Exposure Mapping Framework

• Exposure Mapping Framework is a formal method that transforms raw data into quantitative exposure metrics across imaging, network, and environmental domains.
• It integrates model-driven, deep learning, and hierarchical Bayesian approaches to construct mapping functions and assess uncertainty via tailored evaluation metrics.
• The framework supports applications such as image enhancement, network interference analysis, and risk mapping by systematically modeling spatial-temporal dependencies and exposure aggregation.

An exposure mapping framework provides a rigorous methodology for quantifying, modeling, and fusing measures of exposure—whether physical, environmental, network-based, or visual—by formalizing how input signals, treatments, or activity patterns are mapped to meaningful exposure metrics or synthesized images. Such frameworks are fundamental in fields ranging from image enhancement and computer vision to causal inference under interference, epidemiological risk analysis, remote-sensing vulnerability mapping, and multi-path chemistry. Core principles include formal representation of exposure, construction of mapping functions (deterministic, statistical, or learned), explicit modeling of uncertainty and spatial-temporal dependencies, and rigorous evaluation using tailored metrics. The following sections survey the principal architectures and methodologies of exposure mapping, with technical depth appropriate for academic readers.

1. Formalization and Definitions of Exposure Mapping

Exposure mapping is typically defined as a function $\mathcal{F}$ (deterministic, parametric, or learned) that aggregates, transforms, or fuses raw data into exposure-related quantities. In visual domains, exposure may refer to pixel-wise lightness under varying camera settings; in networks, to the aggregation of peer or environmental treatments; in spatial applications, to the cumulative burden or risk measured across a spatial grid.

Key examples include:

• Image Fusion for Enhancement: Exposure mapping is employed to synthesize multi-exposure images using camera response functions and brightness transfer functions, as in $$I_{\text{synth}}^c(x; \alpha) = g(I_{\text{orig}}^c(x), \alpha)$$ where $g$ is parameterized by empirically derived camera models (Ying et al., 2017).
• Network Interference Modeling: Exposure mapping functions $Z_i = \mathcal{F}(A, D_{-i}, X_{-i})$ summarize how network neighbors’ treatments affect individual outcomes, potentially learned via Graph Convolutional Autoencoders (Huber et al., 9 Jan 2026) or GNN-based architectures (Adhikari et al., 3 Mar 2025).
• Hierarchical Bayesian Modelling: In environmental and chemical risk, exposure mapping is hierarchical, integrating multiple observational layers, covariate calibration, and latent spatial random effects, as in DIMAQ for PM$_{2.5}$ estimation (Shaddick et al., 2016) and Bayesian chemical aggregation (Neucker et al., 22 Sep 2025).

Exposure maps are thus both the intermediate results (quantified exposure) and the algorithmic functions (mapping rules or networks) used to produce them.

2. Methodological Architectures for Exposure Mapping

Exposure mapping frameworks leverage model-driven, data-driven, or hybrid architectures:

• Model-Based and Physics-Informed Methods: Visual exposure mapping frameworks in low-light image enhancement use illumination maps, edge-preserving smoothing (solving $$\min_{T} \|T-L\|^2 + \lambda \|M \circ \nabla T\|_1$$), and parametric camera models to generate synthetic exposures (Ying et al., 2017, Zheng et al., 2020). Retinex decomposition is foundational for separating reflectance and illumination, with advanced frameworks embedding these into Mamba-driven state-space modules for efficient multi-exposure correction (Dong et al., 2024).
• Deep Learning–Based Fusion: CNNs and transformers conduct feature-level exposure mapping, either unsupervised (MEF-SSIM perceptual loss (Prabhakar et al., 2017, Qu et al., 2021)) or in supervised end-to-end models. Self-supervised multi-task training improves generalization in transformer-based architectures (Qu et al., 2021). Pyramid-based fusions with intricate weight schemes preserve local and global exposure fidelity (Zheng et al., 2020).
• Graph Neural Network Exposure: Automatic learning of exposure mapping functions in networks with interference employs multi-layer GNNs. Messages across neighborhood graphs encode peer importance, edge attributes, and complex influence patterns, surpassing hand-crafted motifs for mastering heterogeneous peer effects (Adhikari et al., 3 Mar 2025, Huber et al., 9 Jan 2026).
• Hierarchical Bayesian Integration: Bayesian exposure mapping factors observed data into sources and pathways, models spatial or product-level variation with hierarchical priors or CAR structures, and propagates uncertainty through posterior predictive simulation (Shaddick et al., 2016, Neucker et al., 22 Sep 2025, Dimasaka et al., 2024, Chen et al., 28 Feb 2025). High-dimensional mappings are marginalized over spatial grids or population cells with INLA, MCMC, or variational methods.

