KimberlyNweze/deforestation_transfer_learning

Transferability of an Attention U-Net for Detecting Deforestation in the Congo Basin: A Case Study in South Cameroon # deforestation_transfer_learning Transferability of an Attention U-Net for Detecting Deforestation in the Congo Basin: A Case Study in Cameroon # Datasets ## Baseline - Amazon 4 band Link to download the Amazon.rar file - zenodo.org ## Cameroon raw 4 band satelite imagery Link to Ground Cover - drive.google.com Link to Ground Truth mask - drive.google.com # Repository Structure This repository is organised to support reproducibility, comparison, and documentation of both the baseline and adapted deforestation detection models. ## ChosenContext/ Contains contextual and methodological documentation for the adapted Cameroon study: - Adapted Model Architecture Changes – Details modifications made to the baseline Attention U-Net architecture. - Challenge and SDG Alignment – Discusses project challenges and alignment with relevant Sustainable Development Goals. - Dataset Curation – Describes data sourcing, preprocessing decisions, and ethical considerations. - Cameroon/ – Core implementation for the adapted model, including: * Scripts to generate image tiles * Generated tiles * Training scripts * Trained model weights * Training and evaluation results ## ReplicateBaseline/ Supports full reproduction of the original baseline model: - environment.yml – Conda environment file listing all required dependencies. - Reproduce Baseline – Notebook and scripts to reproduce baseline results from the original repository. - Environment Setup Documentation – Detailed explanation of environment setup and dependency management. - Original Repository – Unmodified files from the original Attention U-Net repository, included for reference and transparency. ## Comparing The Models/ - Contains scripts and outputs used to quantitatively compare the baseline and adapted models, including per-image and aggregate perf …