Collaborative representation has been a popular classifier for hyperspectral image classification because it can offer excellent classification accuracy with a closed-form solution. Collaborative representation can be implemented using a dictionary with training samples of all-classes, or using class-specific sub-dictionaries. In either case, a testing pixel is assigned to the class whose training samples offer the minimum representation residual. The Collaborative Representation Optimized Classifier with Tikhonov regularization (CROCT) was developed to combine these two types of collaborative representations to achieve the balance for optimized performance. The class-specific collaborative representation involves inverse operation of matrices constructed from class-specific samples, and the all-class version requires inversion operation of the matrix constructed from all samples. In this paper, we propose a low-complexity CROCT to avoid redundant operations in all-class and class-specific collaborative representations. It can further reduce the computational cost of CROCT while maintaining its excellent classification performance.
Recent advances in remote sensing technology have made multisensor data available for the same area, and it is well-known that remote sensing data processing and analysis often benefit from multisource data fusion. Specifically, low spatial resolution of hyperspectral imagery (HSI) degrades the quality of the subsequent classification task while using visible (VIS) images with high spatial resolution enables high-fidelity spatial analysis. A collaborative classification framework is proposed to fuse HSI and VIS images for finer classification. First, the convolutional neural network model is employed to extract deep spectral features for HSI classification. Second, effective binarized statistical image features are learned as contextual basis vectors for the high-resolution VIS image, followed by a classifier. The proposed approach employs diversified data in a decision fusion, leading to an integration of the rich spectral information, spatial information, and statistical representation information. In particular, the proposed approach eliminates the potential problems of the curse of dimensionality and excessive computation time. The experiments evaluated on two standard data sets demonstrate better classification performance offered by this framework.
In recent years, oil spill surveillance with space-borne synthetic aperture radar (SAR) has received unprecedented attention and has been gradually developed into a common technique for maritime environment protection. A typical SAR-based oil spill detection process consists of three steps: (1) dark-spot segmentation, (2) feature extraction, and (3) oil spill and look-alike discrimination. As a preliminary task in the oil spill detection process chain, dark-spot segmentation is a critical and fundamental step prior to feature extraction and classification, since its output has a direct impact on the two subsequent stages. The balance between the detection probability and false alarm probability has a vital impact on the performance of the entire detection system. Unfortunately, this problem has not drawn as much attention as the other two stages. A specific effort has been placed on dark-spot segmentation in single-pol SAR imagery. A combination of fine designed features, including gray features, geometric features, and textural features, is proposed to characterize the oil spill and seawater for improving the performance of dark-spot segmentation. In the proposed process chain, a histogram stretching transform is incorporated before the gray feature extraction to enhance the contrast between possible oil spills and water. A simple but effective multiple-level thresholding algorithm is developed to conduct a binary classification before the geometric feature extraction to obtain more accurate area features. A local binary pattern code is computed and assigned as the textural feature for a pixel to characterize the physical difference between oil spills and water. The experimental result confirms that the proposed fine designed feature combination outperforms existing approaches in both aspects of overall segmentation accuracy and the capability to balance detection probability and false alarm probability. It is a promising alternative that can be incorporated into existing oil spill detection systems to further improve system performance.
A one-class sparse representation classifier (OCSRC) is proposed to solve the multitemporal change detection problem for identifying disaster affected areas. The OCSRC method, which is adapted from a sparse representation classifier (SRC), incorporates the one-class strategy from a one-class support vector machine (OCSVM) to seek accurate representation for the class of changed areas. It assumes that pixels from the changed areas can be well represented by samples from this class, thus the representation errors are taken as the possibilities of change. Performances of OCSRC and OCSVM are tested and compared with multitemporal multispectral HJ-1A images acquired in Heilongjiang Province before and after the flood in 2013. The entire image, together with two subimages, are used for overall comparison and detailed discussion. Receiver-operating-characteristics curve results show that OCSRC outperforms OCSVM by a lower false-positive rate at a defined true-positive rate (TPR), and the gap is more obvious with high TPR values. The same outcome is also manifested in the change detection image results, with less misclassified pixels for OCSRC at certain TPR values, which implies a more accurate description of the changed area.
