%A Hedayati,Faraz %A Bahrani,Babak %A Zhou,Aixi %A Quarles,Stephen L. %A Gorham,Daniel J. %D 2019 %J Frontiers in Mechanical Engineering %C %F %G English %K Firebrand,Machine Learing,Gaussian process,image processing,Wildland fire behavior %Q %R 10.3389/fmech.2019.00043 %W %L %M %P %7 %8 2019-July-12 %9 Original Research %# %! Measurement Framework to Characterize Structural Firebrands %* %< %T A Framework to Facilitate Firebrand Characterization %U https://www.frontiersin.org/articles/10.3389/fmech.2019.00043 %V 5 %0 JOURNAL ARTICLE %@ 2297-3079 %X Generation of firebrands from various fuels has been well-studied in the past decade. Limited details have been released about the methodology for characterizing firebrands such as the proper sample size and the measurement process. This study focuses on (1) finding the minimum required sample size to represents the characteristics of the population, and (2) proposes a framework to facilitate the tedious measurement process. To achieve these goals, several firebrand generation tests were conducted at a boundary layer wind tunnel with realistic gusty wind traces. Firebrands were generated from burning structural fuels and collected in 46 strategically located water pans. The statistical analysis showed that the minimum required sample size based on the chosen statistical parameters (standard deviation, confidence interval, and margin of error) is 1,400 for each test. To facilitate characterizing such a large sample of firebrands, an automated image processing algorithm to measure the projected area of the firebrands was developed, which can automatically detect the edges of the background sheet, rotate the photo if its tilted before cropping, detect edges of firebrands, remove erroneous particles (e.g., ash) and finally measures the projected area. To facilitate the weighing process, a Gaussian process regression was performed to predict the mass based on projected area, traveling distance and wind speed. The model can predict the firebrand mass within 5% error compared to the measurement. This framework and model can provide a probabilistic range of firebrand characteristics over the continuous range of the collection region.