Imagine trying to predict regional health outcomes based on air pollution levels. A standard machine learning model assumes that a one-unit increase in $\text{NO}_2$ has the exact same impact on respiratory health whether you are in a dense urban core, a coastal town, or a rural farming community.
This assumption—that the relationship between variables is uniform across an entire study area—is known as the spatial stationarity assumption. While it simplifies mathematics, it frequently fails in the real world. Geographic data is profoundly non-stationary: the underlying statistical properties, correlations, and causal drivers warp, shift, or completely invert depending on where you are on the map.
Overcoming this assumption requires a fundamental shift: moving from global models that treat space as a monolith to localized frameworks that treat location as an active moderator of relationships.
Why Global Uniformity Fails
When a global model forces a single average coefficient across a diverse landscape, it washes out critical local signals, leading to poor predictive accuracy and misleading generalizations.
We see this across multiple domains:
- Public Health: Ishaan Maitra et al. demonstrate that global models fail to capture localized policy anomalies, such as the specific positive impact of London’s Ultra Low Emission Zone on respiratory health (How Your Location Relates to Health: Variable Importance and Interpretable Machine Learning for Environmental and Sociodemographic Data). Similarly, Nengliang Yao et al. show that increasing urologist density significantly reduces prostate cancer mortality in the southern Mississippi River states, but has negligible effects in New England (Geographic Analysis of Urologist Density and Prostate Cancer Mortality in the United States).
- Socioeconomic Dynamics: Susan L. Cutter demonstrates that the drivers of COVID-19 vaccine uptake vary dramatically across the urban-rural divide, meaning a single national strategy fails to address regional concerns (Spatial-temporal differences of COVID-19 vaccinations in the U.S). In urban criminology, Yifei Chen et al. show that the influence of tree canopy on property crime depends heavily on local neighborhood contexts (INVESTIGATING THE INFLUENCE OF TREE COVERAGE ON PROPERTY CRIME: A CASE STUDY IN THE CITY OF VANCOUVER, BRITISH COLUMBIA, CANADA).
- Ecology and Semantics: Even conceptual associations change by location. A study on crowdsourced labels highlights that the tag "energy" correlates with green tech in San Francisco but shifts to oil and gas in Houston (Location-Sensitive User Profiling Using Crowdsourced Labels). In ecosystems, Andrej Aderhold et al. note that species interactions are non-stationary, meaning a predator-prey relationship can completely alter or disappear across geographic zones (Reconstructing ecological networks with hierarchical Bayesian regression and Mondrian processes).
Localizing Parameters via Weighted Regression
The most direct mathematical way to handle spatial non-stationarity is to let the model parameters themselves become functions of space. Instead of fitting a global equation $Y = \beta X + \epsilon$, we fit:
$$Y = \beta(s)X + \epsilon$$
where $s$ represents the spatial coordinates.
This is the core intuition behind Geographically Weighted Regression (GWR). Instead of optimizing a single loss function for the entire dataset, GWR fits a separate regression model for every single point in space, weighting nearby observations more heavily using a spatial kernel.
Researchers have successfully used this to map ground-level $\text{PM}_{2.5}$ air pollution during wildfire seasons, capturing how localized weather patterns and complex topographies alter the relationship between satellite optical measurements and surface pollution (ESTIMATING PM2.5 CONCENTRATIONS...; Yuenan Li et al. in ESTIMATING PM2.5 IN THE BEIJING-TIANJIN-HEBEI REGION...).
To capture non-linear interactions without losing this spatial specificity, Ryan Zhenqi Zhou et al. introduced PyGRF, a Geographical Random Forest model that builds local ensembles to handle spatial heterogeneity efficiently (PyGRF: An improved Python Geographical Random Forest model and case studies in public health and natural disasters).
