channels using average and max pooling operations processed through shared MLPs.
2. Spatial Attention Module: Focuses on "where" the informative parts are located in the feature map, generating 2D spatial attention using aggregated channel information through convolutional layers.
3. Integration with ResNet: CBAM blocks are inserted after each residual block in the ResNet18 architecture, refining feature representations at multiple levels of abstraction.

This dual attention mechanism allows the model to dynamically emphasize salient features in both channel and spatial dimensions, leading to improved performance, especially on images with complex backgrounds or subtle disease symptoms.

Feature Visualization

Experiments

Multiple experimental configurations were evaluated to determine the optimal architecture and training strategy. Key experiments included:

Training Approach

• Mixed Precision Training: FP16 implementation for faster training without sacrificing accuracy
• Adaptive Optimization: Using AdamW optimizer with cosine annealing learning rate scheduling
• Regularization Techniques: Dropout (0.15), weight decay (5e-5), and stochastic weight averaging for enhanced generalization

Model Variants

Two primary model variants were systematically explored:

1. CBAM-ResNet18 v1: The initial implementation with baseline attention parameters, trained for 150 epochs
2. CBAM-ResNet18 v2: An improved version with optimized training parameters and a more efficient training schedule (100 epochs)

Both models were evaluated using a comprehensive set of metrics including accuracy, precision, recall, F1 score, confusion matrices, ROC curves, and precision-recall curves to provide a holistic performance assessment.

Performance Results

Confidence Analysis

The models become increasingly well-calibrated over time, as shown by the decreasing Expected Calibration Error (ECE). This indicates that the predicted confidence levels closely match the actual accuracy of predictions.

By the final epoch, both models show high confidence (>85%) for the vast majority of predictions, with minimal low-confidence predictions. This demonstrates the models' ability to make decisive classifications with high certainty.

Conclusion

The CBAM-augmented ResNet18 architecture demonstrates significant improvements over standard classification approaches for plant disease identification. Key findings include:

• The dual attention mechanism (channel + spatial) significantly improves model performance by helping the network focus on disease-relevant features
• Our approach achieves excellent accuracy (>97%) across 39 diverse plant disease classes
• Attention mechanisms particularly improve performance on challenging cases with subtle symptoms or complex backgrounds
• The model demonstrates strong generalization capabilities, maintaining high accuracy across varied imaging conditions
• Model v2 offers a more efficient alternative with comparable performance but 3x faster training time

These results highlight the potential of attention-enhanced deep learning approaches for practical agricultural applications. By providing accurate, reliable disease identification, this system could significantly contribute to sustainable farming practices through early detection and targeted treatment of plant diseases.

Future Work

Ongoing and planned improvements to the PlantDoc system include:

• Mobile-optimized model variants for field deployment
• Multi-label classification for detecting multiple diseases in a single image
• Disease severity grading capabilities
• Integration with drone/robotic platforms for automated field scouting