VolumeSTCube: A Spatiotemporal Visualization Framework

• VolumeSTCube is a spatiotemporal visualization framework that converts discrete sensor data into a continuous volumetric field for advanced 3D analysis.
• It employs coupled volume rendering and iso-surface extraction to mitigate occlusion and clarify complex spatiotemporal patterns.
• Interactive operations like temporal slicing, spatial spotlighting, and DBSCAN-based cluster selection facilitate focused analysis of large-scale datasets.

VolumeSTCube is a volume-based space-time cube framework for large-scale continuous spatial time series in which discretely distributed spatial time series data are transformed into continuously volumetric data, then analyzed through coupled volume rendering, surface rendering, and direct 3D interaction (Deng et al., 14 Jul 2025). It addresses the space-time cube’s well-known issues such as visual occlusion and depth ambiguity, especially under large-scale spatial time series, by treating time as the vertical axis of a scalar volume and by combining transparency-based volumetric depiction with iso-surface extraction. Within the broader literature, it is most precisely understood as a spatiotemporal visualization framework rather than as an exact geometric volume calculator, although it shares volumetric and Marching-Cubes machinery with several cube-based computational methods.

1. Conceptual definition and problem setting

VolumeSTCube is designed for continuous spatiotemporal phenomena represented initially as discrete sensor time series at fixed spatial locations. Its immediate context is the classical space-time cube (STC), which presents spatial and temporal information synergistically but is “plagued by well-known issues such as visual occlusion and depth ambiguity,” and these issues are “further exacerbated when dealing with large-scale spatial time series data” (Deng et al., 14 Jul 2025). The framework’s response is to replace columnar or sparse event-centered encodings with a continuous volumetric field and to render that field through a hybrid visual pipeline.

The framework is defined by three tightly coupled stages. First, it transforms discretely distributed spatial time series data into continuously volumetric data. Second, it employs both volume rendering and surface rendering: volume rendering is used “to mitigate visual occlusion,” while surface rendering “provides pattern details by enhanced lighting information.” Third, it provides interactions for exploration and analysis from “temporal, spatial, and spatiotemporal perspectives” (Deng et al., 14 Jul 2025).

This organization gives VolumeSTCube a dual role. Analytically, it serves as a multiscale inspection environment for hotspots, cycles, and propagations. Representationally, it reinterprets the STC not as a stack of graphical marks, but as a sampled scalar field over a cube-like domain whose horizontal dimensions are geographic and whose vertical dimension is time. A plausible implication is that the framework shifts the central problem of STC design from mark placement to field reconstruction, transfer-function design, and interactive 3D focus control.

2. Transformation from discrete time series to a volumetric field

The data model starts from a set of discrete sensor time series

$$\{V_i\}_{i=1}^S,\quad V_i = \{v_{i,1}, v_{i,2}, \dots, v_{i,T}\},\; v_{i,t}\in\mathbb{R},$$

with fixed spatial coordinates $s_i=(x_i,y_i)$. VolumeSTCube converts these observations into a regularly sampled 3D scalar field

$$\mathcal{V}:\Omega\subset\mathbb{R}^3\to\mathbb{R},\qquad \Omega=[1,m]\times[1,n]\times[1,T],$$

where $(x,y)\in[1,m]\times[1,n]$ index the geographic grid and $t\in[1,T]$ indexes time (Deng et al., 14 Jul 2025).

The spatial step subdivides the map into an $m\times n$ grid of cells $c_{x,y}$. At each timestamp $t$, only the $S$ scattered samples $\{v_{i,t}\}_{i=1}^S$ are known, so the framework reconstructs the full 2D field $s_i=(x_i,y_i)$0 by “any standard geostatistical or inverse-distance–weighted (IDW) method.” In the IDW example,

$s_i=(x_i,y_i)$1

whereas in Kriging,

$s_i=(x_i,y_i)$2

Stacking these 2D reconstructions over $s_i=(x_i,y_i)$3 yields the volumetric field $s_i=(x_i,y_i)$4 (Deng et al., 14 Jul 2025).

