## Abstract

A combination of computational power, dynamic graphics and geographical information system (GIS) packages creates a powerful platform for advanced visualization tools to explore complex geographical phenomena in an interactive computer environment – known as geovisualization. Geovisualization is a relatively new computer-based approach that refers to a set of methods and techniques to support geospatial data analysis through the use of four-dimensional (4D), multi-variable and interactive visualization. In this paper, we illustrate and discuss the value of several spatialization techniques that are used to perform analysis of spatial structures of climatic and meteorological elements, especially through the geovisualization of characteristics and behaviour of tropical cyclones in the South Pacific Ocean. Preliminary findings are encouraging, allowing patterns and dependencies between chosen cyclone features to be identified, in turn indicating the enormous potential of geovisualization for analysing multi-variate spatial attributes within large tropical cyclone datasets.

Geographical information systems (GIS), originally developed for natural resource inventory and mapping, are now widely used in many scientific disciplines. Land information systems, environmental science, socio-economic studies and many other disciplines now use GIS to analyse and graphically represent geographically referenced processes and phenomena. One of these applications is climatological and meteorological studies, where visualization of spatial objects is important to understanding the underlying process and relations. Modern GIS packages provide powerful tools for advanced geovisualization as one of the steps in exploratory spatial data analysis. Geovisualization refers to a set of tools and techniques supporting geospatial data analysis through the use of four-dimensional (4D), multi-variable and interactive visualization.

Geographical information systems give possibilities to combine different georeferenced variables and parameters in such a way that it should also be possible to give consistently derived estimates of meteorological and climatological variables at any location at any time (Tveito 2007). Many climatological tasks are based on providing information about weather and climate by using observed values at fixed meteorological stations that are then adjusted for representativity, terrain and other effects affecting the local climatology. In many cases, GIS is primarily used as a tool to establish continuous maps of climate reference values of several elements. Many GIS packages offer built-in tools for spatial interpolation, which, being originally developed for stationary objects, are now used more and more for analysis of climatological data, such as tropical cyclones in particular.

The impacts of tropical cyclones on the environments of different types of islands across the tropical South Pacific (TSP) region are not so well understood compared to hurricanes and typhoons in other regions (e.g. the Caribbean Sea and North West Pacific) (Terry 2007). There are many factors associated with the characteristics and behaviour of cyclones, such as their relatively frequent occurrence and the weather effects they bring (intense and prolonged rainfall, storm surge, violent winds driving large erosive waves). Scientists have invested considerable effort to uncover and understand the variability of cyclone characteristics associated with El Niño events, ocean warming and climate change. One important feature is cyclogenesis position and the subsequent spatio-temporal behaviour of cyclones in tropical areas of the South Pacific. Such information is hidden in thousands of records of cyclone tracks, containing various data from satellite measurements for more than 340 cyclones since 1970 until present times.

Analysis of a large dataset (1970–2008) on tropical cyclone characteristics in the TSP region, recently provided by the Regional Specialized Meteorological Service (RSMC) at Nadi in Fiji, requires careful and thorough (time-consuming) checking for errors and accuracy before analysis can begin. This can be done using a combination of manual checks, mathematical tools and GIS. Once corrections have been made, investigation can commence on long-term historical trends in: tropical cyclogenesis positions (i.e. locations of origins); track length; storm duration; maximum intensity produced; and cyclone decay position. This work is amongst the first of its kind on the RSMC data for the TSP region, and thus provides insight into geographical patterns and changing temporal variability, possibly climate-change driven, in tropical cyclone behaviour for this region. Such research is already continuing in other cyclone-prone oceans of the world (e.g. the North Atlantic and North Pacific), but has lagged behind in the South Pacific owing the previous non-availability of (authenticated) data from the RSMC-Nadi. At this stage our research is focused on developing an advanced geovisualization methodology for exploratory analysis of spatio-temporal patterns of cyclones, which can also be used to carry out more demanding mathematical studies of storm movement.

In this paper, we discuss the value of several spatial interpolation and extrapolation techniques (spatialization) that may be used to perform the analysis of spatial structures of climatic and meteorological elements, for the overall purpose of geovisualization of tropical cyclone characteristics and behaviour using GIS.

