GHS-POP Framework: Historical Population Mapping

• GHS-POP Framework is a global gridded population mapping methodology that allocates population counts using remote sensing-derived built-up masks and detailed census data.
• It employs a GEOBIA segmentation workflow with rigorous preprocessing, including geometric correction and radiometric normalization, to ensure high spatial fidelity.
• The integration of declassified KH-9 imagery via the HexaLCSeg dataset enables high-resolution historical reconstructions for data-scarce peri-urban and rural areas.

The GHS-POP framework refers to a global gridded population mapping methodology, in which the spatial distribution of population counts is estimated and allocated to grid cells using remote sensing-derived built-up masks and auxiliary data. Recent advancements have significantly enhanced the framework’s historical reconstruction capabilities via integration of declassified Hexagon KH-9 reconnaissance imagery and detailed, settlement-level census data. Through the introduction of the HexaLCSeg dataset, GHS-POP workflows can now produce high-resolution, historically accurate population grids for previously data-scarce peri-urban and rural contexts, as demonstrated in northern Istanbul for the period 1975–1990 (Gerrits et al., 14 Dec 2025).

1. Data Sources and Preprocessing

The refinement of the GHS-POP framework heavily relies on the integration of reconnaissance satellite imagery and rigorous preprocessing pipelines.

• Imagery Source: The HexaLCSeg dataset utilizes US National Reconnaissance Office (NRO) KH-9 "Big Bird" panchromatic film-based images acquired in 1977, with a ground sampling distance of approximately 0.6–1.2 m per pixel, covering Arnavutköy and Çekmeköy districts of Istanbul.
• Preprocessing Steps:
  • Geometric Correction: Automated tie-point detection aligns KH-9 frames with modern basemaps, supplemented by manual ground control points (GCPs; e.g., road intersections, river bends) and rubber-sheeting adjustments for local distortion minimization.
  • Radiometric Normalization: Histogram matching is applied between adjacent frames to ensure consistent illumination.
  • Mosaicking and Clipping: Frames are mosaicked into a seamless raster and clipped to the World Mollweide equal-area projection (EPSG:54009).

These preprocessing protocols ensure spatial and radiometric fidelity, enabling the robust extraction of built-up land cover at sub-meter precision (Gerrits et al., 14 Dec 2025).

2. Segmentation and Semantic Classification Methodology

The core of HexaLCSeg’s contribution is its GEOBIA (Geographic Object-Based Image Analysis) workflow, implemented in Trimble eCognition, which translates legacy film imagery into meaningful land cover masks suitable for population allocation.

• Multi-resolution Segmentation is executed with parameters (scale=10, shape=0.3, compactness=0.5), generating image objects whose boundaries correspond to spectral and textural patterns.
• Feature Extraction per object includes:
  • Spectral: Mean brightness ($\mu$).
  • Textural: Grey Level Co-occurrence Matrix (GLCM) metrics, specifically contrast: $\mathrm{contrast} = \sum_{i,j} (i-j)^2 p(i,j)$.
  • Morphological Filtering: Speckle removed via opening/closing operations.
• Rule-based Classification: An object $O$ is assigned a built-up label if:

$$D(O) = \begin{cases} 1, & \text{if } \mu_O > \tau_1 \land \mathrm{contrast}_O < \tau_2 \land \mathrm{area}(O) > A_{\min} \ 0, & \text{otherwise} \end{cases}$$

with thresholds $\tau_1=85$ (DN), $\tau_2=20$ (GLCM contrast), $A_{\min}=50\,\mathrm{m}^2$.

• Training and Validation: The classifier is trained using approximately 200 manually labeled objects spanning six classes (e.g., built-up, cropland, shrub) with a 70/30 train/test split (Gerrits et al., 14 Dec 2025).

3. Dataset Characteristics and Output Schema

The HexaLCSeg product provides a semantically segmented, high-resolution built-up mask and associated vector layers designed for seamless integration with GHS-POP’s Pop2Grid workflow.

