A comparative study of pixel level and region level classification of land use types using QuickBird imagery

The availability of high-resolution (HR) remote sensing multispectral imagery brings opportunities and challenges for land cover/use classification. Combination of pixel based spatial feature extraction with spectral feature and region based image classification both are strategies applying to the challenges. However, it is still opening question which strategies can provide higher classification accuracy. Using QuickBird multispectral image in the suburb of Wuhan, China, we evaluate two strategies using the same feature sets and similar classifier. The results show that: (1). A combination of spectral and texture features can improve image classification accuracy; (2). The region level classification method does not necessarily show higher accuracy than the pixel-based method with the same parameter sets and the segmentation result has great influence on the classification accuracy of region based method.