## Mapping tropical cyclones

From mapping and visualization points of view, tropical cyclones (TCs) can be viewed as specific events occurring in particular places, but changing their spatial location and characteristics with time. Cyclones move at variable speeds, forming tracks with certain patterns and directions. Location and other parameters of cyclones are recorded at specific time intervals; thus, a tropical cyclone can be viewed as a single feature represented as sequence of co-ordinates (latitude and longitude) with recorded wind speed and atmospheric pressure. Such presentation makes it appropriate for using discrete GIS vector data models using points and polylines.

Mapping individual cyclone tracks as sequences of points connected by line segments and forming a polyline, is probably the best way to represent and study single occurrences of the phenomena. However, for the analysis of hundreds of cyclones recorded over decades, the entire dataset has to be processed using different spatial data models. Overlaying all cyclone tracks in a single map results in a highly cluttered set of curved lines with little or almost no use for analysis. Alternatively, a point-based data structure can be used. The point data can be analysed as such (density and cluster analysis) or can be used to construct continuous surfaces. In this case, spatio-temporal visualization (geovisualization) can be considered as a valuable method for the systematic study of this hazardous climatological phenomenon. The next section describes the basis of the most commonly used techniques in spatialization. This is followed by a consideration of how these methods might be applied to the geovisualization of changing spatio-temporal characteristics of tropical cyclones, using data and individual examples of several recent storms in the South Pacific for illustration.

## Spatialization and visual analytics

Visual analytics is a general term used to describe the derivation of spatially continuous fields, using spatial interpolation (spatialization), and adding value to such products by combining and using several georeferenced information sources (Tveito 2007; Andrienko *et al.* 2010). The greatest challenge in spatialization of tropical cyclones is to balance the available information (both climatological data and other geographical information) with the specific needs of a particular study. The selection of the proper spatialization and visualization scheme for each individual purpose needs thoroughly worked out specifications concerning the requested spatial and temporal scales, as well as the desired accuracy.

Spatialization of data for tropical cyclones involves several steps: (a) rectification of linear features to a single point according to particular characteristic of a cyclone; (b) preliminary statistical data analysis (test for statistical distribution) and data normalization (if needed); (c) identification of a mathematical method of spatial interpolation, appropriate to a particular phenomenon, and constructing a spatial grid; and (d) validation of the interpolation.

### Rectification of a tropical cyclone track to a point

The source tropical cyclone records, available from Tropical Cyclone Regional Specialized Meteorological Centres (RSMCs), contain three major groups of quantitative information: date, location and wind speed (calculated as a function of atmospheric pressure). These source data can be used for spatialization alone or to derive a variety of secondary parameters, such as cyclone duration, length, sinuosity and velocity (moving speed of the cyclone eye). Whether the source or derivative data are used, the question is what particular location in geographical space should be used as a point to assign particular characteristics of a cyclone? Some points from the source data can be naturally used as such (e.g. genesis and decay of a cyclone are defined by magnitude of wind speed at 35 knots), but which point to use to assign, for example, sinuosity or duration of a cyclone? There might be several alternatives: use start or end (genesis or decay) points of a cyclone track; use one of the existing points from along a track (e.g. with the maximum wind speed); calculate the mean point on a track (e.g. half length or half duration); or calculate a centroid (geometrically weighted location, calculated using co-ordinates of all points of the track). Cyclone tracks are rarely straight lines, and it is assumed that centroids can be positioned off the track (often quite far from the track).

The choice of which point to use should also take into account the nature of the characteristic to be assigned to the point to allow for scientifically correct further analysis (e.g. azimuth of the cyclone can be assigned to the genesis point, while accumulated cyclone energy (ACE) or hurricane destruction potential (HDP) might better be assigned to the point of peak cyclone intensity). If multi-variable analyses are assumed, it is reasonable to make the ‘point convention’, when all cyclone characteristics are assigned to one designated point.

### Data accuracy and completeness

Data accuracy and completeness are important factors to consider while carrying out spatial analysis of cyclone data. Cyclone track records, provided by RSMCs, contain three groups of data: temporal, locational and atmospheric. Location and atmospheric pressure are recorded at fixed time intervals – 12 h for historical records and 6 h for more recent data. While the accuracy of time records is not questionable, the other parameters have a certain level of uncertainty that needs to be taken into account. Ambiguity in recording cyclone location and atmospheric pressure results in ambiguity of derived parameters. For example, the location of a cyclone is defined in geographical co-ordinates, where latitude and longitude values are shown in track records with one digit after the decimal point. This corresponds to approximately 11 km on the Earth's surface in equatorial areas. At the same time, identification of cyclone eye location can be inaccurate to 50 km (*c*. 0.45°) (Prasad 2008 pers. comm. RSMC-Nadi). Locational errors, ranging from ±10 to ±50 km, can significantly skew results of cyclone analysis, especially for those with short lengths (e.g. six cyclone tracks within the RSMC-Nadi database have lengths of 120–180 km, four tracks are 220–280 km in length, and 10 tracks are between 320 and 500 km). Satellite systems, image analysis software and techniques for the identification of the cyclone eye location have changed over the last few decades, and reliability of the source data has improved.

