Non-parametric multiple level set model for efficient image classification in urban areas

Abstract

Multispectral remotely sensing imagery with high spatial resolution, such as QuickBird, IKONOS satellite imagery or Aerial imagery, especially in urban scenes, often perform spectral variations and rich details within a category, resulting in a poor accuracy of classification. To seek an efficient solution, this paper presents a non-parametric and variational multiple level set model by a joint use of Aerial image and two products, digital terrain model (DTM) and digital surface model (DSM), directly or indirectly derived from raw LiDAR (Light Detection And Ranging) 3D point cloud data. Proposed model is to minimize an energy function. The energy includes two terms. First term is mainly image-based energy which introduces Parzen Window density estimation technique in the multiple level set framework. To make up the disadvantages brought by only multispectral image-based classification scheme mentioned above. A novel energy constraint term is added by combining elevation information of objects derived from LiDAR raw point cloud. Thus, a closely integrated and effective classification model under variational level set framework has formed. Finally, comparative experiments on a pair of Aerial image and LiDAR point cloud data have demonstrated the our proposal can obtain more accurate and detailed classification than that of relevant methods only depending on spectral information of image.