Convective overshooting top detection with MSG SEVIRI, Himawari-8 AHI, and CloudSat CPR data

Abstract:

Overshooting Tops (OTs) are the clouds that penetrate into the tropopause and grow to the bottom of stratosphere at the top layer of cumulonimbus with very strong updraft. Severe weather conditions such as ground lightning, large hail, strong winds, and heavy rainfall can cause in the cumulonimbus clouds with OTs, with turbulence and lightning occurring very frequently in the area near OTs. In terms of aviation operations, OTs are a very important risk factor. According to Federal Aviation Administration, 509 cases of 4,326 weather-related events from 1992 to 2001were caused by turbulences. The detection of OTs is important to predict the degree and location of severe weather conditions such as turbulence, lightning, and thunderstorms. There are two methods widely used to detect OTs with multispectral images. One is the Water Vapor-InfraRed window channel Brightness Temperature Difference (WV-IRW BTD), which uses the differences in brightness temperatures at an infrared channel (about 11 ㎛). The other approach is the InfraRed Window texture (IRW-texture) method, which is based on the characteristics of OTs that appear a pixel group with low temperatures. The typical IRW-texture algorithm uses simple thresholds to detect OTs, whereas this research proposes an advanced approach based on machine learning techniques such as decision trees, random forest (RF), and support vector machines (SVM) with various variables from geostationary satellite data such as MSG SEVIRI (over Africa) and Himawari AHI (over East Asia) so as to improve the detection of OTs. OT and non-OT samples (e.g. other types of clouds such as stratus and cirrus) were extracted using the CloudSat cloud profiling radar (CPR) and SEVIRI (and Himawari) imagery. Results show that RF produced the best performance in detection of OTs yielding an overall accuracy of 98.33% and a false alarm rate of 9.01%. The user’s accuracies of OT and non-OT were similar, whereas the producer’s accuracy of non-OT was higher than that of OT. The results from the machine learning approaches were compared with those from the original IRW-texture algorithm (i.e., from GOES-R ATBD) as well as WV-IRW BTD (i.e., used for SEVIRI). Lightning data were also used for quantitative assessment of detected OTs over East Asia region.