Search

As a powerful and promising statistical signal modeling technique, sparse representation has been widely used in various image processing and analysis fields. For hyperspectral image classification, previous studies have shown the effectiveness of the sparsity-based classification methods. In this paper, we propose a nonlocal weighted joint sparse representation classification (NLW-JSRC) method to improve the hyperspectral image classification result. In the joint sparsity model (JSM), different weights are utilized for different neighboring pixels around the central test pixel. The weight of one specific neighboring pixel is determined by the structural similarity between the neighboring pixel and the central test pixel, which is referred to as a nonlocal weighting scheme. In this paper, the simultaneous orthogonal matching pursuit technique is used to solve the nonlocal weighted joint sparsity model (NLW-JSM). The proposed classification algorithm was tested on three hyperspectral images. The experimental results suggest that the proposed algorithm performs better than the other sparsity-based algorithms and the classical support vector machine hyperspectral classifier. [ABSTRACT FROM PUBLISHER]

Point cloud classification is of great importance to applications of airborne Light Detection And Ranging (LiDAR) data. In recent years, airborne LiDAR has been integrated with various other sensors, e.g., optical imaging sensors, and thus, the fusion of multiple data types for scene classification has become a hot topic. Therefore, this paper proposes a Bayesian network (BN) model that is suitable for airborne point cloud classification fusing multiple data types. Based on an analysis of the characteristics of LiDAR point clouds and aerial images, we first extract the geometric features from the point clouds and the spectral features from the optical images. The optimal BN structure is then trained using an improved mutual-information-based K2 algorithm to obtain the optimal BN classifier for point cloud classification. Experiments demonstrate that the BN classifier can effectively distinguish four types of basic ground objects, including ground, vegetation, trees, and buildings, with a high accuracy of over 90%. Moreover, compared with other classifiers, the proposed BN classifier can achieve the highest overall accuracies, and in particular, the classifier demonstrates its advantage in the classification of ground and low vegetation points. [ABSTRACT FROM PUBLISHER]

In tropical biomes such as the Amazon, the cloud cover is frequent. The use of synthetic aperture radar (SAR) sensor systems is important to monitor and study these biomes because they can acquire data under cloud coverage. In this paper, an advanced land observing system phased array L-band synthetic aperture radar fine-beam dual (ALOS/PALSAR-FBD) image was evaluated for land use and land cover (LULC) classification of an Amazon test site. The features extracted from this image were also evaluated. To perform this task, a method for feature selection, considering the desired classes, was proposed. In order to better understand the applicability of this type of data in Brazilian Government projects (such as DETER-B and TerraClass), the results obtained with SAR images were compared with those from LANDSAT5/TM. The results show that the PALSAR-FBD image and the features selected are not suitable for the discrimination of densely forested classes. They presented, however, a good discrimination among the group of forested and agropastoral classes, as well as among nondensely forested classes (i.e., pastures, bare soil, and new regeneration). Therefore, these data present good applicability for mapping and monitoring of both deforestation and LULC and they can be used in the above mentioned projects. The classification of selected features, with eight classes of interest, achieved an increase of about 133% and 69% in the Kappa index (0.32) and an Overall Accuracy (0.54) regarding the PALSAR-FBD classification (0.136 and 0.322, respectively). This result shows the applicability of the proposed method. It is also expected that the features selected in this paper will improve the classification of similar study sites. [ABSTRACT FROM PUBLISHER]

A network of railway infrastructure is one of the most critical infrastructure assets for supporting the national economy and sustainable mobility. Safe and reliable maintenance of railway infrastructure is critical to ensure that rail systems run safely and punctually. Such maintenance requires regular surveying of railway assets, which typically relies on time-consuming and error-prone labor-centric visual inspection. In this paper, we propose a novel supervised method for automatically classifying electrification assets of railway networks using mobile laser scanning data. A hierarchical Conditional Random Field (CRF) was investigated in order to apply both smoothness constraint and spatial regularities, to improve the classification result made by local supervised classifiers. We use a multi-scale line representation of original data, which implicitly combines object geometry cues and makes computation efficient. Our approach focuses on learning the spatial regularities at multiple representation scales to thoroughly understand the railway electrification scene. The spatial regularities are formulated as relative spatial location in a middle range for different line primitive scales and relative displacement in a full range for the final coarsest line primitive scale. The experiment shows that learnt spatial regularities at full range with multi-scales can outperform the model with spatial regularities at limited local ranges. [ABSTRACT FROM AUTHOR]

The Sentinel-1 mission has finally reached its maturity with the launch of the second Sentinel radar. Among the products delivered by the agency, the ground range detected class is raising more and more interest among users due to its reduced computational demand for information extraction and availability on cloud exploitation platforms, like the Google Earth Engine. In this paper, we present a novel multitemporal processing chain, suitable to be applied to Sentinel-1 ground range detected products to obtain RGB images, using a series of single polarization detected images. These products aim at being the equivalent for the recently introduced Level-1 $\alpha$ , exploiting a texture measure instead of the interferometric coherence, to properly render and enhance the presence of built-up areas. The discussion is supported by experiments showing the reliability of this newly introduced class of products in classic synthetic aperture radar applications like image photointerpretation, flood mapping, and long-term urban area monitoring. [ABSTRACT FROM AUTHOR]

