WSTS per-day reliability annotations and observation-aware evaluation protocol

This dataset accompanies the WACV 2027 paper “Unobserved but Burning: How Missing Satellite Labels Blind Wildfire-Spread Models.” It provides per-day reliability annotations and supporting evidence for WildfireSpreadTS (WSTS), covering 607 wildfire events and 13,607 fire-days from 2018–2021. Satellite active-fire maps record detected fire. An empty map can therefore reflect either an absence of fire or a missed observation. This release identifies 1,732 fire-days whose empty maps are considered unreliable evidence of fire absence. It includes 804 interior-gap events, classified as high-confidence unobserved fire (T1), ambiguous (T2), or plausible true pause (T3). T1 and T2 fire-days are flagged; T3 fire-days are not. The release includes daily reliability flags, gap tiers, and attribution reasons; weather summaries; companion VIIRS VNP14A1 evidence; MTBS summaries; and spatial diagnostics. Calendar-date and native-grid information allow users to join the annotations to WSTS. Four year-based splits, target-membership manifests, and standard, flagged-only, and detection-day evaluation rules are also provided. Supporting files include data dictionaries, machine-readable schemas, Croissant metadata, provenance records, checksums, and validation notebooks. The annotations are inferred reliability judgments, not confirmed daily fire observations or reconstructed fire maps. They use information from before and after gaps and are intended for retrospective auditing and offline evaluation. They should not be treated as information already available to a deployed forecasting model. Original WSTS imagery is not redistributed. It is available from the original WildfireSpreadTS dataset on Zenodo, DOI: 10.5281/zenodo.8006177. Please cite both this annotation release and the underlying WSTS dataset. The annotation package is distributed under CC BY 4.0. Underlying datasets and source products retain their respective terms. This work was supported by Google.org research funding to the Societal Computing and Innovation Lab at UC San Diego.