The temporal step applies smoothing independently to each voxel’s time series using a sliding average of window $s_i=(x_i,y_i)$5: $s_i=(x_i,y_i)$6 After this pass, the volume $s_i=(x_i,y_i)$7 has spatial smoothness inherited from interpolation and temporal smoothness from averaging. The domain mapping is then fixed: the horizontal $s_i=(x_i,y_i)$8 plane is the geographic map, the vertical axis $s_i=(x_i,y_i)$9 is time $$\mathcal{V}:\Omega\subset\mathbb{R}^3\to\mathbb{R},\qquad \Omega=[1,m]\times[1,n]\times[1,T],$$0, and the full 3D domain

$$\mathcal{V}:\Omega\subset\mathbb{R}^3\to\mathbb{R},\qquad \Omega=[1,m]\times[1,n]\times[1,T],$$1

becomes the space-time cube itself (Deng et al., 14 Jul 2025).

3. Rendering model: direct volume rendering and iso-surfaces

Once $$\mathcal{V}:\Omega\subset\mathbb{R}^3\to\mathbb{R},\qquad \Omega=[1,m]\times[1,n]\times[1,T],$$2 has been constructed, VolumeSTCube uses two complementary rendering modes. The first is direct volume rendering through front-to-back raymarching. Along each eye ray, with ray parameter $$\mathcal{V}:\Omega\subset\mathbb{R}^3\to\mathbb{R},\qquad \Omega=[1,m]\times[1,n]\times[1,T],$$3, extinction density $$\mathcal{V}:\Omega\subset\mathbb{R}^3\to\mathbb{R},\qquad \Omega=[1,m]\times[1,n]\times[1,T],$$4, emissive color density $$\mathcal{V}:\Omega\subset\mathbb{R}^3\to\mathbb{R},\qquad \Omega=[1,m]\times[1,n]\times[1,T],$$5, background light $$\mathcal{V}:\Omega\subset\mathbb{R}^3\to\mathbb{R},\qquad \Omega=[1,m]\times[1,n]\times[1,T],$$6, and exit distance $$\mathcal{V}:\Omega\subset\mathbb{R}^3\to\mathbb{R},\qquad \Omega=[1,m]\times[1,n]\times[1,T],$$7, the accumulated radiance is

$$\mathcal{V}:\Omega\subset\mathbb{R}^3\to\mathbb{R},\qquad \Omega=[1,m]\times[1,n]\times[1,T],$$8

In discrete form, a transfer function

$$\mathcal{V}:\Omega\subset\mathbb{R}^3\to\mathbb{R},\qquad \Omega=[1,m]\times[1,n]\times[1,T],$$9

maps scalar values to color and opacity, and front-to-back compositing updates

$(x,y)\in[1,m]\times[1,n]$0

A global opacity threshold $(x,y)\in[1,m]\times[1,n]$1 lets users skip all voxels satisfying $(x,y)\in[1,m]\times[1,n]$2 (Deng et al., 14 Jul 2025).

The second mode is iso-surface rendering. The framework extracts the iso-surface

$(x,y)\in[1,m]\times[1,n]$3

through a Marching-Cubes pass, computes normals $(x,y)\in[1,m]\times[1,n]$4 per triangle, and applies standard Phong shading,

$(x,y)\in[1,m]\times[1,n]$5

where $(x,y)\in[1,m]\times[1,n]$6 are ambient, diffuse, and specular coefficients, $(x,y)\in[1,m]\times[1,n]$7 are light intensities, $(x,y)\in[1,m]\times[1,n]$8 is the light direction, $(x,y)\in[1,m]\times[1,n]$9 is the reflected view vector, and $t\in[1,T]$0 is the shininess. In the paper’s formulation, this mode recovers “crisp shape boundaries” and highlights “fine spatiotemporal structure and propagations” (Deng et al., 14 Jul 2025).

The rendering design is therefore explicitly hybrid. Volume rendering is the occlusion-management mechanism; surface rendering is the boundary-articulation mechanism. This suggests that VolumeSTCube treats scalar-field exploration and pattern segmentation as visually adjacent tasks: the transparent field provides context, while the iso-surface provides analyzable structure.

