Investigations on the Effective Use of a Neural Network Classifier for Source Attribution in the Indianapolis Urban Domain

Abstract:

The problem of attributing atmospheric fossil fuel signatures to specific source contributors—i.e. sectors of the economy—from measured greenhouse gas concentrations is an important one to solve so that policy makers can best make educated regulatory decisions towards reducing any desired sector's carbon emissions. One approach for solving this attribution problem involves applying atmospheric inversions to flask measurements collected on communication towers. The precision of the inversions’ attributions can be augmented through the addition of prior knowledge of source locations. Further, by including multiple species’ concentrations in an analysis, unique signatures can be identified to help distinguish between nearby sources with similar characteristics. The analysis presented here expands upon this framework by incorporating a neural network classifier known as a self-organizing map to facilitate the identification of source signatures. Indianapolis is chosen as the urban domain. The Indianapolis Flux Experiment (INFLUX) flask collection towers serve as the measurement sites, and the source locations from Hestia—the high-resolution CO2 emission product—are used to spatially identify source areas of interest. The accuracy and sensitivity of the source attribution model is tested in the theoretical framework using Observing System Simulation Experiments (OSSEs). Based on the OSSEs, the theoretical requirements are presented in terms of the characteristics of the gas species and the relative importance of their uniqueness related to individual sectors of economic activities. A first practical application of the method is presented for the INFLUX flask campaign, analyzing an unprecedented amount of flask samples at multiple site locations available over a single urban area.