基于图神经网络的统计数据精细空间分解

Policy Research Working Paper Fine-Scale Spatial Disaggregationof Statistical Data via Graph Neural Networks Kamwoo LeeBrian BlankespoorDavid Newhouse Policy Research Working Paper11360 Abstract Fine-grained spatial data are critical for informed deci-sion-making in domains ranging from economic planningto environmental management. However, many statisticsare only available for coarse administrative units, necessi-tating techniques for fine-scale spatial disaggregation. Thispaper introduces a graph neural network (GNN) basedframework for disaggregating aggregated indicators to afiner spatial resolution. The GNN approach leverages graph multi-resolution spatial grid well suited to graph-basedmodeling. The paper demonstrates the framework usinggross domestic product (GDP) as a representative example,disaggregating national or regional GDP to fine-resolu-tion cells. The proposed methodology is applicable to abroad class of aggregate indicators, offering a flexible andscalable tool for spatial analysis of economic, social, andenvironmental statistics. The results show that the frame- The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about developmentissues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry thenames of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those Keywords:spatialdisaggregation,graphneuralnetworks,fine-scalestatisticalmapping,regionalGDP,H3spatialindexingJELcode:C45,C55,R12,O18 incorporate local and multi-scale context. The model is trainedunder accounting constraints: predicted intensities are aggre-gated within each administrative unit and matched to observedtotals, ensuring internal consistency with official statistics. Theframework can also incorporate temporal structure via a con-tinuity regularizer that aligns fine-scale estimates across years 1. Introduction Many policy and research questions require information atspatial scales finer than those at which official statistics arereported. Indicators such as economic output, emissions, ser-vice demand, exposure to hazards, and resource use are oftenavailable only for coarse administrative units such as countries(ADM0) or first-level regions (ADM1), limiting subnationalanalysis, targeting, and evaluation. This motivatesspatial dis- We demonstrate the framework using gross domestic prod-uct (GDP) as a representative example, disaggregating na-tional or regional GDP to H3 resolution-6 cells for 2015–2024.GDP provides a concrete and widely studied test case, but themethodological contribution is a reusable template for disag-gregating a broad class of aggregates—including economic, so-cial, and environmental indicators—into fine spatial units in aglobally scalable and internally consistent manner. By bridg-ing modern graph-based representation learning with official This paper proposes a general framework for spatial disag-gregation of indicators. At the spatial level, space is discretizedusing the H3 hierarchical hexagonal indexing system, whichprovides a globally consistent grid with fixed neighborhoodstructure and explicit parent–child relations across resolutions.Building on this discretization, each administrative unit–yearpair is represented as a multi-resolution spatial graph on theH3 hexagonal grid, over which a non-negative latent intensity 2. Background and Related Work 2.1. Spatial Disaggregation Spatial disaggregation refers to the process of taking dataavailable for coarse spatial units (e.g., countries, provinces, counties) and estimating its distribution over finer spatial units(e.g., small administrative areas or uniform grids). This is oftendone for socioeconomic indicators such as population counts,economic output, or resource use (e.g., Wardrop et al., 2018;Leyk et al., 2019; Nordhaus, 2006; Khan et al., 2023). In re-cent years, global research has made significant progress in dis- 2.2.1. Structural Assumptions on Target Components A first class of approaches resolves ill-posedness by impos-ing structural assumptions on the components of the target vari-able. These methods assume that interpretable components ofthe aggregate—such as per-capita rates, per-area intensities (i.e.dasymetric mapping see Wright, 1936), or unit-level productiv-ity—are constant or follow simple rules within each administra- While the aims are similar to small area estimation (SAE),which provides statistical techniques to derive reliable esti-mates for small geographic areas or domains (Rao & Molina,2015), the approach is different. SAE uses clustered or sparsemicrodata (e.g., surveys or censuses) together with auxiliaryvariables that are available for all areas.It is a bottom-supervised prediction or regression problem, in which relation-ships between auxiliary variables and observed outco