In this paper, we propose a new framework for hyperspectral image classification, namely pseudolabel guided kernel learning (PLKL). The proposed framework is capable of fully utilizing unlabeled samples, making it very effective to handle the task with extremely limited training samples. Specifically, with multiple initial kernels and labeled samples, we first employ support vector machine (SVM) classifiers to predict pseudolabels independently for each unlabeled sample, and consistency voting is applied to the resulting pseudolabels to select and add a few unlabeled samples to the training set. Then, we refine the kernels to improve their discriminability with the augmented training set and a typical kernel learning method. Such phases are repeated until stable. Furthermore, we enhance the PLKL in terms of both the computation and memory efficiencies by using a bagging-like strategy, improving its practicality for large scale datasets. In addition, the proposed framework is quite flexible and general. That is, other advanced kernel-based methods can be incorporated to continuously improve the performance. Experimental results show that the proposed frameworks achieve much higher classification accuracy, compared with state-of-the-art methods. Especially, the classification accuracy improves more than 5% with very few training samples. [ABSTRACT FROM AUTHOR]

Effective methods of sea ice classification are crucial to regional ice type mapping using spaceborne synthetic aperture radar (SAR), especially in ScanSAR mode. However, in TerraSAR-X ScanSAR images of ocean scenes, the scalloping and interscan banding (ISB) artifacts are usually visible. Though these two artifacts could be reduced to a certain degree where residual artifacts do not impede the visual interpretation of SAR images, this is not the case when it comes to sea ice classification. The difficulty mainly comes from the involvement of texture features in sea ice classification. Texture features, especially gray level co-occurrence matrix (GLCM) are useful bases for sea ice classification using SAR data in single or dual polarization. When GLCM is applied to TerraSAR-X ScanSAR data, however, scalloping and ISB artifacts are found to be intensified, further affecting the classification result. How to largely eliminate the effects of scalloping and ISB on sea ice classification has not been studied thoroughly yet. In this paper, an approach combining Kalman filter, GLCM, and support vector machine is proposed. An independent testing shows that this approach is effective at the removal of scalloping and ISB's effects on sea ice classification using TerraSAR-X ScanSAR data, with the overall accuracy of 88.26%. Besides, convolutional neural network is implemented to compare with the proposed approach. [ABSTRACT FROM AUTHOR]

Oil spill detection in synthetic aperture radar (SAR) images plays a crucial role in oil slick monitoring on the sea surface. In addition to oil spills, SAR images incorporate a large number of similar phenomena called “lookalikes”. Choosing a method of discriminating between oil spills and lookalikes with high accuracy is a necessity. In this paper, a new algorithm is proposed for discrimination between oil spills and lookalikes, which consists of the following four main parts. 1) Dark spot detection in SAR images. 2) Shape-based feature extraction of oil spills and lookalikes selected from dark spots in the previous stage. 3) Feature transformation using the NMF algorithm. 4) Classification. To evaluate the performance of the proposed algorithm, it was tested on the several real SAR images. This algorithm was able to provide promising results for both dark spot detection and oil spill classification. The average error rate of dark spot detection for a typical image was 1.31% and it could discriminate between oil spills and lookalikes with an average accuracy of 96.31% using feature transformation modes after transforming the features and classifying them using the K-nearest neighbor classifier. [ABSTRACT FROM AUTHOR]

Classification is one of the most important analysis techniques for hyperspectral image analysis. Sparse representation is an extremely powerful tool for this purpose, but the high computational complexity of sparse representation-based classification techniques limits their application in time-critical scenarios. To improve the efficiency and performance of sparse representation classification techniques for hyperspectral image analysis, this paper develops a new parallel implementation on graphics processing units (GPUs). First, an optimized sparse representation model based on spatial correlation regularization and a spectral fidelity term is introduced. Then, we use this approach as a case study to illustrate the advantages and potential challenges of applying GPU parallel optimization principles to the considered problem. The first GPU optimization algorithm for sparse representation classification (SRCSC_P) of hyperspectral images is proposed in this paper, and a parallel implementation of the proposed method is developed using compute unified device architecture (CUDA) on GPUs. The GPU parallel implementation is compared with the serial and multicore implementations on CPUs. Experimental results based on real hyperspectral datasets show that the average speedup of SRCSC_P is more than \mathbf130 \times, and the proposed approach is able to provide results accurately and fast, which is appealing for computationally efficient hyperspectral data processing. [ABSTRACT FROM PUBLISHER]

This paper proposes a spectral-spatial linear discriminant analysis (LDA) method for the hyperspectral image classification. A natural assumption is that similar samples have similar structure in the dimensionality reduced feature space. The proposed method uses a local scatter matrix from a small neighborhood as a regularizer incorporated into the objective function of LDA. Different from traditional LDA and its variants, our proposed method yields a self-adaptive projection matrix for dimension reduction, which improves the classification accuracy and avoids running out of memory. In order to consider the nonlinear case, this paper generalizes our linear version to its kernel version. Experimental results demonstrate that our proposed methods outperform several dimension reduction algorithms. [ABSTRACT FROM PUBLISHER]

Hyperspectral images usually have high-dimensional and abundant spectral information for land-cover types classification. Research shows that multiple kinds of features would be helpful to the classification task. In this paper, a new feature fusion framework—deep multiple feature fusion (DMFF)—is proposed for the classification of the hyperspectral image. First, several different features are extracted for each pixel. Then, these features are fed to a deep random forest classifier. With a multiple-layer structure, the outputs of preceding layers will be used as the inputs of the subsequent layers. After the final layer, the classification probability will be computed. By introducing the information of neighboring pixels, the spectral–spatial information is combined effectively. Besides, the structure of the DMFF is easy to expand. Experimental results based on two widely used hyperspectral datasets (Indian pines image and Pavia University image) demonstrate that the proposed method can achieve a satisfactory classification performance compared with other multiple feature fusion methods. [ABSTRACT FROM AUTHOR]

