NO2 ship-plume segmentation with supervised learning on TROPOMI/S5P satellite data

Starting from 2021, new and more demanding NO_x emission standards came into force for ships entering the Baltic and North Sea waters. All methods that are currently used for ships compliance monitoring are expensive and require close proximity to the ship. As a result, a continuous and global execution of new regulations cannot be performed. 

The unprecedentedly high spatial resolution of the Tropospheric Monitoring Instrument onboard the Copernicus Sentinel 5 Precursor satellite (TROPOMI/S5P) allows the visual distinction of NO₂ plumes produced by individual ships of substantial size. In order to have a scalable method for the estimation of NO_x produced by individual ships, however, the detection of plumes needs to be automated.

Here we propose an automated approach for segmentation of NO₂ plumes from individual ships using supervised learning on TROPOMI/S5P satellite data. For each ship, based on local wind conditions, as well as speed and direction (heading) of the ship, we automatically determine a Region of Interest (ROI) - an area, where the plume of the ship is expected to be located. We standardize the size and orientation of ROI, creating a standardized model of a plume - a plume cone. We then divide the plume cone into predefined subregions so that each subregion has a different probability to contain a plume of the ship. Using this information, we train a machine learning model that separates a plume produced by the analyzed ship from a background and plumes of different origin.

All studied machine learning models significantly outperform a benchmark method based on a globally optimized threshold of NO₂ levels. These promising results enable us to make a next step in developing a tool for continuous NO_x emission monitoring of individual ships above the open sea waters.

This work is funded by the Netherlands Human Environment and Transport Inspectorate, the Dutch Ministry of Infrastructure and Water Management, and the SCIPPER project which receives funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement Nr.814893.