1 Hyperspectral Data Classification Using Extended Extinction Profiles Pedram Ghamisi, Member, IEEE, Roberto Souza, Student Member, IEEE, Jon Atli Benediktsson, Fellow, IEEE, Let´ ıcia Rittner, Member, IEEE, Roberto Lotufo, Member, IEEE, Xiao Xiang Zhu, Senior Member, IEEE Abstract—This paper proposes a new approach for the spectral-spatial classification of hyperspectral images, which is based on a novel extrema-oriented connected filtering technique, here entitled as extended extinction profiles (EEPs). The proposed approach progressively simplifies the first informative features ex- tracted from hyperspectral data considering different attributes. Then, the classification approach is applied on two well-known hyperspectral data sets: Pavia University and Indian Pines, and compared with one of the most powerful filtering approaches in the literature, extended attribute profiles (EAPs). Results indicate that the proposed approach is able to efficiently extract spatial information for the classification of hyperspectral images automatically and swiftly. In addition, an array-based node- oriented max-tree representation was carried out to efficiently implement the proposed approach. Index Terms—Extended multi-extinction profile, Hyperspectral data classification, support vector machines, random forests. I. I NTRODUCTION The classification of hyperspectral images using both spec- tral and spatial information has become a vibrant topic of research recently. In order to efficiently extract spatial infor- mation, there is an extensive number of research works in the literature, based either on crisp or adaptive neighborhood systems. Among these approaches, the ones that are based on mathematical morphology have obtained great attention [1]. In [2], the concept of morphological transformations was considered to form the so-called morphological profiles (MPs). Then, in [3], the concept of MPs was successfully generalized to hyperspectral data leading to the extended morphological profiles (EMPs). Since then, EMPs and their modifications have been enormously used to extract existing spatial in- formation from hyperspectral data. Although MPs and their modifications can produce accurate classification maps, their concepts suffer from a few shortcomings, such as: (i) the shape of SEs is fixed and (ii) structuring elements (SEs) cannot char- acterize information related to the gray-level characteristics of the regions. Pedram Ghamisi and Xiao Xiang Zhu are with German Aerospace Cen- ter (DLR), Remote Sensing Technology Institute (IMF) and Technische Universit¨ at M¨ unchen (TUM), Signal Processing in Earth Observation, Mu- nich, Germany (corresponding author, e-mail: pedram.ghamisi@dlr.de and xiao.zhu@dlr.de). J. A. Benediktsson is with the Faculty of Electrical and Computer Engi- neering, University of Iceland, 107 Reykjavik, Iceland. Roberto Souza, Let´ ıcia Rittner and Roberto Lotufo are with the School of Electrical and Computer Engineering - UNICAMP, Brazil. This research has been partly supported by Alexander von Hum- boldt Fellowship for postdoctoral researchers, Helmholtz Young Investiga- tors Group “SiPEO” (VH-NG-1018, www.sipeo.bgu.tum.de), and FAPESP grants 2013/23514-0, 2015/12127-0 and 2013/07559-3 and CNPq grant 311228/2014-3. To overcome the above-mentioned shortcomings, morpho- logical attribute profiles (APs) were introduced in [4]. The AP is the generalization of the MP, which provides a multilevel characterization of an image using the sequential application of morphological attribute filters (AFs). Although the AP has been recently introduced, there is a considerable number of contributions based on that. As discussed in [1, 4], APs are more flexible than MPs since APs process images, based on different types of attributes. In fact, the attributes can be of any type. For example, they can be purely geometric, or related to the spectral values of the pixels, or on different characteristics, such as spatial relations to other connected components. In [5, 6], automatic frameworks have been proposed for the clas- sification of hyperspectral data, which are able to accurately classify hyperspectral images in an acceptable CPU processing time. A comprehensive survey on APs and their capabilities for the classification of remote sensing data can be found in [1]. In [7], Ghamisi et al. proposed the concept of extinction profiles (EPs), based on extinction filters (EFs), to further improve the classification accuracies of the APs. EPs are extrema-oriented connected filters, which are automatic by nature and in this context, they address the main shortcoming of the conventional APs (i.e., the manual setting of threshold values). Unlike AFs, EFs preserve the height of the extrema kept. In some experiments conducted on benchmark gray scale images and panchromatic remote sensing data [7, 8], the capability of EFs and EPs have been demonstrated through experiments, which confirm that they are better alternatives than AFs in terms of simplification for recognition and ob- tained classification accuracies. In this paper, the concept of EP is generalized for the classification of hyperspectral data sets, entitled as extended extinction profile (EEPs). EEPs simultaneously simplify the input image by discarding unimportant spatial details and preserves the geometrical characteristics of the other regions from the first informative features extracted by a feature extraction approach [e.g., Independent Component Analysis (ICA) or Principal Component Analysis (PCA)] on a hy- perspectral data set. In addition, the proposed approach is fully automatic in the sense that it can adjust the filtering parameters, based on the number of extrema. The output of this step provides a few informative features, which can be fed to a classification approach, e.g., random forest (RF). RF and support vector machines (SVMs) are well-established classifiers in the hyperspectral community since they can handle high dimensional data with a limited number of training