Areal Disaggregation: Methods and Applications

• Areal disaggregation is the process of inferring detailed spatial patterns from coarse aggregated data through structured aggregation and weighting techniques.
• It employs methods ranging from Bayesian spatial models and continuous-domain formulations to machine-learning approaches, integrating fine-scale covariates and ancillary data.
• The approach has practical applications in epidemiology, remote sensing, survey statistics, and urban studies, enabling high-resolution analysis from aggregated observations.

Areal disaggregation is the problem of inferring a latent fine-scale spatial process from outcomes observed only on coarser areal supports. Across the recent literature, it appears as disaggregation modelling, downscaling, areal interpolation/disaggregation, and a change-of-support problem, with applications in epidemiology, remote sensing, survey statistics, urban morphology, and humanitarian mapping. The common construction is to define the scientific quantity of interest on a fine support—pixels, grid cells, districts, mesh elements, or a continuous domain—and then link coarse observations to that latent field through aggregation, usually with population, exposure, or other weights (Nandi et al., 2020, Suen et al., 14 Feb 2025, Wu et al., 4 Mar 2026).

1. Scope and recurring problem formulations

The literature treats areal disaggregation as a support-mismatch problem: responses are observed on polygons, but inference is needed on a finer target support. In epidemiology, the observed response may be district-, province-, or administrative-unit counts or prevalence, while the target is a high-resolution risk map. In survey settings, annual or multi-year estimates may be available only for coarse areas, while the target is a finer administrative geography. In remote sensing and urban mapping, coarse lattice or administrative products are redistributed to finer grids using ancillary spatial information (Nandi et al., 2020).

This suggests a useful distinction between several recurring constructions.

Family	Support relation	Representative papers
Bayesian spatial disaggregation models	coarse polygons to fine square lattice or pixels via aggregation weights	(Nandi et al., 2020, Arambepola et al., 2020)
Continuous-domain coherent disaggregation	polygon counts or point patterns to a continuous latent intensity surface	(Suen et al., 14 Feb 2025, Ripstein et al., 23 Jun 2026)
Weakly supervised and machine-learning disaggregation	parcels or administrative units to pixels or grid cells using imagery, buildings, or aggregate learning	(Archbold et al., 2023, Derval et al., 2019, Wells et al., 19 May 2025, Dimasaka et al., 30 Jul 2025)
Diagnostics and aggregation-error criteria	quantify MAUP, ecological fallacy, or regionalization quality rather than directly downscale	(Duque et al., 2018, Bradley et al., 2015, Daw et al., 2023)

A second distinction separates direct disaggregation methods from adjacent multiscale methods. Some papers explicitly redistribute mass from coarse to fine support; others instead provide diagnostics, regionalization criteria, or multiresolution decompositions that are useful for change-of-support analysis but do not themselves predict on a finer support. This distinction matters because the existence of a multiscale spatial representation does not by itself imply a cross-support disaggregation operator (Suen et al., 14 Feb 2025).

2. Core statistical constructions

A standard count-based formulation defines a latent fine-scale rate or intensity and aggregates it to the observed polygons. In Bayesian spatial disaggregation models, the polygon-level quantity is built from pixel-level rates through an aggregation raster,

$$\textrm{cases}_i = \sum_{j=1}^{N_i} a_{ij}\textrm{rate}_{ij}, \qquad \textrm{rate}_i = \frac{\textrm{cases}_i}{\sum_{j=1}^{N_i} a_{ij}},$$

with likelihoods such as

$y_i \sim \textrm{Pois}(\textrm{cases}_i).$

For disease mapping, an equivalent formulation is

$$y_i \sim \mathrm{Pois}\left(\sum_{j=1}^{M_i} p_{ij}\lambda_{ij}\right), \qquad \log\lambda_{ij} = \beta_0 + \bm{\beta}^T\bm{X}_{ij} + \epsilon_{ij}.$$

The critical modeling choice is that aggregation is applied to latent fine-scale means on the response scale, rather than to polygon-averaged covariates or to the linear predictor alone (Nandi et al., 2020, Arambepola et al., 2020).

