Adaptive spatial reclassification kernels for urban mapping from remotely sensed data: the A-SPARK approach ABBAS ALIMOHAMMADI*† and MEHRAN SHIRKAVAND‡ †Department of GIS, Faculty of Geomatics Engineering, K.N. Toosi University of Technology, Tehran, Iran ‡Department of Surveying and GIS, Exploration Directorate of National Iranian Oil Company (NIOC), Tehran, Iran (Received 4 June 2007; in final form 6 June 2008) The existing spatial reclassification kernel (SPARK) approach provides a simple and practical procedure for discrimination of complex land use classes from the primary land cover components. Previous works have shown that the spatial information extracted from a single kernel size often does not lead to a satisfactory result. Due to the complexity and diversity of most objects of interest, this limitation is more significant in urban dominated landscapes. To overcome this limitation, an adaptive approach for implementation of SPARK based on the automatic evaluation and selection of kernels has been developed in this research. Efficiency of the proposed approach for discrimination of spectrally confused and complex classes such as high- density and low-density residential areas has been evaluated by using SPOT data acquired from part of Tehran’s metropolitan areas, Iran. Results of the practical examination have shown that considerable improvements in the classification accu- racy of different classes such as high-density residential, low-density residential, industrial, orchards and bare lands can be achieved. The overall accuracy of classi- fication has increased from 82.39% to 92% in the best fixed kernel size of 9  9; this is an indicator of the more effective information use in the proposed approach. 1. Introduction Spatial reclassification kernel (SPARK) is a well-known approach for classification of urban dominated landscapes. This approach was proposed by Wharton (1982) and then refined by Gong and Howarth (1992) and Barnsley and Barr (1996) and provides a simple and practical procedure for discrimination of complex land use classes from the primary land cover components (Barnsley et al. 1991, Barnsley and Barr 1992, Eyton 1993, Kontoes et al. 2000, Kuo et al. 2001). In addition to urban areas, SPARK has shown high capability for classification of heterogeneous and complex vegetation types and habitats (e.g. Jaquin et al. 2003, Sluiter et al. 2004), which are very common in woodlands and degraded forests. The land cover data are produced through either traditional supervised or unsupervised classification (e.g. Gong and Howarth 1992, de Jong et al. 2000, Harts et al. 2002, Xu et al. 2003). Reclassification is mainly based on the frequency of land cover types as defined through the rectangular kernels. However, due to the important role of kernels, definition of the suitable kernel size for this purpose is a critical problem. Effects of spatial resolution and texture *Corresponding author. Email: alimoh_abb@kntu.ac.ir International Journal of Remote Sensing ISSN 0143-1161 print/ISSN 1366-5901 online # 2010 Taylor & Francis http://www.tandf.co.uk/journals DOI: 10.1080/01431160902897825 International Journal of Remote Sensing Vol. 31, No. 3, 10 February 2010, 761–774