Image Classification using Deep Learning (CNN)

Rice Leaf Disease Prediction Using CNN

Linkedin: https://www.linkedin.com/posts/harishwaran-vl_deeplearning-cnn-machinelearning-activity-7310412592483573760-h23V?utm_source=share&utm_medium=member_desktop&rcm=ACoAADsBg0IBEc3FDwTSODKTv75sSLL41wgAAiQ

Problem Statement

Rice is a staple crop for a large portion of the global population. However, its productivity is frequently threatened by various leaf diseases such as Leaf Smut, Brown Spot, and Bacterial Leaf Blight. These diseases can lead to significant yield loss if not identified and treated early. Traditional disease detection methods rely heavily on expert inspection and laboratory analysis, which can be time-consuming, costly, and inaccessible to many small-scale farmers.

Project Objective

The objective of this project was to develop an automated deep learning system that accurately classifies rice leaf diseases from image data. The goal was to reduce diagnosis time, support agricultural decision-making, and help farmers take timely preventive measures.

To achieve this, a Convolutional Neural Network (CNN) was trained using a custom rice leaf image dataset. The trained model was then deployed as a web application on AWS EC2, allowing users to upload rice leaf images and get disease predictions in real time.

Dataset Overview

• Dataset Type: Custom image dataset

• Classes:

  • Leaf Smut

  • Brown Spot

  • Bacterial Leaf Blight

    
    

• Preprocessing:

  • Resizing all images to a uniform shape

  • Augmentation (flipping, rotation, zoom) to improve generalization

  • Normalization to scale pixel values

Technical Workflow

1. Data Preparation

• Manually captured and labeled images

• Structured the dataset into training, validation, and test sets

• Applied data augmentation to increase dataset diversity

2. Model Building

• Designed a custom CNN architecture with multiple convolutional and pooling layers

• Included dropout layers to avoid overfitting

• Used ReLU activation and Softmax in the output layer

• Compiled the model using categorical cross-entropy loss and Adam optimizer

3. Model Training

• Trained for multiple epochs with batch processing

• Monitored accuracy and loss on both training and validation sets

• Achieved high accuracy (>90%) in classifying disease types

4. Model Evaluation

• Evaluated on a hold-out test set

• Plotted confusion matrix and classification report

• Verified robustness of predictions with unseen leaf images

Model Deployment

• Framework: Flask for building the web application

• Deployment Platform: AWS EC2 instance (Ubuntu)

• Functionality:

  • Simple UI for image upload

  • Backend uses the trained CNN model to make predictions

  • Displays predicted disease class and confidence score

Results and Performance

• Classification Accuracy: Over 90%

• Low false positive rate, particularly between Leaf Smut and Brown Spot

• Fast inference time: less than 2 seconds per image

• Validated on real-world samples to ensure model generalization

Key Achievements

• Developed an end-to-end image classification pipeline using deep learning

• Created a working diagnostic tool for rice disease identification

• Gained practical experience in:

  • Image preprocessing

  • CNN model design and tuning

  • Flask app development

  • Cloud deployment with AWS EC2

Impact and Applications

• Assists farmers and agronomists in early disease detection

• Reduces dependency on expert visits or lab tests

• Enhances precision agriculture practices through digital diagnostics

• Can be extended to other crops and diseases with retraining

Technologies Used

• Python: NumPy, Pandas, OpenCV, Matplotlib

• Deep Learning: TensorFlow/Keras for CNN modeling

• Web App: Flask

• Cloud: AWS EC2 (Ubuntu)

• Data Handling: ImageDataGenerator, manual annotations

• Others: Gunicorn, Nginx for production deployment