A continuous-domain version replaces the fine grid by an inhomogeneous Poisson intensity. In that construction,

$$N_p \sim \textsf{Pois}\left(\int_{A_p}\lambda(\mathbf s)\,d\mathbf s\right), \qquad \log \lambda(\mathbf s) \approx \eta(\mathbf s) := \beta_0 + \beta_x X(\mathbf s) + u(\mathbf s).$$

The observed areal totals are therefore induced by integrals of a latent continuous surface. This is the paper’s notion of coherent support change: polygon counts and point patterns are linked to the same underlying $\lambda(\mathbf s)$, and fine-scale prediction is obtained by evaluating or reaggregating that same surface (Suen et al., 14 Feb 2025).

Survey-based disaggregation introduces a different observation model because the inputs are often published estimates with design-based variances rather than raw counts. One formulation defines an effective sample size

$$m_t^{*(l)}(A)=\left[\frac{z_t^{(l)}(A)(1-z_t^{(l)}(A))}{\tau_t^{2(l)}(A)}\right],$$

an effective number of cases

$$q_t^{*(l)}(A)=\left[m_t^{*(l)}(A)z_t^{(l)}(A)\right],$$

and then a working likelihood

$$q_t^{*(l)}(A)\mid \pi_t^{(l)}(A)\sim \text{Binomial}(m_t^{*(l)}(A),\pi_t^{(l)}(A)).$$

Spatial and temporal aggregation are then written explicitly, for example through five-year averages and population-weighted tract-to-PUMA aggregation. A related small area estimation formulation writes the coarse mean as

$$\mu_i = \sum_{j:i[j]=i}\frac{N_j}{N_i}\mu_j, \qquad g(\mu_j)=\alpha+\bm{x}_j^\top\bm{\beta}+b_j,$$

so that fine-area parameters are inferred even though the response was observed only at the coarse level (Benedetti et al., 2021, Wu et al., 4 Mar 2026).

Weakly supervised formulations are also explicitly additive. Parcel-to-pixel property-value disaggregation defines

$y_i = V(r_i) = \sum_{p_j\in r_i}v(p_j),$

and learns pixel-level Gaussian outputs whose parcel-level sums are compared with observed parcel totals. Aggregate learning for census redistribution uses

$y_i \sim \textrm{Pois}(\textrm{cases}_i).$0

so the pixel-level model is trained only through its aggregated unit-level predictions. These formulations are top-down and weakly supervised: fine labels are latent, but the aggregation operator is known (Archbold et al., 2023, Derval et al., 2019).

3. Sources of within-area heterogeneity

Disaggregation is only identifiable through additional structure, and the recent literature relies heavily on fine-scale covariates, exposure weights, or built-environment proxies. In epidemiology, the standard inputs are polygon outcomes, fine-resolution covariates, and often a population or denominator raster. In the Madagascar malaria example, environmental covariates included mean land surface temperature, variation in land surface temperature, elevation, and enhanced vegetation index, while population served as the aggregation raster converting rates to counts (Nandi et al., 2020).

A more general covariate taxonomy appears in coherent misalignment models, which distinguish four scenarios: Raster at Full Resolution (RastFull), Raster Aggregation (RastAgg), Polygon Aggregation (PolyAgg), and Point Values (PointVal). For incomplete covariate fields, three strategies are compared: Value Plugin (VP), Joint Uncertainty (JU), and Uncertainty Plugin (UP). In that literature, JU and UP are preferred when covariates are only available as polygon averages or sparse point measurements because they propagate uncertainty from covariate reconstruction into the disaggregated intensity surface (Suen et al., 14 Feb 2025).

Other domains use imagery and built-environment data as the within-area guide. Property-value disaggregation uses 2020 aerial imagery at 1 meter GSD and parcel masks; the learned pixel-level maps assign higher values to building pixels than to lawn pixels. Refugee gridding uses Google Open Buildings and OpenStreetMap Populated Places, with within-admin2 building proportions

$y_i \sim \textrm{Pois}(\textrm{cases}_i).$1

acting as a spatial prior over target grid cells. Deep urban-morphology disaggregation uses Sentinel-1 SAR GRD, Sentinel-2 Harmonized MSI, Dynamic World V1 built-area probability, and multiple footprint datasets, together with expert conditional relationships among wall, roof, height, and macro-taxonomy classes (Archbold et al., 2023, Wells et al., 19 May 2025, Dimasaka et al., 30 Jul 2025).

