Desiderata for Intelligent Maps: A Multiperspective Compilation

Interactive digital maps are useful for illustrating and analyzing geographic data and are used for diverse purposes (e.g., wayfinding, data journalism, data analysis, and citizen engagement). This article discusses the requirements of intelligent maps from three perspectives: the literature, a user survey, and a reverse-brainstorming workshop. The ideas brought forth are relevant to researchers and designers of digital maps as they incorporate innovative features and strive for a good user experience.


Introduction
Digital maps are a central part of our everyday lives for business and personal activities.They support users in visualizing and analyzing geographic data with various functionalities.These maps are often interactive and can be provided with "intelligent" features.The interest in creating and using digital maps from a variety of disciplines has led to diverse uses of the terms "interactive" and "intelligent".Here, interactive maps are defined in line with Roth (2013) as maps that support a dialog with users through a computing device.Following Albus (1991) and adapting his definition of intelligent systems to maps, intelligent maps are maps that have the ability to act appropriately in an 'uncertain' environment 1 .Learning is not required to be intelligent but is necessary to become more intelligent as a result of experience (see Albus (1991a)).Computational approaches to realizing intelligent behavior include (Goel 2022) the use of formalized knowledge and inferencing (a.k.a.symbolic artificial intelligence), the inference of association patterns from observational data (a.k.a.connectionist artificial intelligence), or a combination of both (Hitzler et al. 2022).There have been reviews (Roth 2013;Çöltekin et al. 2017;Griffin et al. 2017; Bartling et al. 2022) and research visions (Degbelo and Kray 2018;MacEachren and Kraak 2001;Thrash et al. 2019;Degbelo 2022) discussing the future of interactive/intelligent maps' research.These reviews were primarily literature-based, that is, they provided a snapshot of the current advances and open issues from the literature on digital geographic maps.Their contributions are valuable, and this article aims to shed some light on the open issues and the future of interactive maps from a complementary angle, namely that of user input.
The main question addressed in this article is: 'what do we want intelligent maps to do for us'? (alternative formulation: 'what are the desiderata for intelligent maps'?), and this is addressed from three complementary perspectives: a literature review, the summary of a user survey, and the analysis of results from a reverse-brainstorming workshop that discussed possible features of intelligent maps.The insights from each perspective in isolation, and all three perspectives in combination can serve as a source of inspiration to researchers and designers interested in incorporating innovative features2 for an improved user experience on digital maps.

Perspective I: Background Literature
This section provides background information for the work.Since the current literature has already provided several reviews that have either explicitly discussed (e.g., Roth (2013); Edsall (2009)) or have useful implications for the design of interactive maps (e.g., Çöltekin et al. (2017); Robinson et al. (2017)), the rest of this section focuses on features (Sect.2.1) and examples (Sect.2.2) of intelligent maps.

Features of Intelligent Maps
The recent literature has brought forth several features of 'intelligent' maps.For instance, Degbelo et al. (2022a) listed four features that apply to intelligent maps more broadly: (i) sensing the application context (e.g., external environment, current events) via appropriate sensors, (ii) integrating knowledge about the external environment into an internal model, (iii) becoming active to achieve goals, and (iv) evaluating how good/bad their previous decisions were.In the more specific context of data visualization and sensemaking, Degbelo and Kray (2018) provided a discussion of features of intelligent geovisualizations.Their list includes: (i) automatic geographic data type recognition, (ii) generation of geovisualization designs, (iii) monitoring of users' understanding of geographic facts, (iv) self-optimization, (v) user activity recognition, (vi) algorithmic transparency, (vii) multidevice portability, and (viii) multimodality support.Intelligent features for geovisualizations mentioned in (Viry and Villanova-Oliver 2021) include: (i) adaptation to a context of use, (ii) graphical portrayals built automatically considering the type of data, and (iii) semiological choices allowing users to better or more quickly perceive, understand, or analyze the depicted phenomenon.One example of a feature in line with supporting users through semiological choices (i.e., choices facilitating the constitution of meaning) is intelligent zooming, mentioned in (Lai and Degbelo 2021).Intelligent zooming consists in automatically directing the users to zoom levels where they can see the most interesting patterns as they explore several spatial regions of the dataset.Finally, Yingjie et al. (2001) proposed a list of four features for intelligent geovisualizations: (i) a geodata acquisition element, (ii) a geodata structuring element, (iii) a user-behavior monitoring element, and (iv) an adaptive map presentation element.Adjectives used as synonyms for "intelligent" in the literature include (Völkel et al. 2020): autonomous, adaptive, interactive, dynamic, context-aware, and multimodal.Hence, it is likely that research on intelligent maps is scattered in the literature under these different labels.For this reason, the examples mentioned in the next subsection include research on adaptive, context-aware, and multimodal maps.

