Detection of tropical cyclone genesis via quantitative satellite ocean surface wind pattern and intensity analyses using decision trees
Introduction
Tropical cyclones (TCs) that originate over warm tropical and subtropical oceans are characterized by surface low pressure systems with organized cloud systems, thunderstorms, strong winds and heavy rainfall. Abrupt (unexpected) TC formation can result in significant human and economic damage (Park, Kim, Ho, Elsberry, & Lee, 2015). Of the hundreds of tropical disturbances that occur over the western North Pacific each year, only a few (~ 20) develop into TCs (Kerns & Chen, 2013). It is important to forecast which disturbances develop into a TC and which ones will just decay. TC formation designation primarily considers the dynamic features (e.g., pattern and strength) of a circulating storm system. Officially, the U.S. Joint Typhoon Warning Center initiates the warning of TC formation when a tropical disturbance develops into the category of tropical depression (TD), in which the maximum sustained surface wind speed (MSW) within a closed tropical circulation meets or exceeds 25 knots in the North Pacific (http://www.usno.navy.mil). MSW is an important metric used to identify TCs at various TC warning centers (e.g., JTWC and National Hurricane Center).
In situ wind measurements are rarely available over the open ocean. Thus, dynamic TC measurements are typically inferred from infrared/visible radiation signatures from the top of clouds. The Dvorak technique (Dvorak, 1972) considers the strength and distribution of TC-organizing circular winds, the degree of distortion in the cloud pattern, and convective vigor using infrared-based cloud-top temperatures (Velden et al., 2006). It has been a very valuable tool in monitoring TCs for more than three decades. However, the infrared technique has an inherent insensitivity to low-level dynamics, which may be obscured by significant convection or cirrus clouds. Microwave (MW) radiation, which is based on the emission and scattering of cloud and precipitation particles, can be used to identify strong convective areas and cloud organization. The difference between the horizontal and vertical polarization MW radiances is related to ocean surface roughness. Therefore, the MW imagery can be used to retrieve ocean surface wind velocity (e.g., 10-meter wind speed; Meissner & Wentz, 2013). Moreover, satellite MW radiometers, such as WindSat (Gaiser et al., 2004), can provide wind angle information via measuring diversely-polarized radiances. Satellite-derived ocean wind vector data provides essential information for detecting and forecasting TC formation.
The Dvorak technique is also limited by the requirement of subjective human input during the formation stage, which can lead to performance variations. Velden, Olander, and Zehr (1998) developed an objective version of the Dvorak technique. However, it cannot be applied during the pre-formation stage when the center (eye) is not well defined. Pineros, Ritchie, and Tyo (2010) proposed a new objective technique for detecting tropical cyclone formation. The Deviation Angle Variation (DAV) technique measures the axisymmetry of a cloud cluster with a predefined radius by performing a statistical analysis based on the orientation of the infrared brightness temperature gradient in the cluster. More recently, Wood et al. (2015) showed the ability of DAV technique in detecting TC formation over the western North Pacific with a hit rate of 96.8%, false alarm rate of 25.6%, and about 18.5 h detection lead time before a tropical disturbance reaches a wind intensity of 30 knots. This technique overcame ambiguities in the conventional Dvorak technique and demonstrated that the single use of the cloud symmetric index (i.e., DAV) could be practically useful for the early detection of TC formation.
While large-scale environmental conditions (e.g., sea surface temperature, large-scale vorticity, and vertical wind shear) as in Gray (1968) can increase the possibility of TC formation, convective clouds within a tropical disturbance are intricately connected to the dynamics of a TC. Some intense convection contributing to cyclogenesis has rotational, deep intense updraft (Houze, 2010). Thus, inner-core cloud and/or dynamic features of a system should be generally focused first when detecting TC formation, followed by large-scale environmental conditions (e.g. Schumacher et al., 2009, Zhang et al., 2015). Schumacher et al. (2009) defined both environmental and convective parameters using multiple geostationary satellite platforms and the National Centers for Environmental Prediction Global Forecasting System. These previous studies represented the convection strength using a cold pixel count and the average brightness temperature from the water vapor channel. The purely satellite-based methods (e.g., the original Dvorak and DAV techniques) focus on cloud pattern recognition and convective intensity. Despite the improved capabilities of satellite-based ocean-surface wind retrieval methods, the identification of dynamic patterns and intensities requires further research. Satellite ocean surface wind data has not been used to quantify dynamical TC genesis.