In this paper, a feature extraction approach based on a three-dimensional shearlet transform (shearlet 3D) is proposed. We aim at exploiting shearlet 3D to highlight the intrinsic properties of hyperspectral images (HSIs), well known by their correlated information and high dimensionality. First, we decompose the HSI to yield coefficients arranged in cubes that help the computing of statistical parameters. Afterward, using a simultaneous orthogonal matching pursuit (SOMP) algorithm, a classification process is carried out. SOMP relies on the powerful sparse representation paradigm, which helps representing data in a low-dimensional space. It is also built over the assumption that the contextual information incorporation into the sparse recovery problem improves the classification performance. To this algorithm, we propose to add an adapted decision rule where a similarity measurement is calculated to well assign the appropriate labels to pixels of interest. Experimental results proved that our proposed method outperformed state-of-the-art classifiers. Thanks to our proposed approach, we succeeded to build discriminative descriptors reaching high overall accuracies for two different HSI datasets, without taking into account all the shearlet 3D coefficients. [ABSTRACT FROM PUBLISHER]

In this paper, a fast and robust self-representation (FRSR) method is proposed to select a proper band subset from hyperspectral imagery (HSI). The FRSR assumes the separability structure of the HSI band set and transforms the problem of separable nonnegative matrix factorization into the robust self-representation (RSR) model. Then, the FRSR incorporates structured random projections into the RSR model to improve computational efficiency. The solution of FRSR is formulated into optimizing a convex problem and the augmented Lagrangian multipliers are adopted to estimate the proper factorization localizing matrix in the FRSR. The selected band subset is constituted with the bands corresponding to the r largest diagonal entries of the factorization localizing matrix. The experimental results show that FRSR outperforms state-of-the-art techniques in classification accuracy with lower computational cost. [ABSTRACT FROM PUBLISHER]

This paper considers methods to classify and discriminate the multidimensional cyclostationary climatological time series. The methods take into account both the seasonal mean cycles and random behavior in the series, simultaneously considering means and all component-by-component autocovariances. This improves classical Hotelling T^2 statistics that classify through mean changes only, and constant-mean speech methods that classify exclusively through sample autocovariance differences. Here, two series are compared by assuming that both follow the same time series model; from this, a test statistic representing a distance between the two series is developed from linear prediction theory. This construction generates a level-\alpha test statistic for Gaussian data that can be used to assess how different the two series are. The derived distances can be used in a clustering algorithm, e.g., to group series with similar behavior. Such information is useful to eliminate or merge the climate stations whose data are redundant to another station or to optimally locate a collection of stations. The techniques are first tested on simulated series with known structures, and then applied to 11 two-dimensional series in the National Oceanic and Atmospheric Administration’s data buoy catalog. Natural clusters emerge which are geographically realistic. Specifically, the methods were able to perfectly group stations in the Gulf of Mexico, the Carolina Coast, the Pacific Ocean, and offshore New England. [ABSTRACT FROM PUBLISHER]

The twenty-seven articles in this special issue is a follow-up to special sessions organized at WHISPERS conferences. Such sessions drew unexpectedly large attendance, signaling the interest and need for a focused platform to exchange knowledge at the intersection of machine learning and remote sensing. The collection of papers in this special issue presents a comprehensive sample of the latest trends in the design of machine learning algorithms for geospatial data. It covers a wide spectrum of remote sensing applications and presents new solutions to answer to the call of the new generation of sensors, covering the electromagnetic range from optical to microwave data. [ABSTRACT FROM PUBLISHER]

Recent advances in sensor design allow us to gather more useful information about the Earth's surface. Examples are hyperspectral (HS) and Light Detection And Ranging (LiDAR) sensors. These, however, have limitations. HS data cannot distinguish different objects made from similar materials and highly suffers from cloud-shadow regions, whereas LiDAR cannot separate distinct objects that are at the same altitude. For an increased classification performance, fusion of HS and LiDAR data recently attracted interest but remains challenging. In particular, these methods suffer from a poor performance in cloud-shadow regions because of the lack of correspondence with shadow-free regions and insufficient training data. In this paper, we propose a new framework to fuse HS and LiDAR data for the classification of remote sensing scenes mixed with cloud-shadow. We process the cloud-shadow and shadow-free regions separately, our main contribution is the development of a novel method to generate reliable training samples in the cloud-shadow regions. Classification is performed separately in the shadow-free (classifier is trained by the available training samples) and cloud-shadow regions (classifier is trained by our generated training samples) by integrating spectral (i.e., original HS image), spatial (morphological features computed on HS image) and elevation (morphological features computed on LiDAR) features. The final classification map is obtained by fusing the results of the shadow-free and cloud-shadow regions. Experimental results on a real HS and LiDAR dataset demonstrate the effectiveness of the proposed method, as the proposed framework improves the overall classification accuracy with 4% for whole scene and 10% for shadow-free regions over the other methods. [ABSTRACT FROM PUBLISHER]

The bag of visual words (BOW) model has attracted a great deal of attention in high-resolution remote sensing image classification. Related studies based on polarimetric synthetic aperture radar (PolSAR) image classification are also still challenging. This paper presents a new method of PolSAR image classification based on the BOW model and supplemented by an affinity propagation clustering algorithm (AP), as well as a spatial pyramid matching (SPM) technique (AP_SPM_BOW). Its classification effect is analyzed compared to those of BOW models with other clustering algorithms (e.g., K-means, meanshift, Gaussian mixture model), as well as the ML classification method by using two types of PolSAR images with different resolutions collected from EMISAR and RADARSAT-2 sensors. These results demonstrate that the proposed AP_SPM_BOW model is not only a stable and effective tool for PolSAR image classification but also more suitable for the analysis of high-resolution PolSAR images. In addition, the use of polarimetric information in BOW models can effectively improve the classification accuracy. [ABSTRACT FROM PUBLISHER]

