A visual analytics process for maritime response,resource allocation and risk assessment

Introduction

As modern datasets increase in size and complexity, it becomes increasingly difficult for analysts and decision makers to extract actionable information for effective decision making. In order to better facilitate the exploration of such datasets, tool sets that allow users to interact with their data and assist them in their analysis are required. Furthermore, such datasets can be utilized to explore the consequences and risks associated with making decisions, thereby providing insights to analysts and aiding them in making informed decisions.

Besides the sheer volume and complexity of such datasets, analysts must also deal with data quality issues, including uncertain, incomplete and contradictory data. Moreover, analysts are often faced with different decisions and are required to weigh up all of the possible consequences of these decisions using these datasets to arrive at a solution that minimizes the associated risks within a given time constraint. Using traditional methods of sifting through sheets of data to explore potential risks can be highly inefficient and difficult because of the nature and size of these datasets. Therefore, advanced tools are required that enable a more timely exploration and analysis. Our work focuses on the use of visual analytics^1,2 in the realm of risk assessment and analysis, and demonstrates the effectiveness of visual analytics in this domain. The work described in this paper is based on the application of visual analytics to analyze historic response operations and assess the potential risks in the maritime environment based on notional station closures. Our work was done in collaboration with the U.S. Coast Guard’s Ninth District and Atlantic Area Commands (See Appendix 1), which are responsible for all Coast Guard operations in the five U.S. Great Lakes. In particular, we focused on the auxiliary stations that are staffed by Coast Guard volunteers and civilians. These auxiliary stations assist their parent stations with their operations and usually operate on a seasonal basis, using a small fleet of boats for conducting their operations. However, the number of auxiliary personnel volunteering their time at these stations has decreased over recent years. This has required Coast Guard analysts to develop possible courses of action and analyze the risks and benefits associated with each option. Several options include seasonal or weekend-only staffing of these units, or, at worst, closure. Closure, however, may involve increased risks to the boating public and a complete analysis of the risks associated with closing an auxiliary station needs to be undertaken. The results of this type of analysis would assist the decision makers in determining the optimal course of action.

In particular, the analysts are interested in determining the spatial and temporal distribution of response cases and their associated sorties (a boat or an aircraft deployed to respond to an incident) for all search and rescue (SAR) operations conducted in the Great Lakes and how closing certain auxiliary stations affects the work load of the stations that absorb these cases. Coast Guard policy mandates the launch of a sortie within 30 minutes and an asset (boat or aircraft) on scene within 2 hours of receiving a distress call. ³ The closure of auxiliary stations implies longer response times, which could potentially translate into the loss of lives and property.

To address these option evaluation challenges, we developed a visual analytics system that supports decision making and risk assessment and allows an interactive analysis of trends, patterns and anomalies among the U.S. Coast Guard’s Ninth District operations and their associated sorties. Our system, shown in Figure 1, allows enhanced exploration of multivariate spatiotemporal datasets. We have incorporated enhanced tools that enable maritime risk assessment and analysis. Our system includes linked spatiotemporal views for multivariate data exploration and analysis and allows users to determine the potential increase or decrease in risks associated with closing one or more Coast Guard stations. The system enables a thorough assessment of all operations conducted by each station. In addition, the system provides analysts with the tools to determine which Coast Guard stations are better suited to assume control of the operations of the closed station(s) by comparing the distances from available stations to all SAR cases previously handled by the closed station(s). Our system features include the following:

OPEN IN VIEWER

Figure 1.

A screenshot of our risk assessment visual analytics system. Here, the user is visualizing all search and rescue (SAR) operations conducted by the U.S. Coast Guard in the Great Lakes region. The main viewing area (a) is a map view with points showing the locations of SAR incidents in the Great Lakes that occurred in July 2008. The right window (b) shows an interactive selection menu for distress types with SAR cases selected in blue. The top window (c) is a time-series view of the SAR incident report data, which shows a monthly distribution of SAR incidents over 5 fiscal years and (d) is an interactive legend of all District Nine maritime zones. The left window (e) shows a calendar view of the SAR incidents in 2008 with the month of July highlighted. An hourly view of the incidents responded to by all stations (to all SAR incidents occurring in July 2008) is shown by the clock view display (f). Finally, the bottom-left window (g) shows the time slider with radio buttons that allow different temporal aggregation levels with the current time frame selected to July 2008.

