Classification of Very High Spatial Resolution Imagery Based on the Fusion of Edge and Multispectral Information

A new algorithm based on the fusion of edge and multispectral information is proposed for the pixel-wise classification of very high-resolution (VHR) remotely sensed imagery. It integrates the multispectral, spatial and structural information existing in the image. The edge feature is first extracted using an improved multispectral edge detection method, which takes into account the original multispectral bands, the linear NDVI, and the independent spectral components extracted by independent component analysis (ICA). Direction-lines are then defined using the edge and multispectral information. Two effective spatial measures are calculated based on the direction-lines in order to describe the contextual information and strengthen the multispectral feature space. Then, the support vector machine (SVM) is employed to classify the hybrid structural-multispectral feature set. In experiments, the proposed spatial measures were compared with the pixel shape index (PSI) and the gray level co-occurrence matrix (GLCM). The experimental results show that the proposed algorithm performs well in terms of classification accuracies and visual interpretation. Introduction At present, commercially available high spatial resolution multispectral images, obtained from QuickBird, Ikonos, etc., can provide a large amount of detailed ground information in a timely manner, thus opening up avenues for new remote sensing applications. However, due to the complex spatial arrangement and spectral heterogeneity even within the same class, traditional spectral methods have proven inadequate for the classification of VHR satellite imagery (Myint et al., 2004). It is generally agreed that combining spectral and spatial information can improve land-use classification from satellite imagery (Dell’Acqua et al., 2004). Therefore, in recent years, many algorithms have been proposed to extract the spatial features, complement the spectral information and improve the pixel-wise classification. One commonly applied statistical procedure is the gray level co-occurrence matrix (GLCM), which is a widely used texture and pattern recognition technique in the analysis of satellite data, and has been successful to a certain extent (Zhang, 1999; Gong et al., 1992; Barber et al., 1991). A method based on straight lines to assess land development Classification of Very High Spatial Resolution Imagery Based on the Fusion of Edge and Multispectral Information Xin Huang, Liangpei Zhang, and Pingxiang Li in high-resolution satellite images was introduced and a set of statistical measures were extracted based on the subwindows showing the regional line distribution (Unsalan et al., 2004). An edge detection method integrating a regiongrowing approach was used to improve classification for the images of the Indian Remote Sensing Satellite 1C (Sun et al., 2003). Benediktsson et al. (2003 and 2005) presented a technique of extended morphological profiles to describe the multi-scale spatial features and to interpret high spatial resolution remote sensing data. Yu (2005) proposed the idea of analyzing lines emanating from a point of interest for segmentation and classification in computer vision. Zhang et al. (2006) proposed a pixel shape index (PSI), which extracted structural features based on contextual spectral similarity. The PSI was calculated by a predetermined number of equally spaced lines radiating from the central pixel (called direction-lines). Based upon the aforementioned work, this paper proposes a methodology that integrates edge and multispectral information to improve the classification accuracy. The flow for the whole process is shown in Figure 1. The salient aspects of our strategy are the following: 1. An improved method of multispectral edge detection is proposed. The original multispectral bands, the linear version of the normalized difference vegetation index (NDVI) (Unsalan et al., 2004), and the spectral independent components extracted by ICA, are combined to produce the edge information. The edge map is a fuzzy image, where each pixel is represented by a value indicating the number of times it has been detected as an edge pixel. 2. The direction-lines are determined based on the fuzzy edge image and contextual spectral similarity in the original multispectral bands. A decision rule is proposed to fuse the edge and multispectral information when direction-lines are being extended. Two effective spatial measures are then proposed to compute the statistical characteristics of the direction-lines for each pixel, called mean and length-width ratio. 3. The proposed structural features are integrated with the multispectral information using the support vector machine (SVM), a relatively new method of machine learning. The notable advantages of SVM include self-adaptability, swift learning pace, and high-dimensional property in feature space. SVM can also reduce the dominance effects of the spectral PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING Decembe r 2008 1585 Photogrammetric Engineering & Remote Sensing Vol. 74, No. 12, December 2008, pp. 1585–1596. 0099-1112/08/7412–1585/$3.00/0 © 2008 American Society for Photogrammetry and Remote Sensing The State Key Laboratory of Information Engineering in Surveying, Mapping, and Remote Sensing, Wuhan University, P.R. China ([email protected]; [email protected]). 06-117.qxd 11/13/08 10:01 PM Page 1585