3. Key Steps in Exposure Mapping Workflows

Exposure mapping frameworks typically comprise:

1. Preprocessing and Signal Decomposition: Raw images, sensor measurements, treatment vectors, or environmental signals are cleaned, normalized, and often decomposed (e.g., illumination/reflection in Retinex-based models (Dong et al., 2024)).
2. Construction or Learning of Mapping Functions: Explicit mapping functions are defined (e.g., brightness transfer functions, kernel smoothing for prevalence (Wu et al., 11 Dec 2025)), learned (CNNs, GNNs), or statistically inferred (factor analysis for high-dimensional nutritional data (Zorzetto et al., 23 Jan 2026)).
3. Exposure Estimation and Fusion: Exposure quantities are synthesized (pixel-wise blends, cumulative integrals, statistically weighted aggregates), frequently segmenting exposure by region, population, or network context.
4. Uncertainty Quantification: Bayesian frameworks propagate uncertainty, compute credible intervals, and aggregate ensemble draws to exposure maps or risk profiles (Shaddick et al., 2016, Neucker et al., 22 Sep 2025). Deep models may report confidence maps.
5. Evaluation and Metric-Based Comparison: Domain-specific metrics assess exposure fidelity (MEF-SSIM, LOE, VIF, DRIM for images), statistical consistency (RMSE, population-weighted RMSE for environmental models), and causal/epidemiologic validity (Type I error, coverage, power) (Ying et al., 2017, Shaddick et al., 2016, Huber et al., 9 Jan 2026, Wu et al., 11 Dec 2025).

4. Representative Applications

Exposure mapping frameworks are broadly applicable:

• Image Enhancement and Correction: Dual-exposure fusion, deep unsupervised fusion, and hybrid CRF-CNN models demonstrably improve visual quality and accurate rendering of low- and high-light regions, surpassing hand-crafted enhancement algorithms (Ying et al., 2017, Prabhakar et al., 2017, Dong et al., 2024, Zheng et al., 2020, Qu et al., 2021).
• Networked Causal Inference: Exposure mapping under interference in networks enables identification and testing of direct/indirect effects, using both user-defined and learned mappings, with robust DML-based conditional-independence tests for mapping validity (Huber et al., 9 Jan 2026). GNN-based mapping architectures generalize peer effect estimation under unknown influence mechanisms (Adhikari et al., 3 Mar 2025).
• Remote Sensing and Risk Mapping: Large-scale exposure mapping aligns multi-year satellite imagery with geocoded vulnerability indices (building counts, material type), fits modified ResNet-50 models, and stitches tile-level predictions into global spatial exposure maps for disaster risk auditing (Dimasaka et al., 2024).
• Environmental and Chemical Risk Quantification: Hierarchical Bayesian exposure mapping enables accurate estimation of atmospheric PM$_{2.5}$ on high-resolution global grids using multi-layered geographic random effects, improving on previous single-slope models (Shaddick et al., 2016). Aggregated chemical exposure frameworks integrate multiple chemical sources and pathways, handling missing data and sparse samples (Neucker et al., 22 Sep 2025).
• Health and Mobility: Statistical and kernel approaches for mapping contextual exposure (e.g. HIV risk from GPS-derived movement data) synthesize activity spaces, time-weighted prevalence, and provide multi-tier demographic and risk stratification (Wu et al., 11 Dec 2025).
• Temperature and Heat Risk Analysis: Dynamic cumulative exposure mapping for urban travelers utilizes second-by-second physiological modeling, MET-aware heat stress metrics, and segment-level risk aggregation, outperforming additive approaches in predicting health risk in transit systems (Fan et al., 2024, Chen et al., 28 Feb 2025).

