A Time-Series Approach to Detect Urbanized Areas Using Biophysical Indicators and Landsat Satellite Imagery

Detecting spatiotemporal dynamics of urbanization is essential to study the patterns and processes of urban ecosystems. However, urban areas are difficult to classify, and migrating static classification algorithms to dynamic detection is also challenging. Here, we develop a time-series approach for the detection, in which urbanization processes are represented by the temporal trajectories of biophysical indicators. According to the spectral analysis of land-cover classes, the indicators of normalized difference vegetation index, automated water extraction index, biophysical composition index, and shortwave infrared index are selected to characterize urbanization. The temporal trajectories of these indicators are calculated using Landsat data on the google earth engine platform; and their long-term trends and abrupt changes are estimated by the Theil-Sen estimator and Mann-Kendall test, as well as the Pettitt test. The urbanized pixel classes are identified from the possible trend combinations of the indicators, in which the falsely alarmed pixels are excluded by the time-series analyses of the indicators. Three Chinese cities-Fuzhou, Nanchang, and Xining-in various physical environments are selected to demonstrate the proposed approach. The accuracy assessments using very high resolution images achieve the overall accuracies above 90% in the three areas, showing comparable performance to supervised techniques whereas no ancillary data are required. By the comparisons with two other urban products, the advantages of the proposed approach in detecting urbanization processes in the early stage and weak impervious situation are demonstrated. Due to little human intervention involved in the proposed approach, it could be readily adapted to other regions.