Wednesday, April 17 2019
9:00am - 11:00am
Masters Presentation
L1 Estimation of Fire Arrival Time Using Satellite Data

The OpenWFM project is designed to provide tools for efficient fire management including a web-based service in which anyone can interact with fire behavior models. Data on fires is gathered from a series of satellites which make ground detections with different temporal and spatial resolutions. The data come as granules (rectangular images aligned with the flight path) and within granules, data may be missing because of cloud cover and other reasons. We are interested in estimating the state of the fire from such incomplete data. The state of the fire is encoded as the fire arrival time; the fire perimeters are the isolines of the fire arrival time.

Earlier methods for this estimation of fire arrival time used spatial statistical interpolation methods such as kriging. We adopt a Bayesian approach with the prior assumption that the fire propagates at the same rate and in the same direction unless we know otherwise. Our earlier investigation used least squares minimization based on this idea, which, however, sometimes resulted in overshoot artifacts when the fire stops and the fire arrival time surface has a sharp bend. This is caused by the fact that L2 methods in effect attempt to distribute the discrepancy across the domain, and are thus not well-suited for estimation of functions with sharp changes.

Here we present a new approach to the estimation of the fire arrival time based on L1 minimization. L1 minimization concentrates larger discrepancies in smaller areas, and thus can accommodate fast changes without smearing or overcorrections. L1 estimation has been used for purposes such as lasso sparsification in regression and image denoising with preservation of sharp details. As a residual advantage, the scripts are entirely written using open source tools, making it a more cost-effective solution to a public and shared project. These tools are also being utilized, along with the finite-difference approximation of second-order derivatives, to reduce run time without sacrificing accuracy with large, dense problems.

The impacts of this research are in the field of wildland fire mitigation and management. This method can be used for standalone estimation from satellite data and for initialization of wildland fire simulations.
Speaker:Lauren Hearn
Affiliation:Department of Mathematical and Statistical Sciences, University of Colorado Denver

Download as iCalendar