High-fidelity environmental monitoring has traditionally relied on stationary, professional-grade regulatory stations. While these stations provide highly precise data, they are prohibitively expensive to install, operate, and maintain. Consequently, ground monitoring networks are inherently sparse, often leaving gaps of tens or hundreds of kilometers between nodes.
This spatial sparsity creates a fundamental data gap: it fails to capture hyper-local variations, street-level pollution hotspots, and microclimates where people actually live and interact. To resolve this sparsity, a shared research direction has emerged that fuses sparse physical measurements with mobile sensing, satellite remote sensing, and cross-domain urban data to reconstruct continuous, high-resolution environmental fields.
The Breakdown of Proximity-Based Interpolation
When attempting to map a continuous environmental field (such as air temperature, precipitation, or pollutant concentration) from discrete points, traditional geostatistical methods rely on spatial proximity. The core assumption is that closer points share more similar values.
However, as De Felice demonstrates in Graph-based Virtual Sensing from Sparse and Partial Multivariate Observations, this assumption collapses in sparse sensor networks. When physical sensors are separated by massive distances, geographic proximity ceases to be a reliable proxy for interpolation. Standard models smooth over localized variations, completely missing transient or hyper-local phenomena like neighborhood-scale ozone fluctuations or the localized plume of a structural fire.
Mobilizing the Sensor Network
One powerful strategy to bypass the constraints of fixed infrastructure is to make the sensors mobile. Rather than multiplying expensive stationary nodes, researchers leverage opportunistic transit networks and wearable devices to sweep through the gaps in the network.
For instance, Li et al. in Sensing the Air We Breathe - The OpenSense Zurich Dataset deployed air quality sensors on public trams. To maintain accuracy without laboratory overhead, the system utilizes opportunistic "on-the-fly" calibration whenever a tram passes a fixed reference station. Similarly, Piedrahita et al. developed low-cost, wearable "M-Pods" in The next generation of low-cost personal air quality sensors for quantitative exposure monitoring that use field-based collocation calibration to map the precise, real-time exposure of individuals during their daily routines.
Going a step further, crowdsensing eliminates dedicated hardware entirely by turning ubiquitous consumer devices into environmental sensors. Yu et al. and Wang et al. developed deep learning frameworks that estimate air quality indices directly from standard smartphone photos. By training neural networks to analyze image haziness and separate sky regions from urban structures, they transform crowdsourced imagery into a dense, real-time proxy for air quality.
Point-Surface Fusion and Remote Sensing
While ground stations offer high precision at isolated points, satellites offer continuous spatial coverage but suffer from lower temporal resolution and atmospheric interference. Modern frameworks resolve sparsity by fusing these two complementary modalities.
Li et al. introduced a point-surface fusion approach that uses a Generalized Regression Neural Network to merge sparse station measurements with satellite observations across China, achieving significantly higher estimation accuracy than traditional linear models.
To map microclimates without dense physical networks, Jihee Kim developed UrbanHeat in Measuring Fine-Grained Urban Air Temperature with Satellite Imagery. This weakly-supervised framework extracts fine-grained urban air temperatures by combining coarse global climate data with high-resolution land surface temperature imagery as proxy labels.
Similarly, in the domain of meteorology, Doyi Kim proposed Station2Radar in Station2Radar: Query‑Conditioned Gaussian Splatting for Precipitation Field, utilizing query-conditioned Gaussian splatting to transform sparse weather station data into continuous, radar-like precipitation fields by treating rainfall as a collection of learnable continuous distributions rather than a rigid grid.
Leveraging Structural Proxies for Virtual Sensing
Why do these sparse reconstruction techniques work? The core intuition is that environmental variables do not vary randomly across space; they are deeply coupled with the structural, topographical, and socioeconomic characteristics of the landscape. By incorporating cross-domain urban data, models can infer conditions in completely unmonitored zones.
For example, Zhang et al. enhanced Land Use Regression (LUR) models by integrating Point of Interest (POI) data, road networks, and meteorology to capture fine-grained PM2.5 variations in Beijing. In topography-driven scenarios, Zhao et al. demonstrated that incorporating elevation data via Co-Kriging significantly reduces interpolation errors in sparse rainfall networks by leveraging the physical relationship between altitude and precipitation.
By anchoring sparse physical measurements to dense structural proxies—such as traffic patterns, land use, elevation, and satellite imagery—modern virtual sensing frameworks successfully map the invisible spaces between our sparse monitoring stations, turning isolated data points into continuous, actionable fields.