I was excited the be invited to present a TED AI talk in Vienna in October 2024 – the first TED conference fully focusing on Artificial Intelligence organised in Europe. In the talk, I presented my DPhil research on using machine learning models for the task of detection of disaster events from space. My last work is focusing on the detection of methane leak events in images captured by new hyperspectral satellites by NASA. The automated detection of these leaks is an important task in the context of Climate Change – after CO2, methane is the second largest contributor to Climate Warming.
However, detecting methane is difficult, as it is invisible in the visible spectrum (and for human vision). Special instruments are usually needed – best technology to do so includes the so-called hyperspectral satellites. In the current landscape of satellite technology, many new hyperspectral missions were launched recently (NASA’s EMIT, Planet’s Tanager-1, the MethaneSat by EDF, planned missions by ESA such as CHIME and others), but the upcoming quantity of data from these missions will be overwhelming (already, the NASA’s EMIT sensor captures up to 600GB of data per day – and the future missions are estimated to produce 10 fold increase).
Current methods for detecting methane leaks often relied on manual annotation by domain experts. With the upcoming deluge of data, this will soon no longer be feasible. Machine learning models I have been working on, are making automated detections of these methane leak events. The research was initially published in the prestigious journal Nature Scientific Reports (https://www.nature.com/articles/s41598-023-44918-6) and since its publication was added into the Journal’s Top 100 of 2023.
Currently, I am finalising my DPhil research in a Thesis titled “Intelligent decision making on-board satellites”, but a pre-print of my last work on methane detection is also publicly available at ArXiv (https://arxiv.org/abs/2410.17248). This work focuses on potentially deploying these machine learning models to run directly on-board of satellites. This would enable much faster detection of super-emitter events (these are very large leaks, where every hour of emitting matters). Current typical processing pipelines take from several days to weeks for an event to be detected (the main bottleneck being the reliance on manual detection and waiting for the data to be downlinked to the ground). Models running on-board of satellites would enable detections only in matter of minutes and would allow much faster reaction on the ground. This technology would not only benefit scientists and environment protection agencies, but also actors inside the Oil and Gas industry selling natural gas (which is primarily composed of methane) – as it’s in their interest not to lose the product they are selling due to technical malfunctions of their pipelines.
I have gained prior experience with deploying machine learning on-board of satellites during my earlier research collaborations with the aerospace company called D-Orbit. In 2023, our model for detection of disaster events on-board of satellites (called the RaVAEn system) was tested on a real satellite by D-Orbit, the ION-SCV 004 (launched during the “Dashing through the Stars” mission). This demonstration also marks the first instance of training deep neural networks on-board of satellites in Space. The insight from this prior work was very valuable for informing the requirements for developing models for future AI powered hyperspectral satellites. Finally, these models are currently being implemented for deployment at the United Nations unit for methane detection, the Methane Alert and Response System (MARS).
To watch the full TED talk, visit https://www.ted.com/talks/vit_ruzicka_how_ai_helps_us_track_methane_from_space
