1. Introduction
Advances in observational technology over approximately the past 70 years have revolutionized the forecasting and understanding of TCs. Innovations including the advent of aircraft reconnaissance in the 1940s, real-time geostationary satellite imagery in the 1960s and 1970s, and remotely sensed microwave data and surface wind field estimates in the past decade (Landsea 2007) have resulted in improved theories regarding the structural underpinnings and fundamental physical processes of TCs (Emanuel 1986). These advances in observational platforms, concomitant with those in numerical weather prediction and data assimilation, have also improved the accuracy of operational forecasting (Rappaport et al. 2009).
However, this technological evolution has introduced observation bias into the TC climatological record, as the proportion of global TCs that were observed and recorded increases dramatically with time. For example, Vecchi and Knutson (2011) found that a substantial adjustment to annual TC counts is necessary in the open Atlantic prior to 1965 to correct for the relative scarcity of ship reports. In some basins with longer historical records such as the Atlantic, the nonhomogeneity of the dataset has made consensus on long-term trends and potential multidecadal oscillations in TC activity elusive (Goldenberg et al. 2001; Mann and Emanuel 2006). In observation-poor regions such as the eastern Pacific, there is simply no formal climatological record prior to the mid–twentieth century. Not only do such uncertainties hamper operational and seasonal forecasting skill, the incompleteness of climatology prior to the satellite era negatively impacts the ability of public and private interests to accurately manage the risks posed by TCs (Emanuel et al. 2012).
With the true number of historical TCs unknown, extending the length and improving the quality of the climatological database is a focus of ongoing research. One such major effort is the Atlantic basin hurricane database (HURDAT) reevaluation project. This undertaking employs a rigorous methodology to systematically revise the track and intensity of existing best-track (BT) cyclones and add previously unknown TCs to the historical record when supported by new observations or when prior observations are reconsidered within the context of new science (Hagen et al. 2012; Landsea et al. 2008). To be considered for inclusion in BT, suspect cases must satisfy the World Meteorological Organization's definition of a TC by demonstrating a closed surface circulation pattern, wind or pressure observations supporting tropical storm intensity, and a nonfrontal structure (Landsea et al. 2008). The HURDAT reevaluation has added dozens of TCs to Atlantic basin climatology through 1940 (with preliminary additions pending approval for 1941–54) and an extension of the project to the eastern Pacific basin is under way (Kimberlain 2012). Separately, Kubota (2012) spliced together three sets of incomplete regional records in the western Pacific basin to construct a quality-controlled and reasonably complete database of TC activity back to 1910, proposing a 35-yr extension to the official record. Despite these efforts, there remain expansive spatial and temporal stretches of TC climatology characterized by either fragmentary records or none at all.
A new and promising means by which TC climatology may be expanded and revised is through the use of reanalysis datasets, a topic first explored in Emanuel (2010). Truchelut and Hart (2011, hereinafter TH11) utilized the second version of the National Oceanographic and Atmospheric Administration/Cooperative Institute for Research in Environmental Science Twentieth-Century Reanalysis (20CR), which begins in 1871 (Compo et al. 2011), to develop a scheme that identified previously unknown Atlantic basin potential TCs. This was accomplished by first compositing reanalysis synoptic fields of historical TCs to ensure that the 20CR represented known TCs with as much fidelity as possible given the resolution of the reanalysis (Walsh et al. 2007). As demonstrated in greater detail in TH11, the 20CR is able to depict broad-scale TC thermodynamic structure correctly to first order. The next step was to manually identify TC-like signatures in the reanalysis that did not correspond to known BT TCs. Observational verification of the resulting candidate events (CEs) using historical ship reports showed the technique identified around 1.5 potential missing TCs per year for the 1951–58 Atlantic basin hurricane seasons. In accordance with the collaborative aim of the work, the list of CEs from TH11 was subsequently shared with the National Hurricane Center (NHC) and is actively being used as a tool to aid suspect case identification in the HURDAT reevaluation project (A. Hagen 2012, personal communication).
This first effort at using reanalysis model output to guide the revision of TC climatology was a successful proof of concept, but due to a method that was dependent on time-intensive manual identification of CEs, only a small fraction of the 138-yr global extent of the 20CR was studied. The intent of our work is not to directly propose additions to climatology, but rather to assist current and future groups involved in revising climatology by producing a high quality dataset of candidate events from which they may draw. To this end, this research extends the TH11 methodology by developing and testing a means of more efficiently and objectively locating TC CEs over the 20CR's entire pre-satellite-era spatiotemporal domain.
2. Experimental design
a. The Twentieth-Century Reanalysis
A crucial component of this study is the reanalysis dataset in which TC-like signatures will be identified. In general, a reanalysis can be defined as a hindcasting numerical weather prediction scheme that assimilates historical observations and returns the most likely atmospheric state at a given time (Thorne and Vose 2010). However, all global reanalyses released prior to 2010 are dependent on assimilating upper-level radiosonde observations to resolve the vertical structure of the atmosphere. Thus, none of these datasets includes synoptic fields for the years prior to 1948, when such observations are rare (Kalnay et al. 1996). This limits their usefulness as tools to improve pre-satellite-era TC climatology.
The second version of the 20CR (Compo et al. 2008, 2011) is the first product to make global reanalysis data available prior to the advent of systematic radiosonde data, providing three-dimensional global fields beginning in 1871. This is accomplished using a technique first described by Whitaker et al. (2004), in which an ensemble Kalman filter (Burgers et al. 1998) is applied to assimilated surface and sea level pressure (SLP) observations taken from the International Surface Pressure Databank (Yin et al. 2008). This yields a best guess of the vertical structure of the atmosphere, along with estimated uncertainty derived from the spread of the 56 ensemble members. The 20CR has a spatial resolution of 2°, 17 vertical levels at or below 200 hPa, and output every 6 h. Using the 20CR ensemble mean as an initial condition, 24-h forecasts of observed SLP demonstrate significant skill against persistence in the Northern Hemisphere (Compo et al. 2011).
b. Prior work
The 300–850-hPa thickness layer demonstrated a statistically significant response of 2.84σ (p < 0.001) to known TC passage that scaled with BT intensity, showing that the 20CR was capable of resolving the coarse thermodynamic structure of TCs. Given that known TCs were shown to have a characteristic signature in the reanalysis (TH11), searching for TC-like events that did not correspond to a known cyclone was therefore possible. The 1951–58 Atlantic Ocean basin hurricane seasons (1 June–30 November) were selected as a test period for this search as they had not yet been subjected to the HURDAT reevaluation process at the time of TH11. Spatial plots of 300–850-hPa thickness anomalies, mean sea level pressure, 850-hPa relative vorticity, and 850-hPa streamlines from the reanalysis were manually searched for signatures consistent with a possible TC each 12 h over the study period.
