Forecasting Post-Fire Vegetation Recovery Using Deep Learning
Topics:
Keywords: Wildfires, Deep Learning, Vegetation Recovery, Resilience, Disturbance
Abstract Type: Paper Abstract
Authors:
Claire Simpson, University of Colorado - Boulder
Morteza Karimzadeh, University of Colorado - Boulder
Nathan Korinek, University of Colorado - Boulder
Rafael Augusto Pires de Lima, University of Colorado - Boulder
,
,
,
,
,
,
Abstract
Wildfires are becoming larger and more frequent in the western U.S., a pattern caused in part by a changing climate. Because of these significant changes to fire regimes, it is becoming increasingly important to better quantify disturbance and recovery (defined as vegetation regeneration and initiation of successional phases) so that we may measure long-term pattern changes and forecast ecological responses. There is a growing body of work that estimates where and when wildfires will occur, however there is little research on quantitatively predicting–at spatial and temporal scales–what happens to landscapes after a fire. Deep learning–specifically sequence modeling–represents an under-studied approach to assessing post-fire recovery. In this presentation, we will demonstrate the potential and challenges of using a Long Short-Term Memory network (LSTM) for forecasting recovery given past landscape characteristics and climate conditions, variables which can be obtained at scale from remote sensing data. Specifically, we will present a pilot application of a time-series recovery-forecasting model developed on forested areas of Colorado that were impacted by fires between 2010 and 2012. This work advances methods for predicting vegetation recovery patterns post-disturbance and contributes to the expanding body of literature that leverages deep learning in the fields of remote sensing and ecology by using time-series forecasting, which has rarely been tested in post-fire contexts.
Forecasting Post-Fire Vegetation Recovery Using Deep Learning
Category
Paper Abstract