A novel density-based super-pixel aggregation for automatic segmentation of remote sensing images in urban areas

Efficient segmentation of remote sensing images needs optimally estimated parameters for any segmentation algorithm. These optimal parameters help algorithms avoid both over- and under- segmentation of image data and provide high-quality inputs for further processing.Recently, the super-pixels method has been introduced as a powerful tool to over-segment the images and replace the pixels with higher-level inputs. Automatic aggregation of super-pixels with image segments is a challenge in the remote sensing and computer programming community. In this paper, a new automated segmentation method, namely density-based super-pixel aggregation (DBSPA), is proposed. This method is based on the spatial clustering algorithm for integrating the obtained super-pixels from the Simple Linear Iterative Clustering (SLIC). The DBSPA algorithm uses a Normalized Difference Vegetation Index (NDVI) and a normalized Digital Surface Model (nDSM) to form core segments and defines the primary structure of geographic features in an image scene. Then, the box-whisker plot was used to analyze the statistical similarity of super-pixels to each core-segment, and spatially cluster all super-pixels. In our experiments, two ultra-high-resolution datasets selected from ISPRS semantic labelling challenge were used. As for the Vaihingen dataset, the overall accuracy was 83.7%, 84.8%, and 89.6% for pixel-based, object-based, and the proposed method respectively. The values for the Potsdam dataset are 85.2%, 85.6%, and 86.4%. The evaluation of results revealed an overall accuracy improvement in Random Forest classification results, while the number of image objects reduced by about 4%.