Strategies for Integrating Information from Multiple Spatial Resolutions into Land-Use/ Land-Cover Classification Routines

With the development of new remote sensing systems, veryhigh spatial and spectral resolution images now provide a source for detailed and continuous sampling of the Earth’s surface from local to regional scales. This paper presents three strategies for selecting and integrating information from different spatial resolutions into classification routines. One strategy is to combine layers of images of varying resolution. A second strategy involves comparing the a posteriori probabilities of each class at different resolutions. Another strategy is based on a top-down approach starting with the coarsest resolution. The multiresolution strategies are tested using simulated multiresolution images for a portion of the rural-urban fringe of the San Diego Metropolitan Area. The classification accuracy obtained from using three multiple strategies was greater when compared with that from a conventional single-resolution approach. Among the three strategies, the top-down approach resulted in the highest classification accuracy with a Kappa value of 0.648, compared to a Kappa of 0.566 for the conventional classifier. Introduction Accurate and timely land-use/land-cover information is essential to many government and private organizations at local, regional, national, and global levels for different applications, such as environmental monitoring and planning, land-use/ land-cover change modeling, transportation planning, urban development planning, and urban modeling. Remotely sensed data have been the major sources of generating land-use/ land-cover maps. Spatial resolution is one of the fundamental considerations when using remotely sensed images. The simplest measurement of spatial resolution is ground resolved distance (GRD), defined as the dimensions of the smallest objects recorded on an image. For a satellite remote sensing system, the spatial resolution is the dimension of the ground-projected instantaneous-field-of-view (IFOV) recorded in an image (e.g., 30 meters per pixel for TM imagery or 10 meters per pixel for SPOT panchromatic imagery). Often the ground sampling distance (pixel size) is used to represent the spatial resolution. But ground sampling distance of an image can be different from the spatial resolution of the sensor that records the image after resampling. For this paper, spatial resolution is used as the ground sampling distance of an image. Strategies for Integrating Information from Multiple Spatial Resolutions into Land-Use/Land-Cover Classification Routines DongMei Chen and Douglas Stow Traditionally, land-use/land-cover mapping with remotely sensed data is conducted at a single resolution by visual interpretation and/or different semiautomated image classification algorithms and strategies, including supervised, unsupervised, and hybrid training approaches (Richards and Jia, 1999); parametric and nonparametric classifiers; segmentation (Conner et al., 1984); artificial neural networks (ANN) (Civco, 1993); fuzzy sets (Wang, 1990; Foody, 1996); and knowledge-based systems (Kontoes and Rokos, 1996). With the development of new remote sensing systems, very high spatial resolution images provide a set of continuous samples of the Earth’s surface from local to regional scales. The spatial resolution of various satellite sensors presently ranges from 0.6 to 25,000 m. Furthermore, high resolution airborne data acquisition technology has developed rapidly in recent years. As an increasing number of high resolution data sets become available, such as Digital Globe (Quickbird), Space Imaging (Ikonos), Orbimage, Indian Remote Sensing (IRS), Digital Ortho Quarter Quadrangle (DOQQ), etc., there is an increasing need for more efficient approaches to process and analyze these data sets. A considerable amount of previous research has been devoted to exploring the magnitude and impact of spatial resolution on image analysis when shifting scale from coarse to fine resolutions (Gong and Howarth, 1992; Cihlar, 2000). Due to the more heterogeneous spectral-radiometric characteristics within land-use/land-cover units portrayed in high resolution images, many applications of traditional single-resolution classification approaches have not led to satisfactory results (Barnsley and Barr, 1996). In general, traditional single-resolution classification procedures are inadequate for discriminating between land-use/land-cover classes where spectral/spatial features and spatial patterns vary as a function of spatial resolution. Based on the concepts of the scene model developed by Strahler et al. (1986) and scale of variance of features in the image by Woodcock and Strahler (1987), scene models may be either high (H) resolution with pixels smaller than objects, or low (L) resolution with pixels larger than objects to be mapped. The ideal situation for image analysis and classification is reached when the pixel size of the image corresponds to the objects in the ground scene (Woodcock and Harward, 1992; Marceau et al., 1994a; Marceau et al., 1994b). Because land-cover/land-use units vary in size, the analysis scale corresponding to one object does not match the others. Therefore, it is difficult to achieve optimal classification performance at a PHOTOGRAMMETR IC ENGINEER ING & REMOTE SENS ING November 2003 1279 D.M. Chen is with the Department of Geography, Queen’s University, Kingston, Ontario K7L 3N6, Canada ([email protected]). D. Stow is with the Department of Geography, San Diego State University, San Diego, CA 92115. Photogrammetric Engineering & Remote Sensing Vol. 69, No. 11, November 2003, pp. 1279–1287. 0099-1112/03/6911–1279/$3.00/0 © 2003 American Society for Photogrammetry and Remote Sensing 02-069.qxd 10/8/03 12:20 PM Page 1279