Sparse representation-based classifier (SRC) is of great interest recently for hyperspectral image classification. It is assumed that a testing pixel is linearly combined with atoms of a dictionary. Under this circumstance, the dictionary includes all the training samples. The objective is to find a weight vector that yields a minimum L2 representation error with the constraint that the weight vector is sparse with a minimum L1 norm. The pixel is assigned to the class whose training samples yield the minimum error. In addition, collaborative representation-based classifier (CRC) is also proposed, where the weight vector has a minimum L2 norm. The CRC has a closed-form solution; when using class-specific representation it can yield even better performance than the SRC. Compared to traditional classifiers such as support vector machine (SVM), SRC and CRC do not have a traditional training-testing fashion as in supervised learning, while their performance is similar to or even better than SVM. In this paper, we investigate a generalized representation-based classifier which uses Lq representation error, Lp weight norm, and adaptive regularization. The classification performance of Lq and Lp combinations is evaluated with several real hyperspectral datasets. Based on these experiments, recommendation is provide for practical implementation.
Sparse representation-based classification (SRC) has gained great interest recently.
A pixel to be classified is sparsely approximately by labeled samples, and it is assigned to the
class whose labeled samples provide the smallest representation error. In this paper, we extend
SRC by exploiting the benefits of using a smoothing filter based on sparse gradient
minimization. The smoothing filter is expected to provide less intra class variability and more
spatial regularity, which eliminating the inherent variations within a small neighborhood.
Classification performance on two real hyperspectral datasets demonstrates that our proposed
method has improved classification accuracy and the resulting accuracies are persistently
higher at all small training sample size situations compared to some traditional classifiers.
Detecting a target with low-occurrence-probability from unknown background in a hyperspectral image, namely anomaly detection, is of practical significance. Reed-Xiaoli (RX) algorithm is considered as a classic anomaly detector, which calculates the Mahalanobis distance between local background and the pixel under test. Local RX, as an adaptive RX detector, employs a dual-window strategy to consider pixels within the frame between inner and outer windows as local background. However, the detector is sensitive if such a local region contains anomalous pixels (i.e., outliers). In this paper, a locality-constrained anomaly detector is proposed to remove outliers in the local background region before employing the RX algorithm. Specifically, a local linear representation is designed to exploit the internal relationship between linearly correlated pixels in the local background region and the pixel under test and its neighbors. Experimental results demonstrate that the proposed detector improves the original local RX algorithm.
In the current remote sensing big data era, the huge amount of data, especially the expansion of spatial dimension, brings challenges to the traditional processing and applications in data transferring, processing and visualization. Model quality assessment is usually employed to evaluate the yielding model and select the optimal simplification method, which is used to improve the efficiency of big data processing. The existing three-dimensional (3-D) terrain model assessment methods mainly exploit the data accuracy or geometric features while ignoring the impact of humans or users. We mainly investigate the quality assessment for the 3-D terrain model. The proposed method provides an integration of structural similarity and the human visual system from the multiview angles for the image quality assessment. The visual impact of observers and the structural information retention of the 3-D terrain model have been fully considered. In the experiments, two kinds of models with different simplification ratios are utilized for the terrain model evaluation. The results confirm that the proposed algorithm has a better performance of terrain model quality assessment than other traditional methods.
Collaborative representation classifier (CRC) has been applied to hyperspectral image classification, which intends to use all the atoms in a dictionary to represent a testing pixel for label assignment. However, some atoms that are very dissimilar to the testing pixel should not participate in the representation, or their contribution should be very little. The regularized version of CRC imposes strong penalty to prevent dissimilar atoms with having large representation coefficients. To utilize spatial information, the weighted sum of local spatial neighbors is considered as a joint spatial-spectral feature, which is actually for regularized CRC-based classification. This paper proposes its kernel version to further improve classification accuracy, which can be higher than those from the traditional support vector machine with composite kernel and the kernel version of sparse representation classifier.
Extreme learning machine (ELM) and kernel ELM (KELM) can offer comparable performance as the standard powerful classifier―support vector machine (SVM), but with much lower computational cost due to extremely simple training step. However, their performance may be sensitive to several parameters, such as the number of hidden neurons. An empirical linear relationship between the number of training samples and the number of hidden neurons is proposed. Such a relationship can be easily estimated with two small training sets and extended to large training sets so as to greatly reduce computational cost. Other parameters, such as the steepness parameter in the sigmodal activation function and regularization parameter in the KELM, are also investigated. The experimental results show that classification performance is sensitive to these parameters; fortunately, simple selections will result in suboptimal performance.
