Classifying urban functional regions by integrating buildings and points-of-interest using a stacking ensemble method

Abstract

The automatic classification of urban functional regions is vital for urban planning and governance. The current methods mainly rely on single remote sensing image data or social sensing data. However, these imagery-based methods have the disadvantage of capturing high-level socioeconomic features, whereas the information from social sensing data alone rarely contains the morphological features. To overcome these limitations, it is necessary to combine multisource data for sensing urban functionalities. This study presents an ensemble classification method that combines vector-based buildings and points-of-interest (POIs). For each block, we constructed an improved graph convolutional neural network (GCNN) to extract morphological features from the constituent buildings. The ‘Word2Vec’ model was used to obtain the socioeconomic characteristics of POIs. On this basis, a stacking ensemble model was designed to combine morphological and socioeconomic features for classifying the functionality of each block. The proposed method was trained and tested in Nanshan District, Shenzhen, China. The results showed a classification accuracy of 86.83%, which was 12.2%–16.1% higher than standalone applications based on single-source data. The trained models were also applied to two other districts, namely Futian and Guangming, achieving accuracies of 85.32% and 68.37%, respectively, which were 3.68%–7.79% and 3.69%–8.94% higher than those obtained using single-sourced data. Moreover, the classification accuracies of the proposed method showed improvements of 2.41%–9.76%, compared with the existing multisource data integration method in the three study areas. These results suggest that our ensemble method can effectively integrate features from different data sources and provide an alternative, higher-accuracy solution for classifying urban functional regions.