MATOPIBA Fire Detection Quality Dataset: Five Labeling Strategies for FIRMS Hotspot Classification

Dataset of 604 hand-annotated FIRMS fire hotspot detections from the MATOPIBA region (Maranhão, Tocantins, Piauí, Bahia, Brazil) in the 2022-2024 period, plus a full panel of 3,698 detections for 2022 with four automated label sources. Supports the convention-dependence framework: each labeling source (weak labels via FIRMS×MCD64A1, dNBR auto via Sentinel-2, MapBiomas Fogo, Painel do Fogo CENSIPAM, blind human annotation) is an independent operationalization of the question "is this detection real fire?". Includes 22 engineered features (FIRMS + ERA5 + Sentinel-2 spectral + MapBiomas land use) and a polygonal shapefile of the MATOPIBA states for Leave-One-State-Out spatial cross-validation. Accompanies the master's dissertation of Bessa (PPGEE/UFAM, 2026).