This article is an updated and expanded version of a paper ⁴ presented at the 2011 Visual Analytics Science and Technology (VAST) Conference. Our work focuses on providing analysts with interactive visual analytics tools that equip them to deal with the risk assessment scenarios associated with closing Coast Guard stations. We emphasize that although our risk assessment toolkit and the examples given in this paper have been based in the maritime domain, these techniques apply equally well to other domains (e.g. criminal offense analysis, syndromic surveillance).

Related work

In recent years, there has been a rapid growth in the development of new visual analytics tools and techniques for advanced data analysis and exploration (for examples see references ⁵ and ⁶ ). From traditional scatterplots ⁷ and parallel coordinate plots ⁸ to tools such as Theme River ⁹ and spiral graphs, ¹⁰ systems incorporate different forms of visualization to provide enhanced analytical tools to users. Although these tools allow users to explore their data and assist them in their decision-making process, researchers have only recently started to employ visual analytics techniques for risk assessment and decision-making domains that allow users to perform a thorough analysis of the risks associated with different decisions.

Migut and Worring ¹¹ propose an interactive approach to risk assessment with a risk assessment framework that integrates interactive visual exploration with machine learning techniques to support the risk assessment and decision-making process. They use a series of two-dimensional visualizations including scatterplots and mosaic plots to visualize numerical and ordinal attributes of the datasets. While the authors demonstrate the effectiveness of using visual analytics in the field of risk assessment, their work is mainly focused on building classification models that can be used interactively to classify data entities and visualize the effects on classification. Gandhi and Lee ¹² also apply visual analytics techniques to the realm of requirements-driven risk assessment. Specifically, they use cohesive bar and arc graphs to illustrate the risks due to the cascading effects of non-compliance with Certification and Accreditation requirements for the U.S. Department of Defense. Sanusi and Mustafa ¹³ introduce a framework to develop a visualization tool that may be used for risk assessment in the software development domain. Their proposed framework allows users to identify the components of the software system that are likely to have a high fault rate. Direct visualizations of risk use tools such as bar graphs and confidence interval charts to visualize measures of risk and are usually constructed using spreadsheet programs such as Microsoft Excel©.^12,14 Although widely used, these techniques fail to work for our purposes, primarily because of the nature of the risk analysis that is required. The Coast Guard SAR dataset is spatiotemporal in nature and the exploration of risk requires domain knowledge that is difficult to incorporate algorithmically.

With respect to the temporal nature of risk assessment, researchers have also developed different visualization systems that allow users to explore the risks associated with financial decisions related to investments and mutual funds, among other financial planning scenarios. Rudolph et al. ¹⁵ propose a personal finance decision-making visual analytics tool that allows users to analyze both short-term and long-term risks associated with making investment decisions. Savikhin et al. ¹⁶ also demonstrate the benefits of applying visual analytics techniques to aid users in their economic decision-making and, by extension, to general decision-making tasks. Both of the previous examples only explore temporal datasets. In this work, we apply visual analytics techniques to explore risks using multivariate spatiotemporal datasets that guide analysts in making complex decisions.

As is the case with most multivariate datasets, data tend to be inherently unreliable, incomplete and contradictory. In order to reach the correct conclusions, analysts must take this into account in their analysis. In this regard, Correa et al. ¹⁷ describe a framework that supports uncertainty and reliability issues in the different stages of the visual analytics process. They argue that, with an explicit representation of uncertainty, analysts can make informed decisions based on the level of confidence of the data. Our system factors data reliability issues in to the risk assessment process and provides confidence levels at each stage of the process, which, in turn, enable analysts to better understand the underlying nature of the data and guide them in making effective decisions.

There are also many geospatial and temporal analytical systems that enable users to explore their spatiotemporal datasets in order to find patterns and provide an overview of the data in a visual analytics platform (for examples see references^18–21). As the needs of our end users are unique, developing a standalone system to address the challenges faced by the Coast Guard analysts is warranted. We plan to further examine these robust geotemporal analysis tools and the degree to which they can be extended to meet the Coast Guard requirements that have been identified in this paper.

There has also been a lot of work done on the visualization of large datasets using interactive cross-filtered and linked views that allow users to explore their datasets. Stasko et al. ⁵ use multiple coordinated views of documents to reveal connections between entities across different documents. Eick and Johnson ²² utilize multiple linked views to visualize abstract, non-geometric datasets in order to reduce visual clutter and provide users with insights into their datasets. Eick and Wills ²³ also demonstrate the effectiveness of linking and interaction techniques in the visualization of large networks. Our system utilizes these practices and allows users to interactively explore their multidimensional and multiattribute datasets using a series of multicoordinated linked views.

