Advances in Detection of Overshooting Cloud Tops Within Satellite Imagery Based on Pattern Recognition Algorithms

Hazardous weather such as heavy rainfall, aviation turbulence, damaging winds, large hail, and tornadoes occur typically close to overshooting cloud tops (OT) regions, which form by deep convective updrafts penetrating through the surrounding cirrus anvil into the lower stratosphere. There has been a strong interest in automated detection of OT occurences from satellite observations with potential applications in weather and climate analysis.

A probabilistic multispectral OT detection approach has been recently developed in support of the GOES-R Aviation Algorithm Working Group. The algorithm improves upon those previously developed by avoiding fixed criteria and incorporating multiple pattern recognition tests to yield an aggregated rating corresponding to a cumulative OT probability. This approach removes the need for fixed brightness temperature (BT) thresholds and does not limit the algorithm's applicability to warm-season and/or tropical events, which makes the new method suitable for generation of a global climate data record of hazardous storm detections. The algorithm has shown a reliable detection not only with high spatial resolution MODIS, VIIRS, and AVHRR imagery, but also with current generation geostationary imagery used in diagnosis and nowcasting of hazardous convective weather.

The proposed detection method uses 1 km/pix resolution imagery in 0.65 um visible (VIS) band and/or the 4 km/pix in the 10.7 um BT band. The algorithm consists of a series of pixel-level tests each producing a relative confidence rating with the final rating being a sum accumulated after all tests. For the BT imagery, the locally coldest spots are analyzed in terms of their shape, prominence of the BT drop, BT uniformity within the anvil cloud, and others. VIS imagery is used as a supplementary field mainly because of the lower OT contrast within the anvil near solar noon and the requirement for the algorithm to also operate during the night time. The visible analysis consists of two major branches, the daytime analysis (in the absence of shadows) and the low-sun conditions (deep shadows). The former exploits the fact that OT regions show up as textured areas surrounded by uniformly bright fields of anvil clouds. The uniformity is verified by histogram analysis and the textures are analyzed by Fourier transformation with subsequent filtering of characteristic spatial frequencies. The low-sun branch is based on detection of shadows produced by protruding OT clouds with subsequent Fourier analysis that targets cloud ripples caused by gravity waves, which appear prominently in VIS imagery at high solar zenith angles.

The obtained cumulative confidence rating is finally combined with atmospheric stability and tropopause information derived from numerical weather analysis data. The final probabilistic OT detection product is then validated using a diverse sample of OT and non-OT anvil regions identified by a human from the 250 m/pix resolution MODIS imagery. The described method has demonstrated a high potential for discriminating between OT and anvil and identified 69% of human OT identifications given a particular probability threshold with a false detection rate of 18%. The false detection rate drops to 1% when OT-induced texture detected within visible imagery is used to filter the IR-based Probability product.