Scalable Image Search and Retrieval in Compressed Earth Observation Data Archives
Earth observation (EO) data archives are explosively growing as a result of advances in satellite systems. This project aims to design and develop learning-based EO data compression and compressed domain analysis systems. We also study the effectiveness of foundation models for these tasks. To this end, we develop a modular LLM agent to automate suitable foundation model selection and adaptation for these tasks. We construct a curated and structured database of EO foundation models, using a schema-guided LLM pipeline to extract relevant metadata from unstructured resources such as papers and repositories. We design an LLM agent capable of interpreting user queries, retrieving and ranking candidate models based on structured constraints, iteratively clarifying requirements, and recommending adaptation strategies. The agent is aimed to dynamically orchestrate the tasks based on user input and resources. This project contributes to more accessible and intelligent use of foundation models, especially in scenarios with evolving requirements, limited resources, or complex trade-offs between performance, modality support, and deployment needs.
Research Group Lead
Postdoctoral Researcher
Doctoral Researcher