HabitAlp 2.0

Official code repository for the paper "Habitat and Land Cover Change Detection in Alpine Protected Areas: A Comparison of AI Architectures".

Paper: arXiv:2511.00073 Dataset: huggingface.co/datasets/JR-DIGITAL/habitalp2.0

Overview

This repository contains the code and notebooks used to develop deep learning models for habitat mapping in alpine environments. The project explores both change detection and semantic segmentation approaches using multi-temporal aerial imagery and foundation models.

Repository Structure

.
├── notebooks/              # Jupyter notebooks for training and data processing
│   ├── data-*.ipynb       # Data download and preprocessing
│   └── train-*.ipynb      # Model training notebooks
├── src/
│   ├── data/              # Data handling and processing
│   │   ├── datamodules/   # PyTorch Lightning data modules
│   │   ├── datasets/      # Custom dataset implementations
│   │   ├── download/      # Data download utilities
│   │   ├── processing/    # Data preprocessing tools
│   │   └── samplers/      # Custom samplers for spatial data
│   ├── inference/         # Inference scripts and configs
│   │   ├── predict_*.py   # Prediction scripts for change detection
│   │   └── configs/       # Inference configuration files
│   └── trainers/          # Training tasks and utilities
│       └── unet_segmentation/ # U-Net training code
└── env.yml                # Conda environment specification

Getting Started

Prerequisites

• Conda or Mamba package manager
• CUDA-compatible GPU (recommended)

Installation

Option 1: Using Docker (Recommended for Reproducibility)

The easiest way to reproduce our research environment is using Docker:

# Pull the pre-built image from Docker Hub
docker pull hkristen/habitalp_2:v1

# Run with GPU support
docker run --gpus all -it hkristen/habitalp_2:v1

See DOCKER.md for detailed Docker setup instructions, including building locally, using docker-compose, and running Jupyter notebooks.

Option 2: Local Installation with Conda

1. Clone this repository:

git clone https://github.com/hkristen/habitalp_2.git
cd habitalp_2

2. Create the conda environment:

conda env create -f env.yml
conda activate habitalp_2

Data

Download the HabitAlp 2.0 dataset from HuggingFace:

• huggingface.co/datasets/JR-DIGITAL/habitalp2.0

The dataset includes:

• Multi-temporal aerial orthophotos (RGB, CIR)
• Digital elevation models and derived terrain features
• Habitat classification labels
• Change detection annotations

Usage

Training

The notebooks/ directory contains training notebooks for different models:

• train-01-unet-post-classification-change.ipynb - Post-classification change detection with U-Net
• train-02-binary-change-unet.ipynb - Binary change detection
• train-03-multiclass-change-unet.ipynb - Multi-class change detection
• train-04-terratorch-CLAY.ipynb - Training with CLAY foundation model

Note: File paths in the scripts are examples from our setup. Update them to match your local data paths.

Inference

Semantic Segmentation (Post-Classification)

The main inference pipeline supports both Clay (terratorch) and U-Net models with weighted blending for large images:

# Run inference with config file (recommended)
python src/inference/inference.py --config src/inference/configs/clay_data_2020_full_roi.yaml

# Override checkpoint
python src/inference/inference.py --config src/inference/configs/clay_data_2020_full_roi.yaml \
    --checkpoint path/to/your_model.ckpt

# Override ROI
python src/inference/inference.py --config src/inference/configs/clay_data_2020_full_roi.yaml \
    --roi path/to/different_roi.gpkg

# Full extent inference (no ROI clipping)
python src/inference/inference.py --config src/inference/configs/clay_data_2020_full_roi.yaml --roi none

Features:

• Weighted tile blending for seamless predictions
• Support for both Clay foundation models and U-Net
• Configurable via YAML files
• ROI-based inference or full extent

Direct Change Detection

For binary and multi-class change detection between time periods:

# Binary change prediction
python -m src.inference.predict_binary_change

# Multi-class change prediction
python -m src.inference.predict_multiclass_change

Note: Update the file paths in these scripts before running.

WandB Integration

The training scripts use Weights & Biases for experiment tracking. Set your API key:

wandb_key = "YOUR_WANDB_API_KEY_HERE"  # Replace with your WandB API key

Key Features

• Direct CD: Binary and multi-class change detection between time periods
• Post-CLassifcation CD (Semantic Segmentation): Direct habitat classification using foundation models (CLAY, Prithvi)
• Geospatial Data Handling: Custom samplers and datasets for working with large rasters
• Multi-Modal Input: Support for RGB, CIR, and terrain features
• Production-Ready Inference: Scripts for predicting on large images with tiling

Citation

If you use this code or dataset in your research, please cite our paper:

@misc{kristen2025habitatlandcoverchange,
      title={Habitat and Land Cover Change Detection in Alpine Protected Areas: A Comparison of AI Architectures}, 
      author={Harald Kristen and Daniel Kulmer and Manuela Hirschmugl},
      year={2025},
      eprint={2511.00073},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2511.00073}, 
}

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

This work was conducted as part of the HabitAlp 2.0 project for alpine habitat monitoring and conservation.

Contact

For questions or issues, please open an issue on GitHub or contact the authors.