How AI and Earth observation can help life on our planet

Earth observation (EO) by satellites, aircraft and ground-based sensors has for decades provided a rich stream of data for governments and businesses to take climate action and ensure sustainable growth. Its uses include creation of early-warning systems, detection of methane leaks and illegal mining, and tracking of afforestation and adaptation efforts.

Thanks to new satellites and advanced sensors, the scale and quality of commercially available EO data have risen exponentially in the past decade. While EO satellites alone provide hundreds of terabytes of data per day – and rising – data from GPS-enabled and internet of things (IoT) devices take the scale of data created way beyond the current capacity to analyse it.

Enter artificial intelligence (AI).

The use of AI coupled with low-cost, high-performance computing is set to do for EO data what large language models (LLMs) have done for text, as a new briefing paper produced in collaboration with Deloitte, The Catalytic Potential of Artificial Intelligence for Earth Observation, shows. This is making EO easier to understand and use, with profound impacts:

By infusing consistent, objective measurements into climate-positive action and environmental disclosures, AI will provide more trust and transparency in actions towards a net-zero economy.

Some use cases of AI in Earth observation include: fusing disparate datasets to enable more holistic situation analysis; creating training data for AI/machine learning (ML) systems; identifying causes (such as those from methane emissions – landfills, farms or pipelines); forecasting events such as cyclones and floods; and monitoring changes using large time series datasets.

The briefing paper provides vital insights for taking forward the Forum’s Earth observation work, which supports organizations to contribute to a nature-positive and net-zero economy, set science-based decarbonization targets, and understand their dependencies and impacts on nature.

Here are some key insights:

Going ahead, the convergence of AI with advanced computing will supercharge the discoverability and usability of EO data. For example, new foundation models could accelerate the analysis of large volumes of EO data “on demand”. Since foundation models are trained on a broad set of data, they can be used across a variety of downstream tasks, obviating the need for new models for each use case. This would reduce costs as a single model would support multiple downstream applications.

An early application helped a team of journalists to discover illicit mining in the Amazon. A customized, deep-learning model would have taken months to build, but with the foundation model, the datasets took days to assemble and milliseconds to curate.

With more accessible UI, novice users will be able to explore and experiment with available Earth data. In turn, greater awareness of EO’s capabilities and the questions it can answer will empower lay users to apply EO insights to use cases in their industry.

Two initiatives of the Massachusetts Institute of Technology (MIT) Media Lab, Earth Mission Control and Climate Pocket, provide such enhanced user interface and user experience. They analyse complex EO datasets, transform them into attractive data visualizations and provide actionable insights.

This provides a vital tool for informed decision-making for industry, policy-makers and communities.

Just as generative AI and LLMs are being tailored for and woven into business models today, AI for EO can provide a range of applications, such as monitoring remote infrastructure, assessing climate risk and adaptively managing supply chains.

EO insights can transform core functions and decision-making processes. For instance, an electricity utility could augment its satellite imagery of electrical infrastructure with geotagged data about its component parts, such as substations and transmission lines. Then, the utility would integrate insights about the status and health of its distributed systems into its workflow. Those insights could enable timely and cost-effective decisions to direct preventive maintenance by the organization’s workforce.

AI models need proper governance and safeguards to avoid negative impacts. For example, farm imagery without farmers’ knowledge could inform costly regulations or insurance premiums that negatively impact the same farmers. Biases in AI models may lead to outcomes that are neither equitable nor inclusive for all regions and populations.

In keeping with the Recommendation on the Ethics of Artificial Intelligence signed by all 193 United Nations Educational, Scientific and Cultural Organization (UNESCO) member-states in 2021, incorporating humans-in-the-loop is a basic step to infuse ethical judgment in both research and implementation of AI for EO. Decisions should be guided not only by AI models of Earth data but also by considerations for the economic, cultural and societal consequences of the models’ results and recommendations.

A focus on governance, standards, open-source solutions and business model innovation would ensure equitable adoption of AI for EO on a global scale. In addition to opening up new frontiers of value creation across sectors and industries, it will ensure a sustainable future for Earth.