HYPERSPECTRAL IMAGE INTERPRETATION BASED ON PARTIAL LEAST SQUARES Andrey Bicalho Santos 1 , Arnaldo de Albuquerque Ara´ ujo 1 , William Robson Schwartz 1 , David Menotti 2 1 Computer Science Department, Federal University of Minas Gerais, Belo Horizonte, Brazil 2 Computing Department, Federal University of Ouro Preto, Ouro Preto, Brazil ABSTRACT Remote sensed hyperspectral images have been used for many pur- poses and have become one of the most important tools in remote sensing. Due to the large amount of available bands, e.g., a few hundreds, the feature extraction step plays an important role for hy- perspectral images interpretation. In this paper, we extend a well- know feature extraction method called Extended Morphological Pro- file (EMP) which encodes spatial and spectral information by using Partial Least Squares (PLS) to emphasize the importance of the more discriminative features. PLS is employed twice in our proposal, i.e., to the EMP features and to the raw spatial information, which are then concatenated to be further interpreted by the SVM classifier. Our experiments in two well-known data sets, the Indian Pines and Pavia University, have shown that our proposal outperforms the ac- curacy of classification methods employing EMP and other baseline feature extraction methods with different classifiers. Index Terms— Remote sensing, Hyperspectral image, Feature extraction, Extended morphological profile, Partial least squares. 1. INTRODUCTION Remote sensed hyperspectral images have been used for many pur- poses and have become one of the most important tools in remote sensing [1]. First termed as imaging spectroscopy [2], and nowa- days known as hyperspectral imaging, this technology has allowed many advances in analysis and interpretation of materials on sur- face of the earth [3]. In the 1980’s, inside NASA Jet Propulsion Laboratory, researchers started developing new instruments, includ- ing the Airborne Imaging Spectrometer (AIS), later called Airborne Visible Infra-Red Imaging Spectrometer (AVIRIS). This sensor cov- ers a range of the electromagnetic spectrum from 0.4μm to 2.5μm, at spectral resolution of 10nm and more than two hundred spectral channels [4]. This technology can identify some materials that can- not be detected with traditional low-spectral resolution analysis [2], such as the Multispectral imaging systems. Hence, one can see Hy- perspectral imaging as a natural evolution of Multispectral imag- ing [1], which acquires data only in four to seven spectral channels. Typically, images are acquired using an airborne or satellite sensor at short, medium or long distance [3] and the analysis and exploitation of such data can bring many advances to the field of land cover clas- sification/interpretation such as urban planning, ecology, precision agriculture, forestry and military [1, 5]. Although the nature of hyperspectral images to produce rich in- formation of an acquired scene, the scarce referenced available data make classification still a challenging task [1, 3, 5]. This challenge inspires new researches and has received attention in the past years The authors would like to thank CNPq, CAPES, FAPEMIG, UFMG, and UFOP for the financial support. for improving classification and/or interpretation. The conventional multispectral analysis techniques may not be suitable for hyperspec- tral imaging, emphasizing the need for more advanced approaches to deal with such high dimensionality [1, 5]. To surmount this problem, kernel based methods have been applied to this context [6]. In par- ticular, the Support Vector Machines (SVM) learning algorithm with Radial Basis Function (RBF) kernel has demonstrated promising re- sults among other traditional kernel methods and classifiers [7]. In terms of spatial information, the majority of hyperspectral systems have resolution between 1 and 30 meters per pixel [1], representing great capability of distinguishing structures and ob- jects, which may be able to increase the classification accuracy. Therefore, many researches aim at integrating both spectral and spatial information in their classification systems have been per- formed [8, 9, 10, 11, 12]. Among them, the Extended Morphological Profiles (EMP) approach [8, 3, 13] has demonstrated to be useful to encode both spectral and spatial information in a vector of features. In this work, we propose a novel classification approach based on the EMP, the Partial Least Squares (PLS) [14, 15], and the SVM learning algorithm, referred to as SpecEmpPLS-SVM. This approach performs the fusion of both spectral and spatial information to take advantage of the ability of the EMP to encode spatial information and the full spectral data given by the raw hyperspectral image, to form a new feature vector using the PLS transformation which em- phasizes the importance of the more discriminative features. Then, the SVM is employed as the classifier. The results obtained on two widely used data sets (Indian Pines and Pavia University data sets) reveal that the spectral and the EMP information fusion using the PLS method and the SVM as classifier (SpecEmpPLS-SVM) presents accuracy indexes greater than those achieved by the EMP feature extraction method alone with Multi- layer Peceptron Classifier [8, 3, 13] and other baselines based on feature selection and kNN (FEGA-kNN) [16] and the raw data (pix- elwise representation) using a SVM classifier. The remainder of this work is organized as follows. Section 2 describes the main concepts of the extended morphological pro- file that is used to encode spectral-spatial information. Section 3 presents the proposed method, a classification approach based on the EMP, the PLS, and the SVM, referred to as SpecEmpPLS-SVM. Ex- periments are described in Section 4 and finally, Section 5 concludes this work. 2. EXTENDED MORPHOLOGICAL PROFILE The Extended Morphological Profile (EMP), a technique based on Mathematical Morphology (MM) [17], has been studied by many researchers [8, 9, 18, 19, 3, 20, 13] and it has shown to be a success- ful method to encode spectral-spatial information for classification purposes. The MM has a set of operations that allows us to process the image to remove undesirable objects or even increase the size