In general, in order to be defined as a CE, an area of interest needed to maintain significant positive thickness anomalies (ΔZ > 1.65σ) of the compact and symmetric presentation broadly consistent with a warm-core cyclone for 24 h. Increased consideration was also given to disturbances that possessed a closed SLP isobar of 1010 hPa or lower, an 850-hPa relative vorticity maximum exceeding 2.0 × 10−5 s−1, or a closed circulation in the analyzed 850-hPa streamlines. Using these criteria, an initial set of 57 CEs was identified in the 20CR for June–November of 1951–58. These CEs were subsequently verified using observations from the International Comprehensive Ocean–Atmosphere Dataset (ICOADS; Woodruff et al. 2011) of ship reports and the National Climatic Data Center's International Surface Database (Lott 2004), which revealed that 12 of these cases could be “missing TCs” consistent with NHC criteria (OFCM 2005).
c. Candidate event identification methodology
The manual methodology was successful in identifying plausible CEs in the test seasons, several of which were later independently proposed by Hagen et al. (2012) for possible addition to TC climatology. However, the technique was time intensive, requiring a manual and inherently subjective inspection of thousands of synoptic maps per year in order to construct the set of CEs. To make the process more expeditious and thus expand its spatiotemporal applicability, this research project endeavored to develop an automated process to identify the initial set of CEs.
This general structure of the CE identification algorithm, shown graphically in Fig. 1, incorporates the synoptic variables used as criteria in TH11, including 300–850-hPa normalized thickness anomalies, sea level pressure, and 850-hPa vorticity, while adding several additional spatial and synoptic criteria. First, annual mean and variance climatology for the thickness of the 300–850-hPa layer over the 1871–1979 period in the second version of the 20CR is calculated according to the same method specified in TH11. Next, beginning with the set of all points in the 20CR from 1871 to 1979, a mask is first applied to filter out grid points that are on landmasses. Next, a second mask is applied, eliminating all points within 10° of an IBTrACS TC for 2 days before and after passage, which was selected because distinct TCs seldom pass closer than within 10° of one another (Schenkel and Hart 2012). A threshold value is then applied for SLP; points that are below the specified critical pressure P1 must also be the minimum value in a centered 12° × 12° box. Similarly, an 850-hPa relative vorticity value exceeding a cutoff V1 must be located within a 12° × 12° box centered on the surface pressure minimum. These spatial thresholds were selected to conform to those in earlier vortex-tracking algorithms (Walsh and Watterson 1997; Cheung and Elsberry 2002), which were adjusted for the 20CR according to the resolution dependency shown by Walsh et al. (2007).
Graphical representation of the process used by the automated candidate event identification algorithm to filter 20CR synoptic fields into a set of distinct and credible cases for possible classification as a suspected missing TC.
Citation: Journal of Applied Meteorology and Climatology 52, 10; 10.1175/JAMC-D-12-0276.1
Graphical representation of the process used by the automated candidate event identification algorithm to filter 20CR synoptic fields into a set of distinct and credible cases for possible classification as a suspected missing TC.
Citation: Journal of Applied Meteorology and Climatology 52, 10; 10.1175/JAMC-D-12-0276.1
Graphical representation of the process used by the automated candidate event identification algorithm to filter 20CR synoptic fields into a set of distinct and credible cases for possible classification as a suspected missing TC.
Citation: Journal of Applied Meteorology and Climatology 52, 10; 10.1175/JAMC-D-12-0276.1
If these criteria are met, normalized 300–850-hPa thickness anomalies are calculated for each grid point in the centered box. The maximum normalized thickness anomaly value in this box must not be over land, associated with an IBTrACS TC, or coincident with negative 850-hPa relative vorticity, which together act as a first-order check for a warm-core thermodynamic structure. Next, an environment-relative thickness maximum is calculated by subtracting the average 300–850-hPa anomaly value of a 12°-square centered on the thickness anomaly maximum from the maximum anomaly value itself. This acts to control for the interannual variability of the region to basin-scale thickness anomalies, which can be caused by global weather pattern drivers like ENSO (Larkin and Harrison 2005). There are two pathways by which a grid point may satisfy the thickness anomaly criterion. First, the absolute normalized thickness anomaly can exceed a specified threshold value T1 while the environment-relative thickness anomaly is greater than a lower limit. Alternatively, the relative thickness anomaly maximum may exceed T1 as long as the absolute maximum is nonnegative. In this way, the algorithm does not discriminate against a potential CE for being located in any particular synoptic environment or interannual variance regime.
The threshold values for SLP, 850-hPa vorticity, and the 300–850-hPa normalized thickness anomaly in the algorithm were experimentally determined using the manual CEs from TH11 as a reference set. Test runs of the algorithm were performed for 3750 combinations of these three threshold values on the 1951–58 Atlantic Basin hurricane seasons. These test values were centered on the subjective constraints used to identify CEs manually in TH11, and were incremented by 0.1σ, 0.5 hPa, and 0.5 × 10−5 s−1, respectively. For each of the synoptic variables, 15 values were tested. Prospective candidate lists for each of these scenarios were compared with the TH11 CE track list, with matches tabulated and broken out by classification type. The threshold values chosen were those that successfully captured all 12 of the potential missing TCs identified in TH11 between 1951 and 1958 with the fewest false alarms, specifically a minimum of type 2 and especially type 1 cases. The thresholds were then fine-tuned, resulting in final values of P1 = 1010.85 hPa, V1 = 2.825 10−5 s−1 for 850 hPa, and T1 = 1.115σ.
These criteria identified 63 events in the 1951–58 period, with new and other types of CEs accounting for 51, or 81%, of those cases. The values of V1 and T1 were used for all global basins, while P1 was adjusted based on the mean sea level pressure within each of five cyclogenesis regions during the peak months of the local TC season. These regions are the North Atlantic basin (NA, P1 = 1010.85 hPa), the eastern Pacific Ocean (EP, P1 = 1007.4 hPa), the western Pacific Ocean (WP, P1 = 1005.4 hPa), the North Indian Ocean (NI, P1 = 1003.5 hPa), and the Southern Hemisphere (SH, P1 = 1005.35 hPa). To address the global applicability of the Atlantic-derived threshold values, a manual sensitivity test of the western Pacific CE data was later performed using the observational verification data described in section 4a(2). In brief, synoptic maps of 20CR output for 850-hPa vorticity and streamlines, SLP, and normalized thickness anomalies for the 1930 and 1931 western Pacific TC seasons were produced at 12-h intervals and manually checked to ensure that the algorithm found most areas resembling a TC without an inordinate number of non-TC events. This process determined that the thresholds used did in fact capture essentially every event that had a well-defined TC-like structure, with an exceedingly small number of clear false alarms. In general, the results of the manual sensitivity test show that there is no possible change to the search thresholds that would dramatically reduce the number of CEs without diminishing the number of type 3 events located, nor are promising candidates being missed. This is evidence that the current thresholds are globally applicable.
After discriminating 20CR grid points based on landmasses, proximity to historical TCs, and the previously discussed synoptic thresholds, two final criteria are applied to reduce the number of baroclinic systems identified. Because of the thickness asymmetry associated with a thermal gradient across the center of a pressure minimum, a maximum difference in average 600–900-hPa thickness between the eastern and western semicircles and between the northern and southern semicircles of the potential candidate was specified in approximate accordance with the procedures described in H03. The maximum 600–900-hPa thickness asymmetry was experimentally determined to be 14 m in both the longitudinal and latitudinal directions. This cutoff both did not eliminate any known potential missing TCs in 1951–58 from consideration and is in good agreement with the threshold between predominantly barotropic and predominantly baroclinic low pressure systems found by H03.
