The large number of spectral channels in a hyperspectral image (HSI) produces a fine spectral resolution to differentiate between materials in a scene. However, difficult classes that have similar spectral signatures are often confused while merely exploiting information in the spectral domain. Therefore, in addition to spectral characteristics, the spatial relationships inherent in HSIs should also be considered for incorporation into classifiers. The growing availability of high spectral and spatial resolution of remote sensors provides rich information for image clustering. Besides the discriminating power in the rich spectrum, contextual information can be extracted from the spatial domain, such as the size and the shape of the structure to which one pixel belongs. In recent years, spectral clustering has gained popularity compared to other clustering methods due to the difficulty of accurate statistical modeling of data in high dimensional space. The joint spatial–spectral information could be effectively incorporated into the proximity graph for spectral clustering approach, which provides a better data representation by discovering the inherent lower dimensionality from the input space. We embedded both spectral and spatial information into our proposed local density adaptive affinity matrix, which is able to handle multiscale data by automatically selecting the scale of analysis for every pixel according to its neighborhood of the correlated pixels. Furthermore, we explored the “conductivity method,” which aims at amplifying the block diagonal structure of the affinity matrix to further improve the performance of spectral clustering on HSI datasets.
Remotely sensed data fusion aims to integrate multi-source information generated from different perspectives, acquired with different sensors or captured at different times in order to produce fused data that contains more information than one individual data source. Recently, extended morphological attribute profiles (EMAPs) were proposed to embed contextual information, such as texture, shape, size and etc., into a high dimensional feature space as an alternative data source to hyperspectral image (HSI). Although EMAPs provide greater capabilities in modeling both spatial and spectral information, they lead to an increase in the dimensionality of the extracted features. Conventionally, a data point in high dimensional feature space is represented by a vector. For HSI, this data representation has one obvious shortcoming in that only spectral knowledge is utilized without contextual relationship being exploited. Tensors provide a natural representation for HSI data by incorporating both spatial neighborhood awareness and spectral information. Besides, tensors can be conveniently incorporated into a superpixel-based HSI image processing framework. In our paper, three tensor-based dimensionality reduction (DR) approaches were generalized for high dimensional image with promising results reported. Among the tensor-based DR approaches, the Tensor Locality Preserving Projection (TLPP) algorithm utilized graph Laplacian to model the pairwise relationship among the tensor data points. It also demonstrated excellent performance for both pixel-wise and superpixel-wise classification on Pavia University dataset.
Image segmentation and clustering is a method to extract a set of components whose members are similar in
some way. Instead of focusing on the consistencies of local image characteristics such as borders and regions in a
perceptual way, the spectral graph theoretic approach is based on the eigenvectors of an affinity matrix; therefore
it captures perceptually important non-local properties of an image. A typical spectral graph segmentation
algorithm, normalized cuts, incorporates both the dissimilarity between groups and similarity within groups by
capturing global consistency making the segmentation process more balanced and stable. For spectral graph
partitioning, we create a graph-image representation wherein each pixel is taken as a graph node, and two pixels
are connected by an edge based on certain similarity criteria. In most cases, nearby pixels are likely to be
in the same region, therefore each pixel is connected to its spatial neighbors in the normalized cut algorithm.
However, this ignores the difference between distinct groups or the similarity within a group. A hyperspectral
image contains high spatial correlation among pixels, but each pixel is better described by its high dimensional
spectral feature vector which provides more information when characterizing the similarities among every pair
of pixels. Also, to facilitate the fact that boundary usually resides in low density regions in spectral domain, a
local density adaptive affinity matrix is presented in this paper. Results will be shown for airborne hyperspectral
imagery collected with the HyMAP, AVIRIS, HYDICE sensors.
KEYWORDS: Data modeling, Detection and tracking algorithms, Principal component analysis, Hyperspectral imaging, Image analysis, Image processing, Sensors, Digital imaging, Data acquisition, Feature extraction
In general, spectral image classification algorithms fall into one of two categories: supervised and unsupervised. In unsupervised approaches, the algorithm automatically identifies clusters in the data without a priori information about those clusters (except perhaps the expected number of them). Supervised approaches require an analyst to identify training data to learn the characteristics of the clusters such that they can then classify all other pixels into one of the pre-defined groups. The classification algorithm presented here is a semi-supervised approach based on the Topological Anomaly Detection (TAD) algorithm. The TAD algorithm defines background components based on a mutual k-Nearest Neighbor graph model of the data, along with a spectral connected components analysis. Here, the largest components produced by TAD are used as regions of interest (ROI's),or training data for a supervised classification scheme. By combining those ROI's with a Gaussian Maximum Likelihood (GML) or a Minimum Distance to the Mean (MDM) algorithm, we are able to achieve a semi supervised classification method. We test this classification algorithm against data collected by the HyMAP sensor over the Cooke City, MT area and University of Pavia scene.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.