Examples of Intelligent Maps/Map-Based Systems
The studies conducted using adjectives associated with 'intelligent' (adaptive, context-aware, and multimodal) in the context of digital maps deserve each, a review on their own.Hence, the examples mentioned below are illustrative only, with no claim of exhaustiveness.The examples come, by and large, from the results of a literature search in Google Scholar using the keywords 'intelligent maps', 'adaptive maps', 'context-aware maps', and 'multimodal maps'.The search was done in March 2023.
Intelligent: A number of tools/systems have been suggested to produce digital maps, based on the encapsulation of Cartography principles as meaningful defaults for users.These tools can be said to feature some kind of intelligence, namely, the semi-automated generation of geovisualization designs.Examples include Descartes (Andrienko and Andrienko 1999), SDG Viz (Gong 2019), Adaptive-Maps (Degbelo et al. 2020), the GAV Toolkit ( Van Ho et al. 2012), the Geoviz Toolkit (Hardisty and Robinson 2011), and commercial solutions, such as ArcGIS Online3 or Carto. 4Next to these systems, Brus et al. (2010) presented an ontology that is used to feed an expert system for Cartography.Intelligent caching of web map tiles was the topic of (García et al. 2013).The authors proposed a neural network that determines the areas that are likely to be most requested by users in the future.These areas are then used to prefetch the tiles that are expected to be most requested.Finally, an early work by Schlieder and Vögele (2002) proposed 'intelligent thumbnails' (i.e., machinereadable indices of the thematic and spatial content of digital maps) to support the reasoning about the relevance of a map to a query.
Adaptive: there are at least three critical elements in the context of adaptive user interfaces and visualization (Chiossi et al. 2022): content, presentation, and interaction.A more elaborated model for map and geovisualization involves (Reichenbacher 2003(Reichenbacher , 2017)): content ( = geographic infor- mation), presentation ( = map layout + map style + map graphics + map text), and user interface ( = map interaction + map functions).
As for presentation, Baus and Wahlster (2002) adapts the presentation of route directions on mobile devices to the quality of location information.Wu et al. (2017) introduced an approach to adapt the positioning of map labels to the zoom level and improve readability.Torres et al. (2019) proposed an approach to relocate points of interest on a map according to various attributes of interest to users, for instance, the true distance when walking from one point to another point (walking distance); the true distance when cycling from one point to another point (biking distance); or the time that it takes when cycling from one point to another point (biking time).Nivala and Sarjakoski (2007) presented and evaluated the GiMoDig prototype, which adapts the visualization of points of interest according to the user's current activity, their language and age group, as well as the time of day/year.They reported that different age groups and different nationalities identified some of the map symbols differently.
As for content, Göbel et al. (2018) used gaze data to adapt both the placement and content of map legends.They reported that users preferred the use of an adaptive approach for the adjustment of the map legend, but not its placement.Mac Aoidh et al. (2009) proposed a method to implicitly learn user profiles based on their location and device interactions.These user profiles are then used to trigger the adaptations of user interface content of a mobile-based application for campus navigation, with the ultimate goal of reducing information overload for users.A similar approach was followed by the RecoMap prototype (Ballatore et al. 2010), which uses information inferred about the users' profiles to dynamically adjust recommendations of items to users of an interactive campus map.Jenny (2012) has introduced an adaptive composite map projection, a projection that adapts the map's geometry to scale, to the map's height-to-width ratio, and to the central latitude of the displayed area.
As for interaction, Kiefer et al. (2017) investigated user preferences during a controlled experiment.They reported that users prefer an adaptation over no adaptation at all, and a togglable adaptation (the adaptation is offered without interfering with the user's task and can be shown/hidden by an explicit user interaction) over a reversible adaptation (the adaptation appears automatically and can be manually reverted).Further examples of adaptive maps can be found in (Sarjakoski and Sarjakoski 2008).
