690 WEATHER AND FORECASTING VOLUME 26 Estimating Tropical Cyclone Intensity from Infrared Image Data MIGUEL F. PIN˜ EROS College of Optical Sciences, The University of Arizona, Tucson, Arizona ELIZABETH A. RITCHIE Department of Atmospheric Sciences, The University of Arizona, Tucson, Arizona J. SCOTT TYO College of Optical Sciences, The University of Arizona, Tucson, Arizona (Manuscript received 20 December 2010, in final form 28 February 2011) ABSTRACT This paper describes results from a near-real-time objective technique for estimating the intensity of tropical cyclones from satellite infrared imagery in the North Atlantic Ocean basin. The technique quantifies the level of organization or axisymmetry of the infrared cloud signature of a tropical cyclone as an indirect measurement of its maximum wind speed. The final maximum wind speed calculated by the technique is an independent estimate of tropical cyclone intensity. Seventy-eight tropical cyclones from the 2004–09 seasons are used both to train and to test independently the intensity estimation technique. Two independent tests are performed to test the ability of the technique to estimate tropical cyclone intensity accurately. The best results from these tests have a root-mean-square intensity error of between 13 and 15 kt (where 1 kt ’ 0.5 m s21) for the two test sets. 1. Introduction estimate the intensity of tropical cyclones was developed by V. Dvorak in the 1970s during the early years of Tropical cyclones (TC) form over the warm waters of satellites (Dvorak 1975). In this technique, an analyst the tropical oceans where direct measurements of their classifies the cloud scene types in visible and infrared intensity (among other factors) are scarce (Gray 1979; satellite imagery and applies a set of rules to calculate McBride 1995). In general, the primary sources of ob- the intensity estimate. The original Dvorak technique is servations for these intense vortical weather systems are subjective, is time intensive, and relies on the expertise from satelliteborne instruments (e.g., Ritchie et al. 2003; of the analyst, but it is still used as the primary intensity Velden et al. 2006b). Although these instruments provide forecasting tool in many tropical cyclone forecasting many observations, including winds at various levels of centers around the world (e.g., Velden et al. 1998, 2006b; the atmosphere and temperature and humidity sound- Knaff et al. 2010). Velden et al. (1998) introduced the ings, among others, none of these include direct measure- difference of temperature between 1) the warmest pixel ments of the maximum wind speed or minimum sea level temperature near the eye of the tropical cyclone and 2) pressure intensity of a tropical cyclone. the coldest of the warmest pixel temperatures found on Because of the lack of direct in situ measurements of concentric rings around the center. This modification is tropical cyclone intensity, several techniques have been known as the objective Dvorak technique, and, although developed to estimate the intensity based on indirect the intensity is objectively calculated, the location of the factors. The most-used technique in operation to eye of the tropical cyclone must still be determined by an expert or by using external sources. Olander and Velden (2007) developed the advanced Dvorak technique (ADT), Corresponding author address: Miguel F. Pin˜ eros, PAS Bldg., Rm. 542, P.O. Box 210081, The University of Arizona, Tucson, AZ which introduces new procedures in making an intensity 85721-0081. estimate from satellite-based imagery rather than sim- E-mail: [email protected] ulating the original Dvorak technique. One of the most DOI: 10.1175/WAF-D-10-05062.1 Ó 2011 American Meteorological Society Unauthenticated | Downloaded 09/25/21 03:36 PM UTC OCTOBER 2011 P I N˜ EROS ET AL. 691 important improvements of the ADT consists of the in- These images are cropped to cover an area from 48 to troduction of regression equations to estimate the tropical 348N and from 1058 to 288W over the northern Atlantic cyclone intensity. Kossin et al. (2007) recently described basin and are resampled to a spatial resolution of 10 km a new satellite-based technique in which the radius of per pixel. Although the period of interest is from 2004 to maximum wind, the critical wind radii, and the two- 2009, tropical cyclones that had the majority of their dimensional surface wind field are estimated from in- trajectory outside the footprint of the cropped satel- frared (IR) imagery. This technique uses 12-h mean IR lite image were excluded from the study. This included imagery and best-track position data to estimate the Hurricane Vincent (2005) and Tropical Storms Beryl two-dimensional wind fields, which are compared with (2006), Chantal (2007), Ten (2008), and Grace (2009). aircraft wind profiles. In addition to visible and infrared As a result, a total of 15 147 