Cross-temporal high spatial resolution urban scene classification and change detection based on a class-weighted deep adaptation network

Multi-temporal urban scene classification and change analysis based on high resolution (HR) remote sensing imagery can provide reliable time-series information for the semantic interpretation of urban land use and the transitional relationships, which is important information for urban planning and sustainable development. However, there are still some difficulties encountered when applying the existing multi-temporal scene classification methods to analyze urban development in China due to the complex urban structure and scene shape. The main reasons for this can be summarized as follows: 1) the multi-temporal data labeling workload caused by the differences of the data distributions among multi-temporal images; and 2) the lack of practical socio-geographical urban unit boundaries resulting from the uniform grid based segmentation. In this paper, a multi-temporal scene classification framework based on a class-weighted deep adaptation network (CWDAN) is proposed. In the CWDAN framework, multi-temporal OpenStreetMap (OSM) road networks are introduced for the scene segmentation at the land parcel level, to build clear and meaningful geographic boundaries for the scene units. The problem of large scale difference of parcels is solved by area-weighted voting (AWV). In order to solve the problems of the high workload of multi-temporal data labeling in the cross-temporal scene classification task, a gradient reversal layer (GRL) is used in the proposed CWDAN to obtain deep features with invariance relative to the shift between the domains. A class-weighted fully connected layer is used to solve the problem of unbalanced proportion of different urban scene classes. Post-classification is finally performed to obtain the scene change information. Experiments with tri-temporal datasets in Chinese areas demonstrated that the proposed framework can obtain a significantly improved performance in the cross-temporal scene classification and change analysis task.