Authors: Mengdi Zhang*, University of Chinese Academy of Sciences
Topics: Remote Sensing
Keywords: object-based image analysis; urban land cover; classification; medium resolution; support vector machine
Session Type: Poster
Start / End Time: 5:20 PM / 7:00 PM
Room: Napoleon Foyer/Common St. Corridor, Sheraton, 3rd Floor
Presentation File: No File Uploaded
Numerous studies have reinforced the perception that the object-based approach outperforms the traditional per-pixel classification algorithms, and this superiority should be most obvious in high-resolution remote sensing images. However, it is still under-explored of the potential of object-based approach in classifying medium resolution images. In order to investigate whether the object-based approach is effective in medium resolution image, two medium resolution images from SPOT-5 and Landsat TM located in a coastal urban city with complex land use/land covers were employed in this study. Five traditional pixel-based classification methods (i.e. Parallelepiped, Maximum Likelihood, Minimum Distance, Mahalanobis Distance and Support Vector Machine) were selected to be compared with the object-oriented and rule-based classification approach and the confusion matrix based accuracy was calculated to evaluate the effectiveness of the object-based approach in these medium resolution images. Experimental results showed that overall accuracy using object-based method was 97.6814% in the SPOT-5 image and 93.1871% in Landsat TM image respectively. The best overall accuracy of traditional per-pixel methods in SPOT-5 and LANDSAT TM images are 86.2138% and 82.3326% both produced by Support Vector Machine, which followed by Maximum likelihood (85.3365% and 79.9076%). These are much lower than the object-based method results. Additionally, we discussed the possible influence caused by different resolutions to figure out the optimal classification parameters and features for each resolution. The results indicated that object-based approach is also effective and superior over traditional per-pixel methods for the classification of medium resolution images, which is significant for regional and global satellite applications.