Various statistical techniques have been applied to relate satellite-derived indices to TC formation. For example, the original Dvorak, 1975, Dvorak, 1984 used an empirical look-up table to derive MSW from the satellite index (i.e., current intensity index and T-number), while the Advanced Dvorak Technique uses a regression-based model (Velden et al., 2006). A machine learning approach, decision trees, has been applied by Zhang et al. (2015) to TC formation detection to classify developing and non-developing systems. Zhang et al. (2015) calculated large-scale environmental indices from coarse-resolution global forecasting systems, such as the Navy Operational Global Atmospheric Prediction System. An 800-hPa relative maximum vorticity, average sea surface temperature, average precipitation rate, divergence ranging from the 1000- to 500-hPa levels and 300-hPa air temperature anomaly were used as input variables for the decision trees. Friedl and Brodley (1997) noted that the advantages of the decision tree include the ability to model nonlinear relationships between the predictors and the final products, and provide rules, thresholds and relative importance values of predictors. Thus, the decision trees method has been widely used for numerous applications, such as land cover change analysis (Im & Jensen, 2005) and convection initiation/absence classification (Han et al., 2015).
This study used WindSat data to quantify the dynamic surface wind patterns and intensities that correlate with TC formation over the western North Pacific. A new TC genesis detection model was developed based on the decision tree approach by quantitatively analyzing representative low-level wind system patterns (i.e., organization and symmetry) and strengths. A TC genesis model development flowchart is presented in Fig. 1. Section 2 describes the data (WindSat and TC precursor track data) and WindSat image sample selection for multiple TC precursor disturbances. Section 3 describes the development of new TC genesis indices based on the satellite images and creation of the TC detection model using the decision tree approach. Section 4 includes a statistical comparison of the indices between developing (Dev) and non-developing (Non-dev) disturbances, and the results of the decision tree approach. Hindcast validation of the new TC genesis model is provided in Section 5. The summary and conclusion are presented in Section 6.
Section snippets
WindSat
WindSat was the first passive MW polarimetric radiometer developed by the Naval Research Laboratory for the U.S. Navy and the National Polar-orbiting Operational Environmental Satellite System Integrated Program Office. The sensor was launched on board the Coriolis satellite mission on January 6, 2003 and is still operating. The satellite travels in an 840-km circular sun-synchronous orbit, with an ascending node local time of 17:59. WindSat records observations during both the forward and aft
Model development procedure and methodology
By analyzing the tropical system-centered Dev and Non-dev samples as in Section 2.2, the TC genesis indices were calculated and an objective model was developed by calibrating and validating the decision tree algorithm.
Satellite-derived index statistics: Dev and Non-dev
Fig. 5 illustrates the box-and-whisker plots of the eight WindSat-derived indices for the Non-dev and Dev samples (973 and 352, respectively). The median value of the Dev samples is consistently larger than that of the Non-dev samples, which corresponds with our original assumption that TC development can be characterized from WindSat images through the quantification of stronger wind and convection intensity, more symmetric cyclonic circulation and better organized wind and convection. The
TC formation detection model: validation results
Hindcast validation was conducted from 2008 to 2009 to evaluate the objective decision tree model performance. According to the JTWC best track data, 55 TCs formed over the western North Pacific from 2008 to 2009 (Table 1). 43 TCs and 253 Non-dev disturbances with at least one WindSat orbit passing over the system and sys_obs_fraction ≥ 60% were used in the validation, which was limited by the number of satellite observations over the precursor systems. Among the 43 TCs, the proposed TC genesis
Summary and conclusion
In this study, an objective technique was developed for detecting TC genesis using Windsat ocean surface wind measurements and a machine learning approach. This is the first study that has quantified various low-level dynamics and precipitation related to TC formation based on 1325 WindSat images from 2005 to 2009. Eight WindSat-derived indices were produced. The average, circular variance and landscape metrics from FRAGSTATS were used to quantify the dynamic and convective TC genesis factors.
Acknowledgments
This work was funded by the Korea Meteorological Administration Research and Development Program under Grant KMIPA2015-1100.
References (44)
- et al.
Decision tree classification of land cover from remotely sensed data
Remote Sensing of Environment
(1997) - et al.
A change detection model based on neighborhood correlation image analysis and decision tree classification
Remote Sensing of Environment
(2005) - et al.
A subjective method for assessing tropical cyclogenesis at the joint typhoon warning center
Tropical Cyclone research and Review
(2012) - et al.
Building type classification using spatial attributes derived from LiDAR remote sensing data
Landscape and Urban Planning
(2014) - et al.
Delineation of climate regions using in-situ and remotely-sensed data for the Carolinas
Remote Sensing of Environment
(2008) - et al.
Evaluation of hurricane ocean vector winds from WindSat
IEEE Transactions on Geoscience and Remote Sensing
(2006) - et al.
Mesoscale structural evolution of three tropical weather systems observed during PREDICT
Journal of the Atmospheric Sciences
(2012) A technique for the analysis and forecasting of tropical cyclone intensities from satellite pictures
NOAA Technical Memorandum NESS
(1972)Tropical cyclone intensity analysis and forecasting from satellite imagery
Monthly Weather Review
(1975)Tropical cyclone intensity analysis using satellite data
NOAA Technical Report NESDIS
(1984)