Determination of burn severity models ranging from regional to national scales for the conterminous United States.

Joshua J. Picotte, C. Alina Cansler, Crystal A. Kolden, James A. Lutz, Carl Key, Nathan C. Benson, Kevin M. Robertson

Research output: Contribution to journalArticlepeer-review

Abstract

Identifying meaningful measures of ecological change over large areas is dependent on the quantification of robust relationships between ecological metrics and remote sensing products. Over the past several decades, ground observations of wildfire and prescribed fire severity have been acquired across hundreds of wildland fires in the United States, primarily utilizing the Composite Burn Index (CBI) plot protocol. These observations have been coupled to spaceborne passive spectral reflectance indices (e.g. Landsat-derived variations of the Normalized Burn Ratio [NBR]) to produce regression models describing their relationship. Here we develop regression models by vegetation type for multiple vegetation classification systems representing a range of spatial scales, and a decision tree framework for evaluating these regression models. Our overall goals were to determine which scale of ecological classifications provided the best estimate of burn severity from Landsat data and how to choose the best regression model. We aggregated a total of 6280 CBI plots for 234 wildland fires that burned between 1994 and 2017 and produced Landsat-derived NBR and differenced NBR (dNBR) values for each plot. We then calculated best fit linear or higher order regression equations between CBI and NBR/dNBR for each landcover classification system from smallest to largest scale: LANDFIRE Biophysical Settings (BPS), National Vegetation Classification macrogroup (NVC) landcover classifications, Omernick III, II, and I ecoregions, LANDFIRE Fire Regime Groups (FRG), and the entire conterminous United States (CONUS) dataset. The CONUS regression model goodness of fit was moderate (R2 = 0.55, P < 0.001) for dNBR and poor (R2 = 0.30, P < 0.001) for NBR. Within landcover classifications, CBI was better fit by dNBR than NBR. Finer scale regional regression models including BPS (dNBR R 2 ¯ = 0.56 and 0.00–0.83 R2 range; NBR R 2 ¯ = 0.43 and 0.00–0.82 R2 range) and NVC (dNBR R 2 ¯ = 0.55 and 0.15–0.78 R2 range; NBR R 2 ¯ = 0.41 and 0.00–0.79 R2 range) were on average the same or better than the CONUS models for dNBR and NBR, with the strongest fit models exhibiting R2 ≥ 0.70, whereas larger scale regional models R 2 ¯ ranged from 0.28 to 0.5. However, variation in accuracy among landcover types indicate that dNBR and NBR regression models could be used to effectively estimate CBI for future fires in certain regions, while for other regions models may require additional field observations or alternative spectral transformations. Our decision tree schema can be used to help users determine which scale is likely to produce the most accurate results using our models. The CBI regression models developed here, paired with the decision tree, provide users with a simple method to estimate burn severity in units of CBI for any fire within CONUS with moderate to high levels of confidence and provide a template for further development of models with new data going forward. • Development of ground-truthed regression models for burn severity. • Regional were generally better than national burn severity models. • Differenced satellite estimates of burn severity outperformed single date.
Original languageEnglish
Number of pages1
JournalRemote Sensing of Environment
Volume263
DOIs
StatePublished - Sep 15 2021

Keywords

  • UNITED States
  • WILDFIRE prevention
  • VEGETATION classification
  • PRESCRIBED burning
  • WILDFIRES
  • BURN care units
  • LINEAR orderings
  • FIRE management
  • dNBR
  • LANDFIRE
  • Landsat
  • Linear regression
  • MTBS
  • NBR
  • Sigmoid regression

Fingerprint

Dive into the research topics of 'Determination of burn severity models ranging from regional to national scales for the conterminous United States.'. Together they form a unique fingerprint.

Cite this