Research in “Multi-Spectral Remote Sensing Image Retrieval Using Geospatial Foundation Models” presented at EGU 2024

For more details on this work and to explore the full paper, visit our Publications section.

The recent paper titled “Multi-Spectral Remote Sensing Image Retrieval Using Geospatial Foundation Models“ marks an advancement in the accessibility and efficiency of satellite imagery analysis. Authored by a team from IBM Research – Europe, including Benedikt Blumenstiel, Viktoria Moor, Romeo Kienzler and Thomas Brunschwiler, this innovative research was presented at the EGU General Assembly 2024.

The paper introduces cutting-edge techniques in the realm of Content-Based Image Retrieval (CBIR) systems that leverage deep learning to efficiently search through vast amounts of satellite imagery and return images similar to a query without the need for annotations. This approach uses Geospatial Foundation Models, which encode multi-spectral satellite data and generalise without further fine-tuning, facilitating a significant reduction in data processing times and improving retrieval accuracy.

Key highlights from the study include:

Implementation of models that maintain retrieval speed and accuracy while reducing the storage and computational demands typically associated with large-scale satellite image analysis.
Achieving an impressive mean Average Precision of 96.62% on BigEarthNet-43 and 44.51% on ForestNet-12, surpassing other RGB-based models in performance.

This research not only enhances the capabilities of remote sensing technologies but also opens new pathways for environmental monitoring, urban planning, and natural resource management, underscoring the potential of AI in transforming how geospatial data is used globally.