Analyzing Wildland Fire Smoke Emissions Data Using Compositional Data Techniques

摘要

By conservation of mass, the mass of wildland fuel that is pyrolyzed and combusted must equal the mass of smoke emissions, residual char, and ash. For a given set of conditions, these amounts are fixed. This places a constraint on smoke emissions data that violates key assumptions for many of the statistical methods ordinarily used to analyze these data such as linear regression, analysis of variance, and t tests. These data are inherently multivariate, relative, and nonnegative parts of a whole and are then characterized as so-called compositional data. This paper introduces the field of compositional data analysis to the biomass burning emissions community and provides examples of statistical treatment of emissions data. Measures and tests of proportionality, unlike ordinary correlation, allow one to coherently investigate associations between parts of the smoke composition. An alternative method based on compositional linear trends was applied to estimate trace gas composition over a range of combustion efficiency that reduced prediction error by 4% while avoiding use of modified combustion efficiency as if it were an independent variable. Use of log-ratio balances to create meaningful contrasts between compositional parts definitively stressed differences in smoke emissions from fuel types originating in the southeastern and southwestern United States. Application of compositional statistical methods as an appropriate approach to account for the relative nature of data about the composition of smoke emissions and the atmosphere is recommended.