Hyperspectral Image Classification Using Denoising of Intrinsic Mode Functions

Demir, Begüm; Ertürk, Sarp; Güllü, M. Kemal

doi:10.1109/lgrs.2010.2058996

Cited by 25 publications

(23 citation statements)

References 12 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…The lower order IMFs, in general the first and second IMFs, include most of the spatial information, whereas higher order IMFs lack of local spatial structure [9]. The highest local frequency and therefore mainly fine spatial detail is included in the first IMF of each hyperspectral image band.…”

Section: Data Set Representations and Their Classificationmentioning

confidence: 99%

“…Narrow-Band Interference for Synthetic Aperture Radar (SAR) data has been suppressed by the utilization of EMD in [6] and the number of bands contained in hyperspectral images has been reduced by exploiting EMD in [7]. Approaches making use of EMD to improve hyperspectral image classification accuracy have been presented in [8][9][10]. Thanks to intrinsic characteristics of EMD, EMD based approaches presented in [8][9][10] resulted in significantly improved SVM classification accuracy.…”

Section: Introductionmentioning

confidence: 99%

“…Approaches making use of EMD to improve hyperspectral image classification accuracy have been presented in [8][9][10]. Thanks to intrinsic characteristics of EMD, EMD based approaches presented in [8][9][10] resulted in significantly improved SVM classification accuracy. While the sum of the first several IMFs of hyperspectral image bands has been exploited for SVM classification in [8], denoising to the first IMFs before summing the first several IMFs has been proposed in [9].…”

Section: Introductionmentioning

confidence: 99%

“…Thanks to intrinsic characteristics of EMD, EMD based approaches presented in [8-10] resulted in significantly improved SVM classification accuracy. While the sum of the first several IMFs of hyperspectral image bands has been exploited for SVM classification in [8], denoising to the first IMFs before summing the first several IMFs has been proposed in [9]. The information contained in the first IMF and second IMF has been combined by composite kernels in [10].…”

mentioning

confidence: 99%

See 3 more Smart Citations

Empirical mode decomposition based decision fusion for higher hyperspectral image classification accuracy

Demir

Ertürk

2010

2010 IEEE International Geoscience and Remote Sensing Symposium

Self Cite

View full text Add to dashboard Cite

This paper proposes a novel Empirical Mode Decomposition (EMD) based decision fusion approach for accurate classification of hyperspectral images. The proposed method consists of three steps. In the first step, EMD, which iteratively decomposes the data into so called Intrinsic Mode Functions (IMFs) in accordance with the intrinsic characteristics of data, is applied to each hyperspectral image band for decomposition. In the second step, the IMFs are assumed as different representations of data, and original hyperspectral data as well as IMF based representations are classified by Support Vector Machine (SVM), independently from each other, to obtain independent decisions. In the final step, these independent decisions are fused by a decision fusion rule to get the final classification result. Provided experimental results demonstrate that the proposed EMD based decision approach results in improved SVM classification.Index Terms-Hyperspectral imaging, Empirical mode decomposition, Decision fusion, Support vector machine INTRODUCTIONObtaining an accurate classification map is an important research topic in the area of hyperspectral imaging in remote sensing applications [1]. Decision fusion techniques are indeed very promising methods in the context of hyperspectral image classification and can significantly improve the classification performance by fusing classification results of individual data sources after each data source has been preliminary classified. Such decision fusion approaches have been proposed to classify hyperspectral data in [2][3][4][5]. Hyperspectral data has been initially decomposed into several sources by taking into account the correlation of spectral bands in [2]. Then, each source is separately classified by Support Vector Machine (SVM) classifiers and the final decision is achieved by considering all outputs as inputs of an additional SVM classifier. The original hyperspectral data and its Extended Morphological Profiles (EMPs) (which are constructed by applying opening and closing operations to the first several principle components obtained by the principle component analysis) are separately classified by SVM classifiers to acquire independent decisions, which are finally fused to get a final classification result in [3]. The advantage of SVM for class separation and the ability of K-means for reducing the impact of spectral variation in homogeneous regions has been analyzed using a decision fusion approach in [4].This paper introduces a novel Empirical Mode Decomposition (EMD) based decision fusion approach for hyperspectral image classification. EMD, proposed in [5], deals with nonlinear and nonstationary data by decomposing signals into Intrinsic Mode Functions (IMFs). EMD has been recently given an increasing interest in the remote sensing community. Narrow-Band Interference for Synthetic Aperture Radar (SAR) data has been suppressed by the utilization of EMD in [6] and the number of bands contained in hyperspectral images has been reduced by exploiting EMD in [7]. Approaches ma...

show abstract

Section: Data Set Representations and Their Classificationmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

mentioning

confidence: 99%

See 2 more Smart Citations

Empirical mode decomposition based decision fusion for higher hyperspectral image classification accuracy

Demir

Ertürk

2010

2010 IEEE International Geoscience and Remote Sensing Symposium

Self Cite

View full text Add to dashboard Cite

show abstract

“…These characteristics can be a priori unknown or known only partly, signal-to-noise ratio can considerably vary from one to another component image (Kerekes & Baum, 2003) and even from one to another data cube of multichannel data obtained for different imaging missions; c. Although there are numerous books and papers devoted to image filter design and performance analysis (Plataniotis &Venetsanopoulos, 2000;Elad, 2010), they mainly deal with grayscale and color image processing; there are certain similarities between multichannel image filtering and color image denoising but the former case is sufficiently more complicated; d. Recently, several papers describing possible approaches to multichannel image filtering have appeared (De Backer et al, 2008;Amato et al, 2009;Benedetto et al, 2010;Renard et al, 2006;Chen & Qian, 2011;Demir et al, 2011, Pizurica & Philips, 2006Renard et al, 2008); a positive feature of some of these papers is that they study efficiency of denoising together with classification accuracy; this seems to be a correct approach since classification (in wide sense) is the final goal of multichannel RS data exploitation and filtering is only a pre-requisite for better classification; there are two main drawbacks of these papers: noise is either simulated and additive white Gaussian noise (AWGN) is usually considered as a model, or aforementioned peculiarities of noise in real-life images are not taken into account; e. Though efficiency of filtering and classification are to be studied together, there is no well established correlation between quantitative criteria commonly used in filtering (and lossy compression) as mean square error (MSE), peak signal-to-noise ratio (PSNR) and some others and criteria of classification accuracy as probability of correct classification (PCC), misclassification matrix, anomaly detection probability and others (Christophe et al, 2005); f. One problem in studying classification accuracy is availability of numerous classifiers currently applied to multichannel images as neural network (NN) ones (Plaza et al, 2008), Support Vector Machines (SVM) and their modifications (Demir et al, 2011), different statistical and clustering tools (Jeon & Landgrebe, 1999), Spectral Angle Mapper (SAM) (Renard et al, 2008), etc. ; g. It is quite difficult to establish what classifier is the best with application to multichannel RS data because classifier performance depends upon many factors as methodology of learning, parameters (as number of layers and neurons in them for NN), number of classes and features' separability, etc.…”

mentioning

confidence: 99%