Hexagon KH-9 Reconnaissance Imagery

• Hexagon KH-9 imagery is a declassified, high-resolution panchromatic satellite record capturing detailed built environment morphology from 1971–1986.
• It employs rigorous digitization, geometric correction, and semantic segmentation techniques to extract built-up areas with high precision.
• Integration with ancillary census data and dasymetric models refines population mapping, enhancing insights into historical settlement patterns.

Declassified Hexagon KH-9 reconnaissance imagery represents a unique, high-resolution panchromatic satellite record captured by the U.S. National Reconnaissance Office during the 1971–1986 period. With a global scope and sub-metre native ground sampling, KH-9 imagery provides direct historical evidence for built environment morphology on a scale absent from other contemporary archives. Integration of this imagery into geospatial population frameworks has demonstrated significant capacity to refine and correct global gridded population products, particularly in reconstructing the spatial patterns of rural and peri-urban settlements in the 1970s for data-scarce regions. The utilization of such material requires sophisticated digitization, geometric rectification, semantic segmentation, and the fusion of ancillary census data, as exemplified in the Istanbul case study by Gerrits et al. (Gerrits et al., 14 Dec 2025).

1. Satellite and Sensor Specifications

The Hexagon KH-9 platform deployed film-based panoramic cameras in a near-circular sun-synchronous orbit at approximately 1,000 km altitude, with a revisit period of roughly 96 minutes and a near-polar inclination. Each overpass delivered film strips measuring 9″×18″, scanned at approximately 8,000 dpi, preserving ground sampling distances of 0.6–1.2 m at nadir. Coverage extends globally (excluding polar extremes), with each pass comprising three overlapping panoramic cameras for a swath width of ~240 km. The imagery is single-band, broad visible panchromatic (grayscale), archived as high-resolution (∼10 cm) film scans. This scalar and spatial configuration enables detection of built-up and non-built-up land cover classes at an unprecedented historical resolution, especially when compared to contemporaneous satellite sources such as Landsat MSS, which was limited by inferior spatial and radiometric resolutions in the 1970s (Gerrits et al., 14 Dec 2025).

2. Image Pre-processing and Geometric Correction Workflow

Effective exploitation of KH-9 imagery demands rigorous digitization and correction workflows. The process initiates with high-resolution scanning of the original film strips, followed by radiometric normalization across frames to resolve vignetting and varying exposure conditions. Automated tie-point detection and the manual geocoding of approximately 30–50 ground control points (GCPs) per frame (taken from stable geometric features such as road intersections or river bends) enable geometric warping and rubber-sheeting to reduce spatial residuals, with root mean square error (RMSE) controlled to under 20 m. Orthorectification is performed utilizing an SRTM-derived digital elevation model to resolve terrain displacement, followed by bundle-adjustment across adjacent frames for seamless mosaicking. The georeferenced, orthorectified raster is clipped to the study area extent and reprojected to World Mollweide (EPSG:54009) at 100 m grid spacing to match the conventions of the Global Human Settlement Layer (GHSL) (Gerrits et al., 14 Dec 2025).

3. Semantic Segmentation for Built-up Land-cover Extraction (HexaLCSeg)

The HexaLCSeg workflow synthesizes object-based image analysis (GeOBIA) with recent advances in GeoAI, specifically adopting a U-Net semantic segmentation architecture (four downsampling/upsampling stages, batch normalization, ReLU activation functions). Training data are derived from approximately 500 km² of hand-digitized polygons marking built-up and non-built-up classes across KH-9 tiles, verified using period-appropriate topographic maps and aerial photography. The network is optimized via a hybrid binary cross-entropy and Dice loss, $$\mathcal{L} \;=\; \alpha\,\bigl[-y\log(\hat y)\!-\!(1-y)\log(1-\hat y)\bigr]\;+\;(1-\alpha)\,\bigl[1 - \tfrac{2\,\sum y\,\hat y}{\sum y + \sum \hat y}\bigr],$$ using $\alpha=0.6$, batch size 8, initial learning rate $10^{-4}$ with cosine decay, and data augmentation (random rotations, flips, contrast jitter) for 50 epochs. Post-processing merges contiguous segments and removes connected speckle below four pixels. Reported accuracy metrics include F₁ of 0.88, intersection-over-union (IoU) of 0.80, and producer’s/consumer’s accuracy of 0.85/0.90 respectively (Gerrits et al., 14 Dec 2025).

