🌾 Crop Row Detection using U-Net

This project implements a deep learning-based semantic segmentation model to detect crop rows in agricultural field images. It uses a modified U-Net architecture trained on annotated crop row images to identify crop rows with semantic precision.

---

✅ Overview

Traditional image processing techniques struggle with occlusions, lighting variation, and background clutter in field imagery. This project leverages supervised deep learning with U-Net to semantically segment crop rows from RGB images.

---

📁 Dataset

• Total Images: 281
• Image Size: 320 × 240 (RGB)
• Label Format: Grayscale masks with pixel values {0, 255}
• Folders:
  • images/ - RGB input images
  • train_labels/ - Ground truth segmentation masks for training
  • test_labels/ - Ground truth masks for testing (optional)
• Data Augmentation: Includes random flips, rotations, brightness, and channel shift

---

🧠 Model Architecture

• Base Model: U-Net (encoder-decoder)
• Activation: ReLU
• Final Layer: Sigmoid (for binary classification)
• Loss Function: Binary Cross Entropy (BCE)
• Optimizer: Adam
• Skip Connections: Used for preserving spatial resolution

You may optionally use a lightweight variant (SqueezeUNet) for faster training on limited hardware.

---

🏋️ Training

• Epochs: 25
• Batch Size: 4
• Validation Split: 10%
• Early Stopping: Enabled (patience = 3)
• Model Checkpoint: Saves best model as model_for_nuclei.h5
• TensorBoard: Logs training under /tensorboard/

---

📊 Evaluation

• Metric: Mean Intersection over Union (mIoU)
• Thresholding: Prediction probabilities are thresholded at 0.25
• Post-processing:
  • Binary mask generation
  • Clustering for crop row shape extraction
• Encoding:
  • Run-Length Encoding (RLE) for mask compression

---

🖼️ Results

• mIoU Achieved: 0.21
• Moderate performance due to:
  • Small dataset size
  • High variability in lighting/backgrounds
  • Possible label noise

Example predictions and masks can be visualized in the results/ directory (optional).

---

🚀 How to Run

# Step 1: Clone this repository
git clone https://github.com/yourusername/crop-row-detection.git
cd crop-row-detection

# Step 2: Install dependencies
pip install -r requirements.txt

# Step 3: Preprocess and augment data
python augment_images.py

# Step 4: Train the model
python train_model.py

# Step 5: Evaluate on test data
python evaluate_model.py

# Optional: Visualize predictions
python visualize_predictions.py