The influence of uncertainty visualization on cognitive load in a safety- and time-critical decision-making task

Abstract

Decisions with spatial visualizations are often made under uncertainty and high time pressure. However, missing or improper representation of uncertainty can hamper the decision-making process. This paper investigates the impact of uncertainty visualization on cognitive load in the context of a safety-critical, time-sensitive decision-making task with a transportation system map. In a controlled experiment (n = 40) with a dual-task paradigm, we compared three different uncertainty visualization techniques and a baseline for different levels of time pressure. Cognitive load was measured using psycho-physiological metrics based on eye tracking and galvanic skin response, as well as self-reported. The results reveal significant differences in cognitive load among different visualization types, with line uncertainty representation techniques leading to lower cognitive load under both low and high-time pressure scenarios ($\alpha <0.05$).

Keywords:

• Uncertainty
• cognitive load
• spatial decision-making
• control rooms

1. Introduction

Decision-making always has some uncertainty associated with it (Newell et al. 2022). Although the significance of uncertainty in transportation is explicitly acknowledged, applications of uncertainty in transportation decision-making have been limited (Ottomanelli and Wong 2011). In most cases, as suggested by Ottomanelli and Wong (2011), the stakeholders (decision makers, planners, and analysts) lack a meaningful representation of uncertainty information to support decision-making adequately. The root of this problem has been traced back to the over-emphasis of visualization paradigms on limiting uncertainty rather than properly communicating it, which often leads to biases in decision-making (Hullman 2020).

Decision makers manage the traffic in road networks through Traffic Management Centres (TMC). One of the main tasks of TMCs is to support frequent traffic incidents such as accidents and car breakdowns. Such events can have cascading effects leading to a widespread traffic jam, such as an incident reported in Delhi where a disabled vehicle caused a traffic jam of about 5 kilometers (Siddharth 2021).

The World Health Organization (WHO) attributes 1.3 million deaths worldwide to road accidents (WHO 2022). In addition, road accidents also result in 50 million non-fatal injuries, a large proportion of which leads to permanent disabilities. Road traffic accidents result in losses equivalent to 3% of the world’s gross domestic product (GDP). Among other factors, inadequate post-crash care is responsible for such an exorbitantly high number of disabilities and mortality from accidents. As stated by the WHO, ‘Delay in minutes can make the difference between life and death’ (WHO (2022), Inadequate post-crash care section). Therefore, decision makers in TMCs need to be mindful of the potential impacts of different events (i.e. change in traffic conditions resulting from the accident), which can have a high degree of spatial and temporal uncertainty.

Goodchild (2008), in his seminal work, described spatial uncertainty as inevitable. As highlighted by Kalamaras et al. (2018) and Weerasekera et al. (2020), advancements in traffic modelling and forecasting allow the system to anticipate future events and potential problems, which helps to make decisions in anticipation instead of dealing with issues as they arise. However, uncertainties inherent in transportation systems add another layer of complexity (Schiewe and Knura 2021). Kinkeldey et al. (2015) emphasize the importance of considering these uncertainties for effective decision-making. Decision makers in TMCs heavily rely on map-based Geographic Information Systems (GIS) for critical spatial decisions (Robinson et al. 2018). Uncertainties in route computation and travel time estimation may arise due to ongoing road changes or errors in geographic data, caused by measurement inaccuracies and network simplification (Schiewe and Knura 2021). Decision makers interpret the information from uncertainty visualizations, and therefore, a thorough study of such visualizations becomes crucial to aid decision makers (Hope and Hunter 2007, Kinkeldey et al. 2015, Schiewe and Knura 2021).

Reducing the cognitive load on decision makers is of paramount importance in traffic control rooms, where the decision makers are under high time pressure (Aerts et al. 2003, Robinson et al. 2018, Kübler et al. 2020). An increase in cognitive load is often accompanied by loss of information from the short-term working memory, thereby decreasing the performance (Barrouillet et al. 2007). Consequently, cognitive load has also been defined as one of the most detrimental factors to the resilience of control rooms (Robinson et al. 2018, Chakraborty et al. 2021). Several methods for assessing cognitive load have been suggested in the literature. However, in the context of control rooms, traditional techniques of assessing cognitive load, such as electroencephalography (EEG) (Keskin et al. 2020) or functional magnetic resonance imaging (f-MRI), can be restrictive (Neuner et al. 2014, Eckstein et al. 2017). We have therefore chosen eye tracking (Kiefer et al. 2016, Duchowski et al. 2018) and galvanic skin response (GSR) (Shi et al. 2007) as physiological measures of cognitive load.

Why one map-based uncertainty visualization is better than others is still an unanswered question. As pointed out by MacEachren et al. (2005) and Borgo et al. (2012), even though empirical studies in uncertainty visualization have been long warranted, research efforts into such studies have been rare. Although researchers have explored several possibilities for representing uncertainty on maps (MacEachren et al. 2005, Kinkeldey et al. 2015), the effects of these representations on the cognitive load, in particular under the influence of time pressure, are yet to be studied.

1.1. Aim of the study

In the current research, we observe the effects of visualizing uncertainty in predicted travel times on decision makers through map-based uncertainty visualizations. Such visualizations will be helpful for stakeholders in devising strategies while taking uncertain information into account. Through empirical evaluations, we aim to answer the question: which uncertainty visualization will best support transportation decision makers under cognitive load and time pressure? Each visualization technique has its pros and cons. The goal is to identify the visualization technique which can aid decision makers while reducing the cognitive load during decision-making. The information from such a study can be used to identify caveats in the current decision support systems as specified by the Federal Highway Authority (FHWA) (Robinson et al. 2018).

Our study targets two hypotheses; the first is that line and point uncertainty representations will incur lower levels of cognitive load than shading (area-based) and baseline (no uncertainty) as they can be easily interpreted owing to high contrast representation of uncertainty levels. The second hypothesis is that cognitive load increases with increased time pressure due to significantly higher working memory requirements. A dual-task paradigm has been chosen because it is a well-established method for systematically inducing cognitive load and working memory requirements as described by Alonso (1998), also in uncertainty visualization studies (Padilla et al. 2020). Furthermore, through a subjective questionnaire and eye tracking metrics, we explore the participants’ decision-making process, thereby gaining further important insights.

The remainder of this paper is structured as follows. The next section discusses related work. The materials and methods for generating stimuli and processing data acquired from the experiment are reported in Section 3. Results and discussions are presented in Section 4 and Section 5 respectively. Finally, the conclusions and future scope are given in Section 6.