Not all methods use this type of ancillary information. Some disease-mapping models work instead with expected counts $y_i \sim \textrm{Pois}(\textrm{cases}_i).$2 as offsets and adjacency structure among observed areas. This supports relative-risk modeling and multiscale decomposition on the observed support, but it does not use fine-scale covariates such as land use, population rasters, or remotely sensed predictors to redistribute counts within regions. This difference separates direct areal disaggregation from adjacent areal count analysis (Flury et al., 2020).

4. Computational frameworks

The computational literature is unusually heterogeneous. A major line of work uses Template Model Builder (TMB) for latent Gaussian disaggregation models with a spatial field defined through an INLA mesh. TMB combines CppAD, Eigen, and CHOLMOD, and uses the automatic Laplace approximation for approximate Bayesian inference. In the Madagascar case study, TMB took 56 seconds whereas NUTS-based MCMC took 94 hours, while yielding very similar parameter estimates (Nandi et al., 2020).

Continuous-domain coherent disaggregation with polygon integrals uses INLA/inlabru together with an iteratively linearised integration scheme. The difficult term

$y_i \sim \textrm{Pois}(\textrm{cases}_i).$3

is approximated by numerical integration over subsets of the domain, then linearised so that INLA can operate on an approximate latent Gaussian model. This makes polygon-to-surface inference feasible while retaining a coherent integral link from the latent intensity to the observed areal counts (Suen et al., 14 Feb 2025).

Areal disaggregation from a small area estimation perspective uses the same basic idea of nonlinear aggregation plus latent Gaussian structure, but emphasizes inlabru for the nonlinear predictor

$y_i \sim \textrm{Pois}(\textrm{cases}_i).$4

The nonlinear predictor is replaced by a first-order Taylor approximation so that INLA can operate with a linearized design matrix and an offset term. Closely related ideas appear in spatio-temporal disaggregation with changing areal boundaries, where the Extended Latent Gaussian Model framework, adaptive Gauss-Hermite quadrature, and a gamma-mixed negative binomial likelihood are used to avoid one latent overdispersion variable per polygon-time pair (Wu et al., 4 Mar 2026, Ripstein et al., 23 Jun 2026).

Machine-learning implementations often hard-wire the aggregation operator into the architecture. The Kedis framework builds a neural network whose latent output is pixel-level risk but whose training target is the polygon-level aggregate

$y_i \sim \textrm{Pois}(\textrm{cases}_i).$5

In the Madagascar experiments, Kedis was more flexible and sometimes faster than disaggregation with a spatial field, but it did not improve predictive accuracy. In spatio-temporal irregular-to-irregular settings, SARN integrates structurally-aware spatial attention (SASA) layers into a GRU, using both global attention and containment-based structural attention to learn from nested coarse and fine geographies (Hall et al., 2023, Han et al., 2023).

5. Application domains and empirical findings

Spatial disease mapping remains the most developed application area. The Madagascar malaria work shows how polygon-level case counts, fine-scale environment, and a spatial field can be combined to produce fine-scale incidence maps. A large simulation study of disaggregation regression then clarifies when such maps are trustworthy: predictive performance improved as the number of observations increased and as the size of the aggregated areas decreased, and when the model was well specified, fine-scale predictions were accurate even with small numbers of observations and large aggregated areas. Under misspecification, performance deteriorated sharply for large polygons, and polygon-level cross-validation was only a moderately good proxy for fine-scale performance, with correlation $y_i \sim \textrm{Pois}(\textrm{cases}_i).$6 between polygon CV correlation and pixel-level correlation (Arambepola et al., 2020).

Survey-based disaggregation extends the same logic to published proportions and small-area estimation. One application disaggregates ACS poverty estimates in Michigan to annual tract-level proportions while accounting for survey design, and another recovers Admin-2 fertility from Admin-1 survey releases in Kenya. The Kenya framework is then applied to the 2021 Kenya Time Use Survey to estimate district-level unpaid care and domestic work and media usage, showing how posterior draws at fine scale can support exceedance probabilities and other fully Bayesian summaries (Benedetti et al., 2021, Wu et al., 4 Mar 2026).

