ISC High Performance 2021 – once again virtual due to the ongoing pandemic – is swiftly approaching. In contrast to last year’s conference, which canceled its in-person component with a couple months’ notice, ISC21 is taking the virtual stage with the benefit of a year’s worth of hindsight on a startlingly changed world. It’s fitting, then, perhaps, that the conference’s first keynote will be delivered by one of the world’s experts in bird’s-eye views: Dr. Xiaoxiang Zhu.
Zhu, a professor of data science and Earth observation at the Technical University of Munich, received the 2018 PRACE Ada Lovelace Award for her work to translate dense Earth observation satellite data into advanced models and rich, descriptive datasets. This line of research will also be the cornerstone of Zhu’s ISC21 keynote, “Artificial Intelligence and Data Science in Earth Observation,” which she described in an interview with HPCwire ahead of the event.
“As the name tells, we are observing the Earth,” Zhu said, describing the various sensors and delivery systems used to accomplish this feat. “As you can imagine, this kind of data could be used to extract geoinformation that could be relevant for a lot of applications – environmental-related issues, urban-related issues. … My work is basically the information retrieval from the acquired data and [providing] the geoinformation that is needed for different applications.”
“Maybe 20 years ago, if we talk about Earth observation, everybody [was] happy if they could have satellite data or certain data of a specific geographic region, right?” Zhu continued. “But the game-changer was [the European Space Agency] ESA’s Copernicus program, which is actually the second-biggest space program of ESA. And they have the Sentinel … satellite fleet with different sensors to basically [act] as a kind of a mapper providing measurements on a weekly basis on a global scale.”
This, Zhu explains, was an “evolutionary change that, more or less, happened in the past few years,” enabling open and free access to dozens of petabytes of Earth observation data (set to become hundreds of petabytes by 2030) and opening the door for new perspectives and opportunities in Earth system modeling. “We are able to provide global geoinformation of the urbanization process or give crucial information for climate change,” Zhu said, “which was not possible before this time.”
Zhu’s team has been leveraging a machine learning-accelerated workflow to process this data, using a technique similar to tomographic reconstruction to build the first-ever global-scale, 3D model of the world’s buildings.
“In order to basically do this kind of high-precision information retrieval, we obviously need the support of high-performance computing,” Zhu said. “One city typically has 10,000 times 5,000 pixels for a region of 10km by 5km. For every pixel, we are solving a very computationally expensive optimization problem. So this is only doable – if we want to really process on a global scale – with high-performance computing.”
“For this reason,” she continued, “we’re actually getting around 46 million CPU hours from the Leibniz Supercomputing Center in order to be able to process large-scale data and finally – hopefully – in a year or two to get this kind of global model open and accessible.”
Zhu’s team’s work extends far beyond the city maps, though: projects including charting population density using a combination of satellite and social media data; working with crisis response teams to assess building damage in the wake of events like 2020’s Beirut explosion; and using satellite data to detect the outlines of glaciers and ice floes.
And, of course, there’s the elephant in the room: COVID-19. Zhu discussed ongoing work to screen social media from around the world, using natural language processing to analyze sentiments and map how people responded to local and regional COVID-19 mitigation policy announcements – and, potentially, to map emerging COVID-19 outbreaks. “By modeling everything that has happened during COVID-19, the purpose of research is probably less in having an impact for the current situation,” Zhu cautioned. “Rather, [it is] to give hints for possible future events. This is where we try to make impacts.”
For much of this work, Zhu has sought – and obtained – access to a variety of HPC resources, including at the aforementioned Leibniz Supercomputing Center (where her team makes use of the 19.5 Linpack petaflops SuperMUC-NG supercomputer) and at the Jülich Supercomputing Centre, where they work with GPU clusters to power their machine learning research.
“Since we are a university, [if] a project is really scientifically interesting and also has a big social impact, it is not difficult to get computing resources,” Zhu said. “Although it is a little bit cumbersome that you always have to have jobs submitted, and the waiting, and so on – and if you really have very large-scale processing you need to turn it into tiny pieces such that the defined framework can work pretty well. But these are the practical things that we can find solutions [for] – generally I’m very happy as a scientist with the resources provided from the HPC community.”
Zhu’s extensive experience has instilled in her an eager optimism for coming advances in the field. “In my lab, we are actually very much excited … to work on the fundamental challenges that are related to machine learning and Earth observation – how to combine domain expertise with a data-driven type of approach,” she said, going on to outline a range of research questions: making these models more explainable, addressing ethical issues around model and data bias, quantifying uncertainty and more.
These data and models, Zhu says, can be a major force for good in the coming era. “We really try to tackle societal grand challenges like urbanization, climate science, [the] UN’s Sustainable Development Goals,” she said. “The whole community should, much more, join forces in order to make really big breakthroughs in these fields – and here, also, experts from the community of HPC are very welcome to join forces with people from other fields.”
Dr. Zhu will expand on her work and the rapidly shifting world of Earth observation data in her 40-minute ISC21 keynote, titled “Artificial Intelligence and Data Science in Earth Observation.” The keynote session will be held the morning of Monday, June 28th and will be accessible online to those who register for ISC21 at this link.