2. Related work

2.1. Uncertainty in GIScience and transportation

Travel time uncertainty significantly impacts the activity scheduling of people, leading to risk-averse behavior (Carrion and Levinson 2012, Zhang et al. 2018). A recent study by Wang and Wu (2021) discovered that providing decision makers with explicit information about uncertainty before making decisions can have significant advantages for the transportation system. The researchers employed a travelling salesman problem with uncertainty as the experimental task and tasked decision makers with making routing decisions for vehicle allocation in delivery services across two cities. The study’s findings highlighted when uncertainty is represented before decision-making, it positively impacts the decisions, ultimately benefiting the efficiency and effectiveness of the transportation system (Wang and Wu 2021).

Therefore, it is essential to provide information about journey time reliability to decision makers. Challenges to modeling of time-varying properties of a transportation network can be primarily attributed to uncertainty in data (such as traffic jams, weather events, and opening hours) and transportation modalities (such as reaching a destination by car, public transport, and on foot) (Lee and Miller 2020). However, due to sporadic supply degradations and demand changes, travel times on metropolitan road networks are unpredictable (Zhang et al. 2018).

From the perspective of Time Geography by Hägerstrand (1970), the available space-time extents of people at a given location are stochastic rather than deterministic due to this fluctuation in trip times (Kuijpers and Othman 2010, Winter and Yin 2011, Long et al. 2014). Such stochasticity in input variables has led GIS researchers to being interested in how humans perceive the GIS representation by integrating the aspects of cognition of geographic space (Mark et al. 1999).

Overall, the long and sustained efforts made by researchers in representing uncertainty to decision makers are a testament to its importance. Although the sources and techniques of modeling uncertainty in GIS are well-studied, the underlying cognitive processes while communicating it have received less attention.

2.2. Uncertainty visualization in transportation

Several researchers have proposed different approaches to visualize uncertainty in the context of transportation (Jeon 2010, Kuijpers and Othman 2010, Lu et al. 2014). Jeon (2010) designed a decision support tool to represent uncertainty to city planners while emphasizing the sustainability of future mobility networks. Uncertainty representation can be used to represent not only future scenarios but also present ones. Lu et al. (2014) demonstrated effective techniques for representing GPS trajectories by including their measurement uncertainty which helped the participants identify errors in current road maps. Significant challenges in depicting uncertainty can be attributed to its dependence on input as well as space, time, and aggregation level (Verstegen et al.2012).

Currently, uncertainty visualization techniques in GIS can address almost all spatial and temporal events as the proposed techniques can represent both spatial and temporal extent and denote point events or events spread across an area. Uncertainty is inherent in transportation systems, affecting factors such as speed and traffic flow (Calvert et al. 2017). The paper examines the feasibility of integrating uncertainty into transportation systems through different visualization techniques and notes that only some uncertainty visualization techniques are suitable for transportation systems. Along the same lines, a previous review article by Kinkeldey et al. (2015) identified the need for an empirical evaluation of user performance through various metrics on different uncertainty representation techniques on maps. This article empirically evaluates the effects of uncertainty representation techniques in a time-critical decision-making task.

Analyzing the effects of uncertainty visualizations on decision-making processes, particularly with maps, is crucial, given their role as decision-making aids. In the context of transportation, isochrones offer a viable method for illustrating travel time uncertainty (Zhang et al. 2018). Research on uncertainty representation has shown that representing uncertainty can facilitate decision-making without significantly increasing the decision-making time. Studies by Riveiro et al. (2014), St John et al. (2000), and Cheong et al. (2016) demonstrate that transparent representations, symbols for ambiguity, and textual cues aid quick decisions, while the impact of cartographic representations varies based on task complexity and decision makers’ objectives. In this study, we concentrate the participants’ efforts towards one goal, i.e., to identify a hospital accessible within 5 minutes. We are looking at how the proposed visualizations can help decision makers under time pressure by analyzing the cognitive load incurred under different levels of time pressure through a dual-task design.

2.3. Dual-task paradigms

Researchers in cognitive science have developed dual-task paradigms for comparing relative differences in cognitive load between two tasks requiring similar cognitive resources (Koch et al. 2018, Rende et al. 2002). Both primary and secondary tasks compete for similar cognitive resources in dual-task experiments. Therefore, the participants’ performance in the secondary task is influenced by the cognitive load induced by the primary task (Padilla et al. 2020). The difference in performance between a single- and a dual-task is known as a dual-task cost (Wajda et al. 2013). Dual-task paradigms have been used to study a variety of practical concepts, such as driving (Strayer and Johnston 2001) or carrying out visualization tasks (Alonso 1998). The underlying constraint in dual-task paradigms is resource overlap, i.e., an increase in the number of elements in the primary task should affect the performance in the secondary task (Padilla et al. 2020).

In this paper, we use mental addition as a secondary task for assessing the cognitive load of the primary task of making decisions based on uncertainty representations. The primary and secondary tasks were chosen to be cognitive in nature following the work by Esmaeili Bijarsari (2021), and the mathematical addition test implemented in this study is considered a hard secondary task (Padilla et al. 2020). This is a promising approach as it simulates ongoing scenarios in TMCs, where the operators have to focus simultaneously on multiple tasks while assisting medical evacuations in case of traffic incidents.

2.4. Decision-making under time pressure

In the case of traffic incidents, time is one of the critical factors that decision makers need to consider. Even with optimal uncertainty visualizations, the decision makers will need time to acquire information for a choice and cognitively analyze it. Because of this, while making judgments in the real world under uncertainty, decision makers nearly always have internal or external time restrictions (Edland and Svenson 1993, Koch et al. 2018). The effects of time pressure on decision-making with uncertainty in maps have been explored by Korporaal et al. (2020). Their study explores the effects of uncertainty depictions on evacuation planning operations. The study considers the participants’ eye tracking patterns and explores the most important aspects of maps explored by the participants under time-pressure scenarios. Their study revealed clear differences in decision-making strategies under time pressure. Also, in our scenario, control rooms, the decision makers operate under uncertainty and cognitive load.

A survey by Zarei et al. (2016) reveals time pressure as a detrimental factor affecting control room decision-making. Therefore, in this study, we are investigating the effect of two different levels of time pressure on decision-making.

2.5. Cognitive load detection

Cognitive load is one of the essential factors that can affect decision makers’ performance in a safety-critical environment. According to recommendations from Robinson et al. (2018), the quality of decisions in the control room depends on the cognitive load on the human factors. Sweller et al. (1998) in his work on Cognitive Load Theory (CLT), categorizes cognitive load into three types: intrinsic, extraneous, and germane cognitive load. The intrinsic cognitive load depends on the task. The presentation of the task influences the extraneous cognitive load, and germane cognitive load is associated with learning (Mitra et al. 2017). In order to control for intrinsic cognitive load, we ensured consistent task complexities by standardizing the number of hospitals and using road networks of comparable complexity (see Section 3.2.2). We tracked the changes in extraneous cognitive load through eye tracking-based measures.

