Geospatial Foundation Models for Efficient Retrieval of Remote Sensing Images

This work explores the potential of content-based image retrieval to enable efficient search through vast amounts of satellite data. Images can be identified across multiple semantic concepts without needing specific annotations. We propose to use Geospatial Foundation Models (GeoFM), for remote sensing image retrieval and evaluated the models on two datasets. The GeoFM named Prithvi uses six bands and outperforms other RGB-based models by achieving a mean Average Precision of 61% on ForestNet-4 and 98% on BigEarthNet-19. The results demonstrate that the model efficiently encodes multi-spectral data and generalizes without requiring further fine-tuning. Additionally, this work evaluates three compression methods: i) binary embeddings, ii) trivial hashing, and iii) locality-sensitive hashing. Compression with binarized embeddings isthe best option for balancing retrieval speed and accuracy. It matches the latency of much shorter hash codes while maintaining the same accuracy as floating-point embeddings.