Charting the Future of Earth Observation 2024

Charting the Future of Earth Observation: Technology Innovation for Climate Intelligence13Photorealistic visualization of AI-generated flood risk FIGURE 5 Technology pipeline: Geospatial AI foundational models Foundation models can use ML to process and analyse large amounts of satellite EO data, which allows them to capture complex patterns within the climate system. These models develop a comprehensive understanding of atmospheric dynamics by learning from diverse datasets, enabling them to perform a wide range of tasks with accuracy and efficiency. They enable localized studies that would be unfeasible through traditional Earth system models and could significantly reduce computational demands while maintaining precision. Developed by a team of researchers at Microsoft, Aurora is a cutting-edge AI foundational model that can forecast global levels of six key air pollutants – carbon monoxide, nitrogen oxide, nitrogen dioxide, sulphur dioxide, ozone and particulate matter – in under a minute. These predictions extend up to five days and are achieved at a significantly lower computational cost compared to traditional models. By using large EO datasets to develop general- purpose representations, Aurora is set to generate high-resolution weather forecasts for up to ten days, surpassing both traditional simulation tools and specialized DL models. According to its developers, Aurora is about 5,000 times faster than state-of-the- art numerical integrated forecasting systems.16 As an example of a public-private partnership (PPP), NASA and International Business Machines (IBM) Research have developed the Prithvi-weather-climate model, an AI foundation model aimed at applications for shorter-term weather and longer-range climate forecasts. The geospatial foundational model, developed with NASA satellite data, will be publicly available on the AI platform Hugging Face. It will be among the largest models uploaded to Hugging Face while being one of the first open-source ML models built in partnership and collaboration with NASA. Trained using Harmonized Landsat Sentinel-2 (HLS) satellite data, the model was originally built for flood and burn scar mapping. However, it can also track deforestation, predict crop yields and monitor greenhouse gases. The MERRA-2 dataset, known for integrating space-based aerosol observations, provided the training foundation for the model, which has since been fine-tuned to enhance climate model resolution twelvefold.17 The model employs downscaling techniques to generate high-resolution outputs from coarse data at low computational costs. Climate ML-based models and geospatial AI foundation models – what, why and how do they relate? Climate ML-based and foundation models serve distinct but complementary roles in climate forecasting, harnessing ML capabilities to address different aspects of climate data processing. Climate ML-based models approximate specific selected physical processes within traditional climate models by identifying and using the statistical patterns in large datasets. They are effective for local studies and quick, high-resolution predictions at low computational costs. They are most useful for applications that need to iterate quickly and often, such as forecasting localized extreme weather. At the other end of the spectrum, geospatial AI foundational models are designed to detect high- level patterns from large amounts of satellite EO data. These models are trained on many different datasets in a self-supervised way, learning patterns from data without labels, enabling the models to understand atmospheric dynamics. geospatial AI models perform well across a wide range of applications and are suitable for both global and local scales, providing high-resolution views of climate events. They can create highly accurate models of global patterns while remaining computationally efficient by incorporating advanced ML methods to process and normalize data. Note: Pre-disaster EO satellite imagery (left) generated flooding post-disaster using physics-informed Generative Adversarial Networks (GANs) (right). Source: Earth Science. (n.d.). Earth Intelligence Engine. https://climate-viz.github.io.