Researchers use either subjective or psycho-physiological measures to assess cognitive load. The psycho-physiological measures of cognitive load can include measures of neuronal activity (electro-encephalogram, magnetic resonance imaging) (Eckstein et al. 2017), eye movement activity (eye-trackers) (Palinko et al. 2010a), or autonomic nervous system activities (galvanic skin response, GSR) (Shi et al. 2007). The subjective measurements, also known as self-reported measures, include self-rating the cognitive load through a questionnaire such as NASA-TLX (NASA Task Load Index) (Hart and Staveland 1988). We employ multiple modalities to measure cognitive load as a multi-modal approach has been proven to be more reliable than measurement through a single modality (Liu et al. 2020). This paper utilizes eye tracking-based cognitive load estimation techniques, GSR-based physiological metrics, and self-reported measures.

2.5.1. Eye tracking based cognitive load detection

Eye tracking is the process of measuring the movement of the eyes with respect to the head of the participant (Duchowski et al. 2018). Modern eye trackers can also identify constriction and dilation of the participants’ pupils. The measurement of small fluctuations in pupil diameter is known as pupillometry (Hess and Polt 1960). Several researchers have used pupillometry for the investigation of cognitive load (Palinko et al. 2010b, Zargari Marandi et al. 2018, Duchowski et al. 2018, 2020, Krejtz et al. 2020, Castner et al. 2020). Kiefer et al. (2016) have demonstrated the differences in cognitive load between map-based tasks using pupillometry.

A review article by Bunch and Lloyd (2006) puts forward the need for using cognitive load theory in GIS research to design better maps and help learners to grasp concepts properly by considering their cognitive load. A recent study by Keskin et al. (2020) explored the differences between novice and expert map users regarding the incurred cognitive load. The article explored differences in learning ability through electroencephalogram and eye tracking evaluation of participants performing different map tasks. The study investigated saccadic and fixational eye tracking metrics to explore the cognitive processes involved. However, the differences in difficulty levels of tasks were unobservable through the eye tracking parameters used in this study. Therefore, we argue that the sensitivity of the eye tracking parameters used to measure cognitive load plays an important role in determining the cognitive load when minute differences are being explored.

Palinko et al. (2010a) have used mean pupil diameter change (MPDC) to assess drivers’ cognitive load, a metric going back to Ahern and Beatty (1979). The authors identified MPDC as a valuable tool for identifying rapid changes in cognitive load. Along similar lines, Marquart and De Winter (2015) identified MPDC as a metric highly sensitive to changes in cognitive requirements. As the differences between uncertainty representations can be very small, in this paper, we are using MPDC as a metric for cognitive load.

2.5.2. GSR based cognitive load detection

It is well-known that Galvanic Skin Response (GSR) can be used as an indicator of cognitive load. For instance, it has been used to investigate the cognitive load while interacting with multi-modal user interfaces (Shi et al. 2007). However, discussions on how to best analyze GSR for cognitive load measurement are ongoing.

GSR measures the number of sweat glands activated by the autonomic nervous systems at any given time by measuring skin conductance response (SCR). GSR signals are a mix of three components: a slow-changing “tonic” component, a fast-changing “phasic” component, and noise (Greco et al. 2015). SCR activity corresponds to the combined interactions of several neural pathways of the human brain (Critchley et al. 2000). Nourbakhsh et al. (2017) demonstrated that several statistical parameters derived from the SCR, such as the number of SCR peaks, can be used to detect cognitive load.

SCRs have been primarily associated with arousal both in response to emotional and cognitive stimuli (Greco et al. 2015). Their study explored the possibility of the deconvolution of SCR peaks from the acquired SCR signals. The authors used a convex optimization-based approach to the recorded signal into individual SCR associated with the stimulus.

This paper uses the GSR processing approach proposed by Greco et al. (2015) to decompose the signal into SCR peaks. We then use the number of SCR peaks as a measure of cognitive load as suggested by Nourbakhsh et al. (2017) and Shi et al. (2007).

2.5.3. Self reported measure (NASA-TLX)

The most widely accepted definition of workload has been proposed by Hart and Staveland (1988) as the cost that a human operator needs to incur to achieve a particular level of performance. NASA-TLX, a questionnaire for the self-reported measurement of cognitive load, as proposed by Hart and Staveland (1988), has been used widely utilized in various research articles. Hart (2006) explored the breadth of different fields where NASA-TLX has been used as a self-reported measure of cognitive load. NASA-TLX considers mental demand (MD), physical demand (PD), temporal demand (TD), performance (P), effort (E), and frustration (F) as the relevant dimensions of cognitive load. As per the suggestions by Galy et al. (2018), we consider each of the dimensions of NASA-TLX in our study separately instead of relying upon a single global score.

3. Methods

The experiment is designed as a dual-task experiment (Figure 1) in which the participants have to perform two tasks sharing the same cognitive resources under time pressure (Baxter 2002). The experiment followed a within-subject design with uncertainty representation techniques (4 levels: baseline, shading, line, and point visualization; refer to Section 3.2.1 and Figure 3) and time pressure (low - 10 seconds/ high - 5 seconds) as independent variables, yielding a total of 8 trials per participant. Counterbalancing between participants was applied to counteract the effect of familiarity or learning.

Figure 1. Dual-task design.

3.1. Participants

Forty participants with an age of $31.025\pm 6.07$ (twenty female participants with a mean age of $31.4\pm 7.1$ and twenty male participants with an age of $30.65\pm 4.9$) from the Singapore-ETH Centre participated voluntarily in this experiment. Out of forty participants, thirty-eight reported regularly using maps, thirty-six considered themselves good at statistics, while seventeen reported a medium to high aversion to uncertain situations. The participants were selected after email advertisement through a mailing list. They came from various backgrounds ranging from researchers to management personnel. All participants were at least 18 years old and signed an informed consent form before the experiment started. Color blindness was used as an exclusion criterion. Each participant received S$ 30 as compensation for their time. The study protocol was approved by the ETH Zurich Ethics Commission (EK 2022-N-74A).

3.2. Setup and materials

3.2.1. Experimental setup

Figure 2 illustrates the experimental setup. The participants were seated in front of a 32-inch standard display equipped with a standard mouse and keyboard. We employed a wireless EDA sensor (shimmer GSR sensor, 120 Hz) and an eye tracker (Tobii Pro fusion, 120 Hz) for collecting the data during the experiment. The eye tracker was magnetically attached to the bottom of the screen, and the EDA sensor was set to wirelessly broadcast the acquired data through Bluetooth.

