Towards Improved Monitoring of Combustion Emissions: An Enhanced Combustion Detection Approach Using Machine Learning and NASAs Black Marble Nighttime Lights Product Suite

Observations from the Visible Infrared Imaging Radiometer Suite (VIIRS) have been utilized to detect combustion events such as gas flaring, biomass burning and estimate the associated greenhouse gas emissions. Monitoring these occurrences and their activity levels globally is necessary to fully quantify emissions with accuracy, assess their impact on Earths climate, and track the climate mitigation efforts towards the Paris Climate Agreement Goal. While estimations of combustion emissions from gas flaring and biomass burning have become mature, assessing the potential errors and uncertainties associated these estimates is a crucial, but missing piece for transparent emission reporting suggested by the IPCC guidelines. However, such assessments of errors and uncertainties has been challenging due to the use of common underlying data and lack of evaluation data. We propose a machine learning approach that leverages NASAs Black Marble nighttime lights product to detect the signature of combustion events, that appear distinctly as anomalies by learning a multispectral model of the daily top-of-atmosphere data. In addition to the thermal bands, our method also includes data from the Day/Night Band (DNB) that captures light emissions from these events, which are observed to be more sensitive to weaker signals. We use an unsupervised learning approach that models the non-anomalous pixels and detects deviations caused by both anomalous thermal and light emissions. Unlike existing VIIRS combustion products that use a thermal-signal only approach, the use of an additional DNB-only signal improves our detection significantly, especially of weak anomalies extending over a small area, such as gas flares. We assess the performance of the approach by comparing our results to VIIRS Nightfire, which has been used for determining the spatial extent of gas flare emissions in EDGAR. By providing independent detections of combustion events, our approach is expected to contribute to error and uncertainty assessment in existing combustion emission estimates. Moreover, the improved detection is expected to enhance the localization of combustion events from daily satellite data, enable more timely tracking of its changes, and relate them to emission datasets for accurately assessing their climate and societal impacts.