Superpixel-based imaging for residential area detection of high spatial resolution remote sensing imagery

Abstract

The precise and efficient location of residential areas using high spatial resolution remote sensing imagery is a popular research area in the field of Earth observation. Most of the existing approaches are supervised or semisupervised and use data training. Among the unsupervised approaches, corner density-based mapping using kernel density estimate has been widely employed to predict the presence of built-up areas. However, it is computationally time-consuming and the statistical threshold segmentation makes it difficult to obtain a stable and accurate output. To overcome this deficiency, a new two-stage object-oriented residential area extraction scheme was designed. First, a set of corners was extracted using the Gabor filter bank with structural tensor analysis to indicate candidate buildings. Then, instead of pixel units, our method takes superpixel-based image partitions as the primary calculation elements, and an object-oriented weighted sparse spatial voting technique was proposed to accelerate the generation of a residential area presence index. It was demonstrated that the superpixel-based voting strategy was not only efficient in accelerating the calculation process, but it also reduced the false negative rate in the final detection result. Second, a graph-cut method was employed to address the residential area segmentation by integrating a density map as a prior cue that preserves the boundary accuracy better than traditional statistical threshold methods. The effectiveness of the proposed method was evaluated using a series of experiments on the sets of high-resolution Google Earth, IKONOS, and GaoFen-2 (GF2) satellite imagery. The results showed that the proposed approach outperforms the existing algorithms in terms of computational speed and accuracy.