Random Forest (RF) is a widely used classifier to show a good performance of hyperspectral data classification. However, such performance could be improved by increasing the diversity that characterizes the ensemble architecture. In this paper, we propose a novel ensemble approach, namely rotation random forest via kernel principal component analysis (RoRF-KPCA). In particular, the original feature space is first randomly split into several subsets, and KPCA is performed on each subset to extract high order statistics. The obtained feature sets are merged and used as input to an RF classifier. Finally, the results achieved at each step are fused by a majority vote. Experimental analysis is conducted using real hyperspectral remote sensing images to evaluate the performance of the proposed method in comparison with RF, rotation forest, support vector machines, and RoRF-PCA. The obtained results demonstrate the effectiveness of the proposed method. [ABSTRACT FROM PUBLISHER]

Although hyperspectral data have very high dimensionality, major information tends to occupy a low-rank subspace and outliers are often found in a sparse matrix. However, due to the mixed nature of hyperspectral data, the underlying data structure may include multiple subspaces instead of a single subspace. In this paper, we propose to use low-rank subspace representation (LRSR) as a preprocessing step for classification in both supervised and unsupervised fashion. In supervised classification, LRSR is shown to improve the performance of various classifiers. In unsupervised classification, both K-means clustering and spectral clustering can be applied on the low-rank matrix to improve the performance. Experimental results demonstrate that the proposed method can increase classification accuracy, particularly for complicated image scenes, and outperform the often-used low-rank representation approach. [ABSTRACT FROM PUBLISHER]

In order to tackle two key issues, e.g., high-accuracy classifier and spatial information model, in the classification of hyperspectral images, a new classification scheme, which combines two powerful techniques, i.e., rotation forest (RoF) and multiscale (MS) segmentation, is proposed in this paper. MS segmentation is used to obtain the spatial information from different levels. Then, the objects produced by MS segmentation are treated as the input of the RoF classifier. Furthermore, multiple classification results generated by the RoF classifiers and MS segmentation are combined using a majority voting rule to generate the final result. Experimental results on two real hyperspectral datasets demonstrate that the proposed method performs particularly well in terms of overall and class-specific accuracies and generates the classification map with much more homogeneous regions than traditional methods. Moreover, the impacts of parameters on the classification performances are also analyzed. [ABSTRACT FROM PUBLISHER]

We propose a semisupervised graph learning (SEGL) method for feature extraction of hyperspectral remote sensing imagery in this paper. The proposed SEGL method aims to build a semisupervised graph that can maximize the class discrimination and preserve the local neighborhood information by combining labeled and unlabeled samples. In our semisupervised graph, we connect labeled samples according to their label information and unlabeled samples by their nearest neighborhood information. By sorting the mean distance between a unlabeled sample and labeled samples of each class, we connect the unlabeled sample with all labeled samples belonging to its nearest neighborhood class. Moreover, the proposed SEGL better models the actual differences and similarities between samples, by setting different weights to the edges of connected samples. Experimental results on four real hyperspectral images (HSIs) demonstrate the advantages of our method compared to some related feature extraction methods. [ABSTRACT FROM PUBLISHER]

The main task of environmental and geoscience applications is efficient and accurate quantitative classification of earth surfaces and spatial phenomena. In the past decade, there has been a significant interest in employing hyperspectral unmixing (HU) to retrieve accurate quantitative information latent in hyperspectral imagery data. Recently, the ground-truth and laboratory measured spectral signatures promoted by advanced algorithms are proposed as a new path toward solving the unmixing problem of hyperspectral imagery in semisupervised fashion. This paper suggests that the sensitivity of sparse unmixing techniques provides an ideal approach to extract and identify dust settled over/upon green vegetation canopy using hyperspectral airborne data. Among the available techniques, this study presents the results of seven selected algorithms: 1) non-negative matrix factorization (NMF); 2) \textL1 sparsity-constrained NMF (\textL_1\\text{NMF}); 3) \textL1/2 sparsity-constrained NMF (\textL1/2\\text{NMF}); 4) graph regularized NMF (\textG\\text{NMF}); 5) structured sparse NMF (\textSS\\text{NMF}); 6) alternating least-square (ALS); and 7) Lin’s projected gradient (LPG). The performance is evaluated on real hyperspectral imagery data via detailed experimental assessment. The results compared with performances of selected conventional unmixing techniques. [ABSTRACT FROM PUBLISHER]

Hyperspectral imagery contains a wealth of spectral and spatial information that can improve target detection and recognition performance. Conventional feature extraction methods cannot fully exploit both spectral and spatial information. Data fusion by simply stacking different feature sources together does not take into account the differences between feature sources. In this paper, a local graph-based fusion (LGF) method is proposed to couple dimension reduction and feature fusion of the spectral information (i.e., the spectra in the HS image) and the spatial information [extracted by morphological profiles (MPs)]. In the proposed method, the fusion graph is built on the full data by moving a sliding window from the first pixel to the last one. This yields a clear improvement over a previous approach with fusion graph built on randomly selected samples. Experimental results on real hyperspectral images are very encouraging. Compared to the methods using only single feature and stacking all the features together, the proposed LGF method improves the overall classification accuracy on one of the data sets for more than 20% and 5%, respectively. [ABSTRACT FROM PUBLISHER]