5. Uncertainty, Validation, and Metric Assessment

Exposure mapping frameworks employ rigorous metric-driven evaluation, including:

Domain	Metrics Employed	Comments
Image Enhancement	LOE, VIF, DRIM, MEF-SSIM, PSNR, SSIM	No-reference and full-reference tailored to fusion distortions
Network Interference	Type I error, Power, Coverage, PEHE loss	DML-based conditional-independence testing, exposure validity
Spatial Mapping	RMSE, Population-weighted RMSE, MSE, MAE	Posterior credible intervals, exceedance probabilities
Epidemiology/Causal	Bias, RMSE, CATE curves, FS-CATE	Posterior summarization, subgroup aggregation
Remote Sensing	MAE, MSE (non-exceedance), visual heatmaps	Multi-scale calibration and error analysis

Frameworks often propagate full posterior uncertainty, conduct cross-validation, and compare performance to state-of-the-art baselines. For instance, DIMAQ reduces RMSE from 17.1 to 10.7 $\mu$g/m$^3$; exposure fusion models achieve higher MEF-SSIM scores than hand-crafted competitors (Shaddick et al., 2016, Prabhakar et al., 2017, Zheng et al., 2020).

6. Generalization, Extensions, and Limitations

Exposure mapping frameworks generalize across domains:

• Multi-Scale and Multi-Source Data: Integration of heterogeneous sources, spatial and temporal variability, multi-path exposure, and mixtures generalizes frameworks to diverse environmental, visual, and epidemiological settings (Shaddick et al., 2016, Dimasaka et al., 2024, Neucker et al., 22 Sep 2025).
• Learning-Based Extensions: Graph-structured and deep networks automate complex exposure mappings, enabling more precise causal inference and robust predictive modeling under unknown structural dependencies (Huber et al., 9 Jan 2026, Adhikari et al., 3 Mar 2025, Qu et al., 2021).
• Uncertainty and Scalability: Bayesian approaches propagate uncertainty, scale via INLA, MCMC, and variational strategies, and account for missing data as latent variables. A plausible implication is better calibration and coverage under sparse or incomplete data regimes (Shaddick et al., 2016, Neucker et al., 22 Sep 2025).
• Limitations: Frameworks may be constrained by computational complexity (global grid predictions), data coverage (lack of ground truth in unmonitored regions), or limitations of the mapping function space (e.g., missing influence mechanisms in network exposure mappings). Evaluation of validity and identifiability is essential, with conditional-independence testing as a methodological safeguard (Huber et al., 9 Jan 2026).

7. Connections Across Research Domains

Exposure mapping frameworks unify principles across computer vision, network science, epidemiology, environmental statistics, and remote sensing:

• Operator-Driven Fusion and Correction: Retinex theory, BTF/CRF modeling, pyramid fusion, transformer attention, and state-space modeling provide a spectrum of technical tools for image-based exposure correction/enhancement (Dong et al., 2024, Ying et al., 2017, Qu et al., 2021).
• Statistical Inference Under Interference: Exposure mapping formalizes the challenge of indirect or spillover effects in networks, underpinning identification and estimation in advanced causal inference research (Huber et al., 9 Jan 2026, Adhikari et al., 3 Mar 2025, Zorzetto et al., 23 Jan 2026).
• Bayesian Hierarchical Models for Risk and Environmental Exposure: Nested random effect structures, CAR priors, factor analysis, and full Bayesian propagation shape modern exposure quantification for global health, chemicals, and disaster risk (Shaddick et al., 2016, Neucker et al., 22 Sep 2025, Dimasaka et al., 2024).
• Spatial and Temporal Generalization: Frameworks accommodate spatially-varying, temporally dynamic, and heterogeneous risk landscapes, enabled by scalable statistical and machine learning tooling.

In summary, exposure mapping frameworks provide technically robust, theoretically justified methodologies for the explicit quantification, fusion, and inference of exposure across a wide spectrum of scientific domains, combining statistical rigor, computational learning, and tailored evaluation for domain-specific applications.