With the improvement of spatial resolution of hyperspectral imagery, it is more reasonable to include spatial information in classification. The resulting spectral-spatial classification outperforms the traditional hyperspectral image classification with spectral information only. Among many spectral-spatial classifiers, support vector machine with composite kernel (SVM-CK) can provide superior performance, with one kernel for spectral information and the other for spatial information. In the original SVM-CK, the spatial information is retrieved by spatial averaging of pixels in a local neighborhood, and used in classifying the central pixel. Obviously, not all the pixels in such a local neighborhood may belong to the same class. Thus, we investigate the performance of Gaussian lowpass filter and an adaptive filter with weights being assigned based on the similarity to the central pixel. The adaptive filter can significantly improve classification accuracy while the Gaussian lowpass filter is less time-consuming and less sensitive to the window size.
An efficient classification framework for mapping agricultural tillage practice using hyperspectral remote sensing imagery is proposed, which has the potential to be implemented practically to provide rapid, accurate, and objective surveying data for precision agricultural management and appraisal from large-scale remote sensing images. It includes a local region filter [i.e., Gaussian low-pass filter (GLF)] to extract spatial-spectral features, a dimensionality reduction process [i.e., local fisher’s discriminate analysis (LFDA)], and the traditional k-nearest neighbor (KNN) classifier, and is denoted as GLF-LFDA-KNN. Compared to our previously used local average filter and adaptive weighted filter, the GLF also considers spatial features in a small neighborhood, but it emphasizes the central pixel itself and is data-independent; therefore, it can achieve the balance between classification accuracy and computational complexity. The KNN classifier has a lower computational complexity compared to the traditional support vector machine (SVM). After classification separability is enhanced by the GLF and LFDA, the less powerful KNN can outperform SVM and the overall computational cost remains lower. The proposed framework can also outperform the SVM with composite kernel (SVM-CK) that uses spatial-spectral features.
Extreme learning machine (ELM) is of great interest to the machine learning society due to its extremely simple training step. Its performance sensitivity to the number of hidden neurons is studied under the context of hyperspectral remote sensing image classification. An empirical linear relationship between the number of training samples and the number of hidden neurons is proposed. Such a relationship can be easily estimated with two small training sets and extended to large training sets to greatly reduce computational cost. The kernel version of ELM (KELM) is also implemented with the radial basis function kernel, and such a linear relationship is still suitable. The experimental results demonstrated that when the number of hidden neurons is appropriate, the performance of ELM may be slightly lower than the linear SVM, but the performance of KELM can be comparable to the kernel version of SVM (KSVM). The computational cost of ELM and KELM is much lower than that of the linear SVM and KSVM, respectively.
In hyperspectral anomaly detection, the dual-window-based detector is a widely used technique that employs two windows to capture nonstationary statistics of anomalies and background. However, its detection performance is usually sensitive to the choice of window sizes and suffers from inappropriate window settings. In this work, a decision-fusion approach is proposed to alleviate such sensitivity by merging the results from multiple detectors with different window sizes. The proposed approach is compared with the classic Reed-Xiaoli (RX) algorithm as well as kernel RX (KRX) using two real hyperspectral data. Experimental results demonstrate that it outperforms the existing detectors, such as RX, KRX, and multiple-window-based RX. The overall detection framework is suitable for parallel computing, which can greatly reduce computational time when processing large-scale remote sensing image data.
New imaging mode has been brought up for collecting multiple scenes in one pass, as is implemented on World View-II.
This greatly helps for acquiring high spatial resolution images that cover urban areas, and is to be adopted in the coming
Chinese satellites. This paper is to discuss the mosaic characteristic and propose a mosaic line generation method by
integrating correlation and the road information. The mosaic line is formed by linking the unique mosaic point on each
line restricted within the road. We position the starting point by connectivity analysis of the road lines, and then locate
the adjacent point along the road with connectivity analysis. A weighed vector, combining correlation and distance to
centre of the road, is used to pick the best point. The points are located on the road unless it is unavoidable, for example,
the road ends or the line touches edge of the image. This method provides instant mosaic line generation for urban areas
with road information available in most cases. By resorting to the road, the mosaic line is more applicable since many
problems for mosaic of high spatial resolution images are solved, for example, tilting of the buildings, the shadows,
motions of the vehicles etc. Experiments have been done with WV-II images and gained favorable results.