half-hourly images from imagery, techniques for estimating the intensity of a tropi- 2004 to 2009 were analyzed, covering the life cycle of 36 cal cyclone have also been developed on the basis of mea- tropical storms and 42 hurricanes. surements from the Advanced Microwave Sounding Unit All samples that were located over land (center (AMSU; Spencer and Braswell 2001; Demuth et al. passed over continents and large islands) were removed 2004). Some of these techniques have been combined to from the database for consistency. Observations show enhance the TC intensity estimation (e.g., Velden et al. that tropical cyclones that make landfall rapidly decay at 2006a). a rate that is inconsistent for overocean tropical cy- A different approach for characterizing the dynamics clones. Thus, a different set of parametric curves will be of tropical cyclones was described in Pin˜ eros et al. required for landfalling TCs and is a topic of future (2008). In that study, a method to quantify the axisym- work. For now, all overland samples are simply removed metry of a tropical cyclone from remote-sensing data from the training set. was introduced. Using 30-min-resolution geostationary The original technique to determine the axisymmetry infrared imagery, the gradient of the brightness tem- of a cloud cluster using the deviation angle is illustrated in peratures was calculated, and the departure of that Fig. 1 (Pin˜eros et al. 2008). First the gradient of the IR gradient from a perfectly axisymmetric hurricane was image at every pixel (in vector form) is calculated. Figure determined. A single value that quantified that depar- 1a shows the pseudo-IR image for an idealized hurricane. ture from asymmetry was calculated, and a time series The associated IR gradient field is shown in Fig. 1b. Next, was built and correlated with the best-track intensity choosing a reference or center pixel, the deviation of the estimates from the National Hurricane Center (NHC). IR gradient vector in a pixel from a radial extending from The technique proved to be quite successful because the the center pixel is determined and stored. This calculation organization of the clouds about the vortex, including of the deviation angle is repeated for every pixel within the cirrus shield, is directly tied to the kinematic orga- 350-km radius of the center pixel. Next, the distribution nization of the vortex, including the organization of the of the deviation angles is plotted (Fig. 1c) and the vari- eyewall, rainbands, and tangential winds. ance of that distribution (the deviation-angle variance or In this paper, an improvement of the tropical cyclone DAV) is determined. The higher the variance of the an- intensity estimation technique described in Pin˜ eros et al. gle distribution is, the more disorganized is the cloud. (2008) is presented. In the next section, a brief review of The lower the variance is, the closer to pure axisymme- the method is presented and the improvement of the try is the cloud pattern. Figures 1d and 1e show the same technique is introduced. Results are shown in section 3. sequence as in Figs. 1a–c but for a single snapshot of Conclusions are discussed in section 4. Hurricane Rita (2005). The calculation is repeated using every pixel in turn as the reference center. The variance values are then plotted back into the reference pixel lo- 2. Method cation to create a ‘‘map of DAVs’’ (Pin˜eros et al. 2010) The study incorporates the 2004–09 North Atlantic that corresponds to the original IR image. In Pin˜eros et al. Ocean hurricane seasons (Franklin et al. 2006; Beven (2010), the map of variances was used to detect tropical et al. 2008; Franklin and Brown 2008; Brennan et al. cyclogenesis. In this study, the map of variances is used to 2009; Brown et al. 2010). The data used in this study are estimate the tropical cyclone intensity by developing longwave (10.7 mm) IR satellite imagery from the Geo- a parameterized curve fitting that relates the DAV values stationary Operational Environmental Satellite-12 with a parameterized function. (GOES-12). The data are available at ;4-km spatial The original DAV technique used a fixed 350-km ra- resolution, but we found previously that reducing the dius for calculation. Here, we improve the system by resolution does not particularly influence the results but using eight different maps for each image in the training does decrease the computational time considerably. set at radii varying from 150 to 500 km in steps of 50 km. Unauthenticated | Downloaded 09/25/21 03:36 PM UTC 692 WEATHER AND FORECASTING VOLUME 26 FIG. 1. (a) Brightness temperature image of an ideal vortex. (b) Gradient vectors of the central section in (a). (c) Distribution of deviation angles for the ideal vortex in (a). (d) Hurricane Rita, 1415 UTC 21 Sep 2005 (intensity: 130 kt and 932 hPa). (e) Distribution of deviation angles in (d), with variance 5 593 deg2.