Crop Disease Detection Using Natural Language Processing
Abstract
Urban farming is expanding, but technologies for assisting novice farmers are limited.
Existing deep learning-based crop disease diagnosis methods primarily rely on CNN-based
detection but fail to provide detailed analyses of disease symptoms by severity. Our solution
integrates Image Captioning and Object Detection to enhance crop disease diagnosis.
The Image Captioning model generates grammatically correct, semantically clear diagnostic
sentences detailing disease symptoms and names, while the Object Detection model
identifies infected areas, enabling farmers to locate damage and verify the captioned
diagnosis. We utilize InceptionV3 as the encoder and Transformer as the decoder for Image
Captioning, and YOLOv5 for Object Detection. This approach significantly improves
accuracy and detail in diagnosis, offering novice farmers reliable, actionable insights to
manage disease spread effectively.
Introduction
Crop diseases such as leaf rust, blight, bacterial spots, and powdery mildew pose significant
threats to agricultural productivity. Early and accurate detection is essential for preventing
disease spread and maintaining crop health. Existing diagnostic technologies often fall short
in providing comprehensive, detailed analyses of disease symptoms.
Our project aims to bridge this gap by combining:
1. Object Detection using YOLOv5 for precise identification of diseased areas.
2. Image Captioning using InceptionV3 and Transformer Decoder for detailed descriptions
of symptoms.
By leveraging deep learning techniques, we aim to empower farmers to take timely actions,
improving disease management accuracy and supporting sustainable farming practices.
Literature Survey
1. Deep Learning for Plant Disease Detection
Objective: Evaluate CNN models for automated crop disease detection.
Findings: Convolutional Neural Networks (CNNs) like AlexNet and ResNet are effective for
disease classification but lack specificity in symptom description.
Limitations: Absence of detailed diagnostics and inability to highlight specific infected
regions.
Inspiration: Our project addresses this gap by integrating Object Detection with Image
Captioning for enhanced detail.
2. YOLO-based Object Detection in Agriculture
Objective: Use YOLO models for real-time object detection in crops.
Findings: YOLOv5 achieves high-speed and moderate accuracy in detecting diseased crop
areas.
Limitations: Low mAP50 scores (mean Average Precision at 50% Intersection-over-Union)
leading to missed detections or false positives.
Solution: Our approach upgrades YOLOv5 to YOLOv8 and incorporates advanced detection
techniques like R-CNN for greater precision.
3. Image Captioning for Medical Diagnostics
Objective: Apply Image Captioning to generate textual explanations of medical images.
Findings: Using Inception-based encoders and Transformer decoders produces clear and
concise diagnostic descriptions.
Limitations: Existing models are domain-specific and lack generalizability.
Inspiration: We adapt this methodology to the agricultural domain, providing
grammatically correct, semantically clear descriptions of crop diseases.
4. Challenges in Agricultural Dataset Development
Objective: Analyze the quality and quantity of datasets for agricultural machine learning
applications.
Findings: Insufficient dataset diversity leads to poor model generalization.
Limitations: Many datasets lack representation of varied crop types, disease symptoms, and
severity levels.
Solution: Our project enhances dataset diversity by including images of different crops,
diseases, and environmental conditions.
Conclusion
By combining advanced deep learning techniques like Object Detection and Image
Captioning, our solution addresses the limitations of existing crop disease detection
systems. It provides:
- Visual localization of diseased areas.
- Detailed, easy-to-understand diagnostic descriptions.
- Support for sustainable farming practices, enabling timely and accurate interventions.
This project supports the growing trend of urban farming and is adaptable to broader
agricultural applications, including pest management and soil health monitoring.