Context-aware: as mentioned in (Nivala and Sarjakoski 2003a), intelligence could be implemented in map-based systems by making them aware of the context.Beyond location, elements of context can include (Nivala and Sarjakoski 2003a, b): the system, the purpose of use, the time, the physical surroundings, the navigational history, the user's orientation, the user's profile, as well as cultural and social elements.Viktoratos et al. (2015) used a knowledge-based approach (i.e., the schema.orgontologies and the RuleML language) to build a context-aware map that delivers targeted offers from POI owners to users.The elements of context taken into account included the user's location and profile.A similar, knowledge-based approach was used by Su et al. (2009), who used the user's location as an element of context to remind users about activities to perform and suggest places for interesting activities.Finally, Kiefer et al. (2013) used the spatiotemporal characteristics of the users' eye movements to identify six users' activities on maps: free exploration, search, route planning, focused search, line following, and polygon comparison.For models of context-aware maps, see Griffin et al. (2017) and Kray and Degbelo (2019).
Multimodal: in the context of Cartography research, the term multimodal is polysemous and has been used, for example, to denote the integration of information about multiple modes of transportation services (Stathopoulos and Tsekeris 2008) or the combination of maps and text for novel readerly experiences in literary texts (Mantzaris 2022).In this section and the rest of the article, 'multimodal' refers to multimodal interaction, i.e., the map responds to inputs in more than one modality or communicates outputs using more than one modality (the definition is adapted from Jaimes and Sebe (2007)).A modality is a mode of communication according to human senses and an input (or output) device.The human senses are sight, touch, hearing, taste, and smell.Examples of input devices include keyboards, pens, mice, microphones, braille-input devices, electronic tongues, electronic noses, and brain-computer interfaces (Augstein and Neumayr 2019).Examples of output devices include monitors, speakers, haptic vests, Electrical Muscle Stimulation (EMS) armbands, braille displays, and vibration motors (Augstein and Neumayr 2019).
The key idea of multimodal interaction is the synergistic combination of modalities instead of their use in separation. 5s for examples related to multiple input modalities, the "put-that-there" system (Bolt 1980) was one of the first to illustrate the opportunities of multimodal interaction for personal and collaborative interaction with maps.The authors used a combination of speech and pointing to add and edit elements on a digital map.The SmartKom system (Jöst et al. 2005) also used a combination of speech input and pointing gestures for map navigation tasks in a tourism scenario.Cheyer and Julia (1995) presented a map-based application for the travel planning domain that uses a synergistic combination of pen/voice interface.The combination of pen-and voice-based input was also used in the Service Transaction System (Oviatt 1997), the Quickset prototype (Cohen et al. 2000) and the CoMPASS prototype (Wilson et al. 2007) to enhance user experience during the interaction with digital maps.Some of these studies presented empirical evidence of the advantages of multimodal interaction over direct manipulation GUIs (Graphical User Interfaces) for map interaction tasks (Oviatt 1997;Cohen et al. 2000).Fuhrmann et al. (2005) presented the DAVE_G prototype that supports interaction with gesture and speech and reported that novice users are able to learn generic GIS interaction functions using both modalities in a short time.Klamka et al. (2015) and Çöltekin et al. (2016) explored how gaze and foot input can be combined to support the interaction with zoomable information spaces and illustrated how this combination can be used for the exploration of the virtual world in Google Earth.In the SleeD prototype (von et al. 2014), users can interact with maps on large displays using two types of input simultaneously: touch interaction with a display wall and touch interaction on a wearable arm display.
As for examples related to multiple output modalities, Kaklanis et al. (2013) provided the 'Open Touch/Sound Maps' tool that uses a combination of haptic interaction, a sonification mechanism and text-to-speech to help users with visual impairments virtually explore a geographic area before going to a specific location.Likewise, Senette et al. (2013) presented a prototype that offers feedback to the users in three possible formats (speech, vibration, or sound alerts) in an attempt to make maps more accessible to blind users.

