Refactoring Data-Driven Model Selection Code for Improvements in Interpretability, Generality, and Computational Expense

Buchholz et al. used observations of total column carbon monoxide (CO) from the Measurements Of Pollution In The Troposphere (MOPITT) satellite instrument to build a record of monthly anomalies between 2001 and 2016, focusing on 7 biomass burning regions in the Southern Hemisphere and tropics. CO anomalies in each of the regions were modeled using climate indices for influential climate modes. A linear modeling approach was used, where de-trended, de-seasonalized, regionally aggregated CO measurements were taken as the response variable, and the climate index anomaly values (at various time lags) were taken as explanatory variables. Initial analyses were completed in MATLAB using serial algorithms carried out over non-functionalized scripts. We sought to refactor this codebase, with 3 specific improvement goals; first, to improve code interpretability in preparation for public release; second, to improve code generality, so that the techniques and code used in this application can be easily adapted for similar problems; and third, to utilize parallel computing to substantially speed up program executions. During the early phase of this refactoring, data structures and algorithms were selected to work with the parallel computing tools in the MATLAB Parallel Computing Toolbox. When the codebase was sufficiently developed, a series of parallel timing studies were performed to assess the extent of realizable time savings; in general, these savings were substantial.