Copyright protection of multispectral images and its integrity become one of the key problems in spatial information service due to rapid development of communication network. This paper depicts a crypto-watermarking scheme by combining watermarking and encryption to protect copyright of multispectral images and to provide security to the watermarked image at dissemination level. We have proposed wavelet-based watermarking and multiplicative-transposition-based cipher for encryption. Experimental results show that the proposed algorithm has not only good robustness against encryption attacks, but also has transparency, strongness, large data hiding capacity, correct extraction of watermark, and strong robustness against JPEG lossy compression, filtering, and noise. Moreover, the classification of original and watermarked data is performed and it is observed that the proposed watermarking scheme leads to less misclassification and it meets all the requirements for copyright protection as well as the security of multispectral images in an effective manner. [ABSTRACT FROM PUBLISHER]

In this paper, a Capsulenet-based framework is proposed for extracting spectral and spatial features for improving hyperspectral image classification. Unlike conventional strategies, the proposed framework simultaneously optimizes both feature extraction and classification. The spectral features/patterns derived at different levels of hierarchies are remodeled as spectral-feature capsules. Consequently, unlike conventional convolutional neural network-based approaches, the relative locations as well as other properties such as depth, width, and position of the spectral patterns are taken into consideration. In addition to learning spectral features/patterns, a convolutional long short-term memory (conv-LSTM) is employed for sequentially integrating the spatial features learned from each band. The integrated spatial-feature representation, thus obtained from the final hidden state of conv-LSTM, forms spatial-feature capsules. The capsule-level integration of spatial and spectral features/patterns yields better convergence and accuracy as compared to both ensemble-based and kernel-level integrations. Along with the margin loss, a spectral-angle-based reconstruction loss is also minimized to regularize the learning of network weights. Experiments over different standard datasets indicate that the proposed approach performs better than other prominent hyperspectral classifiers. Furthermore, in comparison with the recent deep learning models, our approach is found to be less sensitive to the network parameters and achieves better accuracy even with lesser network depth. [ABSTRACT FROM AUTHOR]

The high resolution of the hyperspectral remote sensing images available allows the detailed analysis of even small spatial structures. As a consequence, the study of techniques to efficiently extract spatial information is a very active realm. In this paper, we propose a novel denoising wavelet-based profile for the extraction of spatial information that does not require parameters fixed by the user. Over each band obtained by a wavelet-based feature extraction technique, a denoising profile (DP) is built through the recursive application of discrete wavelet transforms followed by a thresholding process. Each component of the DP consists of features reconstructed by recursively applying inverse wavelet transforms to the thresholded coefficients. Several thresholding methods are explored. In order to show the effectiveness of the extended DP (EDP), we propose a classification scheme based on the computation of the EDP and supervised classification by extreme learning machine. The obtained results are compared to other state-of-the-art methods based on profiles in the literature. An additional study of behavior in the presence of added noise is also performed showing the high reliability of the EDP proposed. [ABSTRACT FROM AUTHOR]

Support vector machine (SVM) is considered as one of the most powerful classifiers for hyperspectral remote sensing images. However, it has high computational cost. In this paper, we propose a novel two-level parallel computing framework to accelerate the SVM-based classification by utilizing CUDA and OpenMP. For a binary SVM classifier, the kernel function is optimized on GPU, and then a second-order working set selection (WSS) procedure is employed and optimized especially for GPU to reduce the cost of communication between GPU and host. In addition to the parallel binary SVM classifier on GPU as data-processing level parallelization, a multiclass SVM is addressed by a “one-against-one” approach in OpenMP, and several binary SVM classifiers are run simultaneously to conduct task-level parallelization. The experimental results show that the solver in this framework offered a speedup of $18.5\times$ over the popular LIBSVM software in the training process for data with 200 bands, 13 classes, and 95 597 training samples, and $81.9\times$ in the testing process for data with 103 bands, 9 classes, 1892 support vectors (SVs), and 42 776 testing samples. [ABSTRACT FROM PUBLISHER]

Feature extraction and image classification using polarimetric synthetic aperture radar (PolSAR) images are currently of great interest in SAR applications. Generally, PolSAR image classification is a high-dimensional nonlinear mapping problem. Sparse representation-based techniques have shown great potential for pattern recognition problems. Therefore, on the basis of the sparse characteristics of the features for PolSAR image classification, a supervised PolSAR image classification method based on sparse representation is proposed in this paper. First, the effective features are extracted to describe the distinction of each class. Then, the feature vectors of the training samples construct an over-complete dictionary and obtain the corresponding sparse coefficients; meanwhile, the residual error of the pending pixel with respect to each atom is evaluated and considered as the criteria for classification, and the ultimate class results can be obtained according to the atoms with the least residual error. In addition, a Simplified Matching Pursuit (SMP) algorithm is proposed to solve the optimization problem of sparse representation of PolSAR images. The verification tests are implemented using Danish EMISAR L-band fully polarimetric SAR data of Foulum area, Denmark. The preliminary experimental results confirm that the proposed method outputs an excellent result and moreover the classification process is simpler and less time consuming. [ABSTRACT FROM PUBLISHER]

This paper proposes a new approach of band selection for classifying multiple objects in hyperspectral images. Different from traditional algorithms, we construct a semisupervised pair-wise band selection (PWBS) framework for this task, in which an individual band selection process is performed only for each pair of classes. First, the statistical parameters for spectral features of each class, including mean vectors and covariance matrices, are estimated by an expectation maximization approach in a semisupervised learning setting, where both labeled and unlabeled samples are employed for better performance. For each pair of classes, based on the estimated statistical parameters, Bhattacharyya distances between the two classes are calculated to evaluate all possible subsets of bands for classification. Second, as our proposed semisupervised framework, the PWBS followed by a binary classifier can be embedded into the semisupervised expectation maximization process to obtain posterior probabilities of samples on the selected bands. Finally, to evaluate the selected bands, all of the binary decisions obtained with multiple binary classifiers are finally fused together. Comparative experimental results demonstrate the validity of our proposed algorithm. The experimental results also prove that our band selection algorithm can perform well when the training set is very small. [ABSTRACT FROM PUBLISHER]