The remaining 20CR grid points that meet all of the synoptic filtering criteria and are thus of interest as locations and times of potential missing TCs are then grouped geographically and temporally into discrete CEs. First, if two candidate points are within 6 and 48 h of each other and no more than 10° apart, they are classified as part of a single CE. If there are no other “hit” grid points between them temporally, linear interpolation is used to determine 6-hourly positions between the hits in order to produce a continuous track. Next, the signature of the CE is tracked for 48 h before the first hit grid point and 48 h after the last hit grid point by following the sea level pressure minimum within an 8° × 8° box centered on the previous position at 6-h time steps. This box size was chosen as 4° represents an extreme upper bound on how far TCs move in 6 h (H03). Finally, because extension of CE tracks often results in overlaps, a final quality control step concatenates these types of cases while also removing duplicate and branching events.
d. Candidate event verification methodology
Once track files have been created for each individual CE, storm-centered maps are produced at 6-h intervals (0000, 0600, 1200, and 1800 UTC) over the lifetime of each case. These plots, an example of which is shown in Fig. 2, show the normalized 300–850-hPa thickness anomaly data and SLP fields from the 20CR in a 12° latitude × 18° longitude box centered on the CE position. Data from ICOADS ship reports within 3 h of the synoptic time are also plotted, including wind speed in knots, wind direction in degrees, SLP in hectopascals, and observation time. The end result is a set of discrete track files, synoptic data, and individual event maps for each of the five cyclogenesis regions in the study.
Normalized 300–850-hPa thickness anomaly and sea level pressure fields from the 20CR, plotted with ship observations of wind speed (kt), wind direction (°), and SLP (hPa) at 1200 UTC 29 Sep 1963. The event plotted is 1963 candidate event 13 located in the central North Atlantic Ocean.
Citation: Journal of Applied Meteorology and Climatology 52, 10; 10.1175/JAMC-D-12-0276.1
Normalized 300–850-hPa thickness anomaly and sea level pressure fields from the 20CR, plotted with ship observations of wind speed (kt), wind direction (°), and SLP (hPa) at 1200 UTC 29 Sep 1963. The event plotted is 1963 candidate event 13 located in the central North Atlantic Ocean.
Citation: Journal of Applied Meteorology and Climatology 52, 10; 10.1175/JAMC-D-12-0276.1
Normalized 300–850-hPa thickness anomaly and sea level pressure fields from the 20CR, plotted with ship observations of wind speed (kt), wind direction (°), and SLP (hPa) at 1200 UTC 29 Sep 1963. The event plotted is 1963 candidate event 13 located in the central North Atlantic Ocean.
Citation: Journal of Applied Meteorology and Climatology 52, 10; 10.1175/JAMC-D-12-0276.1
The resulting plots are subjected to a manual observational verification process identical to the one used in TH11. A key difference from the earlier work is that due to the nearly hundredfold increase in CEs located by the automated methodology relative to TH11, it is well beyond the scope of this study to perform manual synoptic analysis on each CE found by the algorithm. Such an effort is neither practical given available resources nor complementary with the collaborative intent of this project, which is to serve as an advanced starting point to support current and future efforts at revising TC climatology. However, a subset of the CEs was subjected to the observational verification process in order to confirm and quantify the efficacy of the algorithm itself. For these cases, each of the 6-hourly maps is subsequently analyzed in accordance with the criteria for addition to the TC climatological database detailed in Landsea et al. (2008). For cases where there was ambiguity as to the thermodynamic structure of the low pressure system, cyclone phase space diagrams (H03) were made in order to better determine whether or not the CE was likely tropical in nature.
Based on in situ evidence, the CEs were subsequently classified into one of three broad confidence categories depending on the level of observational support. The definition of these categories is the same as in TH11. The first bin, also known as type 1 events, featured those for which the available surface observations were conclusive in showing that no real-world TC could be associated with the CE in the reanalysis. The second variety of CE, type 2 events, included those for which there were too few observations in the vicinity of the feature to reach a meaningful judgment on whether or not a TC was present. This category includes cases for which there is support for a closed circulation and sufficiently strong winds but the thermodynamic structure of the cyclone is ambiguous, as well as cases of low observational density. Finally, type 3 events were those for which surface observations generally support a warm-core thermodynamic structure, a closed circulation, and sustained winds exceeding 33 kt (17 m s−1; 1 kt = 0.51 m s−1) at some point in the CE window, meaning the event is a potential missing TC in accordance with the classification criteria. An example of a type 3 event is shown in Fig. 2. This CE, a warm-core cyclone that evolved from a midlatitude system, was tracked by the algorithm as a strong thickness anomaly signal for several days and shown by the ship observations plotted to have a closed circulation, sustained gale force winds, and a minimum central pressure lower than 1004 hPa. This evidence makes a strong case for the eventual addition of this system to TC climatology.
3. Methodology performance
While the major advantage of using an automated means of identifying CEs is that the scheme can identify events for observational verification more quickly and objectively than manual analysis, it does so at the cost of decreased transparency. Therefore, it is important to test the algorithm against other means of identifying potential missing TCs, including the performance of the manual methodology from TH11. A comparison of the technique's results with those from the HURDAT reevaluation is also made for the 4-yr window of overlap between the two.
The first test of the detection algorithm is how it performs relative to the manual identification of CEs in the 1951–58 NA hurricane seasons. As the algorithm is partially based on the manual search criteria applied successfully to these seasons, ideally the automated and manual methods will produce comparable numbers of CEs. Because the threshold values were selected based on the synoptic characteristics of the most likely manually identified missing TCs, it is known that the algorithm will find all of the type 3 events during this period. However, the proportion of known type 1 and type 2 CEs that are located by the algorithm is not proscribed and is a meaningful test of the algorithm's performance.
After running the “tuned” CE search algorithm on the 1951–58 NA hurricane seasons, the manually and automatically generated lists of CEs were compared for cases that appeared in both sets. Of the 57 CEs located manually, the search algorithm finds and tracks 31 of these cases, for an identification rate of about 55%. The algorithm locates 63 total candidates, about 8 per year, in June–November of 1951–58, for a net overlap of about 49% with the manual set.
The numbers of CEs located by the manual method, the automated method, and both techniques are broken out by observational classification type in Fig. 3. While the automated method classifies fewer than 25%, or 6 of 27, of the type 1 (not missing TC) events identified by the manual method as CEs, it does so for over 70%, or 13 of 18, of the ambiguous type 2 cases. The filtering of weaker type 1 cases demonstrates that the algorithm adds value to the verification process by discriminating in favor of possible TCs while keeping the overall number of CEs low, which is important for end users of the CE database like the HURDAT reevaluation project. It is also worth noting that many of the 32 additional CEs identified by the algorithm but not the manual process appear to be credible cases for further investigation, indicating the subjectivity and imperfection of the manual technique itself. As an additional data point, the mean normalized thickness anomaly over the life of the CE is 1.91σ for type 1 cases and 2.34σ for type 3 events, which is significantly different at a 95% confidence level. This shows that the synoptic assessment of the CE's chances of being an actual TC correlates positively with the amplitude of the thermodynamic signal in the 20CR, a potential key to future applications. In general, these results show that not only does the algorithm yield significant time and manpower savings, it is able to produce a CE set with nearly equal proportions of event types as the manual method.
