📊 statistics

Approximate Bayesian inference for high-resolution spatial disaggregation using alternative data sources

This paper proposes an approximate Bayesian inference framework using a semiparametric spatial regression model and the max-and-smooth approach to integrate satellite-derived alternative data sources for generating high-resolution, pixel-level population estimates, as demonstrated in a case study of Bangalore City, India.

Original authors: Anis Pakrashi, Arnab Hazra, Sooraj M Raveendran, Krishnachandran Balakrishnan

Published 2026-09-09

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Anis Pakrashi, Arnab Hazra, Sooraj M Raveendran, Krishnachandran Balakrishnan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to plan a city's water supply or understand how people impact a local ecosystem, but the only maps you have show population numbers for entire neighborhoods. These neighborhood boundaries, often drawn for administrative convenience, can hide the true story of where people actually live. In many cities, a single district might contain both a dense, bustling downtown and quiet, empty suburbs, yet official records lump them together into one average number. This lack of detail makes it difficult to build infrastructure where it is needed most or to measure the precise footprint of human activity on the environment. Scientists have long sought a way to break these large, blurry blocks down into tiny, clear pieces, a process known as spatial disaggregation. The challenge is that while we have detailed satellite images showing buildings, roads, and trees, we lack the corresponding detailed population counts for those same tiny spots.

A team of researchers has developed a new way to solve this puzzle, using advanced statistics to combine coarse population data with high-resolution satellite imagery. Focusing on Bangalore, India, they took official population counts for 198 administrative wards and used them to estimate the number of people living in over 786,000 tiny squares, each measuring just 30 meters by 30 meters. To do this, they did not just look at the population numbers; they fed the computer a rich diet of alternative data derived from space. They analyzed land cover, such as how much of a square is covered by buildings versus vegetation, the density of streets, the height of buildings, and even the density of drainage networks. By teaching a statistical model to recognize the patterns between these physical features and the known population totals, they could then predict the population density for every single tiny square across the city.

The researchers faced a significant hurdle: the sheer size of the data made standard calculation methods impossible. Trying to compute the answer for nearly 800,000 tiny squares using traditional methods would have overwhelmed even powerful computers. To get around this, they used a clever two-step strategy. First, they used a mathematical shortcut to turn the complex, messy population counts into a smoother, easier-to-handle form. Then, they applied a technique that treats the unknown population density as a continuous, flowing surface rather than a collection of isolated dots. This allowed them to capture the natural way population density changes gradually across a city, rather than jumping abruptly from one neighborhood to the next. They tested this approach on simulated data to ensure it worked correctly before applying it to the real city.

The results revealed a much sharper picture of Bangalore than the official ward-level maps ever could. The model confirmed that the city's center, known as a major technology hub, is indeed packed with people, showing much higher density than the surrounding areas. However, the new map also showed that this density is not uniform; it varies significantly from one small block to the next. The researchers found that the most reliable predictors of where people live were the type of land use, the density of streets, and the amount of vegetation. Interestingly, while the model suggested that areas with more vegetation generally have fewer people, it also showed that the presence of buildings alone was not a perfect indicator of high density, as some built-up areas were less crowded than others. The study also provided a measure of confidence for every estimate, showing that the predictions were most precise in the city center, where the administrative wards are smaller and the data is more detailed.

This work demonstrates that it is possible to create high-resolution population maps without needing to conduct a new, expensive census. By combining existing administrative data with modern satellite observations, the method offers a powerful tool for city planners and environmental scientists. The approach is not limited to Bangalore or to counting people; the same logic could be applied to any situation where researchers need to understand how a specific quantity is distributed across a landscape using available satellite data. The study suggests that with the right statistical tools, we can turn the blurry, large-scale data of the past into the sharp, actionable insights needed to build better, more sustainable cities for the future.

Technical Summary: Approximate Bayesian Inference for High-Resolution Spatial Disaggregation

Problem Statement
The paper addresses the challenge of obtaining precise demographic information at fine-grained spatial scales, a necessity for planning localized public services (e.g., water distribution networks) and understanding local human impacts on ecosystems. While population sizes are typically available for large administrative units (e.g., wards in India), these aggregated data often obscure intricate spatial trends and fail to accurately depict population distribution for applications requiring subdivision into smaller zones. The authors propose a solution to this "spatial disaggregation" problem: recovering original pixel-level information (30m × 30m resolution) from coarser resolution observations (ward-level counts) using alternative data sources.

Methodology
The study employs a semiparametric hierarchical Bayesian latent Gaussian model framework applied to the Bruhat Bengaluru Mahanagara Palike (BBMP) area in Bangalore, India. The dataset comprises 2011 Census population counts for 198 wards and high-resolution (30m × 30m) satellite-derived predictors for 786,702 pixels.