Summary
This section has presented requirements for intelligent maps explicitly brought forth in the literature (Sect.2.1), and examples of tools implementing some of these requirements (Sect.2.2).Several of the examples mentioned were demonstrations that played the role of proofs-of-concept.Evaluations with users have been less frequent, and even less so, evaluations that enable the extraction of general guidelines for designers.Hence, a key takeaway at this point is that the 5Ws+H questions regarding intelligent maps still deserve more attention in Cartography research (i.e., who benefits from (what kind of) intelligent features?Why?, When?, Where?, and How?).Besides, while the different works mentioned in Sect.2.2 have been discussed separately, based on the adjectives the authors initially used to refer to them, it is worth mentioning that the terms overlap naturally to some extent.For instance, García et al. (2013) used 'intelligent system' and on another occasion 'context-aware system' (García Martín et al. 2013) to refer to their work on predictive modeling for the identification of map tiles that are most likely to be requested by users in the future.This use of various, related adjectives on different occasions is not surprising, because authors stress complementary characteristics of their work on different occasions.
Looking at the features of intelligence (Sect.2.1) and the examples reviewed (Sect.2.2), three key traits of intelligent maps stand out: context sensing, adaptation, and self-evaluation (Fig. 1).The starting definition in the introduction-intelligent maps are maps that have the ability to act appropriately in an uncertain environment-can now be made more explicit: An intelligent map is a map that has (i) the ability to sense its context, (ii) adapt its content, presentation, and user interface, and (iii) self-evaluate with respect to a given objective.
The definition is in line with Albus (1991a), who identified sensory processing (sensing), behavior generation (action), and value judgement (decision) as important tasks of an intelligent system.The definition is also in line with the definitions of technological intelligence reviewed in (Völkel et al. 2020).Intelligence as presented above is not an allor-nothing affair, but it varies in degree.Following Albus (1991a), it is a factor of, among other things, the sophistication of the algorithms used for sensing, adaptation, and selfevaluation; the quantity and quality of information stored in memory; and the elegance of the architecture of the mapbased system.The key terms in the definition are taken from existing frameworks from the Cartography literature.
• Context: in the context model from (Griffin et al. 2017), context information emerges through the interaction of four situational components (user, environment, activity, and map).Context sensing denotes means of obtaining information about these different situational components.
The following dimensions were mentioned in (Griffin et al. 2017) for the different situational components: user (spatial abilities, disability, emotional and affective responses, and individual characteristics), environment (characteristics of the physical environment, and distraction), map use activity (tasks, temporal contextual fac-tors, individual vs group usage, and behavioral profiles), and map characteristics (data types, devices, representation strategies, and affordances).• Content: in the model of Reichenbacher (2017), this refers to map features (e.g., amount of data shown, number of classes, boundaries of classes, and geographic area).• Presentation: in the model of Reichenbacher (2017), this refers to the map layout (e.g., title and legend), map style (e.g., map orientation, map generalization, and levels of detail), map graphics (e.g., position of graphical elements, dimension of graphical elements, and symbolization of graphical elements), and map text (e.g., text font, text style, and text symbols).• User interface: in the model of Reichenbacher (2017), this refers to the map interaction (e.g., interaction style and interaction modalities) and map functions (e.g., the availability or not of specific map functions).• Self-evaluation: oftentimes, there is more than one way of adapting a map in response to situational components of context.For instance, there may be more than one way of adapting map legends in response to users' gazes (Göbel et al. 2018).Self-evaluation (i.e., what Albus (1991a) calls 'value judgement') is thus necessary to discriminate between alternatives.This evaluation happens against a design objective.Examples of design objectives include the typical dependent variables of user studies in Cartography: efficiency, effectiveness, satisfaction, perceived usefulness, ease of use, learnability, memorability, understanding, map readability, map usability, and map user experience.What things might a "perfect" geographic map include for you?(e.g., innovative features, unique UX elements, intelligent user support,...).35 people participated in the survey.Since we were interested in ideas at this stage, irrespective of the users' backgrounds, we did not collect demographic information about the participants.We now briefly summarize their answers.