Quantity of each candidate event type found using the manual and algorithmic processes, compared with the quantity found by both methods. As determined by in situ observations, type 1 events are unlikely to be TCs, type 2 cases are ambiguous, and type 3 events are those most likely to be possible missing TCs.
Citation: Journal of Applied Meteorology and Climatology 52, 10; 10.1175/JAMC-D-12-0276.1
Quantity of each candidate event type found using the manual and algorithmic processes, compared with the quantity found by both methods. As determined by in situ observations, type 1 events are unlikely to be TCs, type 2 cases are ambiguous, and type 3 events are those most likely to be possible missing TCs.
Citation: Journal of Applied Meteorology and Climatology 52, 10; 10.1175/JAMC-D-12-0276.1
Quantity of each candidate event type found using the manual and algorithmic processes, compared with the quantity found by both methods. As determined by in situ observations, type 1 events are unlikely to be TCs, type 2 cases are ambiguous, and type 3 events are those most likely to be possible missing TCs.
Citation: Journal of Applied Meteorology and Climatology 52, 10; 10.1175/JAMC-D-12-0276.1
Another check on the performance of the algorithm involves comparing its CEs with those from an independent analysis, the list of suspect events compiled by the HURDAT reevaluation project (A. Hagen 2012, personal communication). While observational verification was performed for the algorithm CE set for January 1951–December 1966, the current progress of the HURDAT reanalysis project is such that a full list of suspect events is only available for January 1951–December 1954. It should be reiterated that the algorithm was neither tuned to nor in any way influenced by the HURDAT results.
In the 48-month window of overlap between the two sets of CEs, the HURDAT reanalysis identified 67 suspect cases, of which 11 were found to potentially be missing TCs (Hagen et al. 2012; Delgado and Strahan-Sakoskie 2012), while the algorithm found 38 CEs, of which 6 were found to be possibly missing TCs upon observational verification. Exactly half of the candidate events identified by the algorithm, or 19, were also found in the HURDAT list of suspect cases. Three of the six possible missing TCs were also found among the HURDAT reevaluation's proposed additions to TC climatology. While the temporal overlap at present between the two methods is limited, the differences between the CE sets indicate that while the algorithm is filtering out some of the higher-latitude suspect cases included in HURDAT, it is also able to locate some CEs that do not appear in the historical synoptic maps used to identify HURDAT suspect cases. This means that the algorithm can find credible CEs apart from those able to be identified by other techniques, making the reanalysis-derived CE set a complementary and supportive tool for any comprehensive effort to revise the climatological record of TCs.
4. The global tropical cyclone candidate event dataset
As the algorithm was shown to be successful in efficiently identifying CEs in the 1951–58 NA hurricane seasons, the process was subsequently applied to all other TC formation areas for the pre-satellite-era length of the 20CR. These include the NA, EP, WP, NI, and SH regions for the period 1871–1979. These runs of the algorithm yielded over 4500 distinct CEs globally over 109 yr of 20CR synoptic fields. Of these, manual observational verification was performed on approximately 300 of the cases, in the NA for 1951–66 and the WP for 1930–37. The results of these synoptic analyses are discussed in section 4a. While it is well beyond the scope of this study to observationally verify each of the 4567 CEs, the spatial patterns and temporal trends of CEs that have not been observationally verified offer insights into both TC climatology and the performance of the method itself. Thus, a discussion of unverified candidate events is presented in section 4b.
a. Observationally verified events
1) Atlantic basin, 1951–66
As stated, TH11 used a manual CE search technique in the NA for 1951–58. Due to the labor-intensive nature of the manual analysis, the CE search was confined to the hurricane season proper, or 1 June–30 November of each year. While the majority of TC activity occurs within this 6-month period, a meaningful portion of the missing activity occurs outside of it. Additionally, as the HURDAT reeevaluation project has completed their proposed revisions to NA TC climatology through 1954 (Hagen et al. 2012; Delgado and Strahan-Sakoskie 2012), there remain 12 more years from 1955 through 1966 that predate reliable satellite coverage that have not yet been reassessed. Therefore, observational verification of all NA CEs from January 1951 through December 1966 is a natural starting point to operationally test the performance of the algorithm.
In accordance with the procedures described in section 2, the algorithm was run with the specified threshold values, resulting in a dataset of 879 distinct CEs in the NA for the 1871–1979 temporal domain. Of these, 151 cases were in the years 1951–66, or about 9.4 CEs per year during the period. CE study maps were generated with ICOADS ship reports for each 6-hourly position, which were then synoptically analyzed in accordance with the Landsea et al. (2008) criteria. Results of the observational verification of these cases are compared with the results from TH11 in Table 1. The somewhat higher number of CEs identified per year and slightly lower proportion of possible missing TCs identified are explained by the algorithm searching the off-season months of December–May, which TH11 did not do. A somewhat lower success rate would be anticipated due to the greater influence of midlatitude westerlies and the increased prevalence of complex nontropical lows during these months, so these results are in line with expectations. Overall, the identification of 25 potential missing TCs over 1951–66, or roughly 1.5 potential missing TCs per year, is consistent with the seasonal activity adjustment proposed for the presatellite era in the NA by Landsea (2007) and Chang and Guo (2007).
Candidate event counts by observational classification bin for manual and automated search methods in the Atlantic basin, 1951–66.
There are a number of interesting patterns in the temporal and spatial distributions of the observationally verified CEs. The annual count of each CE type per year in the NA between 1951 and 1966 is shown in Fig. 4, and maxima, minima, and means for each type of event are found in Table 2. There is a slight decline in the total number of events per year over the study period, with 83 CEs in the first half of the period and 68 CEs in the latter half. Though the trendlines are not statistically significant (in all cases p > 0.05), the decline is concentrated in type 1 and type 2 rather than type 3 events. This makes sense, because though a similar number of missing TCs remains for the algorithm to locate in 1959–66, the increasing density of ship reports assimilated into the 20CR improves the structural resolution of low pressure systems and lowers the number of CEs with few nearby observations. This in turn decreases the quantities of type 1 and type 2 events.
Quantities of type 1, 2, and 3 candidate event cases occurring each calendar year in the North Atlantic basin for 1951–66, subsequent to observational verification using historical ship reports.
Citation: Journal of Applied Meteorology and Climatology 52, 10; 10.1175/JAMC-D-12-0276.1
Quantities of type 1, 2, and 3 candidate event cases occurring each calendar year in the North Atlantic basin for 1951–66, subsequent to observational verification using historical ship reports.
