Ship tracks show the indirect effect of ship emissions on clouds and provide an opportunity to quantify these effects. Generally, ship tracks have been studied in specific regions or for short periods of time as manually delineating them in satellite data is a time-consuming process. However, a team at the University of Oxford developed a method using AI to automatically detect ship tracks. Working with the researchers, NEODAAS were able to apply this to a global archive of satellite data, making detection of over one million ship tracks across the world’s oceans for the past 20 years possible.
Stringent ship emission regulations were introduced in 2020 and this study was able to capture the corresponding reduction in ship tracks (graph below). This research establishes the first clear evidence of a global cloud response to environmental regulations.
The total number of ship tracks by ocean region between 2003-2021 (inclusive), overlain by the global mean shipping emissions of SOx where available.
The ship track dataset generated for this study is the largest dataset (to date) produced by our MAssive GPU Cluster for Earth Observation (MAGEO) system and required processing over 250TB of raw data from the MODIS sensor flown on the NASA Aqua satellite. The dataset is being made available via CEDA for use by other researchers.
MAGEO is a cluster of 5 NVIDIA DGX-1 max-Q nodes, which provided a total of 40 Tesla V100 GPUs (200k CUDA cores), 400 CPU cores and 2.5 TB of RAM. It is operated is part of our NEODAAS Artificial Intelligence Service – which can be accessed through a NERC grant or via direct access.
Lead author Duncan Watson-Parris, a senior researcher within the Climate Processes group at the University of Oxford said “As a climate scientist, it was invaluable having a dedicated facility to help scale up my research analysis to the hundreds of Tbs of data we wanted to process - I couldn’t have done it otherwise. The team worked with me to understand what was needed, parallelise my scripts and manage the resulting deluge of data and made the whole process painless!”
Angus Laurenson, NEODAAS data analyst added “it was great to work with Duncan to deploy his model at scale. We put MAGEO through its paces on a genuinely ‘big’ dataset and I think it is a neat demonstration of how deep-learning can be utilized to conduct research that is otherwise infeasible.”
Computing resources for running the inference and funding for AL and DC was provided through the NERC Earth Observation Data Analysis and AI Service (NEODAAS).