Researchers have also explored different methods to address the challenges posed to maritime security and safety. Willems et al. ⁶ introduce a novel geographic visualization that supports coastal surveillance systems and decision-making analysts in gaining insights into vessel movements. They utilize density-estimated heatmaps to reveal finer details and anomalies in vessel movements. Scheepens et al. ²⁴ also present methods to explore multivariate trajectories with density maps, which allows the exploration of anomalously behaving vessels. Lane et al. ²⁵ present techniques that allow analysts to discover the potential risks and threats to maritime safety by analyzing the behavior of vessel movements and determining the probability that they are anomalous. Some other models for anomaly detection in sea traffic can be found in refereces ²⁶ and ²⁷ . Researchers have also proposed several approaches to maritime domain awareness. For example, Roy and Davenport ²⁸ present a knowledge-based categorization of maritime anomalies built on a taxonomy of maritime situational facts involved in maritime anomaly detection. We observe that these methods and models may help, in risk analysis and understanding the impact of weather and varying speeds of Coast Guard vessels in the Great Lakes, to identify high-risk regions.

In the maritime security domain, Orosz et al. ²⁹ developed a decision-making and planning tool that addresses the security needs based on the port resources. Their focus was on the tactical and strategic tradeoffs that arise because of the different port resource reallocation strategies. The tool provides support in visualizing the effects of the different day-by-day resource allocation strategies and also allow users to forecast the risks and impacts of expansion and changes made by the port security officials. Their definition of risk is based on three components: threat, vulnerability, and consequences. Their system provides users with the ability to model and visualize the risks in specified regions. The system responds primarily to outside threats and hypothetical scenarios, whereas our system is focused on the historic spatiotemporal distribution of SAR incidents. We note that the authors provide an interesting research direction in the analysis of resource allocation strategies, and we plan to explore these further in the future. Ulusu et al. ³⁰ also model the security risks due to maritime traffic accidents. In their approach, they divide the waterways into regions and compute the risks based on potential accidents occurring within these regions. They utilize three-dimensional graphs to visualize the regions and their corresponding risks.

There exists several geographic information systems (GIS) that allow users to visualize the datasets associated with coastal areas. There also exists several GIS systems that incorporate different decision-making processes by visualizing geospatial data in their specified areas. Pelot and Plummer ³¹ focus on traffic modeling for both small and large commercial vessels, and use grid maps that color the regions based on the traffic density in the regions, thereby allowing the visualization of vulnerable spatial regions. Marven et al. ³² analyze Canadian Coast Guard search and rescue operations in the Pacific coast, and follow a spatial and temporal analysis of crime (STAC) model that is based on a clustering algorithm which identifies regions of high spatial incident densities. The authors explore nearest neighbor clustering and kernel density estimation techniques in their work. Abi-Zeid et al. ³³ also developed a geographical decision support system for planning aeronautical search and rescue missions which provides support for resource allocation decisions. Our custom-built GIS system also builds upon the concept of providing analysts with the ability to visualize and analyze their multivariate spatiotemporal datasets in order to make informed decisions.

There has also been a lot of work done to assess and mitigate risks to critical infrastructure and transportation in the maritime domain. Adler and Fuller ³⁴ provide dynamic scenario- and simulation-based risk management models to assess risks to critical maritime infrastructure and strategies implemented for mitigating these risks. Mansouri et al. ³⁵ also propose a risk management-based decision analysis framework that enables decision makers to identify, analyze, and prioritize risks involved in maritime infrastructure and transportation systems. Their framework is based on risk analysis and management methodologies that allow the understanding of uncertainty and enable analysts to devise strategies to identify the vulnerabilities of the system. Furthermore, work has been done to quantify risks in the maritime transportation domain, a summary of which can be found in reference ³⁶ . While these methods facilitate maritime infrastructure risk analysis, our work is focused on assessing maritime risks from multivariate spatiotemporal SAR datasets. In this paper, we present a visual analytics approach to maritime risk assessment and provide examples that demonstrate the advantages of applying visual analytics in this domain.

Visual analytics risk assessment environment

Our visual analytics system provides enhanced risk assessment and analytical tools for users and has been built to operate for SAR incident report data. Our system has been implemented in a custom Windows-based GIS that allows drawing on an Open StreetMap, ³⁷ using Visual C++, MySQL and OpenGL. The system displays geo-referenced data on a map and allows users to temporally scroll through their data. We provide linked windows that facilitate user interaction between the spatial and temporal domains of the data. We also provide advanced filtering techniques that allow users to interactively explore the data. In addition, we have adapted the calendar view presented by vanWijk and Selow ³⁸ and extended it to explore seasonal and cyclical trends of SAR operations, and also as means of filtering data to support advanced analysis.

