🌱 Soil Classification with Deep Learning

NOTE : SOIL-CLASSIFICATION-MODEL.IPYNB CONTAINS THE CODE FOR BOTH TRAINING AND INFERENCE

A high-performance deep learning model for classifying soil types from images with 95%+ F1 score.

VIDEO-2025-05-25-12-12-47.mov

📋 Overview

This project uses computer vision to classify soil samples into four categories:

• Alluvial soil
• Black Soil
• Clay soil
• Red soil

Perfect for agricultural tech, environmental monitoring, and soil science applications!

🔍 Dataset Structure

The dataset contains:

• Training images of soil samples
• Test images for prediction
• Labels CSV with image IDs and soil types

🛠️ Installation

# Clone this repository
git clone https://github.com/punkkkkkkk/soil-classification-model_annam.git

# Install dependencies
pip install torch torchvision albumentations pandas scikit-learn pillow tqdm ranger-adabelief

💻 How It Works

🧠 Model Architecture

• Base Model: ResNet50 pretrained on ImageNet
• Custom Head: Added dropout layers and an intermediate layer for better feature learning
• Training Strategy: Freeze early layers, train later layers for soil-specific features

🏋️ Training Process

Our approach uses:

• 5-fold Cross-Validation: Ensures robust performance across different data splits
• Data Augmentation: Flips, rotations, brightness adjustments to prevent overfitting
• Advanced Optimizer: RangerAdaBelief for faster convergence
• Cyclic Learning Rate: Helps escape local minima and find better solutions
• Early Stopping: Saves the best model based on minimum F1 score across classes

🔮 Prediction with Ensemble

For maximum accuracy:

• Model Ensemble: Combines all 5 fold models for prediction
• Test-Time Augmentation: Uses multiple views of each image for robust predictions
• Temperature Scaling: Calibrates confidence scores

📊 Performance

The model achieves:

• 95%+ F1 Score on the validation set
• High Accuracy across all soil types
• Robust Generalization to unseen images

📝 Code Structure

1. Cell 1: Libraries and dependencies
2. Cell 2: Data loading and fold preparation
3. Cell 3: Model, optimizer, scheduler, and loss function setup
4. Cell 4: Transforms, dataset class, and dataloaders
5. Cell 5: Training and validation functions
6. Cell 6: Full training loop with cross-validation
7. Cell 7: Ensemble inference with TTA and performance evaluation

👨‍💻 Key Features

• Cross-validation strategy for robust results
• Advanced augmentation techniques
• Optimizer tuning for better convergence
• Ensemble prediction with test-time augmentation
• Detailed performance metrics for each soil class

⭐ Results

Soil Type	F1 Score
Alluvial soil	0.97
Black Soil	0.95
Clay soil	0.96
Red soil	0.97
Overall	0.96

---

🚀 How to Run This Project

Prerequisites

• Python 3.7+ installed
• CUDA-compatible GPU recommended (for faster training)
• Access to the soil classification dataset

Setup

1. Clone the repository:

   git clone https://github.com/punkkkkkkk/soil-classification-model_annam.git
   cd soil-classification
   

2. Install dependencies:

   pip install -r requirements.txt
   

3. Prepare your environment:

   • For Jupyter: jupyter notebook
   • For Python script: Make sure your working directory contains the code files

Dataset Structure

Ensure your dataset is organized as follows:

/dataset

/train # Training images

/test # Test images

train_labels.csv # CSV with image_id and soil_type columns

Running the Model

Using Jupyter Notebook

1. Open soil-classification-model_annam.ipynb in Jupyter
2. Execute cells sequentially
3. Modify parameters in cell 6 to adjust training (epochs, batch size, etc.)
4. The final cell will generate submission.csv with predictions

Expected Outputs

• Trained models saved to /models directory as .pth files
• Training metrics printed to console
• F1 scores for each soil class displayed after training
• Final predictions saved to submission.csv

Troubleshooting

• Memory errors: Reduce batch size in cell 4 and cell 6
• CUDA errors: Try torch.cuda.empty_cache() or restart kernel
• Import errors: Ensure all dependencies are installed correctly
• Dataset errors: Verify your image paths and CSV structure match the expected format