2021 journal article

Characterizing past fire occurrence in longleaf pine ecosystems with the Mid-Infrared Burn Index and a Random Forest classifier


By: W. Wall*, M. Hohmann*, M. Just* & W. Hoffmann n

co-author countries: United States of America 🇺🇸
author keywords: Fire modeling; Prescribed fire; Pinus palustris; Pyrodiversity; Rare plant management; Threatened, endangered and at-risk species
Source: Web Of Science
Added: October 12, 2021

Prior to European settlement the longleaf pine (Pinus palustris) ecosystem covered over 92 million hectares in the southeastern United States. Historically, fire was an important driver of species composition in the longleaf pine ecosystem, but fire exclusion since the early 20th century has led to the degradation of longleaf pine communities and has had a detrimental effect on the large number of rare and endemic species found within this system. Thus, accurate estimates of fire history are important for better informed management of longleaf pine communities. Recently, satellite imagery has been used to identify burned areas. However, results have been inconsistent across physiographic regions and vegetation types (e.g. wetlands under high canopy). We developed a model using Landsat satellite imagery, coupled with a Random Forest (RF) machine learning algorithm, to identify burned areas and estimate the fire history from 1991 to 2019 for Fort Bragg, NC, one of the largest contiguous areas of longleaf pine ecosystem remaining. We calculated six spectral indices from the Landsat band values, including the Mid-Infrared Burn Index (MIRBI) and the change in MIRBI through time (ΔMIRBI), and used them as predictors in our RF model. We used the developed RF model to estimate the fire history for all known populations of 24 rare upland and wetland plant species found on Fort Bragg. We compared our results to a recent continental U.S. fire occurrence dataset, as well as the prescribed fire records from Fort Bragg. The overall AUC (area under the curve) for our RF model (0.74) compared favorably to the continental U.S. dataset results for Fort Bragg (0.69), and was able to capture the reduced fire frequency in wetlands. The most important predictor in our RF model was ΔMIRBI. Depending on the model, individual plant species were estimated to have experienced significant differences in fire frequency relative to the prescribed fire records. For our RF model, we estimated that 50% of wetland and 25% of upland species experienced a lower fire frequency relative to that represented in the prescribed fire records. The burn probability and classification tool generated in this paper provides land managers in the southeastern U.S. with a novel approach for accurately identifying burned areas and estimating local fire frequency across landscapes.