Optimising Agricultural Outcomes Through Computer Vision: A Deep Dive into Transfer Learning for Crop Predictions

Welcome to my Final Year Project for my (Hons) Bachelor's in Computing. This project focuses on using advanced computer vision and transfer learning techniques to predict crop yields, specifically aimed at supporting small-scale farms. The project utilizes deep learning models like Convolutional Neural Networks (CNNs) and VGG16 to address challenges related to limited agricultural datasets.

Table of Contents

1. Project Overview
2. Technologies Used
   • Model Architecture
   • Data Preprocessing
3. Dataset Structure
4. Installation
   • Dependencies
   • Installing
   • Executing Program
5. Usage
6. Challenges & Future Work
7. Contributors
8. License
9. Help
10. Acknowledgments

Project Overview

This project explores the use of computer vision and transfer learning for improving crop yield predictions, with a particular focus on small-scale farming operations. By analyzing high-resolution crop images, the project aims to help farmers make informed decisions to improve productivity and resource management.

The project addresses key challenges small-scale farmers face, such as limited datasets and climate variability, by developing a Convolutional Neural Network (CNN) model utilizing VGG16. This pre-trained model is fine-tuned to recognize crop features and predict yields, even with small and diverse datasets.

The project was developed using:

• Python for implementing machine learning models and handling data.
• TensorFlow and Keras for model training and deployment.
• DJI Mini SE Drone for capturing high-resolution field images for analysis.

Technologies Used

Model Architecture

• VGG16 Pre-Trained Model: Utilized for its deep feature extraction capabilities. The model is fine-tuned for agricultural applications, especially in cases with limited training data.
• CNNs: Convolutional Neural Networks, implemented with Keras and TensorFlow, were used for classifying and predicting crop yields from the images.
• Transfer Learning: Transfer learning was applied to adapt the VGG16 model for our specific task, helping improve accuracy with limited agricultural datasets.

Data Preprocessing

• Image Resizing: All input images were standardized to 224x224 pixels to ensure consistency in the dataset.
• Data Augmentation: Techniques such as flipping, rotation, and zooming were used to increase dataset diversity and improve model robustness.
• Normalization: Pixel values were normalized to enhance the model's ability to learn meaningful patterns from the images.

Dataset Structure

The dataset consists of high-resolution images of crops, particularly lettuce, spinach, and grass, collected via a DJI Mini SE Drone. The images reflect various environmental conditions and stages of crop growth, which were essential in training the CNN model.

Key components:

• Raw Images: Unprocessed images, directly captured from the field, include noise from environmental factors like lighting and shadows.
• Processed Images: Standardized and preprocessed images, resized, augmented, and normalized for training and evaluation.

Additionally, external datasets were incorporated:

• Lettuce NPK Dataset: Sourced from Kaggle, containing detailed data on lettuce growth under varying nutrient compositions.
• Spinach Dataset: Specializing in images of spinach crops.
• Grass Dataset: Used to teach the model to distinguish between crops and surrounding flora.

Installation

Dependencies

• Python 3.9+
• TensorFlow 2.8+
• Keras
• Pillow (for image processing)

Installing

1. Clone the Repository:

   git clone https://github.com/DanielGallagher02/Final_Year_Project.git

2. Install Required Python Packages:

   pip install -r requirements.txt

3. Set Up Dataset: Place your drone-captured images in the data/raw folder.

Executing Program

1. Run Preprocessing Script: To resize and normalize the images:

   python preprocess_images.py

2. Train the Model: After preprocessing, train the CNN with:

   python train_model.py

3. Evaluate the Model: After training, the model will evaluate its performance and generate predictions for crop yield:

   python evaluate_model.py

Usage

• Once the model is trained, you can use it to predict crop yields by running the prediction script:

  python predict_crop_yield.py --image_path /path/to/new/image.jpg

• The prediction script will output the crop yield prediction based on the new image.

Challenges & Future Work

Challenges

• Data Scarcity: The project relied on limited data, a common issue for small-scale farms. Transfer learning helped, but gathering larger datasets is necessary for better accuracy.
• Environmental Variability: High variability in crop conditions across different farms presented challenges for model generalization.

Future Work

• Enhanced Model Generalization: Future iterations will focus on improving the model's ability to generalize across diverse crop types and environmental conditions.
• Real-Time Deployment: I aim to integrate real-time predictions using TensorFlow Lite, allowing farmers to monitor crop yields on-site using mobile devices.

Contributors

• Daniel Gallagher - Lead Developer and Researcher
• Dr. Kevin Meehan - Mentor for my thesis document in the 1st semester
• Vini Vijayan - Project Supervisor and mentor for my thesis document in the 2nd semester

License

• This project is licensed under the MIT License - see the LICENSE.md file for details.

Help

If you encounter any issues during the setup or execution of the program, ensure:

• Python and required packages are installed.
• Dataset is correctly placed in the working directory.
• Dependencies are installed via requirements.txt.

Acknowledgments

Special thanks to:

• Atlantic Technological University for supporting this research and DJI Mini SE Drone technology used to capture high-resolution field images.