Refining urban morphology: An explainable machine learning method for estimating footprint-level building height

Abstract

Fine-grained 3D buildings with diverse morphology are a cornerstone of urban physical structures and have profound implications for sustainable city development. However, accurately estimating building height at the footprint-level is a challenge. This study bridged this gap by using random forest models to integrate the elevation, geometry and shape attributes of individual buildings, further refining those with spatial aggregation. It considered over one million buildings across 10 large Chinese cities and trained two-types models that demonstrated commendable performance in city-specific (the mean absolute error (MAE) ranged from 3.43 m to 5.06 m) and combined (MAE = 4.68 m) models. Results revealed that the current dataset had a finer urban morphology compared with existing datasets and showed outstanding generalisability in method transfer and feature ablation tests. By incorporating Shapley values, we explored the features' global and local impacts. The explainable results demonstrated that building area was the most impactful feature, and the elevation-dimension features were particularly beneficial in estimating high-rise buildings. Using the fine-grained 3D buildings, we explored the connections between explicit morphology differences and implicit contexts in cities. Overall, our work is an endeavour to estimate footprint-level building height as fuel for refining urban morphology and enabling sustainable city studies.