What Users Like/Dislike in Current Mapping Products
Q2+Q3 asked users to spontaneously name geographic maps (or map-based applications) they deemed nicely/ poorly usable and the reasons for that.Ten respondents did not have an example of a nicely usable geographic map at hand, while seven could not recall an example of a poorly usable map.The rest mentioned different products and the reasons for their appreciation/frustrations. Since we do not aim to expose any of the providers (positively or negatively), we omit the names of the products mentioned in the answers and only report on the reasons why they were mentioned.Some of the features mentioned by the users are rather unsurprising (e.g., broad range of interaction possibilities, good usability in general, and across devices).Quite surprising here was the fact that the users reported the ability to access the raw data underlying the map (or its absence) as positive/negative.Table 2 shows the results.The likes/dislikes are grouped according to two of the three U's of user interface success, namely, utility and usability (the third U: target user group, see Roth et al. (2015) is left out because implicitly represented through the comments touching on the first two U's in this case).

Users' Wishes Regarding Geographic Maps
The last question collected data about what users would want a 'perfect' geographic map to have.The choice of the adjective was intentional, as we aimed to have the respondents go beyond the current status quo.Here also, a few respondents (N = 5) did not have any spontaneous feature coming to their mind.The rest provided a rich set of ideas as Table 3 shows.While most of the features mentioned fall in the category of information presentation, the table shows also some nice ideas for adaptation ('self-configures itself based on the behavior of other map users'), context-awareness ('notices, when I (consciously) choose a different path from the one suggested'), multimodality ('a zoom factor for touch'), and working with data ('possibility of downloading all data elements').At last, some users mentioned generic concepts they associate with a 'perfect' geographic map (instead of specific features), namely: customization, personalization, and accessibility.This suggests that research on intelligent maps will need to learn from and contribute to ongoing research on personalizing maps (Ballatore and Bertolotto 2015) and accessible maps (Froehlich et al. 2019).

Summary of the User Survey
The user survey asked about concepts users associate with intelligent maps, likes/dislikes of users concerning existing maps, and users' wishes regarding future maps.As for

Perspective III: User Workshop
Similar to the survey, the workshop intended to learn about the issues and wishes of users regarding digital geographic maps.The workshop had the following two goals: (i) identify features of interactive maps that improve or degrade the user experience, and (ii) identify innovative features that can be embedded in digital maps to better assist users in completing their tasks with digital maps (e.g., wayfinding and/or geographic data analysis tasks).

Procedure
The workshop started with two invited talks that gave examples of the use of interactive maps and inspiration for intelligent features of maps.Afterward, the participants were introduced to the method of reverse brainstorming that was used subsequently during the breakout sessions to structure the exchanges.The participants were asked to write down issues they currently face during the interaction with digital maps-in one application domain-as cards on a digital online whiteboard (Miro6 ).Then, they thought about and wrote down solutions cards to the specific issues of their domain individually.The problem-solution pairs were then discussed with all participants of the breakout sessions.
Since the focus was on identifying possible innovative features, the participants were introduced to four application domains that were preselected by the organizers before the session.The topics were chosen, so that they present a mix of both established and emerging application domains for digital maps.The topics were (in alphabetical order): • Data Analysis: the use of maps to facilitate the analysis of geographic data (e.g., reducing the amount of information, saving and sharing insights from a data analysis session, suggesting shortcuts for repetitive tasks) is one of the prime application domains of interactive maps.Several studies looked into the use of maps for data analysis tasks (Roth et al. 2015;Andrienko et al. 2002;Percival et al. 2022); • Data Journalism: the use of maps with the aim of illustrating open geographic data in the field of journalism is increasing.For example, geopolitical maps depicting changes over time are being used more frequently.Recent work on storytelling with maps (Mocnik and Fairbairn 2018;Caquard and Cartwright 2014;Roth 2021) is particularly relevant to that application domain; • Citizen Participation: A promising application context for the use of digital maps is to promote citizen participation (see, e.g., Fechner and Kray (2014)).Previous work relevant to that area has studied, for example, factors of user experience for incident reporting on mobile devices (Einfeldt and Degbelo 2021), the use of speech-based commands to contribute information on mobile maps (Degbelo and Somaskantharajan 2020), and the use of proactive notifications on mobile devices to raise citizens' awareness about engagement opportunities (Fechner et al. 2016); • Wayfinding: wayfinding is also an established application domain for interactive maps with several studies in that context, for example on the use of schematic maps for wayfinding tasks (Galvão et al. 2021), tactile cues to facilitate wayfinding (Pielot et al. 2009), or augmented reality during navigation (Dong et al. 2021).Trends regarding the use of maps for navigation were identified as of particular interest to the UX community in a recent survey (Degbelo et al. 2022b).
Based on the votes from the participants, three topics were selected for the breakout sessions: data analysis, citizen participation, and wayfinding.The workshop was attended by 24 participants and lasted about 120 min.About 14 people took part in the reverse-brainstorming sessions.After the workshop, the outcomes of the breakout sessions were post-processed by the authors (e.g., formulate map-related solutions to the problems identified more sharply) and harmonized across scenarios.The results are presented below.

