🛰️ Deforestation Detection App

This application uses a transformer-based ChangeFormer model to detect deforestation in the Brazilian Amazon using Sentinel-2 satellite imagery. Developed as a final year project, it processes 4-band (RGB + NIR) .tif images from 2020 and 2021 to generate binary change masks and overlay predictions, achieving an F1-score of 0.9986 and IoU of 0.9972 on validation data.

🌍 Overview

Model: Custom ChangeFormer with a Vision Transformer encoder, Feature Difference Module, and Deconv Decoder.
Data: Sentinel-2 Level-2A imagery (10m resolution) and PRODES deforestation labels.
Interface: Gradio-powered app for drag-and-drop uploads and real-time visualization.
Purpose: Scalable and interpretable tool for land monitoring, policy-making, and conservation.

✨ Features

Upload two .tif images (2020 and 2021) with 4 bands: B2 (blue), B3 (green), B4 (red), and B8 (NIR).
Outputs:
- ✅ Raw RGB base image (2021)
- ✅ Binary change mask (black/white)
- ✅ RGB + red overlay for deforested regions
- ✅ Comment on % of detected change
Handles large images with patch-based tiling, normalization, and stitching.

⚙️ Setup & Usage

🔧 Prerequisites

Python 3.8+
Required libraries: torch, torchvision, timm, rasterio, numpy, pillow, gradio

📦 Installation

git clone https://github.com/manuelhorvey/ChangeFormer.git
cd ChangeFormer
pip install -r requirements.txt

Place your trained model checkpoint as:

models/best_model.pth

🚀 Run Locally

python app.py

Then open: http://localhost:7860

🧪 Input & Output

Input

Two 4-band .tif images from the same area:
- One from 2020
- One from 2021
Patches preferred (e.g., 256×256)

Output

RGB image from 2021
Red-highlighted deforestation overlay
Binary change mask
Textual feedback on % change

☁️ Images with <20% cloud cover yield best accuracy.

📊 Project Details

Item	Value
Region	Brazilian Amazon (e.g., APA Triunfo do Xingu)
Dataset Size	19,560 patches (256×256), 4 channels, 2 years
Metrics	F1-score: 0.9986, IoU: 0.9972
Augmentations	Rotations (90/180/270), flips
Future Plans	Web-based monitoring alerts, SAR fusion, forecasting

👥 Authors

Emmanuel Amey
Sammuel Young Appiah
Asare Prince Owusu
Yaaya Pearl Apenu

Based on ideas from Alshehri et al. (2024), IEEE GRSL.

🪪 License

This project is licensed under the MIT License.

---

manuelhorveydaniel
/

ChangeFormer