IN038-06
Fire arrival time estimation from satellite observations using machine learning

Tuesday, 15 December 2020: 16:20
Virtual
Angel Farguell1, James Haley1, Kyle Hilburn2, Adam Kochanski3, Derek V. Mallia4 and Jan Mandel1, (1)University of Colorado Denver, Mathematical and Statistical Sciences, Denver, CO, United States, (2)Cooperative Institute for Research in the Atmosphere/Colorado State University, Fort Collins, CO, United States, (3)San Jose State University, Department of Meteorology and Climate Science, San Jose, CA, United States, (4)University of Utah, Atmospheric Sciences, Salt Lake City, UT, United States
Abstract:
The United States has entered a new era of increasing wildfire frequency and intensity, which has culminated in 14 billion-dollar wildfire disasters over the past two decades. The average cost of these events is $5B, based on data from NOAA National Centers for Environmental Information. The landscape has become more fire-prone due to urban development and climate change, resulting in rising fire-suppression costs. Responding to this challenge requires addressing multiple phases of the disaster management cycle: mitigation, preparedness, response, recovery, and resilience. By pushing the boundary of high-resolution coupled modeling, we are developing improved wildfire modeling tools to address the first three phases of this cycle.

The WRF-SFIRE system is a coupled model that captures the interactions between weather, fire spread, fuel moisture, and smoke. This includes processes such as the dynamical effects of fire-induced winds that are important in driving fire behavior and the radiative effects of smoke shading that play a role in the vertical distribution of smoke downwind of the fire. Accurate simulation of the fire behavior and smoke production requires that fire locations be accurately initialized in the system, and this presentation focuses on that problem.

WRF-SFIRE is capable of initializing fires from dispatch data and satellite data, and we use the concept of fire arrival time to assimilate these observations. This involves calculating a space-time surface that separates burning from non-burning points. Thermal anomaly and active fire satellite products provide a categorical mask, where every pixel is classified as either unknown, non-fire, or fire, and maybe further distinguished by a confidence level. We will evaluate multiple innovative machine learning techniques to estimate fire arrival time using satellite data, and the results are validated using infrared fire perimeters sampled by aircraft for different fire events. We will discuss the pros and cons of low-Earth orbiting (MODIS, VIIRS) versus geostationary (GOES ABI) sensor data for this application and identify gaps in current capabilities. This work demonstrates the benefit of near real time scientific data for wildfire disaster applications.