Advanced Forecasting of Fire Ignitions in Alaska's Boreal and Tundra Regions: A Data-Driven Approach

The increasing occurrence and severity of wildfires in the Arctic, particularly in Alaska's boreal and tundra regions, highlight the critical need for advanced predictive models that accurately forecast fire ignitions. For a long time, we have relied solely on the use of physics-based models to improve our chances of accurately modeling areas with scarce data availability. The recent advances in data-driven weather models have opened an exciting area of research as to how these models can be best leveraged operationally in combination with the already established numerical weather models. Areas such as computational efficiency, dynamic ensembles, among others arise as possible benefits from these data-driven models.

This research aims to develop an integrated operational framework that incorporates a variety of environmental, climatic, and human-induced factors contributing to fire ignition in these remote ecosystems together with a quantified overview of the influence of physics-based and data-driven weather prediction models leveraged as input into our modeling framework. The primary focus of this work is on cloud-to-ground lightning (CGL) as a major ignition source, along with other influential factors such as vegetation type, soil moisture, temperature anomalies, and precipitation patterns. We will discuss the performance of various data-driven and numerical weather models using lightning and fire ignition forecasting as our case studies, with a particular focus on lightning and ignition drivers.

We explore the use of deep learning and physics-based models to predict CGL and fire ignition probabilities over 24-hour, 48-hour, and 10-day intervals. Our workflow involves a GPU-accelerated deep learning pipeline that includes: (1) parallel extraction and computation of features from high-resolution weather model outputs, (2) training and prediction of cloud-to-ground lightning and fire ignition probabilities, and (3) continuous feedback loops to monitor forecasting performance throughout the fire season. The framework is tested across multiple temporal windows, including operational scenarios for the 2023 and 2024 Alaska fire seasons. The findings underscore the importance of continuous monitoring and enhanced predictive capabilities to mitigate the impact of wildfires on the Arctic's delicate environment and the communities that rely on its ecosystem services.