RESEARCH
Prospective challanges froM 2025 tBC
Synthetic aperture radar (SAR) is a powerful Earth observation tool that can monitor ground conditions through clouds and rain. However, the variability and complexity of the SAR data has meant that bespoke ML analysis workflows have been needed each time. Can foundational models, trained on terabytes of SAR data, and tuned using self-supervised learning (SSL), generalise to a wide range of tasks and locations? A true global foundation model that would generalise to broad use-cases, such as mapping wildfires or landslides, requires integrating information from SAR along with optical and infrared (IR) data. This integrated approach could harness the strengths of these diverse data sources to enhance predictive accuracy and reliability.
How can machine learning assist in better preparing for and mitigating natural disasters such as hurricanes, floods and wildfires? Can we improve ML workflows to offer timely warnings of rapid-onset events such as tornadoes, or use ML tools to prepare for gradual changes such as drought? Can we use ML to inform a digital twin of our Earth systems to confidently support long- and short-term decision making?
Large Language Models (LLMs) encode huge amounts of information and are capable of performing useful work, when driven by simple prompts. They are poised to revolutionise science - if we can teach them to produce reliable and safe output. LLMs are now being trained to understand more than just text and images, and being augmented with the ability to call external systems, like web-search, for example. The possibility for LLMs to natively understand geospatial data and drive complex analysis tools could pave the way to build a powerful Earth observation assistant usable by almost anyone. However, with great power comes great responsibility and the assistant must be built with foolproof error checking and validation. Can we train skilled LLM systems to understand Earth observation data, geographic data and Earth science contexts, to deliver an assistant that prioritises explainability, uncertainty and trust?
Urgent and bolder action on climate change is needed. Focus needs to be on reducing harmful emissions rather than applying offsets. Simply put, carbon dioxide, soot, methane and other greenhouse gases must be kept in the lithosphere to keep our planet habitable. Can machine learning help today with the roll-out of clean energy technologies and systems, support enhanced climate modelling, better optimise sequestration or help with monitoring and enforcing emissions standards, using orbital and ground-based sensors?
Networked ML payloads onboard clusters of Earth Observation satellites can collaborate to deliver high-cadence change detection capabilities. This on-orbit technology will enable autonomous satellite tasking and constellation reconfiguration in response to detected events, support very-high-cadence monitoring and provide compressed data products for ground-based Digital Twin Earths (DTEs) in real-time. The DTE will be a large machine learning system offering the ability to predict the future state of Earth under different scenarios - for example, a spreading wildfire, or flooding. This challenge will build on our existing in-orbit ML-Payloads (e.g., RaVAEn) converting them for use in constellations, while in parallel building a specialised downstream model capable of processing the sparse outputs of the in-orbit system alongside more conventional satellite data products.