Mapping burned areas using dense time-series of Landsat data

REMOTE SENSING OF ENVIRONMENT

卷:198

Article

Hawbaker, Todd J.¹  Vanderhoof, Melanie K.¹  Beal, Yen-Ju¹  Takacs, Joshua D.¹  Schmidt, Gail L.²  Falgout, Jeff T.³  Williams, Brad³  Fairaux, Nicole M.¹  Caldwell, Megan K.⁴  Picotte, Joshua J.⁵  Howard, Stephen M.⁶  Stitt, Susan¹  Dwyer, John L.⁶ 

[1] US Geol Survey, Geosci & Environm Change Sci Ctr, POB 25046,MS 980, Lakewood, CO 80225 USA

[2] US Geol Survey, Stinger Ghajjarian Technol, Earth Resources Observat & Sci Ctr, 47914 252nd St, Sioux Falls, SD 57198 USA

[3] US Geol Survey, Core Sci Analyt Synth & Libs, POB 25046,MS 302, Lakewood, CO 80225 USA

[4] Univ Colorado Boulder, Dept Ecol & Evolutionary Biol, Boulder, CO 80309 USA

[5] ASRC Fed InuTeq LLC, Earth Resources Observat & Sci Ctr, 47914 252nd St, Sioux Falls, SD 57198 USA

[6] US Geol Survey, Earth Resources Observat & Sci Ctr, 47914 252nd St, Sioux Falls, SD 57198 USA

关键词: Burned area;    Essential climate variable;    Landsat;    Wildfire;   

DOI  :  10.1016/j.rse.2017.06.027

来源: Elsevier

【 摘 要 】

Complete and accurate burned area data are needed to document patterns of fires, to quantify relationships between the patterns and drivers of fire occurrence, and to assess the impacts of fires on human and natural systems. Unfortunately, in many areas existing fire occurrence datasets are known to be incomplete. Consequently, the need to systematically collect burned area information has been recognized by the United Nations Framework Convention on Climate Change and the Intergovernmental Panel on Climate Change, which have both called for the production of essential climate variables (ECVs), including information about burned area. In this paper, we present an algorithm that identifies burned areas in dense time-series of Landsat data to produce the Landsat Burned Area Essential Climate Variable (BAECV) products. The algorithm uses gradient boosted regression models to generate burn probability surfaces using band values and spectral indices from individual Landsat scenes, lagged reference conditions, and change metrics between the scene and reference predictors. Burn classifications are generated from the burn probability surfaces using pixel-level thresholding in combination with a region growing process. The algorithm can be applied anywhere Landsat and training data are available. For this study, BAECV products were generated for the conterminous United States from 1984 through 2015. These products consist of pixel-level burn probabilities for each Landsat scene, in addition to, annual composites including: the maximum burn probability and a burn classification. We compared the BAECV burn classification products to the existing Global Fire Emissions Database (GFED; 1997-2015) and Monitoring Trends in Burn Severity (MTBS; 1984-2013) data. We found that the BAECV products mapped 36% more burned area than the GFED and 116% more burned area than MTBS. Differences between the BAECV products and the GFED were especially high in the West and East where the BAECV products mapped 32% and 88% more burned area, respectively. However, the BAECV products found less burned area than the GFED in regions with frequent agricultural fires. Compared to the MTBS data, the BAECV products identified 31% more burned area in the West, 312% more in the Great Plains, and 233% more in the East. Most pixels in the MTBS data were detected by the BAECV, regardless of burn severity. The BAECV products document patterns of fire similar to those in the GFED but also showed patterns of fire that are not well characterized by the existing MTBS data. We anticipate the BAECV products will be useful to studies that seek to understand past patterns of fire occurrence, the drivers that created them, and the impacts fires have on natural and human systems. Published by Elsevier Inc.

【 授权许可】

Free