Uncovering temporal changes in Europe's population density patterns using a data fusion approach

The study addresses the gap in understanding Europe’s spatio‑temporal population distribution beyond traditional place‑of‑residence maps, which approximate nighttime population. Population distribution varies strongly over daily and seasonal cycles due to mobility linked to work, study, shopping, and leisure. Mapping daytime population is challenging because there is no single statistical measure comparable to resident counts and multiple activities and locations must be inferred from indirect data. Mobile network data, sensors, and social media offer high spatio‑temporal resolution but face barriers in access, coverage, privacy, and consistency across operators and countries. The purpose is to produce the first EU‑wide, seamless, and consistent representation of intraday (day/night) and monthly population distributions at 1 km² resolution, using a data‑fusion, multi‑layered dasymetric method not constrained by mobile operator data limitations.

Dasymetric mapping has advanced population distribution modeling by disaggregating counts using high‑resolution covariates (e.g., land use/land cover, built‑up density, imperviousness, nighttime lights, user‑generated content). Numerous gridded products exist globally/continentally, with top‑down approaches prevalent where official small‑area data are lacking; bottom‑up address‑based grids are available in Europe (e.g., GEOSTAT 2011). Prior spatio‑temporal efforts include LandScan’s ambient population and various U.S. and European case studies estimating daytime or high‑frequency dynamics, often limited to small regions or single countries. Big geospatial data (Twitter, smartphone location history) have been used to assess mobility and urban structure. Mobile network operator data can map spatio‑temporal densities at high temporal granularity, including for tourists, but systematic large‑area use is hindered by restricted access, privacy concerns, operator market share biases, heterogeneous spatial resolution, antenna switching, and temporal sparsity of call detail records. Recent reviews summarize these challenges and ongoing attempts to harmonize and correct biases. This study expands on top‑down, multi‑source dasymetric methods to deliver an EU‑wide, time‑specific product.

The multi‑layered dasymetric approach proceeds in four phases: (1) estimate monthly, regional stocks of population groups; (2) map land‑use/land‑cover (LULC) and activity features; (3) disaggregate group stocks dasymetrically to 100 m pixels; (4) quality assessment via cross‑comparison. - Population groups and stocks: Sixteen groups were defined based on expected spatial behavior: residents; non‑working and non‑studying (computed as N = U + (R − A − S) at NUTS2, downscaled to NUTS3); students (primary/secondary and tertiary; NUTS2 Eurostat, tertiary downscaled using European Tertiary Education Register); employees in 11 NACE Rev.2 sectors (NUTS3 region of work from Eurostat); and tourists (inbound and outbound). Monthly matrices were compiled for 1311 NUTS3 regions (2011 reference year), incorporating school and academic holiday calendars to shift students to the non‑working/non‑studying group in months with >50% holidays. Tourism: annual nights at NUTS2 were downscaled to NUTS3 via accommodation bed places, split monthly with regional seasonal curves, then divided by days per month to obtain average daily inbound tourists. Country of origin shares (from NSIs/OECD) and a gravity‑type split by distance and GDP were used to estimate outbound tourists by country and region; domestic and intra‑EU outbound tourists were subtracted from origin regions to avoid double counting. - Ancillary geospatial data: A fine‑grained LULC map, enhancing CORINE Land Cover 2012, subdivided 11 artificial classes into 18 detailed classes (e.g., production, commercial/service, public facilities, airport terminals), with minimum mapping units of 1 ha (artificial) and 5 ha (others). Activity density layers were created from POIs and polygons (TomTom Multinet, OpenStreetMap) to represent facilities linked to students and each worker sector; for tourists, a layer of accommodation room density from online booking platforms was built. These POI layers supplemented LULC to better target activity locations. - Dasymetric disaggregation: For each month and population group, NUTS3 stocks were allocated in two tiers: (i) distribute group totals across linked LU types proportional to their occurrence within the region; (ii) allocate from LU totals to 100 m pixels proportional to built‑up density (European Settlement Map 2012). Residents were downscaled from the 1 km GEOSTAT grid using similar logic; non‑working/non‑studying derived by applying NUTS3‑specific ratios to residents at 100 m. Tourists were downscaled separately for nighttime (accommodation density) and daytime (selected LU classes). This produced 204 intermediate 100 m grids (12 months × 17 group layers, counting day/night tourists separately). Monthly nighttime grids equal residents + nighttime tourists; monthly daytime grids equal the sum of the remaining 15 group layers. Final products were aggregated to 1 km² cells, yielding 24 grids (day and night for each month), representing a typical working day, excluding weekend effects and intra‑day transitions beyond day/night frames. - Quality assessment: Cross‑comparison used allocation accuracy (AA = 1 − Σ|P_m − P*_m| / ΣP*_m × 100) at the spatial units of independent references. For Italy, Portugal, and Spain (2011 census), municipal daytime populations were reconstructed via origin‑destination matrices (students and workers commuting) and compared to model outputs without tourists. For Belgium, mobile network operator signaling data (Proximus, ~40% market share) aggregated to Voronoi polygons (>1 km²) for a weekday (10/08/2015) were averaged over 9:30–11:30 a.m. (day) and 3:00–5:00 a.m. (night), rescaled to national totals assuming constant market share, and compared to the grids.

