Data Collection and Labeling
Red Rot Images
We collected a dataset of approximately 800-1000 images of sugarcane plants,
encompassing both healthy and red rot-infected plants. The images were sourced from
multiple agricultural databases and field surveys, covering a variety of environmental
conditions and geographic regions. Expert agricultural professionals labeled each
image to ensure that disease identification was precise, providing accurate
annotations of red rot disease at different stages.
Data Augmentation
To improve the generalization of our models and compensate for the relatively small
dataset, we applied extensive data augmentation. This increased the dataset size to
over 8000 images and added variation to the training data, making the models more
robust. Augmentation techniques included:
Rotation: Random rotations between -40 and +40 degrees to simulate various
orientations of plants in the field.
Flipping: Horizontal and vertical flips to mimic different plant orientations.
Brightness and Contrast Adjustments: Adjusting brightness levels within ±20% to
simulate various lighting conditions during data collection.
Zoom and Cropping: Introducing random zooms and cropping to simulate
differing distances between the camera and plants.
Random Noise Injection: Adding small levels of Gaussian noise to simulate real world image imperfections.
Model Training Pipeline
Contribution:
Our pipeline involved training several deep learning models, including EfficientNet,
ResNet-152, VGG16, and Vision Transformers, as well as a Hybrid CNN + Random
Forest model. The pipeline was designed to compare the performance of these
architectures on the red rot detection task and identify the most efficient and accurate
models for real-world deployment.
Model Architectures and Functionality"
1.Hybrid CNN + Random Forest Model
• Why Chosen: The hybrid model combined the feature extraction capabilities of
Convolutional Neural Networks (CNNs) with the robust classification of Random Forests,
aiming to leverage the strengths of both deep learning and traditional machine learning.
Results: Test Accuracy: 96.15%, F1 Score: 96.11%.
2.EfficientNet Model
• Why Chosen: EfficientNet was selected for its ability to scale network depth, width, and
resolution efficiently, providing a good balance between model complexity and
computational resources.
• Results: Best Validation Accuracy: 99.88%, Test Accuracy: 99.88%, Precision: 100.00%,
Recall: 99.76%, F1 Score: 99.88%.
3.ResNet-152 Model
• Why Chosen: ResNet-152 was chosen for its depth and use of residual connections,
which help mitigate the vanishing gradient problem in deep neural networks, allowing the
model to learn more intricate features.
• Results: Best Validation Accuracy: 100.00%, Test Accuracy: 99.64%, Precision: 99.53%,
Recall: 99.77%, F1 Score: 99.65%.
4.VGG16 Model
• Why Chosen: VGG16 was included for comparative analysis due to its historical
significance in image recognition tasks.
• Results: Best Validation Accuracy: 99.16%, Test Accuracy: 99.16%, Precision: 99.76%,
Recall: 98.59%, F1 Score: 99.18%.
5.Vision Transformers Model
• Why Chosen: Vision Transformers (ViT) were explored to leverage self-attention
mechanisms, potentially capturing global image features more effectively than CNNs.
• Results: Best Validation Accuracy: 97.71%, Test Accuracy: 96.51%, Precision: 96.73%,
Recall: 96.01%, F1 Score: 96.37%
Evaluation Metrics
To assess the performance of the models, we used the following metrics:
Accuracy: The percentage of correctly classified images.
Precision: The ratio of true positive predictions to the total number of positive
predictions.
Recall (True Positive Rate): The percentage of actual positive cases (diseased
plants) correctly identified by the model.
F1 Score: The harmonic mean of precision and recall, balancing these two
metrics.
ROC Curve & AUC: The ROC curve evaluated the trade-offs between the true
positive rate and false positive rate.
Confusion Matrix: A confusion matrix provided detailed insights into model
misclassifications.
Conclusion
These protocols provide a comprehensive approach to building an AI-based system for
red rot disease detection in sugarcane. By following these steps, we were able to
develop models that generalize well across different environmental conditions and
disease stages. Our protocols ensure that this system can be replicated and scaled for
use in other agricultural settings.