Atmospheric data in cyclone track records comprise minimum barometric pressure at sea level and maximum sustained wind speed. These two parameters are indirectly derived using various methods, and the Dvorak technique is the one commonly used in practice. This satellite technique employs image-pattern recognition and empirically based rules to derive an estimate of tropical cyclone intensity in ‘T-numbers’, as a description of a cyclone in terms of cloud characteristics visible in satellite imagery. T-numbers are then used to compute the current intensity (CI number) using a few rules for redeveloping or weakening storms. The CI number is considered the best estimate of the current maximum winds and minimum sea-level pressure of the tropical cyclone; relationships between CI numbers, wind speed and atmospheric pressure are shown in Dvorak look-up tables (Dvorak 1975, 1984).

In this method, thermal infrared (IR) imagery is routinely employed in a subjective way to isolate discrete temperature levels and derive T-numbers from a combination of the temperature at the storm centre (usually a relatively warm eye) and the temperature of the cold convective cloud environment. In certain situations, image enhancement is used with the IR data to isolate discrete temperature levels. While this ‘EIR’ (Enhanced IR) variation of the standard technique yields reasonable estimates of intensity in most cases, analyst judgement on pattern or rule interpretation can occasionally lead to discrepancies between different tropical analysis centres estimating the same storm (Velden *et al.* 1998). The Tropical Cyclone Forecasters Reference Guide (2010) states that mean absolute errors inherent in intensities obtained through the Dvorak technique are large (approximately 15 hPa, where 1 hPa = 100 Pa), but these fixes are frequently treated as ground truth in many cyclone intensity forecast verification studies because these are the only data available (Tropical Cyclone Forecasters Reference Guide 2010). At the same time, atmospheric pressure in the Dvorak table ranges non-linearly from 1009 to 890 hPa, with incremental steps from 4 to 16 hPa for CI=2.0 to CI=8.0, respectively (Table 1 shows the look-up table for Dvorak current intensity (CI) v. maximum sustained wind speed (MWS) and minimum sea-level pressure (MSLP). Compare dMSLP with the mean absolute error of 15 hPa, where dMSLP is calculated as the difference between neighbouring MSLP values.) This introduces a certain level of ambiguity in identification of wind speed, and thus in genesis and decay points of cyclones (see a detailed discussion on evaluation of the Dvorak technique in Knaff *et al.* 2010).

### Exploratory spatial data analysis and geostatistics

Once the source data have been analysed and the ‘point convention’ has been made, the observed data need to be checked for dependency, stationarity and distribution. Spatial dependence is ‘the propensity for nearby locations to influence each other and to possess similar attributes’ (Goodchild 1992, p. 33). Atmospheric pressure, for instance, is more likely to be similar at points 5 km apart than at points 500 km apart. A statistical measure of the similarity of attributes of point locations is called *spatial autocorrelation*. In geostatistics, spatial variation is described in terms of a function (known as a correlogram) that shows how spatial autocorrelation decreases with increasing distance, finally reaching zero, or independence, at a certain distance (de Smith *et al.* 2009).

Spatial stationarity means that the statistical properties of data do not depend on exact locations of sample points. Therefore, the mean (expected value) of a variable at one location is equal to the mean at any other location, data variance is constant in the study area, and the correlation (covariance or semivariogram) between values in any two locations depends only on the vector distance that separates them, not on their exact locations (e.g. do not exhibit spatial autocorrelation). In other words, in spatial stationarity, residuals are identically distributed at all locations and the second order association between pairs of sites is assumed to depend only on the spatial distance between these sites (Perrin & Meiring 1999; Smith 2010).

Geostatistical methods are optimal when data are normally distributed and stationary. If the source data do not fit the normal distribution, then the data have to be normalized in order to obtain the assumption of spatial stationarity. To find the best way to normalize the data is the most challenging task. There are multiple options for dealing with non-normal data. First, the non-normality can be due to a valid reason (real observed data points). Invalid reasons for non-normality include mistakes in data entry and missing data values not declared missing. There are a great variety of possible source-data transformations for normalization, from adding constants to multiplying, squaring or raising to a power, converting to logarithmic scales, inverting and reflecting, taking the square root of the values, and even applying trigonometric transformations such as sine-wave transformations (Osborne 2002).