Citation: Journal of Applied Meteorology and Climatology 52, 10; 10.1175/JAMC-D-12-0276.1
Quantities of type 1, 2, and 3 candidate event cases occurring each calendar year in the North Atlantic basin for 1951–66, subsequent to observational verification using historical ship reports.
Citation: Journal of Applied Meteorology and Climatology 52, 10; 10.1175/JAMC-D-12-0276.1
Mean, maximum, and minimum candidate event counts per season by observational classification bin in the Atlantic basin, 1951–66.
Figures 5a–c plot the track and the classification type of all 151 CEs, broken out by type 1, type 2, and type 3 events, respectively, during the 1951–66 period. Figure 5 is subject to manual quality control in which overland sections of extrapolated tracks are trimmed for clarity. Figure 5 shows that the “genesis” points and interpolated tracks of the CEs, subject to the 2° resolution of the 20CR, are broadly consistent with long-term NA TC climatology. The Cape Verde and Gulf of Mexico regions together account for a majority of the CEs and, generally, events move west or north in accordance with the typical steering currents in the most active months of the hurricane season. Another notable result in Fig. 5 is a distinct type distribution of CEs between the geographic subregions of the NA. As shown in Fig. 5b, over 60% of type 2 events are found in the tropical eastern Atlantic, where observational density remained poor in the middle part of the twentieth century (Vecchi and Knutson 2008). Many of these cases involve African easterly waves with strong signatures in the reanalysis. While the available reports were not inconsistent with the existence of a TC near the 20CR location, there were simply too few nearby observations to make a type 3 classification. For this reason, it is not surprising that a majority of the type 3 cases shown in Fig. 5c are found in the Gulf of Mexico (four), Caribbean Sea (two), or southwestern NA (nine). In these areas, ship traffic and land observations are denser, and the time integration of prior observations increases the representation quality. Conversely, Fig. 5a shows a preponderance of type 1 cases in the Gulf and western Atlantic, where baroclinic cyclogenesis is common, and sufficient in situ observations exist to conclusively determine that the CE is not a warm-core cyclone.
Smoothed track map for all 1951–66 North Atlantic basin candidate events identified in the 20CR. (a) Type 1 events, those unlikely to be tropical cyclones are displayed. (b) Type 2, ambiguous cases, are plotted. (c) The tracks of the type 3 events (those most likely to be missing TCs) are shown. Position data are determined every 6 h from reanalysis SLP fields, and tracks are colored by the highest classification at any point in the existence of the event.
Citation: Journal of Applied Meteorology and Climatology 52, 10; 10.1175/JAMC-D-12-0276.1
Smoothed track map for all 1951–66 North Atlantic basin candidate events identified in the 20CR. (a) Type 1 events, those unlikely to be tropical cyclones are displayed. (b) Type 2, ambiguous cases, are plotted. (c) The tracks of the type 3 events (those most likely to be missing TCs) are shown. Position data are determined every 6 h from reanalysis SLP fields, and tracks are colored by the highest classification at any point in the existence of the event.
Citation: Journal of Applied Meteorology and Climatology 52, 10; 10.1175/JAMC-D-12-0276.1
Smoothed track map for all 1951–66 North Atlantic basin candidate events identified in the 20CR. (a) Type 1 events, those unlikely to be tropical cyclones are displayed. (b) Type 2, ambiguous cases, are plotted. (c) The tracks of the type 3 events (those most likely to be missing TCs) are shown. Position data are determined every 6 h from reanalysis SLP fields, and tracks are colored by the highest classification at any point in the existence of the event.
Citation: Journal of Applied Meteorology and Climatology 52, 10; 10.1175/JAMC-D-12-0276.1
In summary, observational verification of NA CEs from 1951 to 1966 extended the successes of TH11 into the basin's off-season months and to the final eight pre-satellite-era years. The verification process demonstrates that the algorithm is capable of producing a useful dataset, yielding over two dozen “missing TC” events in the test period with similar proportions of type 1, 2, and 3 events as the earlier manual procedure. Analysis of the temporal and spatial patterns of the candidates further confirms that the algorithmic technique can inform future inquiries into improving TC climatology.
2) Western Pacific, 1930–37
As a check on the global efficacy of the algorithm, observational verification was performed in the western Pacific Ocean (west of 180°) for January 1930–December 1937. These years were chosen because they predate the beginning of formal TC records (1945) in the populous and economically important WP, as well as the irregular ship report records associated with wars in East Asia beginning in the late 1930s. As in the NA, neither the algorithm nor the observational verification process was in any way informed by any prior effort to revise or extend WP TC climatology.
Overall, the results of the synoptic analysis process were in line with our expectations. The algorithm located a total of 143 CEs in the 1930–37 period, or 17.9 cases per year. This quantity of CEs is generally consistent with the fact that the WP is climatologically over twice as active as the NA (Gray 1968) but also that the region in the 1930s has a lower density of ship observations that would be likely to reduce the number of CEs resolved in the 20CR. Of the 143 CEs, observational verification yielded final classification of 20 of these cases as type 1 events (14%), 69 as type 2 events (48%), and 54 as type 3 events (38%). Relative to the NA, this is a significantly higher proportion of type 2 and 3 cases and a much lower occurrence of type 1 events. This is likely due to the lower observational density that makes it difficult to demonstrate that a CE is not a missing TC and to the overall higher number of missing TCs “available” to be located in the preclimatology era in the WP.
Figures 6a–c plot the tracks and classification types of the 143 CEs for 1930–37 in the WP, broken out by type 1, 2, and 3 classifications, respectively. As in Fig. 5, the overland segments of tracks are excised prior to plotting. These plots show that the genesis points and tracks of the CEs are once again in agreement with climatological expectations, developing at low latitudes, moving west, and then turning north and northeast with the midlatitude westerlies. One result apparent in Fig. 6 is a strong geographical preference among the various confidence bins. While the handful of type 1 events shown in Fig. 6a are scattered widely, in Fig. 6b almost all the type 2 cases remain to the east of Japan and the Philippines. Likewise, a sizeable majority of the type 3 tracks depicted in Fig. 6c are found in the South China Sea, East China Sea, or Sea of Japan. This low crossover can be explained by the strong gradient between the relatively high observational density close to China and Japan and the much lower density in the open WP. Also interesting to note is that the quantity of type 3 events increases from 20 over 1930–33 to 34 in 1934–37, possibly as a result of the steadily increasing density of ICOADS ship observations.
As in Fig. 5, but for 1930–37 western Pacific Ocean candidate events.
Citation: Journal of Applied Meteorology and Climatology 52, 10; 10.1175/JAMC-D-12-0276.1
As in Fig. 5, but for 1930–37 western Pacific Ocean candidate events.
Citation: Journal of Applied Meteorology and Climatology 52, 10; 10.1175/JAMC-D-12-0276.1
As in Fig. 5, but for 1930–37 western Pacific Ocean candidate events.