Figure 2. Experimental setup: The eye tracker is mounted at the bottom of the screen, and the participant is wearing the Shimmer GSR sensor on his left wrist.

3.2.2. Stimuli: uncertainty visualization and baseline

Each stimulus map consisted of a street network, an incident point marked with a star, and four healthcare centers marked with letters from A to D. The complexity, orientation, and similarity of street networks can have confounding effects. Therefore, we chose eight cities, Glasgow, Kathmandu, Kyoto, Lisbon, London, Oslo, Tehran, and Dublin, as a previous study by Boeing (2019) suggested the street networks of these cities are similar to each other. Following earlier research by Cheong et al. (2016), we removed topographic features such as landmarks and road signs from the maps to ensure participants’ familiarity did not play a role in their decision-making process.

The centroid of the city was chosen as the incident point. The hospital locations were computed from the central incident point as three hospitals (distractors) on the 5-minute travel time and one hospital on the 4-minute 30 seconds travel time (expected choice) using algorithms by Boeing (2017), Schiewe and Knura (2021).

This research relied on publicly available street network data. The data were downloaded, pre-processed, and simplified using the open-source software OSMNX (Boeing 2017). The uncertainties for travel time of 5 minutes as per the recommendations from Wang and Jiang (2003) for all of the eight cities were mapped using the travel time algorithm proposed by Schiewe and Knura (2021). Three levels of uncertainty (Low, Medium, and High) were mapped to the street networks for all three uncertainty visualizations. One example for each of the three uncertainty visualization techniques and for the baseline are depicted in Figure 3:

Figure 3. The four different uncertainty visualization types used in the experiment.

• Baseline visualization was chosen as a visualization of travel time (Figure 3(a)); we used the traditional isochrone realization as proposed by O’Sullivan et al. (2000) as a control visualization using red shading.

• A shading visualization (Figure 3(b)) was used for representing uncertainty. As suggested by Leitner and Buttenfield (2000), we used darker red values to represent more certain information.

• The line visualization (Figure 3(c)) was chosen to be in a similar style as traffic maps currently in use by multiple traffic data service providers such as Google Maps or Mapbox. This representation encodes different levels of uncertainty using different colors following Cheong et al. (2020). It uses a color scheme overlaying the existing road network to visualize uncertainty in travel time. We chose red, yellow, and green to represent high, medium, and low uncertainty as proposed by Schiewe and Knura (2021).

• The point visualization used a three-level color map (Figure 3(d)) to distinguish between three levels of uncertainty. Each intersection on the street network is marked with a point. In this representation, point symbols and colors are used to represent uncertainty following Schiewe and Knura (2021). A palette similar to the line visualization was used to represent uncertainty.

For each of the eight cities, one visualization type was used to represent uncertainty to the participants. One experiment involved eight such trials with all four types of visualizations tested under high and low time pressure.

A pre-study was conducted to check the stimuli used in this study. The pre-study helped us to choose time intervals that are best suited for low- and high-time pressure and check for the number of distractors (Rieger et al. 2021). Based on the pre-study, we chose the time limits for the high and low-time pressure (5s/10s) conditions and implemented a three distractor, one target design for the task.

3.3. Procedure

3.3.1. Pre-experiment briefings

The participants were first given sufficient time to familiarize themselves with an information sheet about the experiment. They were encouraged to ask any questions. After the informed consent form had been signed, demographic information such as age and gender was collected.

3.3.2. Pre-experiment stage

The room’s environmental parameters, for example, light intensity, sound, temperature, and humidity levels, were checked to ensure uniformity across participants. The electrodes from the EDA unit were attached to the participant’s index and middle fingers, and the eye-tracker was checked to ensure both eyes were visible to the tracker. The participants were instructed to try and keep hand and head movement to a minimum during the experiment.

Eye tracker calibration: The eye tracker was calibrated with a 5-point calibration test (Tobii Pro AB 2022). In case of unsuccessful calibration, the calibration process was repeated to ensure the eye tracker records accurate eye tracking data. Before starting the experiment, the wireless link to the EDA unit was checked to ensure proper connectivity.

Demo presentation: At the end of the pre-experiment stage, the participant was given a small demonstration of each of the map-based uncertainty visualizations used in this study using the Tobii pro labs software. The participants were encouraged to ask any questions about the visualizations during this phase.

3.3.3. Experiment stage

The participants had one trial for each uncertainty visualization technique. They were asked to fill in the questionnaire after the completion of each trial. The experiment stage was repeated twice for time pressure (low/high) conditions. The participant received the following task information on the screen: “Your task is to choose the healthcare center (among A, B, C, and D) where the affected individual can be transported fastest from the point marked with a star. In addition, a mental addition task will be presented to you before each visualization, simulating other tasks that decision makers have to perform in control rooms. Finally, please inform us of your choice of healthcare center and the results of the addition. Before we present the next scenario, we will show a cross on the screen; please focus your gaze on the cross and wait for the following scenario to appear. You will be looking at four such scenarios in total. The experiment will be repeated after all tasks are completed”. After the participants read the information, they responded by pressing a button on the keyboard, and the trials began. Each trial began by presenting the participants with a fixation cross at the center of the screen (see Figure 1). Following this, two three digits numbers were presented for five seconds. The uncertainty representation was then displayed. The time pressure was introduced by modulating the stimulus presentation time. The participants had a fixed amount of time (5s/10s) depending upon the low- or high-time pressure criterion, after which the uncertainty visualizations disappeared. After the time ended, the participants responded with their choice of hospital and the sum of the numbers. After each task, the participants were asked to fill out a NASA-TLX questionnaire and open-end questions. All responses were collected on the response sheets. The remote eye tracker and EDA sensor recorded the gaze and skin conductance response throughout the experiment stage.

3.4. Measures

The eye-tracker and the shimmer modules were synchronized using the Tobii pro labs software (Tobii Pro AB 2022). The raw data files containing the participants’ eye tracking and skin conductance values were exported from the software. Finally, the synchronized recordings of pupil diameter and GSR were used to compute different parameters as an objective measure of cognitive load, as detailed in the following. The subjective responses from the participants were collated using a subject id to anonymize the data.

3.4.1. Cognitive load from eye tracking: MPDC

The pupillary activity acquired through the trial was first cleaned by discarding the empty values. It was then processed by filtering the data, then smoothing and averaging the acquired left and right pupil diameter following the recommendations by Kret and Sjak-Shie (2019). The MPDC values were then computed per trial following the recommendations from Marquart and De Winter (2015) and Krejtz et al. (2018).

