A deep information based transfer learning method to detect annual urban dynamics of Beijing and Newyork from 1984–2016

摘要

Mapping activities of urban land change is important for human activity to Earth's dynamic change. To get the detailed information on urban development maps in large area, dense training samples are needed in different area and specific season, which is cost-consuming. To overcome this issue, we provide a transfer learning method based on deep information to extract urban areas in all season and different areas by only parts of training samples from Beijing in 1999. The proposed method, which is based on an improved recurrent neural network model, aims at: 1) learning a novel model to extract urban features with transfer ability in different areas; 2) overcoming the seasonal, annual and spatial variance to extract urban areas in all seasons; 3) learning the annual urban dynamics in two cities over 30 years simultaneously. Experiments are performed on Beijing and New York over the period from 1984 to 2016, and training samples are only used with a part of Beijing images in 1999. The results show good performances on annual urban detection results.