Long-term Urban Mapping And Change Detection Based On Deep Learning

Urbanization is a substantial contributor to global change by influencing on land use and cover changes.Time series of Landsat imagery collected at fine spatial resolution using stable spectral bands over decades are most desirable for monitoring long term urban growth.In practice,however,temporal spectral variance complicates extraction of consistent information on changes in urban land cover.Moreover,the construction and application of effective training samples is time-consuming,especially at continental and global scales.Based on deep learning method,this study proposed an urban change detection method and a long-term urban mapping model.The proposed method and model are promising tools for detecting urban change in massive remote sensing data sets with limited training data.With the proposed long-term urban mapping model and labeled training samples from China,this study generated annual urban maps of all provincial capital cities from 1984-2017.The main results are listed as follows:Firstly,based on recurrent neural network,this study proposed change detection method with transferability.The transfer experiments among KunShan,Taizhou dataset from Landsat and YanChen dataset from EO-1 satellite have demonstrated the effectiveness and robustness of proposed change detection method.We assessed the influence of transfer experiments on various data size and the different complexity of training samples,and the final results show the transfer change detection experiments will be more stable with large and complex training datasets among neighbouring experimental cities.Secondly,this study also proposed a long term annual urban mapping model by applying deep learning method and transfer learning str ategy.We apply the urban mapping model to Landsat observations collected during 1984-2016 and extract annual records of urban areas in four cities in the temperate zone(Beijing,New York,Melbourne,and Munich).The model is trained using observations of Beijing collected in 1999,and then used to extract urban areas in all target cities for the entire 1984-2016 period.The overall accuracy of single-year urban maps is approximately 95 %±3% among the four target cities.The results demonstrate the practical potential and suitability of the proposed urban detection model to minimize seasonal urban spectral variance while enhancing inter-class distance between urban and non-urban pixels.With limited training data,the model is a promising tool for detecting urban change from massive remote sensing datasets in the cities under similar climate conditions.Thirdly,by analysis the urban spectral variances of different regions,Landsat sensors from different time,we labeled a training sample from China.The labeled training samples were applied to all provincial capital cities of China,and then we generated the annual urban maps of all Chinese provincial capital cities with 30 m spatial resolution.The overall urban mapping accuracy is 92 ± 3.5%,and the statistical results of urban region areas of this study is in consistent with the other published papers works.The final urban expansion maps demonstrated that the research of this study is effective to map annual urban regions under different conditions of climatic zone,ecoregion and varied topography with limited training data.This study demonstrates that deep learning is a promising tool for mining valuable information from big remote sensing data by overcoming problems associated with the temporal spectral variance and the scarcity of training samples.The proposed change detection method and long term urban mapping model can be practically apply to different cities,while the annual urban maps of all provincial capital cities will be helpful to other related researches and government decisions.