3.4.2. Cognitive load from GSR: Number of SCR peaks

We have used convex optimization-based processing of SCR proposed by Greco et al. (2015) to remove noise and decompose the acquired signal into phasic and tonic components. The skin conductance response peaks were identified on the clean signal using the neurokits python package (Makowski et al. 2021). We used the number of SCR peaks per second as a measure of cognitive load (Shi et al. 2007).

3.4.3. Questionnaire

Self-reported workload (NASA-TLX): Participants were asked to self-report their workload using a standardized NASA-TLX questionnaire. Each of the NASA-TLX dimensions was individually analyzed, and statistical tests were conducted to check the significance of differences between different visualizations under different time pressure conditions (Galy et al. 2018).

Self-reported preference scores: Participants were asked to rate their preferences for visualizations through the question ‘I will prefer this visualization for making the decision’, using a 5-point Likert scale ranging from ’strongly agree’ to ’strongly disagree’.

Self-reported strategy: To further examine the participants’ perspectives on uncertainty representation techniques, they were asked, ‘How did you reach your conclusion for choosing the hospital? Briefly describe the strategy you used while making the decision’. The inputs from the participants were first coded into multiple categories following suggestions from the streamlined codes-to-theory model (Saldaña 2021). Each of the subtopics was then condensed into four major subcategories or ‘themes’. Finally, each of the themes explored by the participants was studied in detail and will be reported as a part of the discussion section.

3.5. Eye tracking differences in strategies: Gaze transition entropy

We were interested in exploring whether different self-reported strategies (see Section 3.4.3) were reflected in participants’ visual attention. In order to characterize differences in visual attention, Gaze Transition Entropy (GTE) was calculated as a metric that characterizes the complexity and predictability of visual scanning patterns (Shiferaw et al. 2019). The GTE values are computed following Krejtz et al. (2015). We used Tobii Pro AB (2022) to create different areas of interest (with equal dimensions) for the four hospitals and starting points. We calculated the GTE for each visualization type with the gaze data captured during the trial. When individuals employ a systematic and focused scanning strategy, it is characterized by lower GTE values (Krejtz et al. 2015). This often signifies an organized approach, such as sequential examination or targeted exploration of specific areas within a visual scene. Conversely, higher GTE values indicate greater randomness or unpredictability in eye movements, suggesting a more dispersed and erratic pattern of fixations (Shiferaw et al. 2019). Such behavior often corresponds to exploratory or search-oriented strategies where individuals traverse a wider visual space without a defined sequence, potentially reflecting a more comprehensive but less targeted approach to information gathering (Shiferaw et al. 2019).

4. Results

40 participants performed eight trials, yielding a total of 320 trials. The collected data were checked for potential recording errors; none of the participants’ data had to be excluded.

4.1. Cognitive load from eye tracking: MPDC

A comparison between the MPDC levels of different visualizations under different time pressure conditions is shown in Figure 4. A repeated measure analysis of variance (RM-ANOVA) was carried out to assess the combined effects of visualization and time pressure on MPDC. The data satisfied the normality and sphericity conditions required for RM-ANOVA (Muhammad 2023). The test (see Table 1) revealed a non-significant difference in MPDC between the time pressure conditions, while there were significant differences across different visualization scenarios. A pairwise t-test was performed on the MPDC levels to investigate the differences between different visualization types to evaluate differences in cognitive load incurred during the processing of uncertainty visualizations. The results are presented in Tables 2 and 3. All p-values are corrected using a Bonferroni correction.

Figure 4. MPDC for different uncertainty visualizations under varying time pressure.

*denotes significant differences between MPDC values between visualizations, and ns denotes differences that were not significant.

Table 1. RM-ANOVA for different uncertainty representation techniques (UR) and time pressure conditions (TP) for MPDC.

Source	SS	ddof1	ddof2	MS	F	p-unc	p-corr	ng2	eps
UR	0.96736	3	93	0.32245	42.129	$2.30*{10}^{-17}$	$8.02*{10}^{-15}$	0.1987	0.831398
TP	0.018514	1	31	0.018514	1.18429	0.2848	0.2848	0.00472	1
UR * TP	0.0037	3	93	0.0012	0.1753	0.9128	0.8999	0.0009	0.9221

Table 2. Comparison between different uncertainty representation techniques under low time pressure using MPDC.

Visualization 1	Visualization 2	T	dof	p-unc	p-corr	BF10	hedges
line	baseline	−4.008	70.712	0	${0.001}^{*}$	163.76	−0.908
line	point	0.625	68.038	0.534	1	0.282	0.144
line	shading	−3.103	71.991	0.003	${0.016}^{*}$	13.264	−0.692
baseline	point	4.205	69.967	0	$0^{*}$	283.223	0.98
baseline	shading	0.722	73.975	0.473	1	0.298	0.164
point	shading	−3.384	71.999	0.001	${0.007}^{*}$	26.96	−0.775

Table 3. Comparison between different uncertainty representation techniques under high time pressure using MPDC.

Visualization 1	Visualization 2	T	dof	p-unc	p-corr	BF10	hedges
line	baseline	−5.096	75.287	0	$0^{*}$	6438.361	−1.125
line	point	0.539	78.995	0.591	1	0.262	0.119
line	shading	−4.556	80	0	$0^{*}$	993.09	−0.997
baseline	point	5.616	78	0	$0^{*}$	42920	1.244
baseline	shading	0.948	76.409	0.346	1	0.342	0.209
point	shading	−5.13	78.821	0	$0^{*}$	7264.313	−1.128

The results highlight the differences in how the visualization affects the cognitive load of decision makers. Under low time pressure conditions (refer to Table 2 and Figure 4, left), the differences in MPDC between point and line visualization as well as between shading and baseline visualization, were not significant. The MPDC values were significantly lower for line visualization compared to baseline [T=-4.008, p = 0.001] or shading [T= −3.103, p = 0.016] representation. The effect size (hedges = −0.908) indicates a large effect size between the line and the baseline visualizations, while the differences between line and shading (hedges = −0.692) show a medium effect size. For point visualization, the MPDC values were significantly lower than shading [T= −3.384, p = 0.007] and baseline [T = 4.205, p = 0] visualizations. The differences in mean between the point and shading visualizations show a medium effect size (hedges= −0.775), while the differences between the point and baseline visualizations have a strong effect size (hedges = 0.98).

During high-time pressure scenarios (refer to Table 3 and Figure 4, right), the MPDC values followed the same trend as in the low-time pressure condition. However, the effect size was strong in all significant p-value cases.