Information search
Difficulty in downloading data specifically for desired areas/observables The map supports smart downloading of data sets (location-based) Unclear in which repository to search for data The map intelligently derives user profiles to assist during data search Not the desired file extension The map supports the conversion of datasets to different formats Data are not available or cannot be found for my area of interest The map suggests an area that is very similar to my area of interest and where data are available Websites with available local data are in different languages The map has an intelligent translator that specializes in geographic terms Gaps in the (meta)data (Missing) Continuity over time The map supports interpolation for missing values Metadata unclear The map enables contact establishment for exchange between the data provider and the data user Missing/incomplete metadata The map enables metadata enrichment (by providers or automatic merging of datasets from different sources) Data have poor quality (different reference systems, data gaps) The map shows quality assurance workflows for data; the map enables enrichment of metadata (by providers or automatic merging of data sets from different sources) Data creation is not well documented (which method of cloud masking is used) The map enables the enrichment of metadata (by providers or automatic merging of datasets from different sources)

Miscellaneous issues
Data available but not under an open license N/A Poor/no network (mobile) ->missing information N/A Time series have too many outliers N/A

Information search (Missing) Overview of current participation opportunities
The map as an entry point for participation opportunities (Missing) Geographical overview of political candidates The map as an entry point for information about candidates and elected representatives (Missing) Fine-grained spatial distribution of voter turnout The map shows the spatial distribution of voter turnout at different granularities (Missing) Visual idea of spatial consequences of (urban) projects The map as a display medium of links between different data sources (Missing) Information about the feedback on my comments The map informs citizens about the next steps related to their concern Find suggestions from other citizens The map recognizes and shows spatially and/or semantically similar contributions

Working with data
Putting points precisely on the map The map supports defining vague locations Conveniently draw on the map what I want to communicate The map supports drawing of different geometric shapes Upload photos (in addition to free text) The map recognizes and displays text-image relationships (e.g., highlights the same entity, both in text in image as well)

Adaptation
Basic usage difficulties (e.g., browser outdated) The map adapts to different working environments to ensure the best user experience

Information presentation
2D displays must be rethought by the user into 3D, e.g., for complexity estimation The map offers adaptive 2D/3D visualizations Only photos of surroundings are currently used to share information about climbing sites The map is meaningfully combined with other media to share information about climbing sites Mobile displays are very small The map is responsive; the map supports cross-device interaction a Too much unimportant/irrelevant information in the way The map supports the filtering of information; the map 'deliberately' omits information in order to ensure a user-friendly information density

Context-awareness
Spontaneous obstacles in traffic (traffic jam, construction site, accident) The map adapts displayed information to spontaneous changes in the environment Estimating (travel) time considering current information The map shows (travel) time taking into account real-time data from various sources Elevation information difficult to find efficiently on the map The map displays elevation information efficiently and contextually Wayfinding Navigation on a bike with a cell phone difficult because the phone cannot be seen well The map offers alternatives to manual control (e.g., voice control); the map supports auditory control (e.g., via headphones intelligently adjusted to traffic louder or softer) or non-visual (haptic) feedback (e.g., vibration) Orientation in complex buildings difficult, e.g., finding an exit in a parking garage or university buildings The indoor map 'detects' user's orientation difficulties and offers appropriate help Information about paths only known to a few is missing The map gives users the possibility to add information about unknown paths In large sites with several buildings, navigation applications often lead the user only to the site boundary and not to the building they are looking for The map offers users the ability to add paths to reach buildings

