GLAMOUR: GLobAl building MOrphology dataset for URban hydroclimate modelling

Ruidong Li, Ting Sun, Saman Ghaffarian, Michel Tsamados, Guangheng Ni

TL;DR

This paper introduces GLAMOUR, a novel global dataset of building morphology (height and footprint) at a high resolution (0.0009°). It leverages a multi-task deep learning model and publicly available satellite data to provide a more accurate and accessible resource for urban hydroclimate modeling and other applications.

What Problem Does This Paper Address?

Accurately simulating how urban environments interact with climate phenomena like extreme rainfall and heatwaves is crucial for building resilient cities. A key component of these simulations is understanding the 3D structure of buildings – their height and footprint. While efforts have been made to create global building morphology datasets, existing solutions often suffer from limitations:

  • Data accessibility: Many datasets are not publicly available or require specialized computational resources.
  • Accuracy: Existing global datasets can be smoothed or lack detail, especially in complex urban layouts, leading to inaccuracies in derived building heights.
  • Resolution: Lower resolution data can conflate building height with ground elevation, introducing significant uncertainty.
  • Scalability: Generating detailed 3D building data globally using publicly accessible resources has been a significant challenge.

GLAMOUR aims to bridge this gap by providing a globally consistent, high-resolution, and publicly accessible dataset of building morphology.

Methods

The GLAMOUR dataset is generated using a multi-task deep learning (MTDL) model, specifically an enhanced version of the SHAFTS package. The core methodology involves:

  1. Target Area Identification: Urbanized areas are identified using the Gridded Population of the World (GPWv4) and Global Human Settlement-Urban Centre Database (GHS-UCDB), adhering to the UN's Degree of Urbanisation concept.
  2. Explanatory Variable Retrieval: A suite of publicly accessible satellite data is utilized:
  3. Sentinel-1 GRD: VV and VH polarizations for all-weather backscattering characteristics.
  4. Sentinel-2 BOA: Red, green, blue, and near-infrared bands for surface reflectance.
  5. Digital Elevation Models (DEMs): NASADEM for low- to mid-latitudes and Copernicus DEM for other regions. These data are preprocessed and aggregated within Google Earth Engine (GEE) to minimize undesirable effects from non-man-made elements and cloud cover.

  6. MTDL-based Morphology Estimation: The enhanced MTDL model, trained on multi-band image patches and elevation data, simultaneously estimates average building footprint ($\lambda_p$) and average building height ($H_{avg}$). The model was enhanced by increasing the training dataset size and reweighting samples to address potential underestimation in dense areas.

  7. Post-processing: The initial mapping results are refined by masking using the World Settlement Footprint layer to ensure accurate urban boundary representation. The final output is mosaicked into larger tiles for easier accessibility.

Datasets

  • Input Data:
    • Global Human Settlement-Urban Centre Database (GHS-UCDB)
    • Gridded Population of the World, Version 4 (GPWv4)
    • Sentinel-1 GRD (VV/VH)
    • Sentinel-2 BOA (R/G/B/NIR)
    • NASADEM
    • Copernicus DEM
    • World Settlement Footprint (WSF2019) for post-processing.
  • Evaluation Data:
    • GLAMOUR dataset: The newly generated dataset itself, used for internal analysis and comparison.
  • WSF3D dataset: Used as a benchmark for comparison, particularly for its global coverage and similar temporal scope.
  • City-specific reference data: For validation in 18 selected cities, including:
  • EUBUCCO (v0.1) for European cities.
  • Baidu map service (China) where building height is derived from floor counts.
  • High-resolution satellite imagery for specific cases.

Key Insights and Ideas

  • Deep Learning for Global Scale: The paper demonstrates the power of multi-task deep learning models, when combined with extensive publicly available satellite data and cloud computing, to generate high-resolution global datasets of complex urban features.
  • Bridging the Resolution Gap: GLAMOUR's 0.0009° resolution (approximately 100m at the equator) offers a significant improvement over many existing global datasets, enabling more detailed urban hydroclimate modeling.
  • MTDL's Advantage: The MTDL approach, by simultaneously learning building height and footprint, appears to capture interdependencies better than single-task models or traditional methods.
  • Data Fusion is Key: The successful integration of Sentinel-1 (SAR) and Sentinel-2 (optical) imagery, along with DEM data, highlights the benefit of multi-modal data fusion for robust urban morphology mapping.
  • Regional Variations Matter: The joint probability density plots (Fig. 7) reveal significant regional differences in building morphology, with East Asia and Pacific showing higher median building height and footprint, while North America exhibits more sprawl. This underscores the need for localized understanding in urban planning.

Key Results

  • Superior Accuracy: GLAMOUR achieves median Root Mean Square Errors (RMSEs) of 7.5 m for building height ($H_{avg}$) and 0.14 for building footprint ($\lambda_p$) across 18 validation cities. This is a significant improvement over existing datasets like WSF3D.
  • Global Coverage: The dataset covers 13,189 urban areas worldwide at a resolution of 0.0009°.
  • Improved Correlation: GLAMOUR shows a median Pearson correlation coefficient (CC) of 0.54 for building height and 0.52 for building footprint, indicating a moderate statistical correlation with reference data.
  • Performance Nuances: While generally superior, GLAMOUR tends to overestimate building height in lower-rise buildings (0-10m) and underestimate footprint in very densely built-up areas (where $\lambda_p > 0.5$).

Limitations and Future Work

  • Underestimation in Dense Areas: The paper acknowledges that GLAMOUR, like WSF3D, can underestimate building footprint in extremely dense urban areas. This is attributed to the resolution limitations of the input satellite imagery compared to higher-resolution SAR data used in WSF3D.
  • Overestimation in Low-Rise Buildings: There's a tendency for GLAMOUR to overestimate building height for lower buildings (2-3 floors).
  • Data Source Constraints: The reliance on publicly available imagery inherently limits the achievable resolution and detail, especially in complex urban environments.
  • Future Work: The authors suggest integrating GLAMOUR with other vectorized building footprint datasets (e.g., Global ML Building Footprints) and urban climate service tools (e.g., UMEP) for more refined urban hydroclimate simulations and morphometric parameter derivation. They also provide open-source code (SHAFTS) to facilitate further research and application of their methodology.

Summary updated

Abstract

Understanding building morphology is crucial for accurately simulating interactions between urban structures and hydroclimate dynamics. Despite significant efforts to generate detailed global building morphology datasets, there is a lack of practical solutions using publicly accessible resources. In this work, we present GLAMOUR, a dataset derived from open-source Sentinel imagery that captures the average building height and footprint at a resolution of 0.0009<sup>°</sup> across urbanized areas worldwide. Validated in 18 cities, GLAMOUR exhibits superior accuracy with median root mean square errors of 7.5 m and 0.14 for building height and footprint estimations, indicating better overall performance against existing published datasets. The GLAMOUR dataset provides essential morphological information of 3D building structures and can be integrated with other datasets and tools for a wide range of applications including 3D building model generation and urban morphometric parameter derivation. These extended applications enable refined hydroclimate simulation and hazard assessment on a broader scale and offer valuable insights for researchers and policymakers in building sustainable and resilient urban environments prepared for future climate adaptation.
BibTeX
@inproceedings{li2024glamour,
  author = {Ruidong Li and Ting Sun and Saman Ghaffarian and Michel Tsamados and Guangheng Ni},
  title = {{GLAMOUR: GLobAl building MOrphology dataset for URban hydroclimate modelling}},
  booktitle = {Scientific},
  year = {2024},
  doi = {10.1038/s41597-024-03446-2},
}
Figure