Boosted Band Ratio Feature Selection for Hyperspectral Image Classification Zhouyu Fu Terry Caelli Nianjun Liu Antonio Robles-Kelly NICTA * , RSISE Bldg. 115, Australian National University, Canberra, ACT 0200, Australia Abstract Band ratios have many useful applications in hyperspec- tral image analysis. While optimal ratios have been cho- sen empirically in previous research, we propose a princi- pled algorithm for the automatic selection of ratios directly from data. First, a robust method is used to estimate the Kullback-Leibler divergence(KLD) between different sam- ple distributions and evaluate the optimality of individual ratio features. Then, the boosting framework is adopted to select multiple ratio features iteratively. Multiclass classifi- cation is handled by using a pairwise classification frame- work. The algorithm can also be applied to the selection of discriminant bands. Experimental results on both simple material identification and complex land cover classifica- tion demonstrate the potential of this ratio selection algo- rithm. 1. Introduction The development of image sensor technology has made it possible to capture image data in hundreds of bands cov- ering a broad spectrum of wavelength range. The rich in- formation available in hyperspectral imagery has posed sig- nificant opportunities and challenges for feature extraction and classification. Many algorithms have been proposed for this purpose, such as Principle Component Analysis, (Lin- ear) Discriminant Analysis, Decision Boundary , Projection Pursuit, and kernel methods[1]. All these algorithms treat the raw pixel spectra as input vectors in high dimensional spaces and look for linear or nonlinear mappings to the fea- ture space (often with reduced dimensionality) by optimiz- ing certain criterion, leading to statistically optimal solu- tions to classification. An alternative way is to use simple features that are phys- ically meaningful. One such feature that has received much attention in the remote sensing community is the band ratio - the ratio of spectral values between two different bands. The important property of such ratios is that some mate- rials can be identified by simply observing a single ratio. For example, green vegetation can be differentiated from soil and other surface covers by the Normalized Vegetation Index(NDVI) - the ratio between a near infrared band and a * National ICT Australia is funded by the Australian Governments Backing Australia’s Ability initiative, in part through the Australian Re- search Council. visible red band. This has been used extensively for the esti- mation of vegetation coverage over the surface[2]. Another advantage of the band ratio is its invariance to shading, as the geometry factor related to shading is constant for differ- ent bands. This is an attractive feature for terrestrial hyper- spectral imaging, where the surface geometry of the object under study plays a significant role in what is detected by the camera. However, there is still a lack of technical justification for using band ratios. Ratios are typically chosen from empir- ical observations or from domain knowledge. Further, no algorithms have been reported that can automatically de- rive the optimal ratios from spectral data. In this paper, we exploit ratios for feature selection and classification by learning the optimal ratio features. Besides selecting a sin- gle ratio for coarse detection, our algorithm is capable of combining multiple ratios to achieve more accurate classifi- cation. To do this, we adopt a boosting framework to select ratio features iteratively. A robust method is proposed to es- timate the Kullback-Leibler divergence (KLD) between dif- ferent sample distributions and the ratio feature with max- imum KLD is selected at each iteration. Finally, we apply a Support Vector Machine(SVM) to the selected ratio fea- tures for training the classifier. The algorithm can be natu- rally generalized to handle the classification of multi-class samples by casting it into a pairwise classification frame- work. Moreover, the above procedures can also be applied to the selection of optimal bands. The remainder of this paper is organized as follows. Sec- tion 2 describes our algorithm for feature selection and clas- sification. Section 3 presents the experimental results. In the last section we conclude on the work presented here. 2. Algorithm Description The following two sections are focused on binary classi- fication. Generalization to multiclass cases is addressed in Section 2.3. 2.1. Optimal Criterion for Ratio Features Instead of directly dividing the value of one band by the other, we use an alternative definition of the band ratio given by r(λ i ,λ j )= x(λ i ) - x(λ j ) x(λ i )+ x(λ j )+ ǫ (1)