Enhancing Urban Data Exploration: Layer Toggling and Visibility-Preserving Lenses for Multi-Attribute Spatial Analysis

TL;DR

This work introduces Layer Toggling and Visibility-Preserving Lenses to enable scalable, multi-attribute spatial exploration of dense urban data within a single-screen visualization. Built on Kepler.gl, the system organizes nine data layers—ranging from crime and taxi trips to weather and socioeconomic factors—into a cohesive framework that supports overlay comparisons and density-adaptive filtering. The methodology defines clear design requirements and tasks, implements the techniques with a Stream Deck for layer toggling, and validates the approach through illustrative case studies and two user studies, showing improved task times with dynamic filters and no significant gains from additional input devices. The contributions include a comprehensive visualization design, adaptive lens mechanisms, a predictive evaluation layer, and empirical evidence that layered, lens-based interaction reduces cognitive load and enhances exploratory analysis for urban planning insights in São Paulo.

Abstract

We propose two novel interaction techniques for visualization-assisted exploration of urban data: Layer Toggling and Visibility-Preserving Lenses. Layer Toggling mitigates visual overload by organizing information into separate layers while enabling comparisons through controlled overlays. This technique supports focused analysis without losing spatial context and allows users to switch layers using a dedicated button. Visibility-Preserving Lenses adapt their size and transparency dynamically, enabling detailed inspection of dense spatial regions and temporal attributes. These techniques facilitate urban data exploration and improve prediction. Understanding complex phenomena related to crime, mobility, and residents' behavior is crucial for informed urban planning. Yet navigating such data often causes cognitive overload and visual clutter due to overlapping layers. We validate our visualization tool through a user study measuring performance, cognitive load, and interaction efficiency. Using real-world data from Sao Paulo, we demonstrate how our approach enhances exploratory and analytical tasks and provides guidelines for future interactive systems.

Figures (5)

• Figure 1: Multisource urban databases are rendered as spatially aligned layers that are overlaid to create a comprehensive view. Each layer visualizes a specific type of spatial data using a dense representation. On the left, each of the nine layers is associated with a button in the button box, enabling users to interactively toggle its visibility on the screen and explore spatial correlations between layers, leveraging retinal persistence. Access can also be achieved through keyboard shortcuts corresponding to each layer number or via mouse interactions. In the top-right, the Visibility-Preserving Lens dynamically adjusts the brush radius based on the graph link density at the mouse position, thereby enhancing graph legibility and serving as a spatial filtering mechanism. For temporal data, the size of the range sliders adapts dynamically to the local distribution. Animation can be employed to sweep across the entire range while maintaining controlled legibility. The layers interact and overlap, facilitating the analysis of relationships between various types of data.
• Figure 2: The layers under consideration, each corresponding to a distinct database. The first column lists the buttons as they appear in the interface, while the second column displays the associated visual representation. The third column illustrates the data representations based on the available types in Kepler, such as points and arcs. The fourth column denotes the data type within each database, including geolocated data and origin-destination.The fifth column shows how the data is integrated into the spatial domain representation, which is the street a graph in our case. The integration gives rise to feature vectors at each vertex of the graph (street corner) used in the predictive model. The sixth column classifies the data as dynamic or static, depending on update frequency and temporal variability. The seventh and eighth columns outline the tasks associated with each layer and the corresponding potential user. Finally, the last three columns describe how the interaction with different devices for toggling is performed, which can be accomplished using a mouse, keyboard shortcuts, or the button box, with each layer represented by an icon.
• Figure 4: Resources for Analytical Tasks: The figure illustrates the spatial representation of the classification model's behavior, addressing the initial question of where the model underperforms, specifically in the peripheral areas. Additionally, on the right, other resources for analysts are presented, including the correlation matrix and the Shapley values.
• Figure 5: 95% Confidence interval for the difference in means considering all comparisons between groups. For example, the comparison "Button Box - Keyboard" represents the difference calculated as the mean of the "Button Box" group minus the mean of the "Keyboard" group.
• Figure 6: Bar plot of the confidence level demonstrated by the participants in answering the questions about the filters.