Spatial Proximity Classification

• Spatial Proximity Classification is a set of techniques that use spatial relationships like distance and adjacency to predict categorical outcomes.
• It integrates methods ranging from tree-based and graph models to spatially-constrained probabilistic and neural frameworks for robust predictive performance.
• Applications span remote sensing, fire hazard risk, ecological clustering, and medical imaging, demonstrating practical, high-performance classification.

Spatial proximity classification refers to the suite of statistical, machine learning, and deep learning methods for predicting categorical or ordinal outcomes by leveraging the relative spatial locations of entities—either as explicit geometric distances, neighborhood graphs, or through learned spatial representations. Applications span domains ranging from remote sensing, geospatial intelligence, fire hazard risk, and construction safety to ecology, medical image segmentation, audio scene analysis, and personalized comfort modeling. Methodologies include proximity descriptors, tree-based nonparametric models, prototype-based graph approaches, encoded feature representations, and spatially-constrained probabilistic models, unified by the central aim: to learn how spatial relationships—proximity, adjacency, containment, and neighborhood—inform and enhance classification decisions.

1. Proximity Features and Spatial Representations

Spatial proximity metrics underpin most classification frameworks in this domain. The foundational metric is the (Euclidean or geodesic) distance between entities or between entities and landmarks:

• Raw distances: $d_{ij} = \sqrt{(x_i-x_j)^2 + (y_i-y_j)^2}$, or great-circle projections for geographic data.
• Aggregates: minimum distance to landmarks, mean distance to k-nearest features, and binary proximity indicators (within a threshold) (Levin, 2024).
• Proximity-encoding: Multi-Point Proximity (MPP) encodes shapes as vectors of scaled distances to fixed reference points, yielding spatially continuous, shape-centric features useful for regression and classification (Collins, 5 Jun 2025).
• Adjacency matrices: For panoramic or graph-structured scenes, explicit matrices $C^r, C^o, C^{ro}$ count region–region, object–object, and region–object adjacencies, later normalized to probabilities (Xu et al., 2024).
• Proximity catch digraphs (PCDs): Define geometric proximity regions for each sample point and use graph-theoretic constructs (dominating sets, covers) to formalize class membership (Manukyan et al., 2017).

The modeling of spatial relationships may occur at multiple scales (e.g., k-nearest neighbors, multi-resolution graphs) and may incorporate both spatial and temporal dimensions (e.g., events with times and locations, temporally adjacent influences) (Levin, 2024). Encoding approaches such as MPP and Build2Vec furnish continuous, differentiable proximity features compatible with a variety of classifiers (Collins, 5 Jun 2025, Abdelrahman et al., 2021).

2. Algorithmic Methodologies

Approaches for spatial proximity classification encompass both parametric and nonparametric machine learning paradigms:

• Tree-Based Methods: Decision trees (CART), random forests, and Bayesian Additive Regression Trees (BART) allow flexible incorporation of proximity features, discovering both main effects (e.g., distance thresholds) and higher-order interactions with other covariates. These methods can use raw distances or derived proximity attributes, and ensembles (RF, BART) improve robustness and performance over single trees (Levin, 2024).
• Prototype-Based Graph Methods: Proximity Catch Digraphs (PCDs) and their extensions construct class covers from minimum dominating sets in geometric proximity graphs (e.g., PE-PCDs), leading to classifiers that are theoretically Bayes-consistent under separability, robust to class imbalance, and provide interpretable prototype-based decision rules. Construction relies on Delaunay tessellation and barycentric coordinate searches, limiting scalability with dimensionality (Manukyan et al., 2017).
• One-Class Proximity Descriptors: The Average Localised Proximity (ALP) descriptor aggregates multi-scale, multi-neighborhood proximity scores via an Ordered Weighted Averaging (OWA) operator, providing robust, parameter-light classifiers for one-class scenarios (Lenz et al., 2021).
• Spatially-Constrained Probabilistic Models: Regularized mixture models, such as those used in ecology, encourage spatial consistency by penalizing label variability across a structural spatial network using a Laplacian penalty; variational EM is applied to jointly estimate labels and model parameters (Miele et al., 2014).
• Spatial Generalized Linear Mixed Models: Bayesian SGLM/SGLMMs incorporate spatial random effects into latent variable models for binary or multiclass responses; the predictive rule naturally integrates both spatial proximity of observed classes and covariates. Posterior predictive classification leverages kriging-style inference (Berrett et al., 2014).
• Graph and Random Walker Models: Classification using multiresolution feature- and prior-based graphs merges spatial feature similarities (with robust edge-weighting) and aspatial sample priors, using combinatorial Dirichlet energies and hierarchical CRFs to propagate spatial and hierarchical dependencies (Fallah, 2023).
• Neural and Multi-Modal Models: Vision-LLMs, such as Proximity QA, are trained to reason about spatial proximity from image–text pairs by explicit two-phase supervision—object-level depth estimation and relational proximity comparison—enabling spatial QA in rich visual scenes (Li et al., 2024).

