AI for operational methane emitter monitoring from space

TL;DR

A global dataset of thousands of super-emission events is compiled for training and evaluation, demonstrating that MARS-S2L can skillfully monitor emissions in a diverse range of regions globally, providing a 216% improvement in mean average precision over a current state-of-the-art detection method.

Abstract

Mitigating methane emissions is the fastest way to stop global warming in the short-term and buy humanity time to decarbonise. Despite the demonstrated ability of remote sensing instruments to detect methane plumes, no system has been available to routinely monitor and act on these events. We present MARS-S2L, an automated AI-driven methane emitter monitoring system for Sentinel-2 and Landsat satellite imagery deployed operationally at the United Nations Environment Programme's International Methane Emissions Observatory. We compile a global dataset of thousands of super-emission events for training and evaluation, demonstrating that MARS-S2L can skillfully monitor emissions in a diverse range of regions globally, providing a 216% improvement in mean average precision over a current state-of-the-art detection method. Running this system operationally for six months has yielded 457 near-real-time detections in 22 different countries of which 62 have already been used to provide formal notifications to governments and stakeholders.

Figures (22)

• Figure 1: Characteristics of the MARS-S2L dataset In total 53,309 images containing 4,230 emissions over 707 distinct emitters globally are included in the dataset. Maps show the locations of (b) all included sites globally, the Permian Basin (c) and Turkmenistan (d). Examples of plumes in multiband-multipass imagery together with the corresponding hand-annotated masks are shown in (a) with three examples of clear plumes (Samples 1,2,3) and three examples of poorly-defined plumes (samples 4,5,6).
• Figure 2: MARS-S2L model architecture and deployment (a) the architecture and inputs of the MARS-S2L model. The multispectral bands from the current and previous overpass the site together with the MBMP image and wind information are used as inputs. A cloud mask is first generated using the CloudSEN12 model. All data is then fed into MARS-S2L which outputs the probability that each pixel is part of a methane plume. (b) shows the operational deployment process. At 06:30 every morning any new Sentinel-2 and Landsat images are downloaded and predictions generated. These are then shown in the PlumeViewer where analysts inspect each alert and provide details of events over known assets to case managers to issue notifications.
• Figure 3: Model detections during six months of operational deployment. MARS-S2L detections are shown for (a) non-notified plumes and (b) notified plumes. Examples of three notified events are shown in for plumes in (c) the US, (d) Syria and (e) Thailand. MARS-S2L successfully identified 457 emissions in near real time leading to 62 notifications.
• Figure 4: Global performance results Global results showing (a) Performance of MARS-S2L and CH4Net as a function of flux rate. MARS-S2L achieves excellent performance and substantially outperforms CH4Net. An example of successful plume identification for a site in Algeria is shown below, with (b) predicted probability superimposed on RGB imagery, (c) corresponding hand-annotated mask, (d) multiband-multipass image and (e) CH4 enhancement.
• Figure 5: Case study results. Case study results for (a) the Permian Basin, (b) Turkmenistan and (c) offshore platforms. For each case an example time-series of predictions for one site is shown on the left with CH4 enhancement, multi-band multi-pass image, hand-annotated mask and model prediction. Histograms of model predictions over all sites for images with and without a plume are shown on the right. For all three cases MARS-S2L provides skillful identification of plumes.