Remote sensing offers a different validation regime because finer-resolution products sometimes exist. SMAP brightness temperatures were disaggregated from 36 km to 9 km using segmentation plus support vector regression. The disaggregated values were very similar to the SMAP 9 km product, with a mean difference $y_i \sim \textrm{Pois}(\textrm{cases}_i).$7 K, but the standard deviation was lower by 7 K, indicating smoother fine-scale fields than the reference active-passive product (Chakrabarti et al., 2016).

Weakly supervised and deep-learning applications show both the promise and the domain specificity of modern disaggregation. In Hennepin County, the probabilistic parcel-to-pixel property-value model achieved MAE 63,311 and MAPE 27.14\% in the merged-parcel experiment, close to the deterministic weakly supervised baseline while adding uncertainty outputs. In Rwanda, DeepC4 estimated 3,350,277 dwellings against a census total of 3,312,743 for an error of 1.13\%, and 13,246,394 occupants against 13,100,600 for an error of 1.11\%. In humanitarian gridding across 25 Sub-Saharan African countries, a semi-supervised label-spreading workflow achieved 92.9\% average accuracy in placing more than 10 million refugee observations into 0.5-degree grid cells. In irregular spatio-temporal mobility disaggregation, SARN outperformed other neural models by 5\% and 1\% and typical heuristic methods by 40\% and 14\% on two datasets (Archbold et al., 2023, Dimasaka et al., 30 Jul 2025, Wells et al., 19 May 2025, Han et al., 2023).

6. Limits, diagnostics, and adjacent methodologies

Several limitations recur across the literature. First, the support-consistency assumption is strong: the observed coarse-area outcome must genuinely arise from aggregation of a latent fine-scale process. Second, the quality of the fine-scale estimates depends heavily on informative fine-scale covariates, exposure weights, or spatial structure. Third, practical validation is often limited: the disaggregation R package reports in-sample metrics such as RMSE, MAE, Pearson correlation, Spearman correlation, and log-Pearson correlation, but cross-validation tools are not currently included, and the disease-mapping simulation study shows that aggregate-level validation is not a sufficient guarantee of fine-scale accuracy (Nandi et al., 2020, Arambepola et al., 2020).

Some relevant methods are explicitly foundational rather than direct. The multiresolution decomposition for irregular areal count data extends a scale-space framework to counts on adjacency graphs using a BYM-style model and graph smoothing, but its output remains on the original areal support. The paper’s contribution to areal disaggregation is therefore foundational rather than direct: it provides a multiscale representation of count-valued areal data that could be useful inside a disaggregation workflow, but it does not define a cross-support aggregation/disaggregation operator (Flury et al., 2020).

A second set of adjacent papers studies whether aggregation itself is too destructive. The $y_i \sim \textrm{Pois}(\textrm{cases}_i).$8-maup statistic is a nonparametric test of how much the distribution of a spatially intensive variable changes when fine areas are aggregated, and it provides a rejection rule

$y_i \sim \textrm{Pois}(\textrm{cases}_i).$9

for MAUP sensitivity. CAGE and MVCAGE formalize aggregation error as between-scale variance of eigenfunctions in multiscale Karhunen–Loève expansions and use that quantity for regionalization and uncertainty quantification. These are not downscaling algorithms, but they are directly relevant because they measure how much fine-scale structure may already have been lost before disaggregation begins (Duque et al., 2018, Bradley et al., 2015, Daw et al., 2023).

The strongest practical guidance in the recent literature is correspondingly cautious. Point patterns outperform aggregated counts, full-resolution covariates outperform aggregated or sparse covariate fields, and uncertainty-aware covariate reconstruction methods such as JU and UP outperform simple plug-in VP in misaligned continuous-domain disaggregation. This suggests that areal disaggregation is most reliable when fine-resolution covariates, stable support transformations, and uncertainty propagation are all treated as first-class components of the model rather than as preprocessing conveniences (Suen et al., 14 Feb 2025).