4. Integration with Population Data Using Dasymetric Models

Population distribution is modeled using three dasymetric weighting schemes within administrative units (A):

• The standard GHSL baseline models weight as $w_i^{\rm G} = \frac{B_i}{\sum_{j\in A} B_j}$ for pixels $i$ in $A$, with $B_i$ the GHSL-built-up mask. Population allocated as $P_i^{\rm G} = w_i^{\rm G}\,P_A$.
• The Hexagon-enhanced workflow substitutes a high-resolution built-up mask: $w_i^{\rm H} = \frac{H_i}{\sum_{j\in A} H_j}$, where $H_i$ derives from Hexagon segmentation, with $P_i^{\rm H} = w_i^{\rm H}\,P_A$.
• The fully integrated model partitions $A$ into settlements $S_k$ (each with census $P_{S_k}$), allocating population as $w_{i,k}^{\rm C} = \frac{H_i}{\sum_{j\in S_k} H_j}$ and $P_i^{\rm C} = w_{i,k}^{\rm C}\,P_{S_k}$, thus leveraging settlement-level census detail and high-resolution built-up delineation (Gerrits et al., 14 Dec 2025).

5. Quantitative Evaluation Metrics

Evaluation employs standard geospatial prediction metrics computed at both pixel and zonal aggregation levels:

• Bias (%):

$$\mathrm{Bias} = \frac{\sum_i P_{\rm pred,i} - \sum_i P_{\rm ref,i}}{\sum_i P_{\rm ref,i}} \times 100$$

• Mean Absolute Error (MAE):

$$\mathrm{MAE} = \frac{1}{n}\sum_{i=1}^n |P_{\rm pred,i} - P_{\rm ref,i}|$$

• Root Mean Square Error (RMSE):

$$\mathrm{RMSE} = \sqrt{\frac{1}{n}\sum_{i=1}^n (P_{\rm pred,i}-P_{\rm ref,i})^2}$$

• Symmetric Mean Absolute Percentage Error (sMAPE):

$$\mathrm{sMAPE} = \frac{1}{n}\sum_{i=1}^n \frac{|P_{\rm pred,i}-P_{\rm ref,i}|}{|P_{\rm pred,i}|+|P_{\rm ref,i}|}$$

These metrics are calculated at the 100 m pixel level and also aggregated by GHS-SMOD urban/rural typologies to assess peri-urban and rural accuracy (Gerrits et al., 14 Dec 2025).

6. Advantages and Limitations of KH-9 Reconnaissance Imagery

The sub-metre spatial resolution enables detection of fragmented and dispersed settlement typologies (hamlets, roadside villages) that are typically missed by coarser archives such as Landsat MSS. The broad acquisition time frame (1971–1986) fills a critical archival gap in global settlement mapping prior to the widespread operationalization of modern Earth observation satellites. The panchromatic film’s radiometric fidelity is advantageous for isolating anthropogenic features from vegetation or bare soil.

Limitations include the single spectral band, precluding material-level discrimination except by textural and geometric inference; film degradation and artifacting (e.g., scratches, emulsion streaks) potentially degrade segmentation quality; pre-GPS ground control introduces spatial uncertainty in the tens of metres absent careful GCP placement; temporal coverage is irregular and may be spatially incomplete; and high-precision pre-processing and QA steps remain labor-intensive (Gerrits et al., 14 Dec 2025).

7. Impact and Research Significance

Integrated analyses of Hexagon KH-9 imagery with modern GeoAI and census datasets reveal the substantial underperformance of legacy GHSL population grids in peri-urban and rural contexts, where populations were previously misallocated to undeveloped regions. The approach demonstrated by Gerrits et al. sharply increases the spatial fidelity of reconstructed 1970s population grids at 100 m resolution, enabling more accurate diachronic analysis of urbanization and rural transformation where prior data is coarse or missing. Given the global archive of declassified film, this methodology has substantial potential for reconstructing historical population patterns in data-scarce regions worldwide (Gerrits et al., 14 Dec 2025).