Charting the Future of Earth Observation: Technology Innovation for Climate Intelligence12New era in climate forecasting for adaptation and resilience2 Accurate weather and climate predictions strengthen community preparedness for future climate challenges. Satellite EO data, combined with evolving synergistic technologies, offers valuable insights for forecasting future climate scenarios and their global and local impacts. While AI and ML have been in development for over 25 years, their advancements have only accelerated in the last decade, making AI and ML applications increasingly competitive with traditional numerical methods. Systemic data-driven and digital technologies provide critical insights to enable climate adaptation and build resilience in communities and businesses.12 Looking ahead, several key trends are set to transform weather and climate forecasting even more in the years to come. The move from artisanal to industrial-grade data science will make it easier and more efficient than ever for climate scientists to build and deploy large-scale AI models.13 This will allow AI models to enhance the scalability and efficiency of climate predictions. The rise of customized AI models will allow for more tailored and precise forecasting solutions, addressing the requirements of specific regions and sectors. In addition, the integration of multimodal capabilities, like those seen in the most recent generative AI models, will enable the concurrent processing and analysis of diverse data types,14 improving the quality and reliability of climate models. Technology pipeline: Climate ML-based models Climate ML-based models use ML to process and learn from large datasets derived from traditional physics-based models to rapidly assess the uncertainties and risks of climate extremes. These models mimic physical processes by identifying statistical patterns in model inputs and outputs. Once trained, the ML-based models substitute specific parts of traditional models to reduce the computational demand while maintaining the model’s accuracy. Traditional Earth system models are computationally intensive and can be unfeasible for localized studies due to the need to process petabyte-scale datasets. Climate models that integrate physics-informed ML can overcome these challenges, using multiscale ML-based operators to deliver accurate and fast predictions. For example, traditional weather models allocate 30-80% of their computation time to estimating the movement of solar energy through the atmosphere. An ML-based model trained on existing data can produce the same estimates approximately 1,000 times faster than traditional models. DL has emerged as an innovative approach to climate emulations due to its ability to process large datasets and improved emulation of weather forecasts at high spatiotemporal resolutions and low computational costs. Recent developments in generative adversarial networks (GANs) and diffusion models have created photorealistic imagery of faces, animals, satellite views and street-level flood scenes. However, climate disaster planners and responders need more than photorealistic images; they also need physically reliable information.15 The Climate Pocket is an innovative climate education simulation that harnesses fast ML-based climate models to illustrate the localized flood impacts of global climate policy decisions. It integrates physics-informed ML and EO data with physics- based models to provide reliable, physics-consistent predictions. The tool combines GANs with climate science models to produce photorealistic, synthetic and localized climate estimates for any location on Earth for the next 70 years. An ML-based model trained on existing data can produce the same estimates approximately 1,000 times faster than traditional models.

Charting the Future of Earth Observation 2024