Maps Classifier: EuroSAT Deep Learning Pipeline

Welcome to the Maps Classifier project! This repository contains a series of deep learning pipelines for classifying satellite images from the EuroSAT dataset. The project demonstrates the evolution of training strategies, model architectures, and best practices in computer vision for remote sensing.

---

🚀 Project Overview

• Goal: Classify land use and land cover from satellite images using state-of-the-art deep learning models.
• Dataset: EuroSAT (10 classes, RGB, 224x224)
• Framework: PyTorch
• Pipelines:
  • V1: MobileNetV2, basic augmentation
  • V2: EfficientNet-B0, extensive augmentation, better logging & visualization
  • V3: GoogleLeNet (Inception V1), advanced augmentation, stratified split, richer metrics (in progress)

---

🏗️ Pipeline Evolution & Behind the Scenes

V1: Baseline (MobileNetV2)

• Script: trainingV1.py
• Model: MobileNetV2 (pretrained, custom head)
• Augmentation: Random crop, flips, rotation, color jitter
• Split: 80/20 random
• Logging: JSON (logging not fully functional)
• Result: ~96.78% validation accuracy
• Key Logic:
  • Custom FolderImageDataset for loading images
  • Simple training loop with early stopping
  • Model checkpointing

V1 Log (excerpt)

sorry, the logging did not work

the validation accuracy best is 96.78%

---

V2: Stronger Model & Augmentation (EfficientNet-B0)

• Script: trainingV2.py
• Model: EfficientNet-B0 (pretrained, custom classifier, early layers frozen)
• Augmentation:
  • RandomResizedCrop (wider scale)
  • More flips (horizontal, vertical)
  • Stronger ColorJitter, RandomGrayscale, RandomAffine
• Split: 80/20 random (reproducible)
• Logging: CSV, JSON summary, plots
• Visualization:
  • Confusion matrix
  • Training history (loss, accuracy, learning rate)
• Model Saving: Best model + frozen inference model with metadata
• Result:
  • Best Validation Accuracy: 97.39%
  • Final Validation Accuracy: 97.11%
• Key Logic:
  • Modular code (separate functions for training, validation, logging)
  • Extensive data augmentation for robustness
  • Improved logging and visualization for analysis

📈 V2 Training History

• Top Left: Loss curves (train/val)
• Top Right: Accuracy curves (train/val)
• Bottom Left: Learning rate schedule
• Bottom Right: Validation accuracy

📊 V2 Log (excerpt)

Click to expand

epoch,train_loss,train_acc,val_loss,val_acc,learning_rate
1,0.468...,84.87,0.247...,91.78,0.001
...
27,0.080...,97.18,0.091...,97.39,0.0005
30,0.077...,97.30,0.091...,97.11,0.0005

📝 Upgrade Notes: V1 → V2

See upgrade‑notes/V1toV2.md for a detailed breakdown of improvements, logic, and results.

---

V3: Advanced Augmentation & Metrics (GoogleLeNet/Inception V1)

• Script: trainingV3.py (in progress)

• Model: GoogleLeNet (all layers trainable, custom head)

• Augmentation:

  • Even more aggressive: AutoAugment, RandAugment, TrivialAugmentWide, RandomErasing
  • Larger rotation, more flips, stronger color jitter

• Split: Stratified 80/20 (ensures class balance)

• Metrics:

  • Accuracy, precision, recall, F1, macro F1, confusion matrix, classification report

• Logging: CSV, JSON (all metrics), improved plots

• Training: Mixed-precision (AMP), deterministic seeding

• Expected Result: >98% validation accuracy

• Key Logic:

  • Stratified splitting for fair evaluation
  • Richer metrics for deeper insight
  • Modular, extensible codebase

📝 Upgrade Notes: V2 → V3

See upgrade‑notes/V2toV3.md for the full plan and rationale.

---

🧠 How It Works: Code Logic Highlights

• Data Loading: Custom PyTorch Dataset classes for robust, flexible image loading.
• Augmentation: Progressive increase in augmentation complexity to improve generalization.
• Model Selection: Each version upgrades the backbone for better feature extraction and transfer learning.
• Training Loop: Early stopping, learning rate scheduling, and checkpointing for efficient training.
• Logging & Visualization:
  • CSV/JSON logs for easy analysis
  • Plots for loss, accuracy, learning rate, and confusion matrix
• Reproducibility: Deterministic seeds and stratified splits (V3) for fair comparison.

---

📂 References & Artifacts

• Upgrade Notes:

  • V1 to V2

  • V2 to V3

• Logs:

  • V1 Log
  • V2 Log

• Visualizations:

  • V2 Training History
  • V2 Confusion Matrix

---

🛠️ Getting Started

1. Install dependencies:

   pip install -r requirements.txt

2. Prepare the EuroSAT dataset:
   • Place the dataset in data/2750/ with subfolders for each class.
3. Train a model:
   • Run one of the training scripts:

     python trainingV1.py
     # or
     python trainingV2.py
     # or (recommended, in progress)
     python trainingV3.py

4. Check logs and visualizations:
   • See the logs/ and models/ folders for outputs.

---

🤝 Contributing

Contributions, suggestions, and issues are welcome! Please open an issue or pull request.

---

This is part of Shipwrecked