Introduction of Spatio-Spectral Indices for Using Spatial Data in Multispectral Image Classification
Different methods of using spatial information in image classification are presented. One approach is to quantify image texture to produce features for use in classifiers, and there are various methods with adjustable parameters for texture quantification. The produced features are numerous and are in different discriminating image classes. Therefore, there is a need for selecting their optimum combination, or to alternatively create a set of features that abstract their class discernibility. Inspired by spectral normalized difference indices, the concept of the spatio-spectral index is introduced in this article to produce indices from a series of spatial features created from image spectral bands. In the proposed method, the produced spatio-spectral indices for each class are used as the abstract of spatial features. Along with the image spectral bands, they are used as new feature forms for supervised classification. Features with maximum and minimum values in each class were selected after production of the average vector in the feature space, and the removal of features with a small variation range. Next, non-repetitive band pairs were selected and spatio-spectral indices were produced. Using this method, the number of selected spatial features was at most twice the number of classes and was used to produce spatio-spectral indices. Use of the produced features in classification improves classification accuracy significantly (about 30% and 6% in the two test images used here) by enhancing class discrimination and decreasing computational time. This method is also explicit and direct, with no need to use iterative optimization processes.
KeywordsSpatio-spectral index Supervised classification Image texture quantization
The authors would like to thank Telops Inc. (Québec, Canada) for acquiring and providing the data used in this study. We would like to acknowledge IEEE GRSS Image Analysis and Data Fusion Technical Committee and Dr. Michal Shimoni (Signal and Image Centre, Royal Military Academy, Belgium) for organizing the 2014 Data Fusion Contest. We also thank the Centre de Recherche Public Gabriel Lippmann (CRPGL, Luxembourg) and Dr. Martin Schlerf (CRPGL) for their contribution of the Hyper-Cam LWIR sensor, and Dr. Michaela De Martino (University of Genoa, Italy) for her contribution to data preparation.
- Ashoori, H. (2010). Generating image-based features for improving classification accuracy, M.S. Thesis, K.N. Toosi University of Technology.Google Scholar
- Baret, F., Guyot, G., & Major, D. (1990). TSAVI: A vegetation index which minimizes soil brightness effects on LAI or APAR estimation. In 12th Canadian symposium on remote sensing and IGARSS, Vancouver, Canada.Google Scholar
- Castleman, K. R. (1996). Digital image processing. Englewood Cliffs: Prentice-Hall.Google Scholar
- IEEE GRSS Data Fusion Contest. (2014). Online: http://www.grss-ieee.org/community/technical-committees/data-fusion/.
- Leguizamon, S. (1996). Characterization of texture in remotely sensed images by using the wavelet transform. In 4th International symposium on high mountain remote sensing cartography. Google Scholar
- Richardson, A. J., & Wiegand, C. L. (1977). Distinguishing vegetation from soil background information. Photogrammetric Engineering and Remote Sensing, 43, 1541–1552.Google Scholar
- Theodoridis, S., & Koutroumbas, K. (2009). Pattern recognition. Amsterdam: Elsevier.Google Scholar