SHAZAM: Self-Supervised Change Monitoring for Hazard Detection and Mapping

The increasing frequency of hazardous events due to climate change underscores an urgent need for effective hazard monitoring systems. Current approaches rely on expensive labelled datasets, struggle to distinguish hazardous changes from seasonal changes, or require multiple observations for to confirm when a potential hazardous event has occurred. To address these challenges, this work presents Self-Supervised Change Monitoring for Hazard Detection and Mapping—SHAZAM. SHAZAM uses a lightweight conditional UNet model to generate images of a region of interest (ROI) for any day of the year, allowing for the direct modelling of seasonal changes. A modified structural similarity measure compares the generated images with real satellite observations to compute both image-level anomaly scores, and pixel-level hazard maps. In addition, a seasonal threshold is introduced to further reduce the need for dataset-specific optimisation. SHAZAM was evaluated on four diverse datasets that contain wildfires, burned regions, extreme and out-of-season snowfall, floods, droughts, algal blooms, and deforestation, and achieved F1 score improvements of between 0.066 and 0.234 over existing methods. This was achieved primarily through a higher recall over other models, while using only 473 K parameters. SHAZAM demonstrates superior mapping capabilities through a higher spatial resolution and the suppression of background features for both immediate and gradual hazards. SHAZAM offers an effective and generalisable solution for hazard detection and mapping across different geographical regions and a diverse range of hazards, and highlights the potential of further self-supervised method contributions for hazard detection and mapping.