4.2. Cognitive load from GSR: Number of SCR peaks

The number of SCR peaks per second (SCRpeaks) for all four visualizations under both time pressure conditions is depicted in Figure 5. A Generalized Linear Model (GLM) with Poisson regression model (Table 4) was used to explore the relationship between the occurrence of SCRPeaks and the variable time pressure. The model demonstrated a good fit, indicated by a log-likelihood of -572.20 and a deviance of 316.89, suggesting that the model adequately captured the variability in the data. The associated p-value of 0.000 indicates that the effect of time on SCRPeaks is statistically significant. The number of SCR peaks increased with time pressure. However, paired t-tests revealed the number of SCR peaks between visualizations was not significant.

Figure 5. Peaks of SCR (SCRPeaks) for different uncertainty representation techniques between different time pressure conditions. n.s denotes differences were not significant among visualizations.

Table 4. GLM for comparison between time pressure using SCR peaks.

	coef	std err	z	P $>$ | x|	[0.025	0.975]
time pressure	0.5496	0.021	26.088	0.000	0.508	0.591

4.3. Self-reported workload (NASA-TLX)

NASA-TLX values reported by the participants are presented in Figure 6. A one-way multi-variable analysis (MANOVA) was performed to test our hypothesis that time pressure influences cognitive load using mental demand, physical demand, temporal demand, performance, effort, and frustration as indicators of cognitive load. The MANOVA test results are statistically significant for differences between time pressure scenarios [Wilks’ lambda= 0.0937, F(6, 311) = 3.5613, p = 0.002].

Figure 6. Mean NASA-TLX values for Mental Demand (MD), Physical Demand (PD), Temporal Demand (TD), Performance (P), Effort (E), and Frustration (F).

The MANOVA statistics were further used to investigate the differences between different visualization techniques. The test statistics indicated significant differences among different visualizations under high-time pressure scenarios [Wilk’s lambda= 0.8124, F(18, 421) = 1.7865, p = 0.024]. However, in the case of low-time pressure, there were no significant differences [Wilk’s lambda= 0.8867, F(18, 427.57) = 1.0315, p = 0.424]. To further evaluate the differences between visualizations under high time pressure, different NASA-TLX dimensions were then tested with a pairwise t-test to assess the differences. Table 5 shows the NASA-TLX dimensions with significant differences among visualizations in high-time pressure conditions. The significance level was corrected using a false discovery rate correction following Meena et al. (2019).

Table 5. NASA-TLX dimensions as a comparison among different visualizations under high time pressure.

NASA-TLX dimension	Visualization 1	Visualization 2	T	dof	p-unc	p-corr	BF10	hedges
Temporal Demand	baseline	point	−3.42	76.673	0.001	${0.006}^{*}$	30.201	−0.762
Effort	line	baseline	2.333	75.105	0.022	${0.045}^{*}$	2.378	0.521
Effort	baseline	point	−3.458	74.696	0.001	${0.005}^{*}$	33.503	−0.772
Performance	baseline	point	−2.573	74.89	0.012	${0.036}^{*}$	3.266	−0.555
Performance	point	shading	3.15	76.142	0.002	${0.014}^{*}$	2.344	0.517

Under high time pressure, the temporal demand in processing the baseline information was perceived as lower than the point visualization [T= −3.42, p = 0.006]. The participants experienced lower perceived effort while making decisions based on baseline representation compared to point [T= −3.458, p = 0.005] or line visualization [T = 2.333, p = 0.045]. The perceived performance in point visualization was better than baseline [T= −2.573, p = 0.036] or shading representation techniques [T = 3.15, p = 0.014]. We observed a medium effect size (hedges $>0.5$) for all of the comparisons presented above.

4.4. Decision accuracy

The study recorded the choice of hospital for all participants under varying visualization techniques under different time pressure scenarios. The choices from the model by Schiewe and Knura (2021), were seen as the correct responses. The participants’ choices were checked to see if they matched the model’s choices. The results are presented in Table 6. For all representations, the number of correct responses was higher under low-time pressure conditions.

Table 6. Accuracy for uncertainty representation techniques under different time pressure conditions.

Stimulus	low time pressure	high time pressure
Baseline	87.50 %	84.62 %
Shading	87.80 %	51.28 %
Line	97.43 %	72.50%
Point	77.50 %	67.50 %

4.5. Self-reported strategies and visual scanning patterns

The experience of working with the uncertainty depiction techniques was assessed through the following question: ‘How did you reach your conclusion for the hospital? Briefly describe the strategy you used while making the decision’. The participants’ responses included details of the benefits and concerns of working with uncertainty representation techniques. Data analysis (see Section 3.4.3) identified several categories within four major themes: Color, Distance, Straight roads, and Road network density (refer to Sections 4.5.1–4.5.4). The GTE results on visual scanning patterns are presented in Section 4.5.5.

4.5.1. Color

The most common benefit of representing uncertainty through different colors was its ease of interpretation. The benefit of shading and baseline representations was the quick identification of low, medium, or high uncertainty zones under high time pressure. The participants considered the hospitals closer to areas lower in travel time uncertainty by choosing the hospitals closer to the darker colors. As participant number 33 stated, ”I think this map is easier to read in a short period of time”; other participants reiterated these sentiments. This statement highlights one of the significant benefits of shading uncertainty representation under time pressure. However, the participants preferred points and lines representation of uncertainty for their ease of use. As stated by participant 12, ”This map helps me a lot while making the decision. I can notice that B and C are both in the middle uncertainty (yellow), but C is near low uncertainty (green). In the map, I even do not need to consider the complexity of the transportation network. The uncertainty point near the road network really helps.”

4.5.2. Distance

All of the participants used distance as a metric to make decisions. The participants primarily used distance in three ways. The first way was to visually compute the distance between the point of incidence and the hospitals. The second way was to compute the distance of the hospitals from the nearest uncertainty zone. Finally, the participants visually computed the distance of different segments under different uncertainty levels to the hospitals. Primarily, in the case of shading visualization which used different intensity levels of the same color, the participants applied the first strategy. While in cases where multiple colors were used to depict uncertainty, the participants visually computed their distance from the nearest uncertainty zone. The third approach was chosen by fewer individuals; this approach involved the computation of the fractional distance of road segments under different levels of uncertainty from the hospital to the point of incident. These results demonstrate that the participants could interpret and include uncertainty in their decision-making process.