4. Interaction design and analytical operations

VolumeSTCube supports three classes of direct 3D manipulation. The first is volume slicing, described as temporal brushing. Two horizontal clipping planes orthogonal to the $t\in[1,T]$1-axis define a time interval $t\in[1,T]$2. Dragging the bottom map plane upward sets $t\in[1,T]$3, dragging a second plane sets $t\in[1,T]$4, and all voxels outside $t\in[1,T]$5 are clipped away (Deng et al., 14 Jul 2025).

The second is volume spotlight, described as spatial brushing. A circular spotlight with radius $t\in[1,T]$6 and center $t\in[1,T]$7 is specified in the $t\in[1,T]$8 plane. Any voxel whose projection $t\in[1,T]$9 satisfies $m\times n$0 has its opacity set to zero, leaving only the focused cylinder of data. This preserves a direct spatial reference to the map plane while isolating a region of interest in the full 3D cube (Deng et al., 14 Jul 2025).

The third is voxel-cluster selection, intended for spatiotemporal pattern picking. After thresholding at $m\times n$1, the framework runs 3D DBSCAN on the positions of the remaining voxels to detect connected clusters. Each voxel receives a transparent, mouse-clickable “ghost” cube; clicking one voxel selects its entire cluster. The system then automatically performs the corresponding slicing and spotlight operations to isolate that cluster and zooms the camera to focus on it (Deng et al., 14 Jul 2025).

These interactions are not independent widgets but a coordinated analytical grammar. Temporal slicing isolates intervals, spatial spotlight restricts geographic support, and DBSCAN-based cluster selection couples both operations to detected 3D structures. The expert feedback characterizes the slice-and-spotlight combination as a high-flexibility “focus + context” mechanism in 3D, while the cluster selection is described as a powerful “find and zoom” operation (Deng et al., 14 Jul 2025).

5. Empirical evaluation and air-quality case study

The computational profile reported for VolumeSTCube is oriented toward large volumetric datasets. Transformation requires per-slice interpolation with cost $m\times n$2 for each of $m\times n$3 slices, followed by temporal smoothing at cost $m\times n$4. On a PC with Intel i7-13700K and NVIDIA RTX3070, a $m\times n$5 volume, approximately 50 million voxels, can be preprocessed in under 30 seconds. Rendering uses GPU ray-marching at $m\times n$6 screen resolution with 256 samples per ray and runs at 15–30 fps for this 50 M-voxel volume. GPU-accelerated Marching Cubes supports interactive updates of $m\times n$7 in much less than 1 second (Deng et al., 14 Jul 2025).

The controlled user study was a within-subjects comparison between VolumeSTCube and a column-based baseline by Thakur & Hanson ’10, run over two halves of a real AQI dataset. The participants were 12 undergraduates, 6 male and 6 female, with domain-agnostic air-quality knowledge. The task set comprised 18 questions spanning lookup, comparison, and relation-seeking, at both elementary and synoptic levels, across space-oriented, time-oriented, and spatiotemporal-oriented analyses. Accuracy and response time were the two metrics, and paired Wilcoxon signed-rank tests were used for statistical analysis. The reported results are specific: VolumeSTCube yielded at least 25% higher accuracy on Q7, Q8, Q12, Q17, and Q18; it produced significantly lower response times on Q4, Q8, Q11, Q13, and Q17 at $m\times n$8 or $m\times n$9; and there was no task on which the baseline outperformed VolumeSTCube (Deng et al., 14 Jul 2025).

The real-world case study used China’s first-half 2018 hourly AQI data, comprising 448 stations and 8,472 hours, interpolated to a $c_{x,y}$0 grid of approximately 50 million voxels. At $c_{x,y}$1, the full variation was visible but noisy; increasing $c_{x,y}$2 to 150 isolated major pollution episodes. A spotlight on Xinjiang revealed three periodic high-AQI pulses starting in March, whereas a spotlight on Beijing-Tianjin-Hebei showed only short, mild pollution bursts, none lasting beyond April. Slicing the Spring-Festival interval in late January showed AQI below 150 nationwide except for North-Central “fireworks” hotspots. Slicing March 26–29 exposed a large dust-storm patch over Inner Mongolia, the Loess Plateau, and the North China Plain, and lowering $c_{x,y}$3 to 100 outlined secondary-pollution fringes into the Northeast. A late-May storm produced a smaller moderate-AQI footprint around the Taklamakan region. Two propagation processes appeared as tilted iso-surfaces at $c_{x,y}$4: one from the Loess Plateau to the North China Plain and then to the Yangtze Delta as a one-day event, and one as a rapid Southeast sweep from the Henan/Shandong border into the Shanghai region with a flat-headed plume. Click-selecting each surface triggered automatic slicing, spotlighting, and camera fly-to behavior, and the analyst linked the six hourly snapshots to known cold-wave events (Deng et al., 14 Jul 2025).