Figure 1 presents a screenshot of our system. The main viewing window (Figure 1(a)) shows the map view in which the user can explore the spatial distribution of all cases handled by the Coast Guard. We utilize density-estimated heatmaps (Section Geospatial displays) to quickly identify hotspots. Users may draw a bounding box over incident points on the map that generates a summary of all incidents enclosed by the box. We also provide tape measure tools that allow users to measure the distance between two points on a map. The top-most window (Figure 1(c)) shows the time-series view of the data where multiple line graphs can be overlaid for comparison and analysis. Users may visualize time-series plots by department, distress type and Coast Guard captain of the port (COTP) zone to explore summer cyclical patterns. The left-most window (Figure 1(e)) shows the calendar view of the selected Coast Guard cases. The total number of columns on the calendar may be changed as desired to reveal seasonal trends and patterns. The bottom-right window (Figure 1(f)) shows a clock view that allows users to visualize the hourly distributions of SAR incidents. The bottom-left window (Figure 1(g)) shows the time-slider widget that is used to temporally scroll through the data while dynamically updating all other linked windows. The radio buttons beneath the time slider provide several temporal aggregation methods for the data. The center-right window (Figure 1(b)) shows the distress type menu in which all SAR cases (highlighted in blue) have been selected for visualization. Users may select multiple distress types using this menu, which dynamically updates all linked views. We use similar menus to filter cases by other data fields. Users may also interactively search the menu using the search box provided on the top of the menu. Finally, the top-right window (Figure 1(d)) shows an interactive legend of the different Coast Guard District Nine maritime zones. When a user clicks on one of the zones, all cases falling in the zone are highlighted by filling the circles on the map with a solid color and dimming out the other cases displayed on the map.

A key feature of our system is the interactive distress, station and COTP zone filtering component. Users interactively generate combinations of filters that are applied to the data being visualized through the use of menus (such as the one shown in Figure 1(b)) and edit controls. The choices of filters applied affects both the geospatial viewing region and all temporal plots.

Coast Guard SAR data

The SAR data are collected by all U.S. Coast Guard stations and stored in a central repository. When the Coast Guard is called into action, a response case is generated, usually by the maritime zone that has authority in that region and receives the distress call (referred to by the Coast Guard as the search mission coordinator or SMC). Upon receiving the call, this authority will determine if resources will be applied, including which unit will provide the resource, the resource type (e.g. boat, aircraft) and number. Therefore, a response case may generate zero, one or many sorties to respond to an incident. While analyzing the risks associated with the various mitigation options, including station closure, analysts are interested in analyzing the spatiotemporal distribution of both the response cases and their associated sorties.

The SAR data consist of two main components: (1) response cases and (2) response sorties. Each entry in the response case and sortie datasets contains information that provides details of the incidents (e.g. number of lives saved, lost, assisted) and contains the geographic location of the distress.

Uncertainty in decision making

As is the case with most large datasets, anomalies and missing data introduce errors and uncertainty. The SAR data are no exception. We find that many SAR cases do not have an associated geographic location, or have the wrong associated geographic location. These inherent errors in data affect the spatial probability estimates and introduce a certain amount of uncertainty in the decisions that must be considered for an effective risk analysis and assessment. As previously noted, ¹ visual analytics methods help people to make informed decisions only if they are made aware of data quality problems. In this regard, we incorporate uncertainty and confidence levels associated with the SAR dataset into our visualizations by displaying the accuracy of the results at each step of the risk assessment process. We define the accuracy to be the percentage of cases in the dataset with reliable values, which can be used in the decision-making process (Figure 2). This percentage is calculated using the following formula:

A c c u r a c y = N − G N × 100

(1)

Here, N is the total number of cases and G is the number of cases with unreliable values (e.g. unknown geographic coordinates, swapped negative signs). When such errors are not obvious, the data are assumed to be correct and is displayed to the analyst on the map. The analyst can report missed errors in the data and contribute to the data cleaning process.

OPEN IN VIEWER

Figure 2.

Average response time risk assessment when auxiliary station Y is closed. The system automatically chooses the stations (shown in the upper-left window) that are better suited to respond to cases previously handled by station Y, along with a count of cases that each station absorbs. The station Y cases (black circles) to be handled by stations D and E are circled in red. The top graph shows the average response time distribution of these stations to respond to station Y’s cases. The decision makers are provided with accuracy estimates, and, in this case, 93% of the incidents are displayed (7% may not have valid geographic information system coordinates).

