The influence of different building height and density data on local climate zone classification

Abstract

The selection of classification features is significant for local climate zone (LCZ) classification using machine learning methods. Although the sequential utilization of various feature datasets for LCZ classification, the impact of different spatial datasets on the classification outcomes of LCZ remains unclear. This study systematically analyzes the impact of four building height datasets and two building density datasets, combined with spectral data, on LCZ classification using the random forest method. The comparative analyses are performed in three aspects: different building height data, different building density data, and various combinations of the two. The results show that various types and sources of spatial data have distinct impacts on improving the accuracy of LCZ classification. Generally, building density datasets prove more effective in enhancing LCZ classification compared to building height datasets. Among four building height datasets, the digital surface model (DSM) exhibits the most significant improvement in LCZ classification. Additionally, building density extracted from CNBH-10 m (BD1) demonstrates superior improvement in LCZ classification compared to that attained from roof vector data (BD2). Notably, the synergy of DSM and BD2 exhibits the most substantial enhancement, achieving an OA (Over Accuracy) of 89.60% and a Kappa coefficient of 87.70%. This combination of building height and density data simultaneously enhances the classification accuracy of both building area and natural surface types. The result of this study can not only enhance our understanding of the influences of spatial information data on the LCZ classification, but also provide a useful reference to improve LCZ classification.