This paper presents a novel and efficient spectral–spatial classification method for hyperspectral images. It combines the spectral and texture features to improve the classification accuracy. The moment invariants are computed within a small window centered at the pixel to determine pixel-wise texture features. The texture and spectral features are concatenated to form a joint feature vector that is used for classification with support vector machine (SVM). The experiments are carried out on three hyperspectral datasets and results are compared with some other spectral–spatial techniques. The results indicate that the proposed method statistically significantly improved the classification accuracies over the conventional spectral method. The new method has also outperformed the other recently used spectral–spatial methods in terms of both classification accuracies and computational cost. The results also showed that the proposed method can produce good classification accuracy with smaller training sets. [ABSTRACT FROM PUBLISHER]

Sparse representation exhibits good performance in various image processing and has been applied to hyperspectral image (HSI) classification by many researchers. Recently, several new spatial–spectral strategies combined with sparse representation have been proposed to improve classification performance. However, these new strategies rely on spectral reflectance information and its neighborhood, without considering other spectral properties and higher order context information. Thus, in this paper, we present a spatial–spectral derivative-aided kernel joint sparse representation (KJSR-SSDK) for HSI classification. The proposed algorithm includes three novelties: 1) it considers the derivative features of the spectral as well as the original spectral feature; 2) it incorporates higher order spatial context and distinct spectral information; and 3) the l1,2 mix-norm regularization is imposed on the coefficients of spatial–spectral derivative-aided dictionary for KJSR. Based on the rich experimental comparison with the related state-of-the-art algorithms, the effectiveness of the proposed KJSR-SSDK has been confirmed. [ABSTRACT FROM PUBLISHER]

The development of very high spatial resolution remote sensing sensors opens a new era for mapping the earth with submeter level of detail, whereas the increased resolution brings about difficulties for the land-cover classification in terms of intra-class variability and inter-class similarity. This paper presents a novel spatial feature, mean shift (MS) vector-based shape feature (MSVSF), to improve the classification accuracy of very high resolution (VHR) remote sensing imagery. MSVSF is a \bf 3 \times \bf 1 feature vector extracted in per-pixel fashion. It describes the shape of a spectrally homogeneous area surrounding each pixel by measuring the two-dimensional (2-D) image deformation of its local area imposed by the MS vector. The proposed feature is particularly effective to discriminate objects with similar spectral response but different 2-D shapes, such as buildings and roads. Independent component analysis is adopted to extract spectral features and Support Vector Machine (SVM) classifier is adopted to classify the spectral and spatial features and several state-of-the-art spatial/structural features are compared to the proposed feature. A synthetic experiment demonstrates that the proposed feature has good capability to describe 2-D shapes with different scale, two real dataset experiments on QuickBird and IKONOS images show MSVSF has achieved better overall accuracy (OA) than the compared ones. In addition, the MSVSF feature is extended to the object-based classification (OBC), and the result shows that the MSVSF is effective to improve the classification accuracy on high resolution images of the urban area. [ABSTRACT FROM PUBLISHER]

An approach to sea ice classification using dual polarization RADARSAT-2 ScanSAR data is presented in this paper. It is based on support vector machine (SVM). In addition to backscatter coefficients and gray-level cooccurrence matrix (GLCM) texture features, sea ice concentration was introduced as a classification basis. To better analyze the backscatter information of sea ice types, we considered two steps that could improve the ScanSAR image quality, the noise floor stripe reduction and the incidence angle normalization. Then, effective GLCM texture characteristics from both polarizations were selected using the proper parameters. The third type of information, sea ice concentration, was extracted from the initial SVM classification result after the optimal SVM model was achieved from the training. The final result was generated by implementing the SVM twice and the decision tree once. Using this method, the classification was improved in two aspects, both of which were related to sea ice concentration. The results showed that the sea ice concentration parameter was effective in dealing with open water and in discriminating pancake ice from old ice. Finally, the maximum likelihood (ML) was run as a comparative test. In conclusion, the sea ice concentration parameter could play a role in SVM classification, and the whole process provided an effective way to classify sea ice using dual polarization ScanSAR data. [ABSTRACT FROM PUBLISHER]

Currently, the amount of collected Earth Observation (EO) data is increasing considerably with a rate of several Terabytes of data per day. As a consequence of this increasing data volume, new concepts for exploration and information retrieval are urgently needed. To this end, we propose to explore satellite image data via an image information mining (IIM) approach in which the main steps are feature extraction, classification, semantic annotation, and interactive query processing. This leads to a new process chain and a robust taxonomy for the retrieved categories capitalizing on human interaction and judgment. We concentrated on land cover categories that can be retrieved from high-resolution synthetic aperture radar (SAR) images of the spaceborne TerraSAR-X instrument, where we annotated different urban areas all over the world and defined a taxonomy element for each prevailing surface cover category. The annotation resulted from a test dataset comprising more than 100 scenes covering diverse areas of Africa, Asia, Europe, the Middle East, and North and South America. The scenes were grouped into several collections with similar source areas and each collection was processed separately in order to discern regional characteristics. In the first processing step, each scene was tiled into patches. Then the features were extracted from each patch by a Gabor filter bank and a support vector machine with relevance feedback classifying the feature sets into user-oriented land cover categories. Finally, the categories were semantically annotated using Google Earth for ground truthing. The annotation followed a multilevel approach that allowed the fusion of information being visible on different resolution levels. The novelty of this paper lies in the fact that a semantic annotation was performed with a large number of high-resolution radar images that allowed the definition of more than 850 surface cover categories. This opens the way toward an automated identification and classification of urban areas, infrastructure (e.g., airports), geographic objects (e.g., mountains), industrial installations, military compounds, vegetation, and agriculture. Applications that may result from this work can be a semantic catalog of urban images to be used in crisis situations or after a disaster. In addition, the proposed taxonomies can become a basis for building a semantic catalog of satellite images. Finally, we defined four powerful types of high-level queries. Querying on such high levels provides new opportunities for users to search an image database for specific parameters or semantic relationships. [ABSTRACT FROM PUBLISHER]