Geospatial displays

Our system provides analysts with the ability to plot incidents as points on the map and density-estimated heatmaps (Figure 1(a)). In addition, we provide users with the option of filling each incident circle with a color on a sequential color scale ³⁹ that represents its data value. For example, users may choose to visualize the average time taken to respond to an incident for all SAR cases on the map and identify cases with higher response times. Furthermore, to explore the spatial distribution of the SAR cases and quickly identify hotspots, we employ a modified variable kernel density estimation technique (Equation 2) that scales the parameter of estimation by allowing the kernel scale to vary based upon the distance from the point X i to the kth nearest neighbor x in the set comprising N samples. ⁴⁰

f ^ ( x ) = 1 N ∑ i = 1 N 1 mi n ( h , d i , k ) K ( x − X i mi n ( h , d i , k ) )

(2)

Here, N is the total number of samples, d i , k is the distance from the ith sample to the kth nearest neighbor, and h is the maximum allowed kernel width. We choose the maximum kernel width based on asset speed and travel time. Furthermore, we use the Epanechnikov kernel ⁴⁰ (Equation 3) to reduce calculation time:

K ( u ) = 3 4 ( 1 − u 2 ) 1 ( ‖ u ‖ ≤ 1 )

(3)

where the function 1 ( ‖ u ‖ ≤ 1 ) evaluates to 1 if the inequality is true and 0 otherwise.

Time-series displays

Along with the graphical interface, our system provides a variety of visualization features for both spatial and temporal viewing. For temporal viewing, we provide line and stacked bar graphs, calendar and clock views to visualize time-series SAR incident report data.

The line graph visualization allows users to overlay multiple graphs for easy comparison and to visualize trends. Line and stacked bar graph visualizations are both supported and can be interchanged using the radio buttons provided. Users may choose to visualize SAR cases handled by individual stations or maritime zones, or visualize them by distress types. The data are plotted based on a temporal aggregation level that the user selects on the time-slider widget (Figure 1(g)). In Figure 1(c), we show the line graph display of all SAR cases aggregated by month. We can easily observe peaks in the number of SAR cases in the summer months for all maritime zones in the Great Lakes region.

The calendar view visualization was first developed by van Wijk and Selow. ³⁸ This view allows the visualization of data over time, laid in the format of a calendar. In our implementation (Figure 1(e)), we shade each date entry based on the overall yearly trend. Users may interactively change the total number of columns of the calendar, thereby changing the cycle length of the calendar view and enabling users to explore both seasonal and cyclical trends of their datasets. The system also draws histograms for each row and column. This allows analysts to visualize weekday and weekly trends of SAR incidents and further assists them in determining an effective resource allocation scheme. Furthermore, we have modified our calendar view to support an interactive database querying method for easy acquisition of summary statistics from the SAR database.

Finally, the clock view visualization allows users to visualize the hourly distribution of the selected SAR incidents responded to by the Coast Guard stations. Our implementation of the clock view organizes the data in the form of a clock, which is a radial layout that is divided into slices to reflect the number of incidents that occur during each hour of the clock. Each slice of the clock view is colored based on a sequential color scheme to reflect the corresponding number of incidents occuring during that hour. Furthermore, we allow users to visualize the hourly distribution of either all selected stations in a combined clock view display (Figure 1(f)) or individual stations that are drawn as individual clock views over the stations on the map (Figure 3). We note that both these clock view representations are dynamically linked to the time slider (Figure 1(g)), and are dynamically updated as the user interacts with the system. Moreover, many SAR cases do not have an associated hourly time component. In order to depict these inherent errors in the data, we utilize Equation 1 to display the percentage of cases with valid hourly components to the users.

OPEN IN VIEWER

Figure 3.

Visualizing the hourly distribution of search and rescue incidents for three stations in the Great Lakes for fiscal years 2007–2011.

Risk assessment process

In this section, we describe the different methods and techniques that we apply in the Coast Guard risk assessment process.

