Using a deep neural network to detect methane point sources and quantify emissions from PRISMA hyperspectral satellite images

Abstract

Abstract. Anthropogenic emissions of methane (CH4) make up a considerable contribution towards the Earth’s radiative budget since pre-industrial times. This is because large amounts of methane are emitted from human activities and the global warming potential of methane is high. The majority of anthropogenic fossil methane emissions to the atmosphere originate from a large number of small (point) sources. Thus, detection and accurate, rapid quantification of such emissions is vital to enable the reduction of emissions to help mitigate future climate change. There exist a number of instruments on satellites that measure radiation at methane-absorbing wavelengths, which have sufficiently high spatial resolution that can be used for detecting highly spatially localised methane 'point sources' (areas on the order of km2). Searching for methane plumes in methane sensitive satellite images using classical methods, such as thresholding and clustering, can be useful but are time-consuming and often inaccurate. Here, we develop a deep neural network to identify and quantify methane point source emissions from hyperspectral imagery from the PRecursore IperSpettrale della Missione Applicativa (PRISMA) satellite with 30-m spatial resolution. The moderately high spectral and spatial resolution as well as considerable global coverage and free access to data make PRISMA a good candidate for methane plume detection. The neural network was trained with simulated synthetic methane plumes generated with the Large Eddy Simulation extension of the Weather Research and Forecasting model (WRF-LES), which we embedded into PRISMA images. The deep neural network was successful at locating plumes with F1-score, precision and recall of 0.95, 0.96 and 0.92, respectively, and was able to quantify emission rates with a mean error of 24 %. The neural network was furthermore able to locate several plumes in real-world images. We have thus demonstrated that our method can be effective in locating and quantifying methane point source emissions in near real time from 30-m resolution satellite data which can aid us in mitigating future climate change.

STARCOP: Semantic Segmentation of Methane Plumes with Hyperspectral Machine Learning Models

“…There have been very recent works [24][25][26] applying deep learning to hyperspectral data with simulated methane plumes. The workshop paper of 25 frames the detection of methane plumes as semantic segmentation and uses the matched filter product generated using data from the on-demand satellite PRISMA.…”

“…They generate artificial plume shapes using the Large Eddy Simulation (LES) 27 and mix these with the background noise of matched filter outputs from the AVIRIS data. Finally, the preliminary work of 26 combines the tasks of semantic segmentation with regression, by sequentially training several models to first segment and later quantify the methane emissions from the PRISMA satellite images. Similarly, as in the other instances, the annotation is made by methane plume simulations using the LES and mixing the generated signal back into the hyperspectral data.…”

Separating and Quantifying Facility-Level Methane Emissions with Overlapping Plumes for Spaceborne Methane Monitoring

“…To tackle this issue, Bovensmann et al (2010) estimates the XCO 2 backgrounds using the entire observations as inputs; Nassar et al (2017) distinct the plume and the backgrounds with a 1 % density cutoff criteria; Kuhlmann et al (2019) proposes a Z test based plume detection algorithm to mask pixels with statistically higher values as the inputs; Varon et al (2018) combines Student's t test with computer vision (CV) based methods to detect plume pixels. Besides, deep learning methods are employed to perform quantification (Jongaramrungruang et al, 2022) or even end-to-end detection and quantification (Joyce et al, 2022).…”