Multi-source data, either from different sensors or disparate features extracted from the same sensor, are valuable for geospatial image analysis due to their potential for providing complementary features. In this paper, a composite-kernel-based feature extraction method is proposed for multi-source remote sensing data classification. Features from different sources are first fused via a weighted composite kernel mapping, and then projected to a lower-dimensional subspace in which kernel local Fisher discriminant analysis (KLFDA) is used to extract the most discriminative information. We hypothesize that after such a projection, multi-source data would have better class separability between classes, and an efficient linear classification model–multinomial logistic regression (MLR) would be suitable for classification. The efficacy of the proposed method is demonstrated via experiments using two different sets of multi-source geospatial data. For feature fusion, the raw spectral data and extended multi-attribute profiles (EMAPs) derived from the hyperspectral image are used as a testbed for multi-source image analysis. The second multi-source testbed used for validation involves sensor fusion, in which the hyperspectral and light detection and ranging (LiDAR) data are utilized. Experimental results show that composite kernel local Fisher’s discriminant analysis when combined with MLR based classifier (CKLFDA-MLR) is very effective at feature extraction and classification of multi-source geospatial images. [ABSTRACT FROM PUBLISHER]

In this paper, a novel technique known as the Fisher criterion-based nearest feature space (FCNFS) approach is proposed for supervised classification of multisource images for the purpose of landslide hazard assessment. The method is developed for land cover classification based upon the fusion of remotely sensed images of the same scene collected from multiple sources. This paper presents a framework for data fusion of multisource remotely sensed images, consisting of two approaches: 1) the band generation process(BGP); and 2) the FCNFS classifier. We propose the BGP to create a new set of additional bands that are specifically accommodated to the landslide class and are extracted from the original multisource images. In comparison to the original nearest feature space (NFS) method, the proposed improved FCNFS classifier uses the Fisher criterion of between-class and within-class discrimination to enhance the classifier. In the training phase, the labeled samples are discriminated by the Fisher criterion, which can be treated as a preprocessing step of the NFS method. After completion of the training, the classification results can be obtained from the NFS algorithm. In order for the proposed FCNFS to be effective for multispectral images, a multiple adaptive BGP is introduced to create an additional set of bands specially accommodated to landslide classes. Experimental results show that the proposed BGP/FCNFS framework is suitable for land cover classification in Earth remote sensing and improves the classification accuracy compared to conventional classifiers. [ABSTRACT FROM PUBLISHER]

Existing polarimetric synthetic aperture radar (PolSAR) image classification methods cannot achieve satisfactory performance on complex scenes characterized by several types of land cover with significant levels of noise or similar scattering properties across land cover types. Hence, we propose a supervised classification method aimed at constructing a classifier based on self-paced learning (SPL). SPL has been demonstrated to be effective at dealing with complex data while providing classifier performance improvement. In this paper, a novel support vector machine (SVM) algorithm based on SPL with neighborhood constraints (SVM_SPLNC) is proposed. The proposed method leverages the easiest samples first to obtain an initial parameter vector. Then, more complex samples are gradually incorporated to update the parameter vector iteratively. Moreover, neighborhood constraints are introduced during the training process to further improve performance. Experimental results on three real PolSAR images show that the proposed method performs well on complex scenes. [ABSTRACT FROM AUTHOR]

The digital mapping of road environment is an important task for road infrastructure inventory and urban planning. Automatic extraction and classification of pole-like objects can remarkably reduce mapping cost and enhance work efficiency. Therefore, this paper proposes a voxel-based method that automatically extracts and classifies three-dimensional (3-D) pole-like objects by analyzing the spatial characteristics of objects. First, a voxel-based shape recognition is conducted to generate a set of pole-like object candidates. Second, according to their isolation and vertical continuity, the pole-like objects are detected and individualized using the proposed circular model with an adaptive radius and the vertical region growing algorithm. Finally, several semantic rules, consisting of shape features and spatial topological relationships, are derived for further classifying the extracted pole-like objects into four categories (i.e., lamp posts, utility poles, tree trunks, and others). The proposed method was evaluated using three datasets from mobile LiDAR point cloud data. The experimental results demonstrate that the proposed method efficiently extracted the pole-like objects from the three datasets, with extraction rates of 85.3%, 94.1%, and 92.3%. Moreover, the proposed method can achieve robust classification results, especially for classifying tree trunks. [ABSTRACT FROM AUTHOR]

