A Cross-Temporal Benchmark for Weak Gaseous Anomaly Detection in Hyperspectral Imagery

Overview We construct a cross-temporal benchmark for hyperspectral anomaly detection (HAD) under weak gaseous anomalies. Gaseous plumes are semi-transparent and strongly mixed with surface spectra, producing deviations whose amplitude is often comparable to, or even lower than, normal background variations, which makes their detection more challenging than that of conventional solid anomalies. The benchmark comprises an emission scene, Plume_Present, and a matched emission-free control, Plume_Free, acquired by the same sensor over the same area. This paired design separates weak-anomaly response from background false alarms, so a high response on Plume_Present can be attributed to the plume rather than to background fluctuation, while Plume_Free quantifies the false alarms produced by the background alone. Sensor and acquisition Sensor/platform: EnMAP (Environmental Mapping and Analysis Program) hyperspectral imager. Spectral sampling: ~6.5 nm in the VNIR range and ~10 nm in the SWIR range; strong water-vapour absorption bands were removed, and the VNIR and SWIR domains were merged into a single detection task, retaining 206 bands. Spatial resolution: 30 m; scene size after cropping: 100 × 100 pixels. Site: Karaturun East oil and gas field, Mangistau region, Kazakhstan. Plume_Present: acquired on 27 October 2023, during an active well-blowout methane emission event; the reported satellite-retrieved emission rate at this overpass is 28 ± 10 t/h. The plume morphology is consistent with the near-surface wind direction, confirming association with a genuine emission. Plume_Free: acquired on 12 January 2024 over the same area, when the reported emission rate was 0 t/h and no plume structure is present. Emission-event information and emission-rate retrievals are taken from Guanter et al., which documents the 2023 Karaturun East leak using a multi-satellite time series.