Data Layer	Resolution	Schema and Format
Raster (GeoTIFF)	100 m × 100 m	Value: 1(built-up), 0(non), UInt8, EPSG:54009
Vector (Shapefile)	sub-meter polygons	Class (str), Confidence (float 0–1), EPSG:54009

• The raster mask aligns with GHS-POP Pop2Grid inputs, representing built-up status per cell.
• The vector data preserves sub-meter object boundaries and includes class and confidence attributes (membership score from classification function).
• 'NoData' is assigned outside the delineated study area.

This structured schema facilitates both grid-based and object-based population allocation approaches (Gerrits et al., 14 Dec 2025).

4. Accuracy Assessment

Quantitative validation is performed via stratified random sampling (500 points) across built-up and non-built-up strata.

• References for Validation: Manual digitization from 1:25 000 USGS topographic maps (1977) and comparison to high-resolution contemporary orthoimagery.
• Reported Metrics:
  • Precision: 0.89
  • Recall: 0.87
  • F₁-score: 0.88
  • Overall accuracy: 0.90

The accuracy assessment uses standard formulations:

$\text{Precision} = \frac{TP}{TP + FP}$

$\text{Recall} = \frac{TP}{TP + FN}$

$$F_1 = 2\cdot \frac{\text{Precision}\cdot \text{Recall}}{\text{Precision} + \text{Recall}}$$

$$\text{Overall accuracy} = \frac{TP + TN}{TP + TN + FP + FN}$$

These metrics indicate robust performance for the task of built-up area segmentation in historical contexts (Gerrits et al., 14 Dec 2025).

5. Integration with GHS-POP Population Disaggregation

The enhanced GHS-POP workflow incorporates HexaLCSeg in the population allocation (dasymetric mapping) chain:

• Baseline (Standard GHSL): Uses Landsat-derived built-up masks for Pop2Grid.
• Hexagon-enhanced: Replaces the Landsat mask with HexaLCSeg, allocating nonzero weight only to KH-9-derived built-up cells.

Population allocation per cell $i$ is performed via:

$P_i = \frac{W_i}{\sum_j W_j} P_T$

where $P_T$ is the zone total, $W_i$ is 1 if built-up (0 otherwise).

• Fully Integrated Variant (“Hexagon + local census”): Incorporates local settlement-level (LAU-2) census counts ($P_S$), apportioning population to built-up objects within settlements:

$$P_{i} = \sum_{S} \left(\frac{\mathrm{Area}(O_{i}\cap S)}{\sum_{k\in S}\mathrm{Area}(O_{k})} P_{S}\right)$$

This delivers fine-grained, temporally accurate population grids that more precisely reflect historical rural and peri-urban settlement distributions. A plausible implication is improved modeling accuracy in data-scarce regions and periods where only historical reconnaissance imagery and sparse census records are available (Gerrits et al., 14 Dec 2025).

6. Coverage, Scalability, and Access

• Spatial & Temporal Coverage: Current demonstrations apply to Arnavutköy (western) and Çekmeköy (eastern) districts of Istanbul using 1977 imagery. The KH-9 Declass 3 archive (1971–1986) provides nearly global coverage—excluding parts of Canada, Greenland, Australia, and Antarctica.
• Scalability: The methodology is extensible to other regions and epochs for which KH-9 frames are available, enabling replication at continental or global scales.
• Availability: All HexaLCSeg raster and vector products, alongside preprocessing scripts and documentation, are distributed under CC BY 4.0 via GitHub (https://github.com/pjgerrits/hexagon_grid_historical_pop.git). Original KH-9 frames are accessible without charge through USGS EarthExplorer.

By leveraging rigorous GEOBIA segmentation and dasymetric mapping, the revised GHS-POP framework with HexaLCSeg provides one of the first globally scalable, high-resolution built-up datasets for the 1970s–1980s, substantially advancing the reconstruction of historical population patterns in otherwise data-limited contexts (Gerrits et al., 14 Dec 2025).