Classification Using Extended Morphological Attribute Profiles Based On Different Feature Extraction Techniques Stiijn Peeters University of Antwerp, Antwerp, Belgium Dr. Prashanth Marpu Mattia Pedergnana Prof. Jon Atli Benediktsson University of Iceland, Reykjavik, Iceland Dr. Mauro Dalla Mura Prof. Lorenzo Bruzzone University of Trento, Trento, Italy
Overview <ul><li>Background </li></ul><ul><li>Morphological Attribute Profiles </li></ul><ul><li>Classification of Hyperspectral data </li></ul><ul><li>Ongoing work </li></ul><ul><li>Conclusions </li></ul>
Background <ul><li>Morphological profiles (MP) and Morphological attribute profiles (MAP) have been successfully used to fuse spectral and spatial information for the classification of remote sensing data. </li></ul><ul><li>J. A. Benediktsson, J. A. Palmason, and J. R. Sveinsson, “Classification of Hyperspectral Data From Urban Areas Based on Extended Morphological Profiles,” IEEE Trans. Geosci. Remote Sens. , vol. 43, no. 3, pp. 480–490, Mar. 2005 </li></ul><ul><li>M. Dalla Mura, J. A. Benediktsson, B. Waske, and L. Bruzzone, ”Morphological Attribute Profiles for the Analysis of Very High Resolution Images,” IEEE Trans. Geosci. Remote Sens. , vol. 48, no. 10, pp. 3747–3762, Oct. 2010 </li></ul>
Background <ul><li>Traditionally, MPs and MAPs are built using the feature extraction based on principal component analysis (PCA). </li></ul><ul><li>Moreover, the selection of filter parameters is traditionally done manually. </li></ul><ul><li>In this study, </li></ul><ul><li>We analyse the classification results by using various feature extraction techniques (PCA, Kernel PCA, DAFE, DBFE). </li></ul><ul><li>We use a simple method to build the MAPs based on standard deviation attribute automatically. </li></ul>
Morphological Profile (MP) and Morphological Attribute Profile (MAP)
Morphological Profiles When dealing with real images it is difficult to identify a single filter parameter suitable to handle all the objects in the image. Perform a multilevel analysis by using several values for the filter parameters. Build a stack of images with different degrees of filtering. Morphological Profile (MP) M. Pesaresi and J. A. Benediktsson, “A new approach for the morphological segmentation of high-resolution satellite imagery," IEEE Transactions on Geoscience and Remote Sensing , vol. 39, no. 2, pp. 309-320, 2001.
Morphological Profiles Closing Profile Opening Profile Square SE Sizes: 7, 13, 19, 25 Morphological Profiles (MPs) are composed by a sequence of opening and closing with SE of increasing size. MP
X 1 X 1 X 1 X 1 MP X 1 Hyperspectral Image X MP MP X 1 Morphological profile 1 Morphological profile n Feature Reduction F 1 F 2 F n X 1 X 1 Extended Morphological Profile X 1 X 1 X 1 X 1 Extended Morphological Profile
Selection of thresholds for constructing MAP Master Thesis: Mattia Pedergnana (University of Iceland, Iceland and University of Trento, Italy) Optimal Automatic Construction of Morphological Profiles In this study, we only use the attribute profile generated using the standard deviation attribute. The thresholds to build the profile are estimated for every feature separately from the range of standard deviation values of the training samples of all the classes. So, different threshold values are used for diferent profiles. A more general approach to use a big range of attributes has been recently proposed. An entire profile using a wide range of attributes and wide range of thresholds is built and a newly proposed hybrid genetic algorithm is used for feature selection.
Conclusion <ul><li>The results of classifying hyperspectral data using morphological attribute filters with various feature extraction techniques has been studied. </li></ul><ul><li>Better results are obtained with attribute profiles compared to morphological profiles. </li></ul><ul><li>Supervised feature reduction techniques are constrained by the number of available samples and hence do not provide consistent results. </li></ul><ul><li>Linear unsupervised feature reduction techniques such as PCA may not be useful as the features are not always able to distinguish between classes effectively. This is observed in the experiments. </li></ul>
Conclusion <ul><li>However, non-linear techniques such as kernel PCA preserve the information of the independent clusters and hence can be useful in distinguishing between classes. Experiments suggest that EAP with KPCA produces consistent and high quality classification results. </li></ul><ul><li>We are currently investigating the results with a wide range of feature extraction techniques. </li></ul><ul><li>We are also working on methods to automatically identify the thresholds to build profiles. </li></ul>
Experimental Results <ul><li>Pavia Dataset – Second Approach – HML-DBFE-RF </li></ul>Overall Accuracy: 98.3% Average Accuracy: 98.6%
Experimental Results <ul><li>Pavia Dataset – Second Approach – HML-KPCA-RF </li></ul>Overall Accuracy: 90% Average Accuracy: 97 %
Experimental Results <ul><li>Pavia Dataset – Second Approach – HML-KPCA-SVM </li></ul>Overall Accuracy: 93.8% Average Accuracy: 96.9 %
Experimental Results <ul><li>Indian Pine– Second Approach – HML-KPCA-RF </li></ul>Overall Accuracy: 93.8% Average Accuracy: 96 %
<ul><li>Thank You for your attention </li></ul>Software in Matlab can be freely obtained by sending us an email. Prashanth Marpu [email_address] Mattia Pedergnana [email_address] Mauro Dalla Mura [email_address]