Generated from NASA's Your Name in Landsat.

Summary

SHAZAM is a self-supervised change monitoring method demonstrated for hazard detection and mapping in a region of interest (ROI), using multispectral satellite images. SHAZAM is generalisable across different geographical locations, and is able to detect and map a diverse range of hazards without any labelled samples. Some examples of SHAZAM outputs are shown below:

Different hazards captured by the Sentinel-2 satellite (top row) and their corresponding heatmaps made by SHAZAM when monitoring each ROI (bottom row). The hazards are: wildfires, out-of-season snowfall, floods, drought, algal blooms, and deforestation, from left to right.

Library Structure

• shazam/ contains the model architecture (PyTorch Lightning), dataloading and all the helpers.
• preprocessing/ contains a group of sequential scripts that convert the train/test images into train/val/test patches.
• 0_config.py contains all the configuration variables, data dirs, etc.
• 1_train.py trains the model to reconstruct 32x32 patches for a given time of year.
• 2_validate.py re-runs the model on the full-sized training images to output the real image, reconstructed image, and anomaly heatmaps.
• 3_test.py does the same as 2_validate.py, but on the test dataset.
• 4_results.py generates the plots and results tables seen in the paper.

Quickstart (TBC)

Architecture

The SHAZAM method and underlying architecture are shown in the images below - much more detail is provided in the paper.

SHAZAM - An overview of the proposed method. The top left visualises the training stage, the bottom the inference stage for monitoring an ROI, and the top right highlights training data requirements.

SIU-Net model architecture.

Datasets

The Sequoia National Park (SNP) dataset is currently available on Zenodo at this link. Contains wildfires and extreme, out-of-season snowfall.

Sample images from the SNP dataset. The leftmost image is from the training dataset, showing the region in August. the middle image shows wildfires, and the right image shows out of season snowfall in April.

Cite this Work (arXiV)

@misc{garske2025shazam,
      title={SHAZAM: Self-Supervised Change Monitoring for Hazard Detection and Mapping}, 
      author={Samuel Garske and Konrad Heidler and Bradley Evans and KC Wong and Xiao Xiang Zhu},
      year={2025},
      eprint={2503.00348},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.00348}}