4.5.3. Straight roads

Participants chose the hospitals having straight roads with the point of incidence. They considered the possibility of the emergency vehicles traveling at increased speed on straight roads. However, when uncertainty information depicted in the maps contradicted the straight road approach, participants chose the path nearest to the low uncertainty zones. As participant number 40 states: “A had the lowest uncertainty and the least amount of medium and high uncertainty in the path. Therefore, I chose it over other centers despite them having a straight path.”

4.5.4. Density of road network

The density of road networks was an essential factor for the participants. Most commonly, participants assumed denser road networks to be detrimental to the travel time of emergency vehicles and generally preferred sparse road networks. Road network density was viewed as an indicator of human activity in the city, and the participants assumed that denser road networks would mean increased traffic. This, in turn, would increase the probability of traffic jams and lower traffic flow rates. However, nine participants among the 40 concluded that having a denser city network could result in multiple options for the same hospital and, therefore, might be a safer alternative. Participant 38 states, “many smaller streets beside the main road which could be used for reaching the hospital when there is too much traffic on the main road” as a reason for choosing the hospital on the denser side of the city.

4.5.5. Gaze transition entropy

While the participants generally agreed on how color and distance influenced their decision-making, noticeable differences emerged in their responses on the influence of road network, particularly the presence of straight roads and varying density of roads. Therefore, we decided to explore the differences between the visual scanning strategies. We divided the participants into two groups based on their self-reported search strategy. We assigned people who focused on straight roads to group $G_{\mathit{\text{sr}}}$ (29 participants) and the participants who preferred a denser street network to group $G_{\mathit{\text{dsn}}}$ (11 participants). A GLM with Gamma distribution assessed the relationship between GTE and the groups. The model demonstrated a good fit with a log-likelihood value of infinity and a deviance of 1448. The associated p-value of 0.000 indicates significant differences in GTE (Table 7) for different groups. The GTE values increased from group $G_{\mathit{\text{sr}}}$ to group $G_{\mathit{\text{dsn}}},$ indicating greater randomness or unpredictability in eye movements, suggesting a more dispersed and erratic pattern of fixations by group $G_{\mathit{\text{dsn}}}.$

Table 7. GLM for comparison between groups ($G_{\mathit{\text{sr}}}$ and $G_{\mathit{\text{dsn}}}$ - based on self reported strategies) and GTE.

	coef	std err	z	P $>$ | x|	[0.025	0.975]
Groups ($G_{\mathit{\text{sr}}}$ and $G_{\mathit{\text{dsn}}}$)	13.7584	0.636	21.623	0.000	12.511	15.005

Furthermore, we investigated the relationship between the uncertainty representation techniques and GTE using a GLM. We found a statistically significant positive relationship (p-value < 0.000, Table 8). This suggests that uncertainty depiction techniques increased the randomness of fixations compared to a baseline with no uncertainty represented on it.

Table 8. GLM for comparison between Stimulus and GTE.

	coef	std err	z	P $>$ | x|	[0.025	0.975]
Stimulus(Baseline, Shading, Line, Point)	7.2072	0.356	20.230	0.000	6.509	7.905

5. Discussion

Through empirical evaluations, we aim to answer the question of which uncertainty visualization will best support transportation decision makers under time pressure, with focus on cognitive load? The goal is to identify the visualization technique which can aid decision makers while reducing their cognitive load. Four different visualizations (baseline, shading, lines, and points) for representing uncertainty were studied in this experiment (n = 40).

In line with expectations (see Section 1), physiological and self-reported measures pointed toward increased cognitive load during high-time-pressure scenarios compared to low-time-pressure scenarios. The results highlight the importance of selecting the optimal uncertainty representation technique when decision makers are forced to make decisions under time pressure.

Again, in line with our expectations, point and line representation techniques generally incurred a lower cognitive load in low- and high-time pressure scenarios than shading and baseline representation techniques. The MPDC values for shading and baseline visualization techniques were significantly higher than those for line and point visualization techniques, suggesting higher cognitive load requirements for area-based uncertainty and travel-time visualization. However, no significant differences in MPDC were found among the point and line visualizations, nor among the shading and baseline visualizations. The differences in MPDC values across time pressure conditions were not significant, indicating a consistent cognitive load across time pressure from the visualizations.

Contrary to expectations, the physiological measure of the number of SCR peaks and self-reported measures of NASA-TLX were unable to differentiate between the differences in cognitive load under low-time pressure conditions. One reason for this can be attributed to the measures’ sensitivity. The SCR parameters and the self-reported measures are not sensitive enough to track the small changes in cognitive load incurred while processing different visualizations. In the case of higher time pressure, the variation of the cognitive load was compounded with the increase in time pressure and, therefore, could be tracked through less sensitive measures such as NASA-TLX.

Under high time pressure, the participants demonstrated higher perceived effort requirements for incorporating line and point uncertainty depictions into decision-making. Although the perceived effort requirement for incorporating line and point representation was higher, the decision makers perceived higher performance while using these uncertainty representation in decision-making. Ultimately the perceived usefulness of uncertainty representation can lead to higher information integration into decision-making as hypothesized by Hullman (2020). One possible explanation for decision makers perceiving similar workload requirements for all visualizations under low time pressure is that the decision makers were able to incorporate all of the data available into their decision-making process. This is in line with previous research by Cheong et al. (2016) and Korporaal et al. (2020).

The participants used the uncertainty information encoded in color to consider the time spent on each route under different uncertainty levels. Uncertainty representation helped the participants to explore multiple options (routes, turns, density).

The accuracy of responses reveals that the participants gave fewer correct responses under the high time pressure condition than the low time pressure condition for all visualizations. The low time pressure condition helped the participants by giving them more time to collect relevant information from the stimulus, thereby improving the accuracy of responses. Although the current state-of-the-art visualization (baseline) performed well under both time pressure scenarios, the cognitive load incurred was much higher, as evident from the physiological (MPDC) and self-reported (NASA-TLX) values. The line visualization, on the other hand, had the highest accuracy in the low time pressure conditions. Furthermore, the cognitive load and effort requirements for line visualization were lower compared to the baseline visualization under low-time pressure.

In line with cognitive load, the empirical data collected during the experiment demonstrate the participants’ perceived usefulness of the uncertainty representation techniques (see Figure 7). The participants strongly preferred line visualizations. Preference for this line representation could be attributed to the higher contrast between the uncertainty zones, making it easier for the participants to extract information from the visualization.

Figure 7. Likert scale ratings demonstrate the participants’ perceived usefulness of the representation techniques in response to the statement “I will prefer this visualization for making the decision.” under low and high time pressure.

