The data that are typically used in trend analyses of tropical cyclones are the so-called ''best track'' data (Neumann et al. 1993). The process of compiling the best track data involves a review of the available tropical cyclone data by tropical cyclone forecasters, usually at the end of the tropical cyclone season, using all data sources available at the time that the review is performed. Thus for climate analysis there are immediate issues regarding the homogeneity of such data, particularly for less well-estimated variables such as tropical cyclone intensity, as the best available techniques for estimating this have changed over time (Landsea et al. 2006).
There are really two questions that need to be addressed in a reanalysis of the best track data, depending on the ultimate use of the data. The data can be made as accurate as possible for each storm, using all data available at the time and our present-day improved knowledge of tropical cyclones to update earlier estimates of variables contained in the data sets. Nevertheless, a data set that was reanalyzed in this fashion would not be homogeneous, as observational data and techniques have improved over time, thus potentially introducing spurious trends into the data. An argument can therefore be made for the creation of a ''degraded'' but uniform data set, one in which only a base level of data is used, combined with present-day analysis techniques (e.g. Kossin et al. 2007).
There are good reasons to believe that inhomogeneities have been introduced into the tropical cyclone best track record (e.g. Harper 2006; Kepert 2006). A very obvious change was the introduction of weather satellites in the 1960s; before this, many storms would have gone unrecorded. By the 1970s, these polar-orbiting satellites were providing regular, twice-daily visible and infrared images. By the 1980s, geostationary satellites provided 3-hourly coverage. The introduction of passive microwave sensors, followed by scatterometer data and cloud drift winds in the 1990s, provided improved delineation of tropical cyclone structure. Finally, in recent years, 3-axis stabilized geostationary satellites have provided rapid interval scans of tropical cyclones.
Moreover, analysis techniques have themselves changed. The gradual introduction and evolution of the Dvorak (1975, 1984) technique (Velden et al. 2006) of estimating tropical cyclone intensity from the appearance of the satellite image will have led to changes in estimated tropical cyclone intensities. This is particularly important in regions of the globe where there is no ground truthing of this technique as there is in the Atlantic ocean. More recently, objective techniques (Olander and Velden 2006) have further improved our ability to estimate tropical cyclone intensity.
It can be easily argued that even recent tropical cyclone records are not free from data inhomogeneities. The most telling example of this so far is the analysis of Kamahori et al. (2006) and Wu et al. (2006). Wu et al. (2006) examine trends in severe tropical cyclone numbers in different competing "best-track" data sets in the northwest Pacific region, those of the Joint Typhoon Warning Center, the Hong Kong Observatory and the Japanese Regional Specialized Meteorological Center (RSMC). The JTWC analyzed substantially greater numbers of intense cyclones than the other two forecast offices even in very recent times, when the data should be best. Wu et al. (2006) ascribe this result to the different analysis techniques used in the rival data sets; Kamahori et al. (2006) attribute these differences to modifications made by the Japan Meteorological Agency (JMA) to the original Dvorak technique so that it agreed better with surface observations. At present, it is not clear which data set best represents reality. Other best track data sets have similar issues. In the Australian best data set, three different versions of the Atkinson and Holliday (1977) wind-pressure relationship have been used at various times (Harper 2002; Kepert 2006).
In recognition of these problems, reanalyses of the tropical cylone record have been performed (Landsea et al. 2004; Harper 2006). Recent partial reanalyses of the tropical cyclone record have shown substantial corrections in trends compared with studies that have analysed existing best-track data. Kossin et al. (2007) use geostationary satellite images degraded to a consistent horizontal resolution over the period 1983 to 2005 to remove time-dependent biases, finding that detected changes in a measure of cyclone intensity in basins other than the Atlantic are smaller than in previous analyses. A recent reanalysis of the record in the western Australian region (Harper 2006) has also found that increases in severe tropical cyclone numbers are less than previously estimated using best track data. Landsea et al. (2006) use modern intensity estimation methods applied to satellite images of non-Atlantic tropical cyclones from the late 70s and early 80s to show that the intensities of the storms are likely significantly underestimated in the best track data that were compiled at that time.
There is a limit, however, to the ability of such studies to recreate completely the tropical cyclone record. Clearly, cyclones that have never been observed are lost forever and only estimates can be made of the numbers of storms that have been missed from the record. Landsea (2007) make such an estimate for the Atlantic basin, noting that although this region has been monitored by aircraft reconnaissance since 1944, such observations would not have covered the portion of the Atlantic east of 55W. Landsea (2007) estimates that about 2.2 storms per year would have been missed over the period 1900-1965, before the advent of routine satellite monitoring. In contrast, Holland and Webster (2007) estimate considerably smaller numbers of missing storms. Moreover, Holland (2007) questions Landsea's (2007) assumption that the ratio of landfalling storms to oceanic storms has been constant over this period, showing that this ratio may have changed due to cyclical variations in the formation locations of tropical cyclones. This conclusion was reinforced by Chang and Guo (2007), who estimate about 1.2 missing storms per year over the same period. In any event, the magnitude of the detected trend in Atlantic tropical cyclone numbers appears only to be reduced, not eliminated entirely when missing storm numbers similar to those assumed by Landsea (2007) are included in the data record (Mann et al. 2007). Any trend analyses of best track data would need to consider this and other data issues.
Was this article helpful?