Predicting 100-Year Storms Requires Extrapolation past Historical Weather Records

As rapidly changing climates defy forecasts grounded in traditional historical weather patterns, specialized geospatial artificial intelligence is becoming critical for assessing extreme risk. Quebec-based Geosapiens provides deep learning solutions to the insurance and municipal planning sectors, helping clients quantify risks associated with flooding and wildfires.

This approach also addresses systemic risks associated with general generative AI. The company notes that common deep learning models can suffer from 'hallucinations' because they lack control over the uncertainty of the input data. Geospatial AI aims to quantify this uncertainty, providing a more rigorous basis for high-stakes financial and infrastructural decisions.

The application of these tools differs significantly between sectors. For insurance carriers, the model’s immediate utility is calculating annual premiums based on next year's predicted risk. Conversely, municipal planning requires assessing long-term resilience, demanding an understanding of what a 100-year storm would mean in a future climate change scenario.

While Environment Canada plans to incorporate AI into its forecasts, experts caution that most current models are fundamentally reliant on historical data and may not accurately capture non-stationarity, the fact that physical mechanisms governing weather variables shift over time. For these tools to become reliable for long-term planning, they must develop robust mechanisms to understand how climate variability will change rather than simply predicting what has happened before.