Short-term fire hazard forecasting
Orion Lab has developed deep learning methods for daily wildfire danger forecasting. We approach daily fire danger prediction as a machine learning task, using historical Earth observation data from the last decade to predict next-day’s fire danger. Our work aims at producing fire hazard forecasts at large scales and with better spatial resolution compared to other existing methods and operational tools.
We implement a variety of Deep Learning (DL) models to capture the spatial, temporal or spatio-temporal context and compare them against a Random Forest (RF) baseline.
Our DL-based proof-of-concept provides national-scale daily fire danger maps at a much higher spatial resolution than existing operational solutions.
We use explainable Artificial Intelligence methods on top of our DL models to gain more insights about the predictions. This allows us to answer important questions, for example, which are the main drivers? are there meaningful space/time patterns that increase fire risk? are there ways to know how sure the model is about a prediction? what is the impact of a change of a predictor on the fire risk?
Seasonal fire forecasting
Our team is developing a first-of-its-kind prototype system that predicts the seasonal burned areas sizes for Europe using global environmental variables combined with Terrestrial Ecosystem Modeling, and simulates their impact on local fire regimes. We represent the Earth as a graph and we model the underlying relations that exist between fire drivers in different areas of the planet - this is an innovative approach in the field of Earth System Science.