### Spatial interpolation

Spatial interpolation is used widely in the environmental sciences to estimate a spatial random field at unmonitored locations or to interpolate data onto a regular grid of points for use in subsequent analyses. Some geostatistical models are based on simplifying assumptions such as spatial stationarity (Perrin 1999; Smith 2010). There is a wide choice of spatial interpolation methods to interpolate (or predict) spatially distributed data; among them are inverse distance, kriging, and various polynomial and spline techniques (Lam 1983; Li & Heap 2008).

The inverse distance method uses a ‘simple’ distance-weighted averaging method to calculate grid node values. Inverse distance-weighted methods are based on the assumption that the interpolating surface should be influenced most by the nearby points and less by the more distant points. The interpolating surface is a weighted average of the scatter points, and the weight assigned to each scatter point diminishes as the distance from the interpolation point to the scatter point increases.

Kriging is a group of geostatistical techniques to interpolate the value of a random field (e.g. the elevation, *z*, of the landscape as a function of the geographical location) at an unobserved location from observations of its value at nearby locations. Kriging is a set of linear regression routines that minimize estimation variance from a predefined covariance model. Kriging is based on the assumption that the parameter being interpolated can be treated as a regionalized variable. A regionalized variable is intermediate between a truly random variable and a completely deterministic variable in that it varies in a continuous manner from one location to the next and therefore points that are near each other have a certain degree of spatial correlation. However, points that are widely separated are statistically independent (the data being estimated are stationary) (Davis 2002). Over the past several decades kriging has become a fundamental tool in the field of geostatistics.

The local polynomial gridding method is most applicable to datasets that are locally smooth (i.e. relatively smooth surfaces within the search neighbourhoods) (Bresnahan & Dickenson 2008). Local polynomial interpolation can be seen as a combination of (global) polynomial methods and the moving average procedure. As with global polynomials, a least-square polynomial fit to the data is applied, with options for Order 1, 2 or 3 equations. However, instead of fitting the polynomial to the entire dataset, it is fitted to a local subset defined by a window, as in the moving average model. The size of this window needs to be large enough for a reasonable number of data points to be included in the process (de Smith *et al.* 2009).

A spline is a special function defined piecewise by polynomials. In interpolating problems, spline interpolation is often preferred to polynomial interpolation because it yields similar results, even when using low-degree polynomials. Spline uses an interpolation method that estimates values using a mathematical function which minimizes overall surface curvature, resulting in a smooth surface that passes exactly through the input points (Franke 1982).

### Validation of interpolation

It is a common strategy to carry out an independent validation procedure when comparing the estimates with the observed values can assess the precision of spatialization. There are a few common similarity measures that can be used to validate estimates from almost any spatialization method, and there are basically two practical approaches for it. In the first method, the major part of the source dataset is used to construct a spatial interpolator, and the rest is used to estimate the accuracy of interpolation. This is probably the most accurate method, and can be applied when the original data sample is large and regularly sampled, and fulfils the criterion of stationarity.

The other approach is cross-validation, which considers all the data in the validation process. In a cross-validation procedure, one data point is left out of the data sample at a time. An estimated value for this point is derived by using all the other data points. This procedure is repeated until a value is estimated for all of the original data points. One possible drawback of using cross-validation is that the whole data sample is often used to define the interpolation model and, therefore, the validation might not be considered to be totally independent. However, this consideration is negligible when the dataset used is fairly large, and, therefore, cross-validation can be considered as an objective method of assessing the quality of interpolation. It is also used to compare the relative quality of two or more candidate methods. Results of cross-validation can also be used to assess the spatial variation in interpolation quality and to guide data sampling.

## Geovizualization of tropical cyclone attributes

Spatial data have a complex structure involving space, time and a number of thematic attributes, which poses significant challenges to geovisualization. The geovisualization of spatial data requires the use of maps or three-dimensional (3D) displays where at least two display dimensions are utilized to represent the physical space, which is different from information visualization dealing with abstract data spaces. This restricts the possibilities for the representation of the temporal and thematic components of the data. In modern geovisualization software, such data are represented using both traditional cartographic techniques based on the use of colours, textures, symbols and diagrams; and using computer-enabled techniques such as map animation and interactive 3D views. Moreover, maps are used in combination with non-geographical visualization techniques such as scatterplots or parallel co-ordinates. The use of multiple interactively linked views providing different perspectives into the data has become a kind of standard in geovisualization. However, a number of problems have yet to be solved, such as the scalability of geovisualization tools and their usability.

