Detection of Tropical Overshooting Cloud Tops Using Himawari-8 Imagery
Overshooting convective cloud Top (OT)-accompanied clouds can cause severe weather conditions, such as lightning, strong winds, and heavy rainfall. The distribution and behavior of OTs can affect regional and global climate systems. In this paper, we propose a new approach for OT detection by using machine learning methods with multiple infrared images and their derived features. Himawari-8 satellite images were used as the main input data, and binary detection (OT or nonOT) with class
... ty was the output of the machine learning models. Three machine learning techniques-random forest (RF), extremely randomized trees (ERT), and logistic regression (LR)-were used to develop OT classification models to distinguish OT from non-OT. The hindcast validation over the Southeast Asia and West Pacific regions showed that RF performed best, resulting in a mean probabilities of detection (POD) of 77.06% and a mean false alarm ratio (FAR) of 36.13%. Brightness temperature at 11.2 µm (Tb11) and its standard deviation (STD) in a 3 × 3 window size were identified as the most contributing variables for discriminating OT and nonOT classes. The proposed machine learning-based OT detection algorithms produced promising results comparable to or even better than the existing approaches, which are the infrared window (IRW)-texture and water vapor (WV) minus IRW brightness temperature difference (BTD) methods. Satellite remote sensing data have been used for detecting and monitoring OTs. Researchers have proposed various OT detection methods using visible and/or infrared images. Most of the OT detection models are based on infrared images, since infrared imagery can be utilized regardless of image acquisition time, while visible images are only available during the daytime    . There have been two methods widely used for detecting OTs with infrared images. One method is the dual channel difference approach, which uses the brightness temperature (hereafter Tb) difference between water vapor and window channels, documented as the Water Vapor-InfraRed Window channel Brightness Temperature Difference (WV-IRW BTD) in this study (also known as the Dual Channel Difference method)    . However, the method may be inappropriate, since the horizontal advection of stratospheric water vapor not associated with OTs increases the false alarm of OT detection [16, 17] . Furthermore, the threshold used in the method varies depending on the characteristics of the satellite data used, such as spatial resolution and spectral wavelengths [17, 18] . To address this limitation, the second method, the InfraRed Window texture (IRW-texture) algorithm, was developed based on the characteristics of OTs that appear as a group of pixels with low Tb [17, 19] . The IRW-texture method overcomes the over-detection of OTs in the WV-IRW method, as it does not depend on the water vapor distribution in the lower stratosphere and the Tb of a water vapor absorption band  . However, the fixed thresholds used in the IRW-texture method for the characteristics of OTs, such as their size and Tb, would still be insufficient to cover the characteristics of various OT cases  . More recently,  developed a new satellite-based probabilistic OT detection algorithm, producing OT detection results as the probability of occurrence of OTs. Compared to the previous texture-based method, the probabilistic OT detection approach is relatively less affected by the issue of fixed thresholds, but does require additional processing. In , a series of pattern recognition analyses were used to define the anvil cloud extent and assign final OTs with a stepwise rating evaluation method, including a score test, OT shape test, and anvil roundness analysis. Logistic regression was then applied to produce OT probability with numerical weather analysis data. Although the carefully designed processing methods can enhance algorithm performance, the multilevel analyses are time consuming and often produce false alarms. With the rapidly growing volumes of satellite data, a simple but robust detection of OTs is desirable. The goal of this study is to propose machine learning approaches for binary OT detection ("OT occurrence" vs. "no OT occurrence") with class probability, which are dedicated for operational use on Himawari-8 satellite data. The key idea is to find rules or patterns to differentiate OT pixels from nonOT, based on the characteristics of various channels and their spatial features. The objectives of this study are: (1) to develop OT detection models based on machine learning methods including Random Forest (RF), Extremely Randomized Trees (ERT), and Logistic Regression (LR), producing OT results as both binary output (i.e., OT or nonOT) and class probability using various input variables related to the characteristics of OT extracted from Himawari-8 infrared imagery; (2) to evaluate the performance of the OT detection models and examine the contributing input variables to OT classification; and (3) to perform hindcast validation of the OT detection models to assess the reliability of the proposed models. While  used LR to identify OTs from MODIS data for model development and GOES-14 for model evaluation, we propose here to use machine learning-based algorithms as well as LR to detect OTs from Himawari-8 data. As the Himawari-8 Advanced Himawari Imager (AHI) is a geostationary (GEO) imager, it has higher temporal resolution than sun-synchronous or low earth orbiting (LEO) sensors such as MODIS, and has more channels than existing GEO imagers such as the Geostationary Operational Environmental Satellite (GOES), except for GOES-R/GOES-16 and the Multi-functional Transport Satellites (MTSAT) series. To the best of our knowledge, this is the first study where Himawari-8 data have been used for OT detection. Multiple channels from Himawari-8 data could be useful in providing valuable information about the characteristics of OTs. Since Himawari-8 data were directly used to develop OT detection algorithms in this study, the results demonstrate the applicability of the OT detection algorithms to other planned GEO imagers, such as GOES-R/GOES-16 and GEO-KOMPSAT 2A (GK-2A), that have similar characteristics to Himawari-8.