Risk profile

Our system also provides users with the ability to interactively generate risk profiles, which can be used to identify regions with little SAR coverage by the Coast Guard stations in the Great Lakes. Figure 4 illustrates risk profile heatmaps that present an overview of the Coast Guard SAR coverage in the Great Lakes. Selected filter settings affect the visual output, and, in this case, we are looking exclusively at small boat station coverage. When areas of low coverage exist, resources with additional capability (e.g. aircraft) are often provided to ensure coverage of all areas. Figure 4 (left) shows the time (in minutes) that the Coast Guard stations would take to respond to a specific SAR incident in the Great Lakes, assuming a transit speed of 15 nautical miles per hour. This profile is generated assuming that the station closest to the incident responds. The regions in the Great Lakes that take the longest time for the Coast Guard to respond to an SAR case can be clearly seen in this figure. Users may interactively close stations, filter on a different resource type (e.g. boat, aircraft) or change the asset speed, updating the risk profile interactively. This further enables the analysts to visualize the increase or decrease in risk when a station is closed. Moreover, analysts can set the lower threshold of the color scale to 120 minutes (or any arbitrary time), thereby allowing them to easily identify regions that may take more than 120 minutes to respond to. We plan to incorporate contour lines into our system to demarcate the regions that may take more than the set threshold response time to reach.

OPEN IN VIEWER

Figure 4.

Risk profile. (left) A heatmap showing the time taken (in minutes) by the Coast Guard stations to deploy an asset to the Great Lakes to respond to an SAR incident, assuming a speed of 15 nautical miles per hour. (right) A heatmap showing the Coast Guard SAR coverage (i.e. the number of stations that respond to a particular region) in the Great Lakes. The squares along the coast show the locations of the stations.

Figure 4 (right) provides another risk profile visualization that allows officials to identify regions with low Coast Guard coverage for SAR operations in the Great Lakes. Regions with high SAR coverage by the Coast Guard stations are shown by darker colors. This further allows analysts to identify stations where resources may be reallocated without increasing maritime risk.

Dynamic assignment of station asset speeds

In order to assist analysts with the different resource allocation strategies, we note that not all stations are allocated resources uniformly. Stations that respond to a high number of incidents per year are typically assigned more assets and may be assigned assets with higher speeds and performance. This necessitates analysts to factor in the station asset allocation strategy in addition to the station case load when they determine the stations that could viably absorb the incidents of a hypothetically closed station(s). In order to facilitate this process, our system allows users to dynamically assign asset speeds to different Coast Guard stations. The system then factors these asset speed assignments into both the average response time analysis (section Average response time for SAR incidents) and the risk profile heatmaps (section Risk profile). An example of the user changing the asset speed for a station is shown in Figure 5, in which the user has changed the asset speed for one of the stations from 15 to 30 nautical miles per hour. The resulting heatmap clearly shows that the regions that previously took more than 120 minutes to respond to are now reachable within a tolerable time limit.

OPEN IN VIEWER

Figure 5.

Dynamic assignment of asset speed reflected in risk profile heatmaps. (left) A heatmap showing the time taken (in minutes) by the Coast Guard stations to deploy an asset, assuming an asset speed of 15 nautical miles per hour for all stations. (right) Updated heatmap in which the asset speed of station J has been changed to 30 nautical miles per hour.

Exploring risk using spatiotemporal linked views

While examining which auxiliary stations are most suitable for closure, analysts need to weigh up all of the options and analyze the potential increases or decreases in associated risks. They must also consider the increase in work load of the stations that absorb the closed stations’ cases to effectively determine the optimal response of available resources. In this section, we describe a scenario in which a group of analysts are trying to quantify and assess the risks and effects of closing multiple stations in the Great Lakes. An example that describes a complete analysis of a hypothetical single station closure scenario can be found in our earlier paper. ⁴

When analyzing the risks associated with station closures, analysts often have to consider several resource allocation scenarios and analyze all possible mitigation strategies in order to reach solutions that minimize the overall risk. This type of analysis requires them to focus on all aspects of possible strategies and extensively use their domain knowledge to reach these solutions. Moreover, analysts are required to consider the different types of resources available at each station and the possible resource redistribution strategies once some stations have been closed. Finding solutions to these challenges often proves to be extremely laborious and, as such, requires advanced techniques to aid analysts in this process. We also note that a domain expert who can assess all these different possibilities is an integral component in this visual analytics process.

In order to assist analysts with these challenges, we utilize several techniques in our visual analytics framework. Here, we describe a hypothetical scenario in which a group of analysts are assessing the risks associated with closing two stations at the same time in the Great Lakes (stations M and N in Figure 6) and are analyzing different resource allocation strategies among the stations that absorb the closing stations’ incidents. We note that this type of analysis can be extended to the closures of more than two stations at the same time. In this analysis, we assume that stations M and N have two types of assets that must be redistributed after they are closed. These assets include two types of boat: the first set have typical boat speeds of 15 nautical miles per hour and the second set are the hypothetical new generation boats with speeds of 30 nautical miles per hour. To simplify this scenario, we assume that there are equal numbers of 15 and 30 nautical miles per hour speed boats at stations M and N. We also assume that the analysts have a budget of purchasing a limited number of new generation speed boats and distributing them among the stations that absorb the work load of the closing stations.