Geovisualization is a set of innovative methods and tools for visualizing geospatial data, processes, analyses and models for synthesizing and understanding geospatial phenomena. Geovisualization is one of the main tools of advanced exploratory visual analysis, widely used by geographers and environmental scientists to outline strategy and tactics for further processing using numerical statistical and data-mining methods. The main objective of this section is to investigate methods and tools of geovisualization for exploratory visual analysis of spatio-temporal patterns of tropical cyclones, to uncover general tendencies and trends in selected storm features.

While tropical cyclones in the South Pacific region are observed during a certain period of year (the normal cyclone season spans from November through to May), their occurrences in space and time are mostly random. Thus, the cyclone dataset can be generally characterized as set of records of single events (rarely concurrent within the same region at the same time), irregularly distributed in 2D space (the spatial component) with certain cyclic patterns in time (the temporal component). Cyclones rarely occur at exactly the same location or move along exactly the same path, thus there is no repeatability of observations at the exact same location over time. Another issue is the spatial density of measurements. All of these issues result in uncertainty of the data both in space and time, which will have consequences for the homogeneity of gridded (interpolated) time series. One of the consequences of such biased input data is a high risk of performing extrapolation instead of interpolation. Thus, the representativity of the dataset is probably the most serious problem within the spatialization of tropical cyclones.

### Comparison of interpolation methods

Modern GIS and surface modelling packages provide a wide range of interpolation methods. Several interpolation functions were tested to spatialize various source and derivative tropical cyclone characteristics in the tropical South Pacific (TSP) region. Cyclone data were derived from the cyclone database, provided by RSMC-Nadi, comprising more than 340 cyclone track records from 1970 up to 2008 (Fig. 1).

Figure 2 shows the spatial distribution of cyclones and their duration in days (shown at cyclogenesis origins), with the general trend as background. Figure 3 illustrates results of spatialzation of cyclone duration using the following methods: inverse distance weighted, kriging, natural neighbour, triangulation with linear interpolation, minimum curvature, nearest neighbour and radial basis function.

While visual comparison of plots is useful to comprehend the entire picture of how duration of cyclones relates to their locations and how different techniques represent the phenomenon, it is useful to have a quantitative measure of how well one or another spatialization method reconstructs the source data. Table 2 shows cross-validation statistics for all methods employed, providing estimated and residual values (including root mean square error, RMSE). According to the cyclone dataset, the duration of cyclones range from 0.25 to 39.0 days, giving a range of 38.75 days. The nearest neighbour interpolation gives the most accurate presentation of the real range of the value (see estimated *Z*-value). Yet, at the same time, the residuals (maximum, minimum, range and RMSE) are not in agreement with the nearest neighbour method. Kriging is the most applied method, and has the advantage of being based on a spatial structure function that is founded on stochastic theory and not on trigonometric or curve-fitting techniques.

### Cyclogenesis

There are many features associated with the nature and behaviour of cyclones that make them an important natural hazard in affected regions, such as their relatively frequent occurrence and the weather effects they bring (intense and prolonged rainfall, storm surge, and violent winds driving large and erosive waves). As mentioned earlier, the place and time of cyclogenesis (origin) is one of the critical characteristics of tropical cyclones that is often investigated, and is therefore worthy as a focus for applying geovisualization. Several geovisualization techniques have been used to explore different aspects of cyclogenesis in the TSP region based on cyclone track analysis from the RSMC-Nadi database.

### Spatial and temporal distribution and seasonality

Tropical cyclones in the TSP region have a clear seasonal pattern – the cyclone season starts in November and lasts until May (and sometimes until June). Figure 4 illustrates the occurrences of tropical cyclones from 1970 up to 2008 during the cyclone season. To explore temporal patterns of cyclone characteristics over a number of years, this seasonal nature of the phenomena has been taken into account and used as a focus for geovisualization (Fig. 5). Figure 6 illustrates the spatial distribution of selected tropical cyclone characteristics.