Results
The same categories used in Table 3 to describe the results of the user survey (context-awareness, adaptation, information search, information presentation, working with data, and multimodality) are reused here to group the results of the workshop.Tables 4, 5, and 6 present the outcomes for the three application contexts.Since all items in Table 4 revolve around the topic of 'working with data', the topic is not explicitly mentioned in the table.If map-based features cannot directly address a problem identified during a breakout session, this is signaled by a 'N/A' in the table.
The tables suggest the following key takeaways for the three application contexts: • Data Analysis: the issues talked about touched upon the topics of information search and data/metadata incompleteness more broadly.While some of the issues cannot be addressed by maps directly (e.g., the lack of an open license for a dataset) and some are already addressed by existing solutions (e.g., map-based solutions supporting the conversion to different formats 7 ), some ideas could be worth exploring.In particular, the topics of intelligent assistance of users during geographic data search, multilingual support, and support for metadata enrichment still deserve some attention from current research on the design of map-based systems.• Citizen Participation: the issues talked about can be grouped under the themes of information search, working with data, and adaptation.Here also, some systems have been proposed/implemented, which address some of the issues (e.g., A 'Dialog Map' that displays engagement opportunities for sustainability projects8 ).Nonetheless, some ideas point at directions worth investigating in future work, notably helping citizens visualize/ understand the spatial consequences of urban projects through maps, the use of maps as an entry point to visualize information about candidates/elected representatives, and helping citizens find other citizens with similar ideas through recommendations of spatially and/or semantically similar contributions.• Wayfinding: the issues talked about can be grouped under the themes of information presentation, context-awareness, and wayfinding-specific matters.Here, the exploitation of contextual information (e.g., user profile) to adapt the display (2D vs. 3D) during task completion, the adaptation of the information shown on the map to spontaneous changes in the environment, and techniques to learn about users' orientation difficulties to provide adequate help were mentioned as examples that can inspire future research.

Desiderata for Intelligent Maps
We now revisit the question mentioned at the outset of the article: 'what do we want intelligent maps to do for us'.The section connects the ideas from the three perspectives (literature, user survey, and user workshop).We first mention adjectives relevant to the design of intelligent maps, before moving on to the features of these.While compiling the features, a key assumption was that intelligent maps embed automation.This unique angle on intelligent maps is what sets this article apart from the research agendas that have previously discussed the future of digital maps (MacEachren and Kraak 2001;Roth 2013;Griffin et al. 2017;Çöltekin et al. 2017).
Adjectives: These describe the qualities of products and hence give insights into possible characteristics of intelligent maps.The literature review discussed three adjectives associated with 'intelligent': context-aware, adaptive, and multimodal (Sect.2).The user survey (Sect.3) has suggested four additional adjectives: interactive (Table 1), and customizable, personalized, accessible (Table 3).Hence, an intelligent map is interactive, context-aware, adaptive, multimodal, customizable, personalized, and accessible.These seven characteristics are not mutually exclusive.
Features: The list below illustrates automation features that map research and design could strive to (increasingly) embed into existing maps.This embedding should integrate findings from user experience research, for instance, that users welcome automation but prefer to remain in control (i.e., decide to accept/reject the suggestions of the computer system; see Kiefer et al. (2017)).The source of the ideas is documented using the letters L (literature review), S (user survey), and W (user workshop).Features relevant to more than one thematic category are not counted twice.

Automation features during information search
1. automatic generation of thumbnails that summarize the maps' content (L) 2. automatic translation of content during search (SW) 3. automatic suggestion of spatially/thematically similar items (W) 4. automatic recognition of users' intent/activity (LW) 5. automatic recognition of the settings of the working environment (W).

Automation features during information presentation
6. automatic adjustment of the information items relevant to users (S) 7. automatic adjustment of the map projection (L) 8. automatic identification of tiles most likely to be requested by users (L) 9. automatic adjustment of the map legend (L) -automatic recognition of the settings of the working environment (W).

Automation features during the work with data
10. automatic recognition of the type of geographic data at hand (L) 11. automatic suggestion of meaningful geovisualization designs (L) 12. automatic monitoring of users' understanding of geographic facts (L) 13. automatic reconfiguration of the map user interface to maximize user understanding (L) 14. automatic redirection to zoom levels with the most interesting patterns (L) -automatic recognition of the settings of the working environment (W) -automatic recognition of users' intent/activity (LW).

Additional features specific to wayfinding
15. automatic adjustment of the navigation support according to the familiarity with places (S) 16. automatic adjustment of suggestions according to realtime changes in the traffic (W) 17. automatic recognition of users' orientation difficulties (W).
The 17 features present opportunities for research in Cartography, Software Engineering, and Human-Computer Interaction.9

