From Pixels to Language: The Evolution of AI for Earth Observation

Kurzfassung

Earth Observation (EO) provides an unprecedented amount of information about our planet, but transforming these observations into meaningful and actionable information increasingly relies on Artificial Intelligence. This talk provides an overview of the evolution of AI for Earth Observation, from task-specific machine learning and deep learning approaches for classification, segmentation, regression, and object detection to the emergence of large-scale foundation models.

Particular attention is given to recent developments in multimodal and generative EO foundation models, including Prithvi-EO and TerraMind, and to the growing integration of vision and natural language. The talk introduces the key concepts behind Vision-Language Models (VLMs), from vision-language alignment to visual instruction tuning, and discusses the LLaVA architecture as an example of a general-purpose visual assistant.

The potential and challenges of adapting VLMs to Earth Observation are illustrated through a case study on flood understanding, showing how domain-specific adaptation can support interactive interpretation of satellite imagery. Finally, the talk discusses emerging research directions, including hyperspectral Vision-Language Models and agentic AI systems for Earth Observation.