https://watch.eeg.cl.cam.ac.uk/videos/watch/0431b465-1b45-433e-bb0e-de1c3eb9d7db Abstract: Geospatial Foundation Models such as TESSERA enable large-scale geospatial analysis through general-purpose embeddings, but they lack spatial context and are costly to store due to their high dimensionality. We propose a framework that incorporates spatial information while efficiently compressing TESSERA embeddings using a Masked Autoencoder (MAE). By treating embedding patches as images, we evaluate Vision Transformer, Swin Transformer, and UNet architectures as encoder backbones. Across multiple configurations and datasets, compressed latent embeddings match the performance of original representations in classification on TreeSAT-AI-TS. The MAE achieves up to a 10× compression ratio, with reconstructions retaining an average of 85% performance across segmentation, classification, and regression tasks. Furthermore, MAE pretraining improves downstream UNet performance. These results demonstrate that integrating spatial context via MAE yields more compact and effective GFM embeddings, supporting scalable Earth observation applications. Bio: Zejia Yang is a Part II undergraduate student in Computer Science at the University of Cambridge, supervised by Frank Feng and Prof. Srinivasan Keshav. Her Part II dissertation focuses on MAE pretraining of spatial feature extractors for TESSERA embeddings. Fri, 15 May 2026 20:22:43 GMT https://validator.w3.org/feed/docs/rss2.html PeerTube - https://watch.eeg.cl.cam.ac.uk https://watch.eeg.cl.cam.ac.uk/client/assets/images/icons/icon-512x512.png https://watch.eeg.cl.cam.ac.uk/videos/watch/0431b465-1b45-433e-bb0e-de1c3eb9d7db All rights reserved, unless otherwise specified in the terms specified at https://watch.eeg.cl.cam.ac.uk/about and potential licenses granted by each content's rightholder.