Two improved local-region filters, adaptive weighted filter (AWF) and collaborative representation filter (CoRF), are proposed for feature extraction and classification in hyperspectral imagery. The local-region filters generate spatial-spectral features of a hyperspectral pixel by incorporating its surrounding pixels. The work of this paper is an extension of our previously introduced local average filter (LAF). Unlike LAF, which gives the surrounding pixels the same weight, AWF and CoRF explore the internal similarity in the local region with an adaptive weight. More specifically, AWF is set up considering the spatial distance to the central pixel, and CoRF is constructed with spectral similarities adopting the idea of collaborative representation. The two improved local-region filters adaptively extract spectral-spatial features from neighboring pixels and are proven to be effective in many aspects, such as edge information preservation and classification performance, with experiments on two real hyperspectral datasets.
Recently, representation-based classifications have gained increasing interest in hyperspectral imagery, such as the newly proposed sparse-representation classification and nearest-regularized subspace (NRS). These classifiers provide excellent performance that is comparable to or even better than the classic support vector machine. However, all these representation-based methods were originally designed to be pixel-wise classifiers which only consider the spectral signature while ignoring the spatial-contextual information. A Markov random field (MRF), providing a basis for modeling contextual constraints, has currently been successfully applied for hyperspectral image analysis. We mainly investigate the benefits of combining these representation-based classifications with an MRF model in order to acquire better classification results. Two real hyperspectral images are used to validate the proposed classification scheme. Experimental results demonstrated that the proposed method significantly outperforms other state-of-the-art approaches. For example, NRS-MRF performed with an accuracy of 94.92% for the Reflective Optics System Imaging Spectrometer data with 60 training samples per class, while the original NRS obtained an accuracy of 81.95%, an improvement of approximately 13%.
Simultaneous orthogonal matching pursuit (SOMP) has been recently developed for hyperspectral image classification. It utilizes a joint sparsity model with the assumption that each pixel can be represented by a linear combination of labeled samples. We present an approach to improve the performance of SOMP based on a priori segmentation map. According to the map, we first build a local region where within-segment pixels are preserved while between-segment pixels are excluded. Hyperspectral pixels in the preserved region around the test pixel are then simultaneously represented by a linear combination of training samples, whose weights are recovered by solving a sparsity-constrained optimization problem. Finally, the label of the test pixel is determined to be the class that yields the minimal total residuals between the test samples and the approximations. Experimental results demonstrate that the proposed adaptive SOMP (ASOMP) is superior to some existing classifiers, such as the original SOMP and the recently proposed weighted-SOMP (WSOMP). For example, the ASOMP performed with an accuracy of 95.53% for the ROSIS University of Pavia data with 120 training samples per class, while SOMP obtained an accuracy of 87.61%, an improvement of approximately 8%.
In this paper, a feature extraction method using a very simple local averaging filter for hyperspectral image classification is proposed. The method potentially smoothes out trivial variations as well as noise of hyperspectral data, and simultaneously exploits the fact that neighboring pixels tend to belong to the same class with high probability. The spectral-spatial features, which are extracted and fed into a following classifier with locality preserving character in the experimental setup, are compared with other features, such as spectral only and wavelet-features. Simulated results show that the proposed approach facilitates superior discriminant features extraction, thereby yielding significant improvement in hyperspectral image classification performance.
A wavelet-based nearest-regularized-subspace classifier is proposed for noise-robust
hyperspectral image (HSI) classification. The nearest-regularized subspace, coupling the nearest-
subspace classification with a distance-weighted Tikhonov regularization, was designed to
only consider the original spectral bands. Recent research found that the multiscale wavelet features
[e.g., extracted by redundant discrete wavelet transformation (RDWT)] of each hyperspectral
pixel are potentially very useful and less sensitive to noise. An integration of wavelet-based
features and the nearest-regularized-subspace classifier to improve the classification performance
in noisy environments is proposed. Specifically, wealthy noise-robust features provided
by RDWT based on hyperspectral spectrum are employed in a decision-fusion system or as
preprocessing for the nearest-regularized-subspace (NRS) classifier. Improved performance
of the proposed method over the conventional approaches, such as support vector machine,
is shown by testing several HSIs. For example, the NRS classifier performed with an accuracy
of 65.38% for the AVIRIS Indian Pines data with 75 training samples per class under noisy
conditions (signal-to-noise ratio ¼ 36.87 dB), while the wavelet-based classifier can obtain
an accuracy of 71.60%, resulting in an improvement of approximately 6%.