Citation: Journal of Applied Meteorology and Climatology 52, 10; 10.1175/JAMC-D-12-0276.1
While observational verification found an average of seven missing TCs per season in the early and mid-1930s, the fact that very few of the CEs in the open Pacific were found to be type 1 cases and that many type 2 examples had very long paths reminiscent of classic typhoon tracks and strong normalized thickness anomalies in the 20CR means that the algorithm likely identified between around 15 credible TC-like events per season. While this is fewer than the 20–25 TCs in an average WP typhoon season (Gray 1968), the CE dataset nevertheless is a significant resource upon which a future effort to formalize climatological records in the WP prior to 1945 could be built in conjunction with other extant records, such as Kubota (2012) and media landfall reports. Therefore, the results from observational verification of the reanalysis-derived CE set in the WP for 1930–37 are regarded as a successful test of the technique's ability to inform and expand TC climatology globally.
b. Unverified candidate events
While these results show promising indications that the CE locator algorithm can successfully identify credible missing TC candidates in global tropical basins, observational verification was not performed for all the identified CEs due to the large number of events. It should be mentioned that the detection of unverified CEs is dependent on the characteristics of the 20CR itself, including the model physics, assimilation scheme, and resolution, as well as the specific observational data ingested into the model. These intrinsic qualities of the 20CR influence its ability to resolve TC-like features differently in different basins and eras. However, the tracks in time and space of the unverified CEs still provide a wealth of information that makes for an intriguing comparison with known TC climatology.
The first test of whether or not the algorithm is finding realistic CEs across the full span of the dataset is to check whether the annual distribution of TCs resembles the climatological annual distribution. Figure 7 shows the basin-relative frequency of CEs per respective calendar month in the top panel, compared to the relative frequency of known TC climatology in the bottom panel. The scales on the top and bottom panels of Fig. 7 are the same to allow for direct comparison. In general, the annual cycles for all five basins conform quite closely to long-term climatology, with the NA, NI, EP, and WP regions seeing the same peaks in activity in the late summer and early fall as has been observed in known TCs. Sharp declines in CE activity are noted in Northern Hemisphere winter and spring. Conversely, the SH sees a 3-month plateau in activity between January and March, again with few CEs found during local winter and spring in accordance with climatology (WMO 2008). In general, the temporal distributions of each basin's CEs are flatter curves than that basin's TC climatology, but the same seasonal cycles are captured. Because there is no seasonal preference specified in the algorithm, this result is organic and increases our confidence that the algorithm is predominantly identifying realistic candidates during the climatologically most active months as possible missing TCs.
Relative frequency of all candidate event occurrences for each analysis region by calendar month of the year in which (top) the suspect case develops and (bottom) known cyclogenesis events occur. Upper and lower scales are the same to allow comparison.
Citation: Journal of Applied Meteorology and Climatology 52, 10; 10.1175/JAMC-D-12-0276.1
Relative frequency of all candidate event occurrences for each analysis region by calendar month of the year in which (top) the suspect case develops and (bottom) known cyclogenesis events occur. Upper and lower scales are the same to allow comparison.
Citation: Journal of Applied Meteorology and Climatology 52, 10; 10.1175/JAMC-D-12-0276.1
Relative frequency of all candidate event occurrences for each analysis region by calendar month of the year in which (top) the suspect case develops and (bottom) known cyclogenesis events occur. Upper and lower scales are the same to allow comparison.
Citation: Journal of Applied Meteorology and Climatology 52, 10; 10.1175/JAMC-D-12-0276.1
With this first-order reality check in hand, the next step is to assess trends in the number of annual CEs occurring in each basin with time. Figure 8 shows a 10-yr moving average of gross CE count for all five regions, along with known TC counts from IBTrACS over the same period. Though complex, the five CE time series behave more or less as would be expected, given their relative levels of climatological TC activity, the different starting points of their respective formal records as shown in Fig. 8, and the inconsistent density of ship reports and other observations between basins and with time. As an example of the interplay among these factors, the NA is initially the most active basin for CEs due to greater observational coverage until the beginning of the climatological record in 1886, at which point the average number of cases falls to a low and stable level until the mid-1930s. This point is as far as the HURDAT reevaluation recommendations have been incorporated into the formal climatological record, so the number of CEs rises thereafter and continues rising in the 1940s and early 1950s as wars expand the global weather observation network. As the operational coverage and real-time reporting of ship observations improve in the late 1950s and 1960s, the CE count once again begins to decrease until the start of the satellite era in 1967 in the NA and EP and 1980 elsewhere.
The 10-yr moving average of total candidate event counts identified by the search algorithm in the 20CR per year in each of the five analysis regions for 1871–1979.
Citation: Journal of Applied Meteorology and Climatology 52, 10; 10.1175/JAMC-D-12-0276.1
The 10-yr moving average of total candidate event counts identified by the search algorithm in the 20CR per year in each of the five analysis regions for 1871–1979.
Citation: Journal of Applied Meteorology and Climatology 52, 10; 10.1175/JAMC-D-12-0276.1
The 10-yr moving average of total candidate event counts identified by the search algorithm in the 20CR per year in each of the five analysis regions for 1871–1979.
Citation: Journal of Applied Meteorology and Climatology 52, 10; 10.1175/JAMC-D-12-0276.1
Results for each of the other basins are similarly explicable. The EP has the lowest annual average number of CEs because of the extreme scarcity of ship traffic in the basin's main development region and the fact that TCs tend to track away from landmasses into the open ocean. Other than an increase in the 1940s that is curtailed by the advent of formal climatology in 1949, the mean number of CEs is generally fewer than 5 per year. The NI, where climatology begins in 1880 (Knapp et al. 2010) and most TC development occurs close to well-observed landmasses, is not surprisingly the steadiest series with only a slow increase from low single-digit to the high single-digit numbers of CEs observed over the period. Dramatic increases in observational density and quality are also responsible for a fairly monotonic increase in the CE count in the WP, which is the most active TC basin in the world with an annual average of around 25 TCs per year (Elsner and Liu 2003). The slope of this statistically significant positive trend is steeper in the preclimatology era prior to 1945, at about three additional CEs per decade (p < 0.01), but continues to be approximately 1.2 additional CEs per decade through to 1979 (p < 0.01). This continued increase may be due to better resolution in the 20CR of the monsoon trough, which likely accounts for some of the additional CEs found by the reanalysis. Finally, the SH is the most volatile time series. The South Indian and South Pacific regions collectively are as active as the WP (Landsea and Delgado 2011), but due to the severe scarcity of observations over much of the period and spotty operational coverage, the Southern Hemisphere lags the WP in CE count over the period. This region also shows the greatest sensitivity to temporary increases in observational density in the 1880s and 1940s and the steepest decline following the advent of formal climatological records in 1945. Overall, while the regional density of surface pressure observations assimilated into the 20CR predominantly determines the trends in CE count rather than interannual variability, it nevertheless is an intriguing glimpse into the preclimatological era of TC history.