Statistical Model:
- Likelihood: The population counts ( $Y_i$ ) for each ward are modeled as independent Poisson variables with an intensity function dependent on the ward area ( $|A_i|$ ) and a spatially varying log-intensity ( $\lambda^*_i$ ).
- Intensity Function: The log-intensity is modeled as a linear combination of covariates (land cover, land use, street density, building height, vegetation, drainage) plus a spatially dependent error term. The error term is assumed to follow a zero-mean Gaussian Process (GP) with an exponential covariance kernel, avoiding sparsity-based approximations (like SPDE) or low-rank approximations to maintain a "true" dense GP structure.
- Hierarchical Structure: The model includes hyperpriors for regression coefficients ( $\beta$ ), marginal variance ( $\sigma^2$ ), and the spatial range parameter ( $\phi$ ).
Approximate Bayesian Inference (Max-and-Smooth):
- Given the high dimensionality (786,702 pixels), exact Bayesian inference is computationally infeasible. The authors utilize the Max-and-Smooth approach (Hrafnkelsson et al., 2021).
- Laplace Approximation: The non-Gaussian Poisson likelihood is approximated by a Gaussian likelihood using the Laplace approximation. This relies on the Central Limit Theorem, noting that ward-level counts are sums of many pixel-level counts, making the maximum likelihood estimator of the log-intensity approximately normal.
- Gibbs Sampling: This Gaussian approximation allows for conjugacy between the prior and the approximated likelihood, enabling the use of Gibbs sampling (avoiding Metropolis-Hastings steps) to update high-dimensional parameter vectors.
- Computational Strategy: To handle the massive dimensionality without exhausting memory on standard workstations, the authors:
  - Use a discrete uniform prior for the range parameter $\phi$ to facilitate parallel computation of covariance matrices.
  - Avoid calculating the full $786,702 \times 786,702$ covariance matrix. Instead, they compute marginal posterior means and standard deviations for the pixel-level intensities using conditional mean and covariance formulas derived from the joint normal distribution of ward and pixel intensities.
  - Process computations ward-by-ward to manage storage constraints.

Key Contributions

Methodological Framework: The paper proposes a semiparametric Bayesian spatial regression model that integrates alternative data sources (satellite imagery, deep learning outputs) for high-resolution spatial disaggregation.
Scalable Inference: It demonstrates the feasibility of applying a "true" dense Gaussian process prior to a dataset of nearly 800,000 pixels using the Max-and-Smooth approach combined with Gibbs sampling, circumventing the need for sparse matrix approximations (e.g., SPDE) which introduce approximation errors dependent on mesh construction.
Computational Architecture: The authors provide specific implementation strategies (parallel computing, discrete priors for range parameters, and conditional expectation calculations) that make high-dimensional Bayesian inference feasible on standard workstations.

Results

Simulation Study: The proposed Laplace-GP model was compared against a standard Bayesian GLM, a Laplace model without spatial dependence, and a Laplace model with white noise.
- The Laplace-GP model demonstrated superior empirical coverage (near 100%) compared to non-spatial models, particularly when the data generating mechanism included smooth spatial surfaces.
- It achieved the lowest Deviance Information Criterion (DIC) and Watanabe-Akaike Information Criterion (WAIC) in scenarios with spatial smoothness, outperforming the standard Bayesian GLM.
- While computationally more intensive than non-spatial models, the inference time (approx. 200 seconds for posterior estimates) was deemed reasonable for the dimensionality involved.
Bangalore Case Study:
- The model successfully identified significant predictors of population density, including Land Use, Street Density, and Vegetation Count (negatively related).
- The estimated spatial range parameter ( $\phi$ ) corresponded to an effective range of approximately 837 meters, which the authors note is realistic for a rapidly developing city like Bangalore.
- The resulting pixel-level maps revealed that population densities are higher in the city center compared to the periphery, a nuance that was less apparent in the raw ward-level data. The uncertainty estimates (posterior standard deviations) were generally low, with slightly higher variability in the central regions due to smaller ward definitions.

Significance and Claims
The paper claims that its methodology offers a robust alternative to existing spatial disaggregation techniques that rely on sparse approximations (like SPDE) or fail to handle the scale of modern high-resolution datasets. By utilizing the Max-and-Smooth approach, the authors achieve a balance between computational feasibility and statistical rigor, allowing for the use of dense Gaussian processes without the approximation errors associated with mesh-based methods.

The authors state that while the case study focuses on demographic applications in Bangalore, the developed methodology is readily applicable to count-type spatial datasets across various scientific disciplines where high-resolution alternative data sources are available. They emphasize that the approach can be extended to general non-Gaussian likelihoods and different covariance kernels (e.g., Matérn, squared exponential) depending on the specific data characteristics.

Technical Summary: Approximate Bayesian Inference for High-Resolution Spatial Disaggregation

More like this