1. Abstract

Federated Crop-Disease Detection Across Farms is an AI-based agricultural project that uses Federated Learning and deep learning to identify crop diseases while protecting farm data privacy. Traditional crop disease detection systems often rely on centralized image collection, requiring farmers to upload crop images to a central server for analysis and model training. The proposed system enables multiple farms to train local disease detection models using their own crop images without directly sharing their raw datasets. The local model updates are shared with a central server and aggregated using Federated Averaging (FedAvg) to create a global model. A Convolutional Neural Network (CNN) can classify crop images into healthy and diseased categories. The system aims to support collaborative agricultural intelligence, improve disease detection across diverse farming environments, and promote privacy-aware precision agriculture.

2. Objectives

1. To develop a federated learning framework for crop disease detection across multiple farms.
2. To identify crop diseases from leaf images using deep learning.
3. To protect farm data privacy by keeping raw images on local systems.
4. To train local models using farm-specific crop datasets.
5. To aggregate model updates using Federated Averaging.
6. To improve disease detection across different crop types and farming conditions.
7. To evaluate the model using accuracy, precision, recall, and F1-score.
8. To develop a user-friendly dashboard for uploading crop images and viewing prediction results.

3. Existing System

Existing crop disease detection systems commonly use centralized machine learning or deep learning models trained on crop images collected from public datasets or individual farms. Farmers may upload leaf images to a centralized application for disease classification. However, these approaches can require data sharing and may not adequately represent differences in crops, lighting conditions, regions, and farming practices.

Disadvantages

1. Dependence on centralized crop image datasets.
2. Privacy and ownership concerns when sharing farm data.
3. Limited collaboration between farms.
4. Potentially reduced accuracy under unfamiliar field conditions.
5. Difficulty incorporating new farm-specific data.
6. Dependence on internet connectivity for centralized services.

4. Proposed System

The proposed system uses Federated Learning to train a shared crop disease detection model across multiple farms without directly transferring their raw crop images to a central server. Each farm trains a local CNN model using its own images. The server aggregates the local model updates using Federated Averaging and distributes the improved global model to participating farms.

Farmers can upload crop leaf images through a web application to identify whether a crop is healthy or affected by a disease represented in the training dataset. The application can also display confidence scores and basic disease information.

Key Features

1. Federated learning across multiple farms.
2. CNN-based crop disease classification.
3. Local training using farm-specific images.
4. FedAvg-based global model aggregation.
5. Privacy-aware agricultural data processing.
6. Disease prediction with confidence scores.
7. Web-based interface for farmers.
8. Model performance monitoring and visualization.

5. Implementation Procedure

1. Dataset Collection: Collect crop leaf images from public datasets such as PlantVillage or suitable farm-specific datasets.

2. Data Preprocessing: Resize images, normalize pixel values, remove corrupted files, and prepare the images for model training.

3. Farm Client Simulation: Divide the dataset into separate partitions representing multiple farms with different crop distributions.

4. Model Development: Build a CNN model or use transfer learning with architectures such as MobileNetV2 or ResNet50.

5. Local Model Training: Train the selected model independently on each farm's local dataset.

6. Federated Learning Setup: Use the Flower framework to coordinate training between farm clients and a central aggregation server.

7. Model Aggregation: Apply Federated Averaging to combine local model updates into a global crop disease detection model.

8. Iterative Training: Distribute the global model to the farms and repeat local training and aggregation for multiple rounds.

9. Disease Prediction: Allow farmers to upload leaf images and classify them into supported healthy or diseased categories.

10. Performance Evaluation: Measure accuracy, precision, recall, F1-score, and confusion matrices across different farms.

11. Dashboard Development: Build a web application using Streamlit or Flask to display uploaded images, disease predictions, confidence scores, and model metrics.

12. Testing: Evaluate the model on unseen images and compare federated training with local-only or centralized training baselines.

6. Software Requirements

7. Hardware Requirements

Recommended: Intel Core i5 or equivalent, 16 GB RAM, SSD storage, and a GPU for faster deep learning experiments.

8. Advantages of the Project

1. Farm Data Privacy: Raw crop images can remain within individual farms.
2. Collaborative Learning: Multiple farms contribute to a shared disease detection model.
3. Early Disease Identification: Helps identify supported crop diseases from leaf images.
4. Better Data Diversity: Different farms can contribute images from varied growing environments.
5. Reduced Data Transfer: Avoids routinely transferring entire image datasets to a central server.
6. Scalability: Additional farms and supported crop categories can be added.
7. Accessible Interface: Farmers can obtain predictions through a simple web application.
8. Performance Monitoring: Enables comparison of model performance across farms.
9. Precision Agriculture Support: Can contribute to data-driven crop monitoring and management.