Dock extraction from China's Gaofen-2 multispectral imagery based on region-line primitive association analyses

Object-based image analysis (OBIA) is the mainstream technique for the analysis of high-spatial-resolution (HSR) images. However, routine OBIAs often follow a region-based technical framework, which limits their performance in remote sensing information extraction. In this study, a more flexible OBIA technical framework and methods are designed to extract a man-made object, i.e., docks, from HSR images. The proposed method includes the following steps. 1) Waters are extracted by object-based land/water classification, and buffer zones around the shorelines are built to limit the dock searching. 2) Edge line primitives (ELPs) for dock extraction are obtained from the shorelines by edge scanning and are closed into region primitives (RPs) for region-based OBIA. 3) Straight line primitives (SLPs) in contact with the RPs are extracted and spatial relationships between the RPs and SLPs are built based on the region-line primitive association framework (RLPAF). 4) Docks are then detected and refined by RLPAF features. Different with routine OBIAs, RPs are not simply obtained by image segmentation. The proposed line-to-region conversion prevents the influence of segmentation errors and imprecise segment boundaries and makes RPs accurate in morphology. In addition, synergetic analyses involving multiple region and line primitives make a flexible OBIA and improve its performance. The proposed method is tested using China's Gaofen-2 multispectral images with spatial resolution of 4 m, and compared with the results obtained with eCognition's rule-based classification. Experiments show that the proposed method can extract docks from HSR images with much better accuracy than routine OBIAs. In the three experimental areas, accuracy measures such as precision, recall, F-measure and boundary recall are more than 0.90, 0.95, 0.95, and 0.85, respectively.