OPEN IN VIEWER

Figure 6.

Spatial distribution of cases responded to by stations M and N in fiscal years 2007 to 2011. The incidents are represented on the map as circles and are color coded by their owner station.

In order to begin the analysis process, the analysts select the two stations for closure using the interactive menu. The analysts first use the system’s calendar view to determine the maximum number of cases that the two stations respond to in a single day. They find that the maximum number of cases responded to by stations M and N in a day was nine, occurring on the Independence Day weekend of the year 2007 on July 8. They now visualize the hourly distribution of these nine incidents using the clock view, in order to determine a better suited resource allocation strategy, and determine that only seven of the nine cases have an associated time component, and that these seven incidents are uniformly spread out over a 24 hour period. This gives them an indication of a worst case daily resource distribution scenario when stations M and N are closed.

Restricting analysis to stations F, G and H

The analysts now perform an average response time risk analysis over a period of 5 fiscal years from 2007 to 2011. They first perform a visual analysis of the spatial incident distribution of the cases responded to by stations M and N and observe that the better suited stations to absorb the cases of these two closing stations are stations F, G and H (Figure 6). This is indeed confirmed when the analysts run the average response time analysis with stations M and N closed, which reveals that the three stated stations absorb 723 of the total 730 cases previously handled by stations M and N (assuming a uniform asset speed of 15 nautical miles per hour for all stations in the Great Lakes). They therefore restrict their analysis to these three stations and allow only stations F, G and H to absorb the cases previously handled by the closing stations M and N.

Now, in order to determine a good resource allocation strategy, the analysts perform a series of iterative steps in which they assign different types of resources to the three stations. They start their analysis by assigning equal asset speeds of 15 nautical miles per hour for stations F, G and H and visualizing the average response time redistribution of cases previously handled by the two closing stations M and N. The resulting average response time distribution is shown in Figure 7(a). As can been seen from the Figure, the analysts determine that the three absorbing stations yeild a median average response time of around 91 minutes, with stations F, G and H absorbing 142, 407 and 181 cases respectively. The system also determines that 97% of the total cases fall within the tolerance limit of 120 minutes (shown as a percentage value below the graph). From a visual inspection of the cases responded to by stations M and N (Figure 6), the analysts note that the resulting case distribution among the three stations are valid as station G is located in between stations M and N. This allows them to formulate their initial hypothesis that station G may need to be assigned more assets than stations F and H combined after stations M and N have been closed.

OPEN IN VIEWER

Figure 7.

Average response time analyses after stations M and N have been closed and the resources have been reallocated equally among stations F, G and H.

Visualizing station coverage risk profile heatmaps

Finally, the analysts decide to visualize the risk profile heatmaps for the three cases in which the new generation boats are assigned to stations F, G and H respectively. This conforms to the average response time case distributions shown in Figure 7(b) to (d). The risk profile heatmaps obtained are shown in Figure 8, with the left image showing the assignment of the new generation boats to station F, the middle image to station G and the right image to station H. The analysts note that assigning the new generation boats to only stations F or H introduces a potential high risk region in the Great Lakes, and that assigning station G the allocation of new generation boats pays off well in terms of station response coverage when stations M and N have been closed. This further increases the confidence in their hypothesis of assigning the cases of the closed stations M and N to station G, and hence allows them to make an informed decision.

OPEN IN VIEWER

Figure 8.

Station coverage risk profile heatmaps after stations M and N have been closed and new generation boats with speeds of 30 nautical miles per hour are assigned to stations F (left), G (middle) and H (right). Note that assignment to station G yields an optimal coverage.

Domain expert feedback

Our system was assessed by two analysts at the U.S. Coast Guard’s Ninth District and Atlantic Area Commands who are currently using the system to determine the potential risks in the maritime domain associated with the hypothetical allocation of Coast Guard resources, which include both planned day-to-day and contingency operations. In this section, we summarize the feedback that we received from them after conducting several informal interviews. The analysts emphasized the need of such systems, which allow them to quickly and easily process large datasets in order to derive actionable results, in the maritime domain. One of the analysts noted that processing the desired queries took him a fraction of the time when using our system as compared with other software (e.g. Microsoft Excel©), which he had previously been using in his analysis. He was impressed by the fact that the system is intuitive to use and requires little user training. He observed that the system’s ability to process large datasets allows him to quickly filter the data into manageable subsets while providing interactive spatiotemporal displays that further aid him (and ultimately the senior level decision makers) in making a decision using the best information available.