Global Model: [All Data Points] ---> Single Average Parameter (β)
Local Model: [Data Point A] ----> Weights nearby points heavily ----> β(Location A)
[Data Point B] ----> Weights nearby points heavily ----> β(Location B)
To avoid the computational burden of fitting a unique model for every single coordinate, Wala Draidi Areed et al. proposed Bayesian Cluster GWR. This framework uses Dirichlet process mixtures to automatically segment space into discrete, statistically uniform neighborhoods, allowing parameters to vary across clusters rather than individual points (Bayesian Cluster Geographically Weighted Regression for Spatial Heterogeneous Data). Similarly, Yanxiu Jin et al. developed Spatially Clustered Factor Analysis (SCFA) to allow dimensionality reduction structures to shift across regions (Clustered Factor Analysis for Multivariate Spatial Data).
Designing Non-Stationary Kernels
For spatial interpolation and probabilistic mapping, Gaussian Processes (GPs) are the standard tool. However, traditional stationary kernels (like RBF or Matérn) assume that the correlation between two points depends solely on the Euclidean distance between them ($|s - s'|$), regardless of their absolute location.
This assumption breaks down when the physical environment changes abruptly. For instance, a major transportation hub or a city center has a much wider sphere of influence than a rural area. To address this, Khurram Yamin leverages deep neural networks to learn non-stationary kernels for tracking COVID-19 hotspots (Early Detection of COVID-19 Hotspots Using Spatio-Temporal Data).
In climate science, Vidhi Lalchand et al. utilize a "Gibbs kernel" where the lengthscale parameter varies continuously across space. This allows the GP to adapt its smoothness to the rugged, highly variable topography of mountainous basins (Kernel Learning for Explainable Climate Science).
Stationary Kernel: ( Point A ) <--- Same Smoothness ---> ( Point B )
Non-Stationary Kernel: ( Point A / High Complexity ) <--- Short Lengthscale ---> ( Point B / Smooth Plain ) <--- Long Lengthscale --->
For robotics, Weizhe Chen et al. developed "Attentive Kernels" in their POAM framework. This allows autonomous agents to dynamically adjust mapping resolution between featureless plains and jagged, complex structures in real time (POAM: Probabilistic Online Attentive Mapping for Efficient Robotic Information Gathering).
To scale these non-stationary fields to massive datasets, Kangrui Wang et al. proposed Stochastic Process Mixing (SPM) to capture varying space-time couplings (Non-separable Non-stationary random fields), while Mitchell Krock et al. combined basis function expansions with a fusion penalty to handle sharp transitions across geographical boundaries like land and sea (Modeling Massive Highly Multivariate Nonstationary Spatial Data with the Basis Graphical Lasso).
Spatial Moderation in Deep Learning
Standard deep learning architectures assume data is independent and identically distributed (i.i.d.), which completely blinds them to spatial context.
To bridge this gap, the STAR framework by Yiqun Xie et al. dynamically branches a neural network based on geographic location. It uses a "spatial moderator" to share structural insights across regions while preserving local nuances, significantly outperforming global neural networks in high-dimensional tasks like crop monitoring (Statistically-Guided Deep Network Transformation to Harness Heterogeneity in Space (Extended Abstract)).
However, physical coordinates alone do not tell the whole story. Two distant locations might behave similarly because they share environmental attributes, while two adjacent locations separated by a mountain range might behave differently. To capture this, Sihan Ni et al. introduced the SANNWR model. This framework constructs a unified "spatial-attribute" distance metric, allowing neural network weights to adapt based on both geographic proximity and inherent entity features (Spatial and Attribute Neural Network Weighted Regression for the Accurate Estimation of Spatial Non-Stationarity).
The Intuition of Local Context
Ultimately, modeling spatial non-stationarity works because it replaces a rigid, global average with a flexible, context-aware mapping. By allowing parameters, lengthscales, and network weights to adapt to local contexts, these frameworks prevent localized signals from being washed out by global averages. Instead of treating spatial heterogeneity as noise to be smoothed away, these models treat location as an active, vital component of the data's underlying truth.
Key Foundational Work
Interpolating Earth-science Data using RBF Networks and Mixtures of Experts (1996) — contributed 1 concept elements to this direction.
Gaussian process modeling of large-scale terrain (2009) — contributed 1 concept elements to this direction.