A spatial comparison may also be made between patterns in which the CEs and known TCs form and move. Figure 9 shows the count of the first algorithm “hit” point in the tracks of CEs (top) and the cyclogenesis point from IBTrACS of known TCs (bottom) between 1871 and 1979 in a 2° box, which was chosen to match the resolution of the 20CR. In general, the two plots are highly similar to one another. Regions that are known to be prime development regions for historical TCs also tend to be those that produce the greatest numbers of CEs, including the eastern tropical Atlantic, the South China Sea, just east of the Philippines, and in the Bay of Bengal. Overall, the spatial patterns are strikingly similar, including in regions like the SH and NI, where the CE and TC genesis morphologies are nearly exact matches for one another. In other areas where the observational density is very poor, such as the EP, the geographical bounds of the development region are climatologically correct although the density of CEs is somewhat lower than would be expected.
Counts of (top) candidate event identification points and (bottom) tropical cyclone genesis points, globally, for 1871–1979. The resolution of the plot is 2° in both latitude and longitude to conform to the resolution of the 20CR. Genesis points are defined as the first entry in the IBTrACS database for which the system is a tropical cyclone with winds exceeding 19 m s−1. The scales in the top and bottom panels are the same to allow for direct comparison.
Citation: Journal of Applied Meteorology and Climatology 52, 10; 10.1175/JAMC-D-12-0276.1
Counts of (top) candidate event identification points and (bottom) tropical cyclone genesis points, globally, for 1871–1979. The resolution of the plot is 2° in both latitude and longitude to conform to the resolution of the 20CR. Genesis points are defined as the first entry in the IBTrACS database for which the system is a tropical cyclone with winds exceeding 19 m s−1. The scales in the top and bottom panels are the same to allow for direct comparison.
Citation: Journal of Applied Meteorology and Climatology 52, 10; 10.1175/JAMC-D-12-0276.1
Counts of (top) candidate event identification points and (bottom) tropical cyclone genesis points, globally, for 1871–1979. The resolution of the plot is 2° in both latitude and longitude to conform to the resolution of the 20CR. Genesis points are defined as the first entry in the IBTrACS database for which the system is a tropical cyclone with winds exceeding 19 m s−1. The scales in the top and bottom panels are the same to allow for direct comparison.
Citation: Journal of Applied Meteorology and Climatology 52, 10; 10.1175/JAMC-D-12-0276.1
There are two major spatial differences between the CE and TC plots worth noting. The first is the presence of several “hot spots,” or localized regions favored for CE identification where there is no similar preponderance of genesis cases in the known TC data. Specific examples of these hot spots include the far western Gulf of Mexico, just to the south of Papua New Guinea, and near the island of Hainan in the South China Sea. The observational verification process in the NA for 1951–66 showed that many of the CEs in the western Gulf of Mexico were associated with frontal boundaries moving rapidly south in the spring and late fall. Almost all of these CEs were determined to be type 1 cases, or verifiably not missing TCs. While the reason for hot spots is unclear, they are mostly located near the boundaries between a continental landmass and a warm body of water. High variance between the 20CR's ensemble members regarding the speed of baroclinic wave generation combined with local geographic factors may be resulting in these false positives (Schenkel and Hart 2012). Alternatively, they may simply be cases of spurious cyclogenesis in the 20CR (TH11).
The second notable difference is that there is a poleward extension of the CE identification zones relative to the TC genesis areas, which is most prominent in the WP and the NA. This is likely due to the algorithm being more likely to identify CEs in the higher observation density regions near the Japanese home islands and the U.S. mid-Atlantic than the lower-density regions to the south; it also is likely due to the structural misdiagnosis of baroclinic lows in the Gulf Stream and Kuroshio by the 20CR. When accounting for these differences, the set of CE genesis points is remarkably similar to known cyclogenesis points, which increases the credibility of the CE dataset.
5. Conclusions
Overall, the algorithmic methodology demonstrated significant successes in locating credible TC CEs in global reanalysis model data. Major results include the creation of tracks for an average of 42 CEs per year globally for the 1871–1979 presatellite era years, and the finding of many possible missing TCs among these CEs during the observational synoptic analysis of a subset of these cases. While the intent of this work is not to directly propose additions to climatology, extrapolating the sampled success rate of the algorithm suggests that at least several hundred heretofore-unknown missing TCs are contained within the CE dataset awaiting verification. These results indicate that automated methods can generate a set of CEs that will assist current and future groups involved in the field-wide effort to extend the scope of global TC counts to decades prior to the start of the current climatological record by producing a high quality dataset of candidate events from which they may draw. Preliminary results from this research have been shared with the NHC and are being used to support both the Atlantic and East Pacific HURDAT reevaluation projects.
In the collaborative spirit of this work, all data from this project, including full track files and 6-hourly synoptic observation maps for the 4567 CEs, have been made available to the TC community (see http://moe.met.fsu.edu/tcce/). As the observational analysis of all CEs is inconsistent with the goals of this project, the complete CE dataset is offered for current and future studies of TC climatology to freely use. Due to the breadth of the results, there is also great potential for further insights into missing TCs to follow from the refinement and extension of the method, including the use of individual members of the 20CR ensemble or ensemble spread or considering additional observations like wave height and swell direction from the ICOADS database and microfilm records of mid-twentieth-century synoptic maps. We also plan on pursuing a rigorous examination of the performance of the CE identification algorithm in the satellite era for 20CR. In conclusion, the findings of this study suggest that reanalysis models, when used in conjunction with observations, effectively add new information to TC climatology and therefore are a promising basis upon which to dramatically increase the comprehensiveness of the TC historical record.
Acknowledgments
The authors are grateful for the support this research received from the National Science Foundation (ATM-0842618) and the Risk Prediction Initiative of the Bermuda Institute for Ocean Studies. It has also benefited from discussions with and feedback from Chris Landsea and Todd Kimberlain of the National Hurricane Center, as well as from the suggestions of the three anonymous reviewers. The authors are appreciative of Gil Compo, Jeff Whitaker, and NOAA/CIRES for the development and availability of the Twentieth-Century Reanalysis products. Support for the Twentieth-Century Reanalysis Project dataset is provided by the Department of Energy, Office of Science Innovative and Novel Computational Impact on Theory and Experiment (DOE INCITE) program, the Office of Biological and Environmental Research (BER), and by the National Oceanic and Atmospheric Administration Climate Program Office. Finally, we thank the Young Scholars Program of The Florida State University for providing our group a talented guest researcher for the summer of 2012.
REFERENCES
Burgers, G., P. J. van Leeuwen, and G. Evensen, 1998: Analysis scheme in the ensemble Kalman filter. Mon. Wea. Rev., 126, 1719–1724.
Chang, E. K. M., and Y. Guo, 2007: Is the number of North Atlantic tropical cyclones significantly underestimated prior to the availability of satellite observations? Geophys. Res. Lett., 34, L14801, doi:10.1029/2007GL030169.
Cheung, K. K. W., and R. L. Elsberry, 2002: Tropical cyclone formations over the western North Pacific in the Navy Operational Global Atmospheric Prediction System forecasts. Wea. Forecasting, 17, 800–820.
Compo, G. P., J. S. Whitaker, and P. D. Sardeshmukh, 2008: The 20th Century Reanalysis project. Proc. Third WCRP Int. Conf. on Reanalysis, Tokyo, Japan, World Climate Research Programme. [Available online at http://wcrp.ipsl.jussieu.fr/Workshops/Reanalysis2008/abstract.html.]