The subjective evaluation in both study and case analysis is also precisely reported. Subjects rated the combined volume-and-surface rendering as highly intuitive, cited near-zero occlusion at high thresholds, and praised voxel-cluster click selection as a “one-step hotspot finder.” The expert feedback emphasized intuitive 3D mapping of time to height and the 2D map to $c_{x,y}$5, minimal occlusion through volume transparency, crisp iso-surfaces for pattern articulation, and the flexibility of slice-and-spotlight operations (Deng et al., 14 Jul 2025).

6. Relation to cube-based volumetric computation and naming scope

VolumeSTCube uses volumetric fields and a Marching-Cubes pass, but it should be distinguished from research whose central objective is exact or accelerated volume computation inside cube-based discretizations. Nguyen-Le et al. develop an algorithm for efficient computation of intrinsic volume from CT or MR volumetric data by combining multivariate calculus, the marching cube algorithm, and a binary indexed tree data structure. Their method reduces the $c_{x,y}$6 binary corner patterns, via symmetries and complements, to 30 distinct volume-computation cases; computes a per-cube volume in $c_{x,y}$7 from a lookup table; builds a 3D Fenwick tree in scan-line order; supports update and prefix-sum operations in $c_{x,y}$8; answers arbitrary axis-aligned subregion queries in $c_{x,y}$9; has overall build time $t$0; requires $t$1 memory; and reports maximum deviation within $t$2 on simple shapes and complex cardiac models (Nguyen-Le et al., 2024).

A closely related antecedent computes partial volumes inside a Cartesian cube by the divergence theorem rather than tetrahedralization. In that method, each cube corner is classified as inside or outside relative to the isovalue, the 256 patterns are reduced to 23 unique triangulation templates, and the volume is obtained from a uniform flux sum over the triangles closing the partial cell. The per-cube complexity is $t$3, the method is reported as more robust and roughly 20–50% faster than tetrahedralization in a real Front-Tracking / finite-volume code, and a unit-sphere mesh-convergence study shows second-order convergence in both volume and area (Wang, 2013).

For interface meshes given directly as triangles rather than as scalar fields on cube corners, the Front2VOF algorithm addresses the problem of computing the volume of the reference phase inside a cube. It clips each triangle against the six cube faces, assembles the clipped polygons, and evaluates the enclosed volume through the divergence theorem with an explicit treatment of the $t$4 face contribution. The reported behavior is purely local to a single cell, with each incident triangle clipped in $t$5 time and contributing in $t$6 time with $t$7; in tests on single-plane, multiple-plane, and spherical configurations, it recovers analytical volumes to machine precision, handles arbitrarily complex interface-cell topologies, and converges second-order as the triangulation is refined (Pan et al., 8 Jan 2025).

The shared vocabulary of “volume,” “cube,” and “surface extraction” can obscure the fact that these are separate problem classes. VolumeSTCube in the strict sense is a visualization framework for continuous spatiotemporal scalar fields. It is likewise distinct from graph-based cubic-template segmentation for vertebral MRI in Cube-Cut (Schwarzenberg et al., 2014), mass-preserving volumetric mapping of genus-zero solids onto a solid cube (Sandhu et al., 2012), and deterministic quasi-polynomial approximation of clipped-cube volumes in high dimensions (Costandin, 2024). A plausible implication is that VolumeSTCube occupies the visualization end of a broader cube-centered methodological spectrum whose other ends include geometric integration, segmentation, mapping, and high-dimensional volume approximation.