Resolve the DFC2020-S1 geographic-overlap assessment

Establish whether geographic proximity between DFC2020-S1 test samples and SSL4EO-S12 v1.1 pretraining tiles affects reported performance by obtaining or otherwise recovering coordinate metadata for the DFC2020-S1 dataset.

Background

The paper tests whether geographic overlap between the SSL4EO-S12 v1.1 pretraining corpus and downstream evaluation sets inflates transfer-performance gains. The analysis identifies geographically proximate pretraining tiles for EuroSAT-S1, BigEarthNet-S1, and Flood-S1, then repeats pretraining after removing the overlapping tiles.

DFC2020-S1 cannot be assessed directly because its released dataset lacks coordinate metadata. Although the de-collided model produces essentially unchanged DFC2020-S1 performance, the authors describe this only as indirect reassurance and explicitly retain the missing-coordinate issue as an open limitation. A direct geographic-overlap assessment therefore remains unresolved.

References

DFC2020-S1 remains an open limitation for the overlap analysis itself, since its dataset release provides no coordinate metadata to check test-set proximity directly; its unchanged performance under the de-collided checkpoint above offers indirect reassurance only.

— Data-Efficient ConvNeXt Pretraining for High-Resolution Sentinel-1 GRD SAR Representation Learning  (2610.21993 - Natsume et al., 15 Oct 2026) in Section 4.8, “Geographic Data Overlap Analysis” (Section~\ref{sec:leakage})