Detecting global urban expansion over the last three decades using a fully convolutional network

Abstract

The effective detection of global urban expansion is the basis of understanding urban sustainability. We propose a fully convolutional network (FCN) and employ it to detect global urban expansion from 1992–2016. We found that the global urban land area increased from 274.7 thousand km ^2 –621.1 thousand km ^2 , which is an increase of 346.4 thousand km ^2 and a growth by 1.3 times. The results display a relatively high accuracy with an average kappa index of 0.5, which is 0.3 higher than those of existing global urban expansion datasets. Three major advantages of the proposed FCN contribute to the improved accuracy, including the integration of multi-source remotely sensed data, the combination of features at multiple scales, and the ability to address the lack of training samples for historical urban land. Thus, the proposed FCN has great potential to effectively detect global urban expansion.

Keywords

• fully convolutional network
• global urban expansion
• deep learning
• nighttime light data
• vegetation index
• land surface temperature