The analysts noted that the system had been utilized throughout its inception and development to assist decision makers in making time-critical resource allocation decisions as well as support long-term alternative analysis for force structure, coverage and allocation. As an example, they noted that the system was used during the Coast Guard’s planning cycle for Hurricane Irene. Forecast models had Hurricane Irene striking the Eastern Seaboard of the United States over the Labor Day holiday weekend in 2011. Senior Coast Guard officials expected a need for additional resources, and wanted to pull resources from the Ninth District to aid in the relief efforts. Our system was used, and output was provided to these senior level decision makers that nullified the request to move additional resources from the Ninth District. The system displayed an usually large response need in the Great Lakes over this holiday weekend. The potential increase in risk in the Great Lakes was deemed too high and the resources were pulled from elsewhere. The analysts said that to perform this analysis by leveraging Microsoft Excel© would have taken hours, whereas this system allowed analysis to be done in minutes. They further noted that the system was built for time critical analysis, but could be leveraged to assist in longer term planning and development of hypothetical alternative analysis to find efficiencies in their current resource pool. They noted that these efforts have been noticed by other Coast Guard District Commanders as well, and that the system was currently being extended for the Fifth Coast Guard District that encompasses the Mid-Atlantic U.S. states.

We note again that the discussion presented so far in this section is a result of our informal discussions with our end users to determine the merits and usability of the system. We have adopted a user-centred design approach ⁴¹ by involving the users in the design process of the system. This process has included conducting initial interviews with a focused end user group to collect data related to their needs and expectations, after which an initial working prototype was produced. This prototype was then delivered to the end users and their feedback was acquired by conducting several on-site interviews. Thereafter the system underwent a series of iterative refinements in which the data and results were validated and verified throughout the development process.

We also plan to perform a formal user evaluation of our system to assess its benefits and limitations. Our evaluation goals include a human–computer interaction flavored evaluation that would look at errors using the system’s user interface. Other evaluation goals include evaluating the system in terms of the problem that our end users intend to solve using the system. This approach requires a comparison of the results against established ground truths about the consequences of a decision, and, as such, we plan to devise test scenarios after consulting with our end users. We note that designing this type of analysis-based evaluation strategy to evaluate the likelihood that our end users have achieved their objectives and minimized the risks after station closures and resource allocations is especially challenging because of the lack of solutions that act as a control. Thus, domain knowledge and expertise become critical in these types of decision-making processes.

Conclusions and future work

Our current work demonstrates the benefits of visual analytics in analyzing risk and historic resource allocation in the maritime domain. Our visual analytics system provides analysts with a set of tools for analyzing risks and the consequences of taking major decisions that could cause the loss of lives and property. Our results show how our system can be used as an effective risk assessment tool when examining various mitigation strategies to a known or emergent problem.

Before this system was developed, Coast Guard officials explored possible mitigation strategies, including the implementation of seasonal or weekend-only auxiliary duty stations, but the sheer volume of data and information inhibited the efficient processing of the data. However, using our system, the decision makers were quickly made aware of the fact that most response cases happened on Mondays/Tuesdays at some of the units. This further demonstrates the benefits of the use of visual analytics in the maritime domain.

In addition to performing risk analysis on the Coast Guard SAR cases, our system can also be used to conduct a thorough review of the operations (i.e. non-distress cases) conducted by different Coast Guard stations. Users may choose to visualize different datasets and analyze how each station performs in terms of factors including average response times, average distance to target, lives saved, lives assisted and lives affected. Hence, the officials may analyze the efficiency of each Coast Guard station and identify problem areas that may require further attention.

Future work includes deploying our system to assist in the analysis and optimization of all operations conducted by the U.S. Coast Guard Ninth District and expanding the use of our system to other Coast Guard districts. Our plan includes:

Finally, we note that although the initial feedback received so far by our end users has been positive, we plan to conduct a formal user evaluation of our system in the future to better understand and evaluate the impacts and limitations of our system in the decision-making process of our end users. We plan to develop several evaluation and performance metric schemes that will consider the different aspects of our visual analytics system (e.g. visual data cognition, system interactivity and utility from a user perspective, validation of data and results). ⁴² We also intend to perform qualitative assessments of the system to assess the merits of the system in the decision-making process of our end users by performing a qualitative video analysis along with providing our end users with questionnaires after user training and another after they become adept with the system.