A book from 1984 bears testimony to Dr Carsten Brockmann’s long interest in artificial intelligence (AI).
Today he is applying this knowledge at an ever-increasing pace to his other interest, climate change. “What was theoretical back then is now becoming best practice,” says Brockmann, who believes AI has the power to address pressing challenges facing climate researchers.
Orbiting our planet with sensors pointing Earthwards are over 700 Earth observation satellites, transmitting hundreds of terabytes each day to downlink stations. Processing and extracting useful information is a huge data challenge, with volumes rising quasi-exponentially. And it’s not just a problem of the data deluge: our climate system, and environmental processes more widely, work in complex and non-linear ways. AI, and in particular machine learning, is helping to meet these challenges, as accurate knowledge about global climate change becomes more urgent.
ESA’s Climate Change Initiative (CCI) provides the systematic information needed by the UN Framework Convention on Climate Change. By funding teams of scientists to create world-class accurate, long-term, datasets that characterise the Earth’s changing climate system, the CCI is providing a whole-globe view. Derived from satellites, these datasets cover 21 GCOS-defined ‘Essential Climate Variables’ from greenhouse gas concentrations to sea levels and the changing state of our polar ice sheets. Spanning four decades, these empirical records and underpin the global climate models that help predict future change.
AI algorithms – computer systems that learn and act in response to their environment – can improve detection rates in Earth observation. For example, it is common to use the ‘random forests’ algorithm, which uses a training dataset to learn to detect different land cover types or areas burnt by wildfires. In machine learning, computer algorithms are trained, in the statistical sense, to split, sort and transform data to improve dataset classification, prediction, or pattern discovery.
“Connections between different variables in a dataset are caused by the underlying physics or chemistry, but if you tried to invert the mathematics, often too much is unknown, and so unsolvable,” says Brockmann. Thanks to Earth observation projects like the CCI programme, we have huge, multi-dimensional datasets to explore. “For humans it’s often hard to find connections or make predictions from these complex and nonlinear climate data,” he says.
Neural networks are used to take account of cloud cover by the ESA Climate Change Initiative Ocean Colour project when generating global monthly composite map of chlorophyll concentration (August 2018)
AI helps by building up connections automatically. Exposing the data to AI methods enables the algorithms to “play” with data and find statistical connections. These so-called ‘concolutional neural network’ algorithms have the potential to resolve climate science problems that vary in space and time. For example, in ESA’s CCI programme, scientists in the aerosol project need to determine changes in reflected sunlight due to the presence of dust, smoke and pollution in the atmosphere, called aerosol optical depth.
Dr Thomas Popp, who is science leader for the aerosol project, thinks there could be further benefits by using AI to retrieve additional aerosol parameters, such as their composition or absorption from several sensors at once. “I want to combine several different satellite instruments and do one retrieval. This would mean gathering aerosol measurements across visible, thermal, and ultraviolet spectral range, from sensors with different viewing angles.” He says solving this as a big data problem could make these data automatically fit together and be consistent.
‘Explainable AI’ is another evolving area is the field which could help unveil the physics or chemistry behind the data, says Carsten Brockmann, who is in the CCI’s Ocean Colour science team.
“In AI, computer algorithms learn to deal with an input dataset to generate an output, but we don’t understand the hidden layers and connections in Neural Networks: the so-called black box,” says Brockmann. “We can’t see what’s inside this black box, and even if we could, it wouldn’t tell us anything. In explainable AI, techniques are being developed to shine a light into this black box to understand the physical connections.”
This post was sponsored by the European Space Agency. See our editorial guidelines for what this means.