Optimal tiling strategy for neurally compressed Earth Observation embeddings

Determine an optimal tiling scheme for creating and serving neurally compressed Earth Observation embeddings—including tile sizes and boundaries—that balances downstream-task granularity with computational and storage efficiency, given that large standard EO tiles are too coarse for neural models and smaller tiles increase the number of embeddings to compute and store.

Background

The implementation section discusses practical challenges for deploying neurally compressed embeddings in EO platforms. Standard EO tile sizes (e.g., 110×110 km for Sentinel‑2) are generally too large for neural network inputs, and embeddings are treated atomically, limiting subdivision options. Conversely, smaller tiles increase the number of embeddings to compute and store, raising operational overhead.

The authors explicitly state that the best tiling strategy is not yet determined, underscoring a concrete operational open problem for platform design and data delivery of neural feature compression.