RelMap: Reliable Spatiotemporal Sensor Data Visualization via Imputative Spatial Interpolation

Juntong Chen, Huayuan Ye, Zhu He, Siwei Fu, Changbo Wang, Chenhui Li

Abstract

Spatiotemporal sensor data is critical for monitoring environmental phenomena such as air quality and temperature. However, visualization via heatmaps often suffers from irregular sensor coverage and a lack of transparency regarding data uncertainty. This paper presents RelMap, a novel visualization pipeline that enhances the reliability of spatiotemporal heatmaps. RelMap utilizes Graph Neural Networks (GNNs) to perform imputative spatial interpolation, effectively "filling in" gaps in sparse sensor networks. By introducing virtual sensors and employing extrinsic visualization techniques—specifically hatch patterns and reliability glyphs—RelMap provides a more accurate spatial representation while clearly communicating the trustworthiness of the data to the end-user.

1. Introduction

Heatmaps are the standard medium for visualizing geospatial sensor data. Despite their ubiquity, they face two significant challenges: irregular coverage and hidden uncertainty. In many real-world scenarios, sensors are concentrated in urban centers, leaving rural or remote areas with sparse data. Traditional interpolation methods, such as Kriging or Inverse Distance Weighting (IDW), often produce speculative "blobs" in these gaps without indicating the lack of underlying evidence. RelMap addresses these issues by densifying the sensor network and providing a visual language for uncertainty.

2. The RelMap Pipeline

The RelMap architecture is divided into three distinct stages designed to transform raw, sparse sensor readings into a reliable, high-fidelity visualization.

2.1 Adaptive Sensor Densification

To mitigate the issues caused by irregular sensor distribution, RelMap first identifies "holes" in the network using Kernel Density Estimation (KDE). Once these gaps are identified, the system places "virtual sensors" to act as anchors for interpolation. To ensure these virtual sensors are distributed uniformly and do not cluster, the pipeline utilizes Centroidal Voronoi Tessellation (CVT). This densification provides a more stable framework for subsequent spatial analysis.

2.2 Imputative Spatial Interpolation

At the core of RelMap is a Graph Neural Network (GNN) trained to predict values for the newly created virtual sensors. The model incorporates several advanced features:

  • Principal Neighborhood Aggregation (PNA): This allows the model to capture complex spatial relationships across the sensor graph.
  • Geographical Positional Encoding (GPE): Unlike standard GNNs that are location-blind, GPE provides the model with absolute coordinates. This allows the system to learn semantic geographical patterns, such as the difference between coastal and inland climatic behaviors.
  • Temporal Convolutions: By modeling time-series patterns, the system can perform "temporal super-resolution," generating smooth animations from data recorded at infrequent intervals.

2.3 Uncertainty Visualization

RelMap introduces an extrinsic layer to the heatmap to communicate data reliability without distorting the primary color-coded data. This is achieved through two methods:

  • Hatch Patterns: Diagonal lines are overlaid on regions with low sensor density, signaling to the user that the data in these areas is primarily interpolated.
  • Reliability Glyphs: These arrow-like shapes function similarly to box plots. The height of the primary arrow represents the average deviation from expected values, while auxiliary "wings" indicate variance. The width of the glyph reflects the distance to the nearest physical sensors.

3. Experimental Evaluation

3.1 Datasets

The authors evaluated the RelMap pipeline using four real-world geospatial datasets:

  • ushcn: Precipitation data from the U.S. Historical Climatology Network (1,218 sensors).
  • uspm25: PM2.5 air quality data from the EPA's AirData (749 sensors).
  • catmp: Temperature data from California (141 sensors).
  • dtaqi: Air Quality Index data from the Yangtze River Delta (294 sensors).

3.2 Quantitative and Qualitative Results

RelMap demonstrated superior performance over baselines such as IGNNK and Ordinary Kriging. As the data mask rate (the percentage of missing data) reached 50%, RelMap maintained significantly lower RMSE and MAE. Furthermore, using the Structural Similarity Index (SSIM), the study found that adding 20-40% densified virtual sensors resulted in heatmaps that were substantially closer to the ground truth.

In a user study, the uncertainty glyphs received an understandability rating of 4.33/5. Participants noted that the "sharper" glyphs provided an intuitive signal regarding which regions of the map should be viewed with caution.

4. Key Insights

  • Virtual Sensors as Anchors: The strategy of creating a densified network before applying Radial Basis Function (RBF) interpolation allows the final map to reflect complex local patterns that traditional interpolation would smooth over.
  • The Importance of GPE: Geographical Positional Encoding is a critical component, as it allows the GNN to understand that "North" or "South" carries specific semantic weight in environmental contexts.
  • Dynamic Movement: The temporal modeling capabilities of RelMap allow it to capture dynamic movements, such as smoke plumes, more effectively than simple linear transitions between time steps.

5. Conclusion and Future Work

RelMap provides a robust framework for making spatiotemporal visualizations more accurate and transparent. While currently limited to scalar, univariate data, future iterations aim to incorporate multivariate inputs—such as satellite imagery—to improve predictions. Additionally, the authors plan to address the $O(n^2)$ computational complexity of dense graph modeling to scale the system for national-level sensor networks.

Summary updated

Abstract

Accurate and reliable visualization of spatiotemporal sensor data such as environmental parameters and meteorological conditions is crucial for informed decision-making. Traditional spatial interpolation methods, however, often fall short of producing reliable interpolation results due to the limited and irregular sensor coverage. This paper introduces a novel spatial interpolation pipeline that achieves reliable interpolation results and produces a novel heatmap representation with uncertainty information encoded. We leverage imputation reference data from Graph Neural Networks (GNNs) to enhance visualization reliability and temporal resolution. By integrating Principal Neighborhood Aggregation (PNA) and Geographical Positional Encoding (GPE), our model effectively learns the spatiotemporal dependencies. Furthermore, we propose an extrinsic, static visualization technique for interpolation-based heatmaps that effectively communicates the uncertainties arising from various sources in the interpolated map. Through a set of use cases, extensive evaluations on real-world datasets, and user studies, we demonstrate our model's superior performance for data imputation, the improvements to the interpolant with reference data, and the effectiveness of our visualization design in communicating uncertainties.
BibTeX
@article{chen2026relmap,
  author = {Juntong Chen and Huayuan Ye and Zhu He and Siwei Fu and Changbo Wang and Chenhui Li},
  title = {{RelMap: Reliable Spatiotemporal Sensor Data Visualization via Imputative Spatial Interpolation}},
  journal = {IEEE Trans. Vis. Comput. Graph.},
  year = {2026},
  url = {https://openreview.net/forum?id=VHwBWmsp2m},
}
Figure