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

TL;DR

SHAZAM introduces a self-supervised change monitoring framework that jointly models normal seasonal dynamics and hazard detection in satellite image time series. It uses a Seasonally Integrated UNet (SIU-Net) to translate baseline ROI patches to any day of the year, enabling direct separation of hazardous changes from seasonal variations, with anomaly scoring via a SSIM-based SDIM measure and a seasonally adaptive threshold. The method achieves improved hazard detection (higher recall) and high-resolution hazardmaps across four diverse datasets while remaining lightweight (≈473K parameters), outperforming cVAE, RaVAEn, and COLD in most metrics. This approach offers a practical, generalisable solution for real-time hazard monitoring and mapping in varied geographical contexts, with potential extensions to multi-region coverage and onboard processing.

Abstract

The increasing frequency of environmental hazards due to climate change underscores the urgent need for effective monitoring systems. Current approaches either rely on expensive labelled datasets, struggle with seasonal variations, or require multiple observations for confirmation (which delays detection). To address these challenges, this work presents SHAZAM - Self-Supervised Change Monitoring for Hazard Detection and Mapping. SHAZAM uses a lightweight conditional UNet to generate expected images of a region of interest (ROI) for any day of the year, allowing for the direct modelling of normal seasonal changes and the ability to distinguish potential hazards. A modified structural similarity measure compares the generated images with actual satellite observations to compute region-level anomaly scores and pixel-level hazard maps. Additionally, a theoretically grounded seasonal threshold eliminates the need for dataset-specific optimisation. Evaluated on four diverse datasets that contain bushfires (wildfires), burned regions, extreme and out-of-season snowfall, floods, droughts, algal blooms, and deforestation, SHAZAM achieved F1 score improvements of between 0.066 and 0.234 over existing methods. This was achieved primarily through more effective hazard detection (higher recall) while using only 473K parameters. SHAZAM demonstrated superior mapping capabilities through higher spatial resolution and improved ability to suppress background features while accentuating both immediate and gradual hazards. SHAZAM has been established as an effective and generalisable solution for hazard detection and mapping across different geographical regions and a diverse range of hazards. The Python code is available at: https://github.com/WiseGamgee/SHAZAM

Figures (9)

• Figure 1: This Sentinel-2 SITS example shows a bushfire (wildfire) event in Sequoia National Park, California, progressing from normal conditions (left), to active bushfires (middle), and to burned regions (right). The corresponding hazard detection heatmaps were created by the proposed method, SHAZAM.
• Figure 2: 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.
• Figure 3: SIU-Net model architecture.
• Figure 4: Network blocks.
• Figure 5: Visualising the seasonal encodings for the day of the year, $t$.