Compo, G. P., and Coauthors, 2011: The Twentieth Century Reanalysis Project. Quart. J. Roy. Meteor. Soc., 137, 1–28.
Delgado, S., and D. Strahan-Sakoskie, 2012: A reanalysis of the 1954 Atlantic hurricane season database. Abstract, 30th Conf. on Hurricanes and Tropical Meteorology, Ponte Vedra Beach, FL, Amer. Meteor. Soc., P1.49. [Available online at https://ams.confex.com/ams/30Hurricane/webprogram/Paper205376.html.]
Elsner, J. B., and K. B. Liu, 2003: Examining the ENSO–typhoon hypothesis. Climate Res., 25, 43–54.
Emanuel, K. A., 1986: An air–sea interaction theory for tropical cyclones. Part I: Steady-state maintenance. J. Atmos. Sci., 43, 585–604.
Emanuel, K. A., 2010: Tropical cyclone activity downscaled from NOAA-CIRES Reanalysis, 1908-1958. J. Adv. Model. Earth Syst., 2, 1–12.
Emanuel, K. A., F. Fondriest, and J. Kossin, 2012: Potential economic value of seasonal hurricane forecasts. Wea. Climate Soc., 4, 110–117.
Goldenberg, S. B., C. Landsea, A. M. Mestas-Nunez, and W. M. Gray, 2001: The recent increase in Atlantic hurricane activity. Science, 293, 474–479.
Gray, W. M., 1968: Global view of the origins of tropical disturbances and storms. Mon. Wea. Rev., 96, 55–73.
Hagen, A. B., D. Strahan-Sakoskie, and C. Luckett, 2012: A reanalysis of the 1944–53 Atlantic hurricane seasons: The first decade of aircraft reconnaissance. J. Climate, 25, 4441–4460.
Hart, R., 2003: A cyclone phase space derived from thermal wind and thermal asymmetry. Mon. Wea. Rev., 131, 585–616.
Hart, R., and R. H. Grumm, 2001: Using normalized climatological anomalies to rank synoptic-scale events objectively. Mon. Wea. Rev., 129, 2426–2442.
Kalnay, E., and Coauthors, 1996: The NCEP/NCAR 40-Year Reanalysis Project. Bull. Amer. Meteor. Soc., 77, 437–471.
Kimberlain, T., 2012: Re-analysis of the eastern North Pacific HURDAT. Abstract and Recorded Present., 30th Conf. on Hurricanes and Tropical Meteorology, Ponte Vedra Beach, FL, Amer. Meteor. Soc., 2B.2. [Available online at https://ams.confex.com/ams/30Hurricane/webprogram/Paper204625.html.]
Knapp, K. R., M. C. Kruk, D. H. Levinson, H. J. Diamond, and C. J. Neumann, 2010: The International Best Track Archive for Climate Stewardship (IBTrACS): Unifying tropical cyclone best-track data. Bull. Amer. Meteor. Soc., 91, 363–376.
Kubota, H., 2012: Shifts in typhoon tracks over the western North Pacific during the 20th century based on the recovery of historical data. Abstract and Recorded Present., 30th Conf. on Hurricanes and Tropical Meteorology, Ponte Vedra Beach, FL, Amer. Meteor. Soc., 1B.7. [Available online at https://ams.confex.com/ams/30Hurricane/webprogram/Paper205199.html.]
Landsea, C. W., 2007: Counting Atlantic tropical cyclones back to 1900. Eos, Trans. Amer. Geophys. Union, 88, 18.
Landsea, C. W., and S. Delgado, cited 2011: What are the average, most, and least tropical cyclones occurring in each basin? Hurricane Research Division, Atlantic Oceanic and Meteorological Laboratories. [Available online at http://www.aoml.noaa.gov/hrd/tcfaq/E10.html.]
Landsea, C. W., and Coauthors, 2008: A reanalysis of the 1911–20 Atlantic hurricane database. J. Climate, 21, 2138–2168.
Larkin, N. K., and D. E. Harrison, 2005: On the definition of El Niño and associated seasonal average U.S. weather anomalies. Geophys. Res. Lett., 32, L13705, doi:10.1029/2005GL022738.
Lott, N., 2004: The quality control of the integrated surface hourly database. Preprints, 14th Conf. on Applied Climatology, Seattle, WA, Amer. Meteor. Soc., 7.8. [Available online at https://ams.confex.com/ams/pdfpapers/71929.pdf.]
Mann, M. E., and K. A. Emanuel, 2006: Atlantic hurricane trends linked to climate change. Eos, Trans. Amer. Geophys. Union, 87, 233, 238, 241.
OFCM, 2005: National hurricane operations plan. Office of the Federal Coordinator for Meteorology Publi. FCM-P12-2005, 166 pp.
Rappaport, E. N., and Coauthors, 2009: Advances and challenges at the National Hurricane Center. Wea. Forecasting, 24, 395–419.
Schenkel, B., and R. E. Hart, 2012: An examination of tropical cyclone position, intensity, and intensity life cycle within atmospheric reanalysis datasets. J. Climate, 25, 3453–3475.
Thorne, P. W., and R. S. Vose, 2010: Reanalyses suitable for characterizing long-term trends. Bull. Amer. Meteor. Soc., 91, 353–361.
Truchelut, R. E., and R. E. Hart, 2011: Quantifying the possible existence of undocumented Atlantic warm-core cyclones in NOAA/CIRES 20th Century Reanalysis data. Geophys. Res. Lett., 38, L08811, doi:10.1029/2011GL046756.
Vecchi, G. A., and T. R. Knutson, 2008: On estimates of historical North Atlantic tropical cyclone activity. J. Climate, 21, 3580–3600.
Vecchi, G. A., and T. R. Knutson, 2011: Estimating annual numbers of Atlantic hurricanes missing from the HURDAT database (1878–1965) using ship track density. J. Climate, 24, 1736–1746.
Walsh, K. J. E., and I. G. Watterson, 1997: Tropical cyclone–like vortices in a limited area model: Comparison with observed climatology. J. Climate, 10, 2240–2259.
Walsh, K. J. E., M. Fiorino, C. Landsea, and K. McInnes, 2007: Objectively determined resolution-dependent threshold criteria for the detection of tropical cyclones in climate models and reanalyses. J. Climate, 20, 2307–2314.
Whitaker, J. S., G. Compo, X. Wei, and T. M. Hamill, 2004: Reanalysis without radiosondes using ensemble data assimilation. Mon. Wea. Rev., 132, 1190–1200.
WMO, 2008: Tropical cyclone operational plan for the southeastern Indian Ocean and the South Pacific Ocean. WMO Rep. TCP-24, TD 292, 81 pp.
Woodruff, S. D., and Coauthors, 2011: ICOADS release 2.5: Extensions and enhancements to the surface marine meteorological archive. Int. J. Climatol., 31, 951–967.
Yin, X., and Coauthors, 2008: The International Surface Pressure Databank land component version 2.2. National Climatic Data Center, Asheville, NC, 12 pp.