Discussion
The ideas from the literature and the outcomes of both the user survey and workshop were consolidated into a tentative definition (Sect.2.3), seven adjectives that describe the qualities of these maps (Sect.5), and features of intelligent maps (Sect.5).A consensus-based definition of 'intelligence' is still a work in progress, but as mentioned above, the stance of this work is that automation is central to intelligent maps.Ergo, the ethical issues and requirements posed by autonomous systems in general apply to intelligent maps as well, most notably, privacy threats (Such 2017), vulnerability to bias (Mehrabi et al. 2022), and demands for transparency (Bryson and Winfield 2017).These issues and requirements are acknowledged, though not discussed indepth in this article.Several standards were brought forth to address ethical considerations of autonomous systems (e.g., IEEE 7000-2021: consideration of ethical values; IEEE 7001-2021: transparency; IEEE 7002-2022: data privacy).
The literature on intelligent maps will benefit from case studies documenting their use in specific scenarios as well as the lessons learned.Finally, it is worth briefly mentioning some ongoing work that may share some similarities with the current work, but have a different emphasis.For instance, Kang et al. (2022) recently presented a review of GeoAI research for Cartography.While artificial intelligence may be useful as one builds more intelligent maps, the focus of the work reviewed in (Kang et al. 2022) is not on context sensing and adaption, but instead on map creation from the existing data.The term 'AI-generated maps' is occasionally used to refer to that line of work (Kang et al. 2023).Likewise, because intelligent maps may be multimodal, research on these may learn from existing research on audiovisual Cartography (Edler et al. 2019), or multimedia Cartography (Cartwright and Peterson 1999) more generally.Here also, it is the distinct focus on context sensing, adaptation, and self-evaluation that would set an intelligent multimedia map from a 'non-intelligent' or less intelligent one.

Limitations
This study has a few limitations, notably, that the data collected both in the user survey and the user workshop were collected from German-speaking participants.Hence, the features mentioned cannot be said to be representative of all users of maps, nor can they be said to be exhaustive.An additional limitation induced by the way the data was collected is that the results in the tables were translated from German to English and thus mirror the gist of the participants' ideas, not the original phrasings used to mention their issues and wishes.Furthermore, some of the items mentioned by users (Tables 2, 3, 4, 5 and 6) can be assigned to different/many categories, but the presentation has chosen one that appears most relevant.Finally, the problems collected were at different levels of granularity due to the data collection methods (user survey and workshop discussion from participants with a broad range of backgrounds).While this was helpful at this stage to get a preliminary idea of features for intelligent maps more broadly, more specific data collection activities will be useful to further sharpen the insights (e.g., according to further application domains or the devices used for cartographic interaction).

Future Work
There are opportunities for further research that were identified in the process of writing this article.First, a more elaborate formalization of 'intelligence' in the context of map design is needed, to relate and compare different approaches that make digital maps more intelligent.That formalization would need to clarify the scope of associated adjectives, beyond the preliminary listing that was done in this work, based on the existing literature and the user survey.Related to this, a second opportunity for future research touches upon state-of-the-art reviews to document lessons learned and open areas for future research on intelligent maps.This article has provided a foundation to build upon, but interdisciplinary reviews on intelligent maps and systematic reviews using each of the related adjectives would increase our understanding of the achievements of past research related to that area.At last (and from a more practical point of view), there is a need for a systematic account of the different types of intelligence and their implications (i.e., benefits and tradeoffs) for user experience in the context of interactive maps.This type of work will inform designers about the expected consequences of innovative features integrated into digital maps as they work toward improving the existing solutions.

Conclusion
This article has discussed the requirements of intelligent maps from three perspectives: the literature, a user survey, and a reverse-brainstorming workshop.The requirements extracted from these three perspectives were compiled into 17 features that present opportunities for research on information presentation, information search, and working with geographic data.These 17 features stand for research directions that can inspire researchers and designers of digital maps interested in incorporating innovative features to improve the user experience on digital maps.

Fig. 1
Fig.1Three key traits of intelligent maps: sense the context, adapt map elements accordingly, and evaluate themselves with respect to design objectives

Table 1
When I think of intelligent geographic maps, the first thing that comes to mind is.

Table 2
Likes and dislikes of users regarding existing geographic maps, grouped according to utility and usability aspects

Table 3
Features of the 'perfect' geographic map mentioned by users grouped according to nine categories: context-awareness, adaptation, information search, information presentation, multimodality, working with data, miscellaneous features, non-functional requirements, and general concepts associated with a 'perfect' map

Table 4
Current issues and possible map-based features addressing them for the application context 'data analysis'

Table 5
Current issues and possible map-based features addressing them for the application context 'citizen participation'

Table 6
Current issues and possible map-based features addressing them for the application context 'wayfinding' Brudy et al. (2019)2019)for a recent review of cross-device researchProblemPossible Map-based Features