RELIFLOOD source data and processed benchmark for reliability-aware satellite flood mapping

This dataset supports the manuscript “Observability limits of satellite flood maps for emergency water-risk decisions”.

The archive contains the processed data and figure-specific source data used to evaluate RELIFLOOD, a reliability-aware satellite flood-mapping framework. The study analyses a multi-continental multimodal benchmark of 446 georeferenced flood scenes and 2,230 modality images, together with 494 external validation units from three independent sources. The released files support analyses of scene-level observability, residual cross-sensor misregistration, boundary-sensitive flood mapping, decision-critical object metrics, external emergency-mapping validation and failure-aware deployment.

The archive includes processed multimodal tiles where redistribution is permitted, flood labels, train/validation/test split files, scene-level metadata, event and external-validation inventories, quality-control flags, residual-shift diagnostics, ΔNMI alignment diagnostics, modality-reliability scores, decision-critical object metrics for settlement edges, roads, riverbank/levee-adjacent corridors and critical infrastructure, release-tier outputs, risk–coverage results and source-data tables for the main and extended figures.

Raw satellite imagery and restricted third-party geospatial layers are not redistributed where their licences do not permit redistribution. For such sources, the archive provides a machine-readable manifest with provider names, product identifiers, acquisition dates, spatial footprints, preprocessing parameters and access links, together with scripts or instructions for reconstructing the processed inputs where permitted by the original data terms.

This archive is intended to enable verification and extension of the results reported in the manuscript. The accompanying code is available separately in the RELIFLOOD code repository described in the manuscript Code availability statement.