3. Application Domains

Spatial proximity classification methods are applied across a range of settings:

• Fire Hazard Risk and Safety: Enhanced YOLOv8-powered frameworks segment fire and smoke and, using dual-model detection, quantify real-world proximities to at-risk objects via pixel-to-meter conversion. Risk scores combining fire severity, object vulnerability, and distance-based exposure inform situational alert tiers. Validated performance achieves precision, recall, and F1 exceeding 90% (AlMhdawi et al., 10 Mar 2026).
• Construction Site Proximity Monitoring: Monocular 2D cameras, combined with anchor-free monocular 3D detection and a post-processing discretization, assign categorical risk labels (Dangerous, Potentially Dangerous, Concerned, Safe) based on continuous proximity and predicted orientation. F1 scores reach approximately 0.77 overall and 0.9 for the most critical tier within a 50 m radius (Ding et al., 2023).
• Ecological Clustering: Joint model-based clustering and spatial cohesion via label-smoothing over graph Laplacians produces spatially coherent ecological partitions (e.g., biogeographic provinces) consistent with ecological similarity networks and real-world spatial barriers (Miele et al., 2014).
• Geospatial Machine Learning: MPP encoding enables high-fidelity regression and classification on spatial shape properties (length, area, orientation, complexity), and pairwise spatial relationships (intersects, contains, borders) with ROC AUC >0.95 (Collins, 5 Jun 2025).
• Personal Comfort in Buildings: Build2Vec integrates BIM-extracted spatial context and EMA feedback into attributed graph embeddings, yielding a 14–28% improvement over conventional comfort classifiers, demonstrating the explanatory power of structured spatial proximity information (Abdelrahman et al., 2021).
• Medical and Scientific Imaging: Multi-layer hierarchical random walker-CRF models achieve significant increases in per-class recall and Dice scores for anatomical segmentation tasks by encoding spatial and hierarchical dependencies (Fallah, 2023).
• Vision-and-Language Navigation: Hierarchical spatial proximity reasoning modules learn adjacency probability matrices (region–region/object–object/region–object), improving navigation success and grounding in standard VLN datasets (Xu et al., 2024).
• Audio and Spatiotemporal Sensing: Joint direction-and-proximity neural architectures achieve F₁ ≈ 84% for near/far source discrimination, with multi-label designs enabling scene-level spatial parsing from binaural audio (Krause et al., 2021).

4. Evaluation Protocols and Empirical Results

Spatial proximity classification methods are assessed using metrics tailored to the classification task:

• Accuracy, Precision, Recall, F1 Score: Standard for categorical tasks; for fire detection and proximity risk, F1 >0.86 is reported (AlMhdawi et al., 10 Mar 2026); for camera-based safety, F1 ≈ 0.77 within 50 m (Ding et al., 2023).
• Mean Squared Error (MSE), AUC: For continuous or ordinal outcomes and pairwise spatial relation prediction (MPP encodings), R² up to 0.99 and AUC ≈ 0.99 are achieved (Collins, 5 Jun 2025).
• Permutation-Based Importance: Random forests and BART expose the marginal predictive value of proximity features; in real data, proximity to mass shootings exerted a modest influence on gun control support, whereas proximity to border crossings had negligible independent effects (Levin, 2024).
• Robustness and Consistency: Prototype-based and spatially-constrained methods are theoretically consistent under separability, robust to class imbalance, and, in high dimensions, sensitive to appropriate dimension reduction (Manukyan et al., 2017, Miele et al., 2014).
• Domain Adaptation and Generalizability: Some camera-based and neural vision approaches acknowledge the potential real-vs-synthetic data gap, and propose augmentation or multi-view fusion as future remedies (Ding et al., 2023, Li et al., 2024).

5. Model Selection, Computational Complexity, and Theoretical Guarantees

• Model Selection: Integrated Classification Likelihood (ICL) and cross-validation are employed to choose cluster number Q and spatial penalty λ in regularized EM frameworks. Fold-based tests support the statistical significance of spatial feature inclusion (Miele et al., 2014, Abdelrahman et al., 2021).
• Computational Complexity: Varies by paradigm—tree-based and random forest methods scale efficiently; prototype-based covers are polynomial in n but exponential in d unless dimension reduction is used; graph-embedding and random walker models exploit sparsity and conjugate gradient solvers for scaling (Manukyan et al., 2017, Fallah, 2023).
• Bayesian Inference: MCMC sampling delivers posterior predictive classification in spatial GLMMs; SGLM/SGLMM distinction has minimal effect on error rates, but impacts neighbor influence (Berrett et al., 2014).
• Consistency: Cover-based PCD classifiers are Bayes-consistent under strict separability, with further evidence of adaptive robustness in the presence of class imbalance (Manukyan et al., 2017).
• Effect of Spatial Penalties: Regularization enforces spatially coherent clusters, but overly large penalties in the presence of high spatial discordance can collapse clusters to a single class (Miele et al., 2014).

6. Limitations and Future Directions

Limitations are domain-dependent. Principal constraints include:

• Curse of dimensionality for prototype-based proximity covers (mitigated by PCA or spatial embedding) (Manukyan et al., 2017).
• Sensitivity to accurate metric parameterization (e.g., scale factors for pixel-to-meter conversion, reference point grids for MPP) (AlMhdawi et al., 10 Mar 2026, Collins, 5 Jun 2025).
• Performance trade-offs when spatial proximity is weak or collinear with dominant predictors (e.g., party affiliation in political opinion) (Levin, 2024).
• Domain gap between synthetic and real-world vision datasets; efforts to incorporate temporal cues and multi-view geometry are outlined as plausible next steps (Ding et al., 2023, Li et al., 2024).
• For neural architectures, reliance on synthesized supervision for depth or adjacency can limit fine-grained spatial inference in complex scenes (Li et al., 2024).
• Computational cost for large multiresolution graphs, though polynomial algorithms and efficient solvers are mature (Fallah, 2023).

Overall, spatial proximity classification methodologies provide versatile, interpretable, and often theoretically grounded frameworks for leveraging spatial relationships in classification. Their integration with modern representation learning, probabilistic modeling, and domain-specific engineering makes them central to a growing set of applications across disciplines.