Pattern Recognition 42 (2009) 395--408 Contents lists available at ScienceDirect Pattern Recognition journal homepage: www.elsevier.com/locate/pr Spectral derivative feature coding for hyperspectral signature analysis Chein-I Chang a,b,c, ∗ , Sumit Chakravarty a , Hsian-Min Chen b,d,e , Yen-Chieh Ouyang b a Remote Sensing Signal and Image Processing Laboratory, Department of Computer Science and Electrical Engineering, University of Maryland Baltimore County, Baltimore, MD 21250, USA b Department of Electrical Engineering, National Chung Hsing University, Taichung, Taiwan, ROC c Environmental Restoration and Disaster Reduction Research Center, National Chung Hsing University, Taichung, Taiwan, ROC d Department of Medical Research, China Medical University Hospital, Taichung, Taiwan, ROC e Department of Radiology, China Medical University Hospital, Taichung, Taiwan, ROC ARTICLE INFO ABSTRACT Article history: Received 16 December 2006 Accepted 14 July 2008 Keywords: Spectral analysis manager (SPAM) Spectral derivative feature coding (SDFC) Spectral feature-based binary coding (SFBC) This paper presents a new approach to hyperspectral signature analysis, called spectral derivative feature coding (SDFC). It is derived from texture features used in texture classification to dictate gradient changes among adjacent bands in characterizing spectral variations so as to improve better spectral discrimination and classification. In order to evaluate its performance, two known binary coding methods, spectral anal- ysis manager (SPAM) and spectral feature-based binary coding (SFBC) are used to conduct comparative analysis. Experimental results demonstrate that the proposed SDFC performs more effectively in capturing spectral characteristics than do SPAM and SFBC. © 2008 Elsevier Ltd. All rights reserved. 1. Introduction A hyperspectral signature provides significant spectral informa- tion for signature discrimination and classification due to the use of hundreds of contiguous spectral bands. Over the past years, two general approaches have been investigated for hyperspectral sig- nature characterization. One is a coding-based approach which en- codes spectral signatures as code words and spectral analysis is then conducted by using the Hamming distance as a spectral similarity measure. Two such methods are notable. One is called spectral anal- ysis manager (SPAM) developed by Mazer et al. [1] which encodes an L-dimensional signature as a (2L - 2)-dimensional binary code word composed of the first L binary values used to encode the sign of the difference between a signature and its signature mean, and additional L - 2 binary values used to encode the sign of the differ- ence in spectral values between a band and its adjacent band. The SPAM binary coding was further extended to the so-called spectral feature-based binary coding (SFBC) by Qian et al. [2] who introduced additional L - 2 binary values to encode a signature as a (3L - 4)- dimensional binary code word. The new added L - 2 binary values are used to dictate whether the deviation of a spectral variation from ∗ Corresponding author at: Remote Sensing Signal and Image Processing Lab- oratory, Department of Computer Science and Electrical Engineering, University of Maryland Baltimore County, Baltimore, MD 21250, USA. Tel.: +1 410 455 3502; fax:+1 410 455 3969. E-mail address: cchang@umbc.edu (C.-I. Chang). 0031-3203/$ - see front matter © 2008 Elsevier Ltd. All rights reserved. doi:10.1016/j.patcog.2008.07.016 the signature mean is greater than a prescribed threshold. Both binary coding methods have demonstrated some success in spectral signature coding. Another is a signature estimation-based approach which estimates spectral profiles for signatures and spectral analysis is then carried out by using the commonly used least squares error as a criterion for optimality. There are also two such methods devel- oped along this line such as wavelet [3] and Kalman filter [4]. This paper takes the coding-based approach and presents a new signa- ture coding method, referred to as spectral derivative feature coding (SDFC) that improves both SPAM and SFBC in the sense of signature characterization. It is a spectral texture feature-based coding method and can be viewed as a generalization of both SPAM and SFBC by en- coding gradient changes in adjacent bands in a signature as spectral texture features. The idea is derived from texture analysis based on a recently developed texture feature coding method (TFCM) [5–7] which has a promising effect in medical imaging applications such as liver disease classification in sonograms [5,6], and mass detection in mammograms [7]. Instead of dealing with spatial texture features in an image as was done [5–7], the proposed SDFC converts image- based texture features to spectral derivative features as spectral textures in terms of spectral variation. It encodes a signature as a bi- nary code word while keeping track of gradient changes in spectral variation as spectral derivatives in three adjacent bands. In doing so it re-interprets SPAM and SFBC as binary encoders which use memory to record sign changes in comparison with spectral mean and deviation. More specifically, SPAM uses a 1-bit memory for each of L - 2 bands excluding the first and the last bands to remember a sign change in difference between a band and its adjacent band.