### Spatial v. temporal presentation

The plot of the temporal distribution of cyclone wind speed at origins (Fig. 7, vertical plot, right) shows some difference in wind speed patterns between the periods 1970–1980 and 1981–2008. This is an interesting problem. As the point of cyclone origin is defined as the place where wind speed reaches 35 knots, if RSMC-Nadi officially names cyclones at this value, then the plot should have a uniform colour (light green, in our case). In reality, the plot shows clear evidence of non-uniformity of wind speed at designated origins, for which there might be several reasons. For example, some cyclones may have developed rapidly, quickly gaining intensity from 30 to 40–55 knots within a 6 h interval; that is, between one track fix and the next. Another explanation is that until the 1980s, cyclone tracks were recorded at 12 h intervals, resulting in the official designation of some cyclones with wind speeds already over 35 knots at their origins. There may also be discrepancies in the data owing to technological limitations, methodology used, human factors or various other reasons. Visualization of temporal patterns therefore identified important inconsistencies in the source data, which, in turn, allowed for a clear differentiation of analysis of the spatial distribution in wind speed before and after 1980 (Fig. 7, plots at the bottom and in the middle).

### Volumetric geovisualization

Volumetric geovisualization is another powerful tool for exploratory spatial and temporal data analysis. Modern GIS packages provide several techniques for 3D visualization, including volumetric rendering and iso-surfaces. Figure 8 shows the result of volumetric interpolation of wind speed at cyclone origins. The attempt was made to visually explore variations of wind speed in space, with the purpose of identifying clusters at different thresholds of wind speed value. The plot represents the spatial location of the major cluster of wind speed at 65 knots, outlined as an iso-surface and a vertical cross-section, with contour lines representing the variation in wind speed along the longitude or latitude axis.

The plot in Figure 9 is an example of simultaneous 3D visualization of two cyclone characteristics: wind speed at the origin and cyclone duration. Wind speed is shown as volumetric density, calculated by 3D spatial interpolation. Cyclone duration is shown as an iso-surface. Comparing such plots of iso-surfaces drawn from several viewpoints not only shows the spatial distribution of the chosen characteristic in geographical space (latitude and longitude) but also allows the study of dependencies between cyclone location, duration and wind speed at the origin. Owing to the obvious limitations of still pictures, it is not possible in this paper to illustrate the dynamic changes of shape and location of iso-surface in space (which would be possible by watching a film of time-sequenced images). However, the two sets of plots shown in Figure 9 nonetheless provide some idea of the tremendous functionality, power and scope of volumetric geovisualization.

### Individual track analysis

While the previous work was aimed at representation of the entire set of available data to uncover general spatio-temporal patterns in cyclone activity in the TSP region, in other analysis the individual tracks were targeted for the exploration of cyclone behaviour. Figure 10 illustrates the results of iterative polynomial fitting of the track of Tropical Cyclone (TC) Percy, a category 4 intensity system that severely affected parts of the Cook Islands in early 2005. In this case, an attempt was made to use polynomial coefficients as one of the attributes of cyclone trajectories to be used for data mining using decision tree algorithms. Figure 11 serves as an illustration of 3D visualization of the track of TC Percy in 2005, when another attribute (wind speed) was used as the third dimension. These analyses incorporate the temporal nature of the phenomena but do not illustrate the time component explicitly.

## Conclusions

Considering the dangers associated with the occurrence of tropical cyclones in affected ocean basins around the globe, it is incumbent on the scientific community to investigate the behaviour of these severe storms, and to assess whether future projections associated with El Niño events, ocean warming and climate change will influence cyclone numbers, frequencies, seasonality and other characteristics, such as duration, intensity and track shape. In spite of the enormous advances that have been made in understanding tropical cyclone processes over recent decades, a neglected step so far has been the geovisualization of cyclone spatio-temporal attributes. In response, the goals of this research were to illustrate and provide preliminary exploration of the wide range of methods and tools that may be of enormous value for the geovisualization and analysis of tropical cyclone attributes, using the available dataset for the tropical South Pacific region. The main conclusion is that, although geovisualization is certainly not a trivial task in terms of either processes or interpretation of imagery produced, the first results have been encouraging. In particular, findings have allowed the uncovering of certain patterns and dependencies in cyclone behaviour, which helps to outline further strategy in advanced data analysis using appropriate statistical and data-mining methods.

## Acknowledgments

The authors thank the Fiji Meteorological Service for providing archived cyclone track data, and anonymous reviewers who recommended improvements on the original manuscript.

- © The Geological Society of London 2012