Streamlined carbon footprint computation : case studies in the food industry

Abstract

One of the greatest barriers in product Carbon Footprinting is the large amount of time and effort required for data collection across the supply chain. Tesco's decision to downsize their carbon footprint project from the original plan of 70,000 house brand products to only a small fraction of them exemplifies the tradeoff between cost and good intention. In this thesis, we have merged salient characteristics from several recent works in this area to develop a fast and cheap method to calculate food carbon footprint accurately. We defined sources of uncertainty as data quality, data gaps and cut-off error, and quantified them. Firstly, quick judgment uncertainty was applied to assess data quality, reducing the time and the expertise needed. Secondly, we showed that it is feasible to use averaged proxies in a preliminary carbon footprint calculation to select the inputs with high impact. The analysis was streamlined by getting specific data only for a subset of high impact inputs while leaving the insignificant inputs represented by low resolution averaged proxies. Monte Carlo simulations and analytical solutions were introduced to account for the total variance of averaged proxies. We applied hierarchy structures to organize the existing emission factors to facilitate proxy selection, but found that the hierarchy required either expert knowledge for design or large numbers of emission factors to average out the inconsistencies within the same input types. Lastly, by integrating uncertainty calculation with iterative carbon footprint calculation, we demonstrated convergence of the calculated carbon footprint and its uncertainty results, providing firm support for our techniques of leaving less significant inputs represented by low resolution averaged proxies. The novel contribution of this work is the application of test sets to 1) prove that carbon footprints calculated using the streamlined approach converged quickly to a stable estimate even when the true values were beyond the range of the proxies, and 2) show an adaptive and justifiable way to select the minimal number of high impact inputs for further analysis.