Characterizing the spatial pattern of annual urban growth by using time series Landsat imagery

摘要

Previous studies of urbanization have largely focused on the irreversible urban growth process and the conversion of non-urban lands into impervious surfaces, but less on the conversion from impervious surfaces to green space, also referred to as deurbanization. However, urbanization and deurbanization are both typical urban renewal process, which may happen simultaneously during the urban renewal. In this study, we proposed a new method to retrieve and map annual impervious surface percentage (ISP) and to characterize urban growth patterns using time series medium-resolution images. The method is implemented by employing the Cubist tree model for annual ISP inversion (AoCubist), optimizing multi-temporal Landsat composite images to minimize the impact of phenology and inter-year climate variation, and developing the C5.0 decision tree algorithm with temporal-spatial filtering rules to improve the space-time continuity and separability of patterns derived by unsupervised K-means classification. The method was applied to investigate the urban renewal in Guangzhou, China, between 2000 and 2010. The results demonstrate that the use of ISP slope series can capture the spatial variations and temporal trends of urban growth. Validation by fieldwork and comparing with Google Earth imagery indicates that our classification yielded a reasonable overall accuracy, ranging from 88.32% to 90.85%. Annual urban expansion rate remained between 4% and 10%, while annual deurbanization rate varied from 1% to 5%. In addition, the total pixels of rapid deurbanization surpassed that of rapid urban expansion. This finding suggests that various change directions occurred in the urban renewal process and that deurbanization was a way to counter-balance the rapid urbanization. This study provides a solid methodology for ISP change detection and fresh insight into the characteristics of urban growth in terms of timing, duration, and magnitude. (C) 2019 Elsevier B.V. All rights reserved.