Line and point representation techniques, in general, incurred lower cognitive load both in physiological and self-reported measures. The cognitive load and accuracy of the decisions were related to the nature of the uncertainty visualization technique (area-based, line-based, and point-based visualizations). With respect to CLT proposed by Sweller et al. (1998), we can assume that differences in cognitive load between the three types of uncertainty visualizations are related to differences in extraneous cognitive load, since we modified the way the information is presented but not the nature of the task. The cognitive load for line and point visualization was lower compared to baseline because they explicitly represent uncertainty, thus providing different information that helps in making the decision. Furthermore, the exploration of the self-reported decision-making strategies through GTE metrics indicates a significant variation between participant groups, suggesting differences in their cognitive strategies. The results align with what GIS experts expected in Dübel et al. (2017). Combining these factors (lower number of visual elements and the color contrast) led to a lower perceived effort in NASA-TLX, enabling the decision makers to multi-task efficiently. The participants’ choice of visualizations and accuracy scores suggests that line uncertainty visualization is the best possible alternative to represent uncertainty to decision makers under low time pressure; under high time pressure, baseline or shading visualizations appear most promising.

In summary, the study provided important insights for visualizing uncertainty to decision makers under time pressure with emphasis on cognitive load. Through the eye tracking metrics (MPDC), we can successfully observe the differences in cognitive load between different visualizations. The decision makers also demonstrated an affinity towards shading or baseline visualizations under high-time pressure. They focused more on incorporating uncertainty in their decision-making when they had more time to interpret the presented data (low-time pressure conditions). The participants could incorporate the complex street network and uncertainty information into their decision-making process. However, line representation was favored by the participants owing to their ability to reduce the complexity of considering street networks in decision-making, helping the decision makers and thereby reducing the perceived effort.

6. Conclusions and future research directions

This research explored the effects of different map-based representations of uncertainty on the participant’s ability to decide under low and high time pressure. Thus, our experimental design incorporates multiple measures of cognitive load (eye tracking-based, GSR based, and self-reported) as a critical parameter for analyzing the effects of uncertainty representation on decision makers. On the other hand, this research contributes to the decision makers’ perspective of uncertainty representation and aims to identify which visualizations decision makers like to use while making decisions.

The experiment compared three uncertainty depictions and one travel-time representation in a safety-critical decision-making scenario under low-time and high-time pressure. The results show strong evidence of a link between the uncertainty visualization technique and cognitive load. We employed eye tracking-based cognitive load detection measures, which have been tested independently by multiple researchers. The line-based uncertainty depiction technique incurred the lowest cognitive load, followed by the point-based uncertainty depiction technique considering the eye tracking-based cognitive load measures. From the subjective ratings by the participants, the line-based representation was found to be the most preferred representation technique, indicating a direct relationship between ease of use and cognitive load. Moreover, the line representation technique demonstrated superior accuracy under low-time pressure compared to other visualization techniques. The cognitive load perceived by individuals when using line visualization was similar to that of other visualizations.

Dual-task paradigms have been widely used to study working memory requirements for an action; although such experimental designs are widely used in psychology, they are quite infrequent in the visualization and cartography domains. Empirical examinations of decision-making under uncertainty should also examine the effects of additional secondary load on the decision makers. It is common for a decision-maker to have additional sub-tasks while assisting in an emergency. Results are arguably less likely to reflect decision makers’ authentic behavior in emergency decision-making settings without such incentives. We hope the benefits of dual-task design will inspire future researchers to evaluate spatial visualizations accordingly.

The results of this research can be applied to GIS-based decision support systems, especially in risk-laden domains such as emergency response. Emergencies are usually accompanied by periods of high stress and cognitive load. The decision-maker needs to focus on many goals and sub-goals, which can lead to missed signals and eventually deteriorate the resilience of the monitored system. For example, during a traffic accident, the decision-maker might need to contact multiple agencies (hospitals, police, or firefighters) while closely monitoring the situation of traffic on the roads. This study limits the scope and interactions required while performing multiple tasks using a dual-task paradigm. However, the complexity might be even higher in a real use-case scenario, as an operator has to multi-task.

In summary, in this paper, we have examined one aspect of improving decision-making: representing uncertainty to decision makers. Researchers can use the experimental findings from this study to represent uncertainty in networks, thus allowing the stakeholders an improved understanding of the scenario, including the impreciseness of the information (predictions, spatial or temporal information).

Future work should extend this research in several ways. Cartographic representations with uncertain information will aid decision makers in making better decisions by reducing the biases related to missing information. In control rooms, decision makers can leverage advances in uncertainty communication in interactive interfaces and augmented reality to foster a deeper understanding of ongoing situations. This study uses three levels of uncertainty, which can be open to interpretation from the decision makers. Furthermore, the color schemes used in this paper can be made colour blind safe in order to enhance the usability of uncertainty maps. We explore different decision-making strategies employed by decision makers through GTE, and future research can use such metrics to train machine learning classifiers, making them capable of detecting the decision-making strategies and, in turn, to be be used to create adaptive decision support systems. In dynamic systems, hypothetical outcome plots (HOPs) can portray uncertainty’s spatial and temporal effects, making decision makers aware of possible threats or opportunities (Hullman et al. 2015). In addition, modern interaction techniques, such as changing the density of information displayed based on cognitive load or assistance through highlighting critical elements/information on maps, can also help decision makers.

Acknowledgement

This work is an outcome of the Future Resilient Systems project at the Singapore-ETH Centre (SEC) supported by the National Research Foundation, Prime Minister’s Office, Singapore, under its Campus for Research Excellence and Technological Enterprise (CREATE) program.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data and codes availability statement

The data supporting the findings from this study are available at https://doi.org/10.6084/m9.figshare.23616738.

Additional information

Notes on contributors

Suvodip Chakraborty

Suvodip Chakraborty is a doctoral student at the Swiss Federal Institute of Technology (ETH Zurich), affiliated with the Singapore ETH-Centre (SEC) in Singapore. His research interests lie in Human-Computer Interaction, human cognition, and decision-making. He contributed to conceptualization, software development, writing, reviewing, editing and visualization.

Peter Kiefer

Peter Kiefer is a Senior Scientist at the Chair of Geoinformation Engineering at the Swiss Federal Institute of Technology (ETH Zurich, Switzerland), where he is leading the geoGAZElab (https://geogaze.ethz.ch/). His research activities are located at the intersection of Geographic Information Science, Human-Computer Interaction, eye-tracking, Spatial Cognition, and Artificial Intelligence. He contributed to the conceptualization, review and editing.

Martin Raubal

Martin Raubal is a Professor of Geoinformation Engineering at the Swiss Federal Institute of Technology (ETH) Zurich. His research interests lie in human mobility, spatial cognitive engineering, and mobile eye-tracking. He contributed to the conceptualization, review and editing.