🌲 Amazon Deforestation Detection & Land-Cover Mapping

An end-to-end deep learning and remote sensing pipeline that benchmarks convolutional neural networks (CNNs) on the EuroSAT dataset and deploys the best classifier to map regional land-use and detect deforestation in Rondônia, Brazil using multitemporal Sentinel-2 imagery.

---

📊 Workflow Overview

graph TD
    A[EuroSAT Dataset] -->|Train & Benchmark| B[CNN Model Zoo]
    B -->|Dynamic Accuracy Scan| C[Best Classifier: ResNet-18]
    C -->|Sliding-Window Inference| D[Sentinel-2 Composites]
    D -->|Stitch Grid Predictions| E[Land-Cover Maps]
    E -->|Temporal Comparison| F[Change Detection Matrix]
    F -->|Filter Forest -> Non-Forest| G[Deforestation Overlay]
    G -->|Grid-Downsampling| H[Hansen Reference Map]
    H -->|Quantitative Assessment| I[Validation Reports]

Loading

---

🏆 CNN Benchmarking Results (EuroSAT)

We evaluated seven CNN architectures on the EuroSAT RGB dataset. The table below highlights the models that matter most for the final story, showcasing the selection of ResNet-18 as our deployment model:

Model Architecture	Test Accuracy	Parameters	Model Size (Disk)	Key Characteristics
ResNet-18	96.04%	11.2M	43 MB	Optimal balance of accuracy, speed, and size (Selected)
EfficientNet-B0	94.10%	4.0M	16 MB	High efficiency, low footprint
GoogLeNet	90.10%	6.0M	23 MB	Multi-scale inception processing
AlexNet	84.10%	57.0M	218 MB	Historical architecture, heavy fully-connected layers
LeNet-5	74.20%	0.06M	0.25 MB	Extremely lightweight, limited capacity

---

🗺️ Notebook Map

The repository is structured as a clear, sequential step-by-step workflow:

1. 01_EDA.ipynb: Explores class balance and spectral properties of the EuroSAT dataset.
2. 02_Preprocessing.ipynb: Formulates train/val/test splits and data augmentation.
3. 03_LeNet.ipynb - 08_EfficientNet.ipynb: Trains and evaluates individual CNN architectures.
4. 09_Model_Comparison.ipynb: Benchmarks parameters, disk space, and inference speed.
5. 10_Sentinel2_Inference.ipynb: Performs sliding-window land-cover mapping on Ji-Paraná (Region 1).
6. 11_Deforestation_Detection.ipynb: Detects changes between temporal pairs in Porto Velho Frontier (Region 2).
7. 12_Validation.ipynb: Validates predicted deforestation against Hansen reference masks in Porto Velho (Region 2).
8. 13_Project_Demo.ipynb: End-to-end showcase executing the complete pipeline in Ariquemes Corridor (Region 3).

---

📁 Repository Structure

├── data/                  # EuroSAT raw files and region composites (gitignored)
├── src/                   # Core pipeline modules
│   ├── models/            # CNN architecture implementations
│   ├── dataset.py         # EuroSAT datasets and PyTorch loaders
│   ├── training.py        # Trainer loops and early stopping
│   ├── evaluation.py      # Validation and testing metric loops
│   ├── inference.py       # Sliding-window patch generation and stitching
│   ├── change_detection.py# Temporal comparison, transition matrix, and deforestation masks
│   ├── regions.py         # Bounding boxes and regional composite fallbacks
│   └── utils.py           # Plottings heatmaps and export helpers
├── notebooks/             # Exploratory and deployment notebooks
├── train.py               # CLI tool to train CNN architectures
├── evaluate.py            # CLI tool to test checkpoints and dump metrics
├── run_demo.py            # CLI tool running the complete Ariquemes pipeline
└── download_region.py     # CLI tool to download Sentinel-2 imagery via GEE

---

🚀 Getting Started

1. Installation

# Clone the repository
git clone https://github.com/your-username/deforestation-detection.git
cd deforestation-detection

# Set up virtual environment and install dependencies
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

2. Training a Model

python train.py --model resnet18 --epochs 15 --lr 0.001 --batch_size 32

3. Evaluating a Checkpoint

python evaluate.py --model resnet18 --checkpoint outputs/checkpoints/resnet18/best_model.pth

4. Running the End-to-End Deforestation Pipeline

python run_demo.py --model resnet18 --checkpoint outputs/checkpoints/resnet18/best_model.pth

---