Remote sensing image segmentation using geodesic-kernel functions and multi-feature spaces. (August 2020)

• Record Type: Journal Article
• Authors: Zhao, Xuemei; Wang, Haijian; Wu, Jun; Li, Yu; Zhao, Shijie
• Is Part Of: Pattern recognition. Volume 104(2020:Aug.)
• Publication Date: 2020-08
• Subjects: Remote sensing -- Image segmentation -- Riemannian manifold -- Manifold projection -- Kernel function; 00-01 -- 99-00; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2020.107333
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: Using features expressed in the Riemannian manifold space, the spectral space and the label field to propose four remote sensing image segmentation algorithms which considers the Riemannian manifold space as baseline. Compared the feature expression ability of Riemannian manifold space, spectral space and the label field. It turns out that feature expression ability of the Riemannian manifold space is the strongest. Explore the complementation among features expressed in different feature spaces. Experimental results show that there exists complementary information among different feature spaces and combining all the features spaces improves the segmentation accuracy significantly. Combing features in the Riemannian manifold space and the spectral space outperforms combining features in the Riemannian manifold space with the label field a little, and obtains much better results than combining the spectral space and the label field. Abstract: Image representation is the key factor influencing the accuracy of remote sensing image segmentation. Traditional algorithms rely on the pixel-wise characteristics exhibited in the feature space. They introduce spatial information by establishing the connections between neighboring pixels in the neighborhood system. But the spectral-spatial features cannot be well expressed. In this paper, a Riemannian manifold space is introduced to express the contextual information by jointly mapping the spectral features of a pixel and its neighboring ones on to it. To benefit from the expression ability and geometric properties of the Riemannian manifold, a data submanifold and a parameter submanifold are established to depict the characteristics of the detected image and all possible segmentation results. On the parameter submanifold, only points representing objects of the current segmentation are active. Then distance between a point on the data submanifold and an active point on the parameter submanifold is measured by a geodesic-kernel function. Consequently, four geodesic-kernel function-based manifold projection criteria are proposed to explore the complementation between features expressed in different feature spaces. Experiments on synthetic and real remote sensing images demonstrated that the proposed geodesic-kernel function-based manifold projection algorithm outperforms traditional ones and features expressed in different feature spaces did contain some complementary information.