Narrowing the gap for city building height predictions

TL;DR

This paper shows how high-resolution satellite data can accurately predict building heights in cities of the Global South, helping with urban planning and sustainability.

Contribution

The study introduces a method using high-resolution digital elevation models and deep learning to estimate building heights in data-scarce regions.

Findings

Building heights were predicted with less than 1 meter mean absolute error using digital elevation models.

A deep learning model achieved 2.2–7.0 meter mean absolute error in predicting building heights from satellite imagery.

Google’s Open Buildings dataset improved predictions but tended to overestimate building heights in some cities.

Abstract

Understanding the 3D evolution of urban environments at high resolution through space and time is crucial for targeting sustainable development and enhancing resilience to hazards but usually requires expensive commercial satellite or aerial imagery. This leads to data scarcity and analytical biases in countries without access to these capabilities. Here we use high (1.5 m) resolution digital elevation models (DEMs) derived from satellite imagery to measure the vertical component of three cities in the Global South (Nairobi, Kathmandu and Quito), which we evaluate against published datasets of modelled heights. Building heights could be determined to < 1 m mean absolute error (MAE) using the DEMs, and 2.2–7.0 m MAE using a deep learning model trained to predict heights using high-resolution satellite imagery. Google’s Open Buildings 2.5D Temporal Dataset further improved on our deep learning models for two of the three cities, although tended to overestimate building heights. Constraining the building-scale vertical dimension of urban growth creates new opportunities to quantify population distributions, assess natural hazard exposure and vulnerabilities, and evaluate material consumption for sustainable development. Deep learning derived building heights begin to address global inequalities in data availability but should be evaluated locally alongside reference data to determine biases.

Figures