SEMANTIC SEGMENTATION KNOWLEDGE BASED MMRF OPTIMAL METHOD FOR FINE-GRAINED URBAN INFRASTRUCTURE CLASSIFICATION

Automatically understanding pixel-level semantic relation in urban area from very high resolution (VHR) aerial images is the kernel way of urbanization development and planning. However, due to aerial images covering a large urban area with VHR, manually annotated pixel-level multiple instances from urban area is unimaginable. Then, when the low quality pixel-level labelling dataset used for semantic segmentation network supervised learning, these methods cannot provide the fine-grained urban infrastructure classification mapping results. Related to this problem, we proposed the semantic segmentation knowledge (SSK) based multi-scale Markov random field (MMRF) optimal method for fine-grain urban infrastructure classification mapping. First, proposed feature ensemble way is employed to final prediction layer to fuse multiple inputs. These multiple inputs can leverage network to produce the better SSK. Second, the multi-scale wavelet decomposition and SSK are used for modelling of MMRF to produce finer urban classification results. Finally, several experiments based on ISPRS dataset is used for demonstrate that proposed method can produce fine-grained land cover classification results than the state-of-the-art methods.