Reconstructing savannah wildfire events using a deep learning framework

Le résumé fourni par la source

This research presents a method to produce fire reconstructions with a high spatial and temporal resolution by downscaling geostationary satellite observations using a deep learning segmentation framework. This was achieved by training a U-Net on low earth orbit active fire detections paired with synchronous observations from the geostationary satellite Himawari-9/Advanced Himawari Imager (AHI) resampled to the resolution of the 500 m 0 . 64 μ m RED channel. Geostationary satellites provide a means for continuous monitoring of fire behaviour but are constrained by the low spatial resolution of the infrared channels used to detect active fires. Localising fire activity using information from the higher spatial resolution solar reflective channels of geostationary satellites enables detailed fire progression mapping with a comparable spatial resolution to low earth orbit satellite systems. Six case study fires in the northern Australian savannahs are reconstructed with their lifecycles compared to the burn scar mapped by the Northern Australia and Rangelands Fire Information (NAFI), with F1-scores ranging from 0.80 to 0.96. Model predictions synchronous to VIIRS active fire detections during the selected case study fires were used to test performance during case study events. The results indicate a high positive detection rate (75%) for detections with a fire radiative power (FRP) above 12.4 MW during day-time and 3.1 MW at night-time which degraded with decreasing FRP as the fire signal becomes increasingly difficult to detect within the 2 km instantaneous field of view (IFOV) of the AHI infrared channels against the high day-time background temperatures within the northern Australian savannahs. • Reconstructions of savannah fires produced with a high spatial and temporal resolution. • Geostationary and low earth orbit satellite data combined using deep learning. • 500 m Himawari/AHI RED channel (0.64 μ m) used to resample all inputs to 500 m.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.