Leading the Next Generation of Data-Driven Discoveries (Renewal)

This fellowship seeks to address some of the most significant data challenges we face across many fields in science, and it will do that by combining insights from astrophysics and earth observation, using both public engagement via citizen science, and machine learning/AI aspects of data science, as the unifying threads which make this a coherent and innovative programme.

Astrophysics is at the beginning of a data flood, a proliferation of data from upcoming large sky surveys, one of which will survey the sky repeatedly -- every few nights -- for the next decade. This is a data advance when comparing the deepest data "stack" to our current next best thing. When considering the changing sky, it is a monumental leap forward in the volume, variety, and velocity of data we will have access to.

If we can accurately classify this data, we can make equally major advances in our understanding of the science itself. But first, we need to know if, for example, a galaxy contains a buried, feeding supermassive black hole; or whether a new point of light in the sky is a supernova, and what kind.

The available data following a crisis such as a natural disaster may be a flood, or a trickle, but either way, responders and decision makers need to know where roads are blocked, where buildings are damaged, and where survivors may be sheltering. Often the best data for this task come from satellites, which can survey a large area of the planet at a resolution adequate for the needs of decision makers and first responders.

These "stakeholders" are often most interested in change detection, of a type that bears significant resemblance to astronomers eager to determine what type of distant, transient phenomenon was just observed. The imaging data is technically similar, and some of the algorithms used in one discipline can cross over to the other.

This fellowship has so far capitalised on several areas of symbiosis between astrophysics and earth observation, including those facilitated by the use of citizen science and machine learning as analysis tools. For example, the project has found that the algorithms used to combine the inputs of multiple human data labellers can be applied to combine the predictions of various AI algorithms.

Each algorithm may have its strengths and its limitations: by measuring those tendencies and weighting them accordingly, we can make better overall predictions with both humans and machines (or both, together).

The extension of this project will continue to develop tools for efficient and accurate data classification. This includes labelling data that arrives rapidly but may also be sparse and of varying quality, which is highly relevant to both disciplines. The project will contribute to improved humanitarian aid outcomes in future deployments, advance our knowledge of the evolution of galaxies and supermassive black holes, and facilitate significantly more accurate and uniform measurements of the type of exploding star that serves as a "standard candle", allowing us to more accurately measure the accelerating expansion of the Universe.