- Product: 24 EU‑wide population grids (EU‑28 as of 2019) at 1 km² resolution for 2011, providing daytime and nighttime estimates for each month, representing a typical working day. - Urban patterns: Visualization for Milan and difference maps for Paris, Lisbon, and Milan demonstrate strong day–night shifts (central gains by day, residential belt losses) and distinct August–January nighttime seasonal patterns (e.g., Paris center gains in August; Lisbon and Algarve gains; Milan metro losses with gains near lakes Maggiore, Como, Garda). - Largest EU cities (n = 34; day or night population >1 million): Average day/night population ratio = 1.097 (σ = 0.098). Highest ratios: Budapest and Warsaw (1.31–1.32), followed by Brussels (1.24). Low ratios: Madrid, Barcelona, Valencia, Athens, Stockholm (0.94–0.99). Daytime composition (yearly averages): 48.9% employees (σ = 7.0%), 22.7% students (σ = 3.0%), 1.2% tourists (σ = 0.7%), remainder 27.2% non‑working/non‑studying residents. - Density profiles: Rescaled concentric profiles show exponential decay with distance to city center for both periods. Fitted exponentials across 34 cities yielded R² = 0.844 (daytime) and R² = 0.754 (nighttime), n = 1360 measurements each, p < 0.0001. The mean day/night ratio peaks at 1.9 at the city center, declining rapidly to just above 1 beyond 5 km, with greater variability beyond 15 km. - City typologies: K‑means clustering of day/night ratio profiles (first 15 km) identified four clusters, including a distinct group (Madrid, Barcelona, Valencia, Lyon) with predominantly residential centers where daytime densities surpass nighttime only toward the periphery. - Validation: Allocation accuracy vs. census‑based municipal data—Italy: 99.2% (night), 92.8% (day); Portugal: 99.6% (night), 92.6% (day); Spain (selected municipalities): 99.3% (night), 92.3% (day). Pearson correlations were near 1.0 in all census cases. Against Belgian mobile operator data: AA = 79.8% (night), 78.0% (day); correlations 0.866 (night) and 0.849 (day). Spanish cities’ day/night ratios from census: Madrid 1.014, Barcelona 1.006, Valencia 0.963 vs. model 0.981, 0.954, 0.948, corroborating low ratios though model slightly underestimates daytime within these cities.

The study demonstrates that a multi‑layered dasymetric fusion of official statistics with detailed LULC and POI/activity layers can produce consistent, large‑area spatio‑temporal population grids without relying on mobile operator data. The results capture known urban dynamics: strong daytime concentration in city centers, exponential decay of densities with distance, and distinct seasonal shifts. The city‑center day/night peak ratio (~1.9) and the average metropolitan ratio (~1.10) quantify diurnal concentration across Europe’s largest cities. Validation against census‑derived day and night totals shows high allocation accuracy, particularly at night, and good agreement during day, supporting the plausibility of the approach. Differences with mobile operator data reflect conceptual and measurement disparities, yet similar day/night accuracies suggest comparable quality across frames. The dataset enables improved assessments of human exposure to hazards, transport and infrastructure planning, and comparative urban analysis across borders, providing a consistent baseline for 2011 and a reference for future updates.

This work delivers the first EU‑wide, freely accessible set of 1 km² population grids that distinguish day/night and monthly variations, created via a multi‑layered dasymetric approach modeling 16 population groups with rich geospatial covariates. It quantifies diurnal centralization in major cities and provides validated, policy‑relevant inputs for risk assessment, planning, and urban studies. Future research should: (i) update to newer census years and allow rescaling for interim use; (ii) increase temporal resolution by integrating temporal signatures (e.g., from mobile phone data) while addressing access and bias; (iii) stratify grids by demographic/socioeconomic attributes; and (iv) explore synergies with synthetic population methods to enrich activity‑based representations.

- Temporal and conceptual scope: Grids represent a typical working day per month with only two frames (day/night). Weekend effects, intra‑day transitions (commuting peaks, pre/post‑work activities), and non‑workday variability are not modeled. - Reference year: 2011 baseline; population change since then may affect absolute levels, though internal structures may be relatively stable. Interim rescaling is suggested but introduces uncertainty. - Spatial resolution: Public release at 1 km² (originally modeled at 100 m) favors lower uncertainty and avoids false precision but may be coarse for highly localized applications. - Data integration and uncertainty: Heterogeneous sources (statistics, LULC, POIs) with differing definitions and errors lead to propagation of uncertainty. Links between population groups and LU types rely on expert judgment. - Tourism and seasonality: Monthly tourism modeled from nights and accommodation capacity; assumptions on origin/destination shares and proportional outbound subtraction may introduce biases. - Validation constraints: Lack of fully comparable spatio‑temporal population records means cross‑comparisons (census daytime reconstructions, single‑operator mobile data) are not definitive; conceptual differences limit strict error interpretation. - Geographic generalization: Approach is transferable but performance depends on availability/quality of input statistics and ancillary geospatial data in other regions.