This paper aims to assess the potential of radar data combined with optical data to support local administrations in the knowledge of the land use and land cover at regional scale. The work starts from the actual available thematic maps owned by two different regional administrations in Italy to assess at what extent they can be improved or reproduced by Earth Observation data. In particular, the contribution of data available in the future through the Sistema Italo-Argentino di Satelliti per la Gestione delle Emergenze (SIASGE) project, combining L-band and X-band radar imagery, is assessed in order to produce thematic maps of the regions. Moreover, the further contribution brought by C-band and especially by optical bands has been evaluated. The classification problem is driven by the legend of already existing maps and quality checked against the same maps in order to tackle the real needs of the land managing authorities. As the combination of data from optical imagery is fundamental to achieve good thematic accuracy, the work has exploited the support vector machine (SVM) learning technique, which is more suitable than standard statistical parametric approaches in this respect. Concerning the classification steps, some algorithmic issues have been faced to improve the results, such as training set selection strategy and data fusion techniques. The work has proved that the multisource dataset (radar and optical) is fairly suitable to produce thematic maps comparable to what is already in use at local administrative level, allowing one to achieve classification accuracy in the order of 90%. [ABSTRACT FROM PUBLISHER]

In this paper, we have applied supervised probabilistic principal component analysis (SPPCA) and semi-supervised probabilistic principal component analysis (S^2PPCA) for feature extraction in hyperspectral remote sensing imagery. The two models are all based on probabilistic principal component analysis (PPCA) using EM learning algorithm. SPPCA only relies on the labeled samples into the projection phase, while S^2PPCA is able to incorporate both the labeled and unlabeled information. Experimental results on three real hyperspectral images demonstrate the SPPCA and S^2PPCA outperform some conventional feature extraction methods for classifying hyperspectral remote sensing image with low computational complexity. [ABSTRACT FROM PUBLISHER]

This paper presents a probability-based improved binary encoding algorithm (PIBE) for classification of hyperspectral imagery. In the proposed PIBE method, the spectral, texture and shape information from hyperspectral images as well as height information from digital elevation models (if available) are combined to form a binary code. Based on this, a probability-based approach is further introduced to match the constructed binary code to the corresponding one obtain from target classes (or training data set). Some experiments on a pair of hyperspectral images confirm the effectiveness of the proposed PIBE method. [ABSTRACT FROM PUBLISHER]

In this paper, a new denoising method for hyperspectral images is proposed using First Order Roughness Penalty (FORP). FORP is applied in the wavelet domain to exploit the Multi-Resolution Analysis (MRA) property of wavelets. Stein's Unbiased Risk Estimator (SURE) is used to choose the tuning parameters automatically. The simulation results show that the penalized least squares using FORP can improve the Signal to Noise Ratio (SNR) compared to other denoising methods. The proposed method is also applied to a corrupted hyperspectral data set and it is shown that certain classification indices improve significantly. [ABSTRACT FROM PUBLISHER]

Various sparsity-based methods have been widely used in hyperspectral image (HSI) classification. To determine the class label of a test sample, traditional sparsity-based frameworks mainly use the sparse vectors to compute the residual error for classification. In this paper, a novel sparsity-based framework is proposed, which adopts the max pooling operation for HSI classification. Compared with the traditional sparsity-based frameworks using residual error, sparse vectors in our proposed framework are utilized to generate the feature vectors using max pooling operation. Experimental results demonstrate that our proposed framework can achieve the state-of-the-art classification performance. [ABSTRACT FROM PUBLISHER]

Semisupervised learning has shown its great potential in land cover mapping. It exploits the information of unlabeled training samples and converts those samples to labeled training samples to enhance classification. In this paper, the spatial information extracted by a two-dimensional (2-D) Gabor filter was stacked with spectral information first, and then the spatial neighborhood information of labeled training samples was combined with active learning (AL) algorithm to select the most useful and informative samples, which were used as the unlabeled set to aid the probability model-based supervised support vector machine (SVM). Experiments on two hyperspectral datasets showed that the spatial-spectral information-based semisupervised classification algorithm (\bfS^\bf2 \bfISC) can produce high classification accuracy with a small number of labeled samples, and outperformed the compared semisupervised algorithms. [ABSTRACT FROM PUBLISHER]

Extreme learning machine (ELM) has attracted attentions in pattern recognition field due to its remarkable advantages such as fast operation, straightforward solution, and strong generalization. However, the performance of ELM for high-dimensional data, such as hyperspectral image, is still an open problem. Therefore, in this paper, we introduce ELM for hyperspectral image classification. Furthermore, in order to overcome the drawbacks of ELM caused by the randomness of input weights and bias, two new algorithms of ensemble extreme learning machines (Bagging-based and AdaBoost-based ELMs) are proposed for the classification task. In order to illustrate the performance of the proposed algorithms, support vector machines (SVMs) are used for evaluation and comparison. Experimental results with real hyperspectral images collected by reflective optics spectrographic image system (ROSIS) and airborne visible/infrared imaging spectrometer (AVIRIS) indicate that the proposed ensemble algorithms produce excellent classification performance in different scenarios with respect to spectral and spectral–spatial feature sets. [ABSTRACT FROM PUBLISHER]

Hyperspectral data analysis has been given a growing attention due to the scientific challenges it raises and the wide set of applications that can benefit from it. Classification of hyperspectral images has been identified as one of the hottest topics in this context, and has been mainly addressed by discriminative methods such as SVM. In this paper, we argue that generative methods, and especially those based on manifold representation of classes in the hyperspectral space, are relevant alternatives to SVM. To illustrate our point, we focus on the recently published PerTurbo algorithm and benchmark against SVM this generative manifold learning algorithm in the context of hyperspectral image classification. This choice is motivated by the fact that PerTurbo is fitted with numerous interesting properties, such as 1) low sensitivity to dimensionality curse, 2) high accuracy in weakly labelled images classification context (few training samples), 3) straightforward extension to on-line setting, and 4) interpretability for the practitioner. The promising results call for an up-to-date interest toward generative algorithms for hyperspectral image classification. [ABSTRACT FROM PUBLISHER]

In this paper, we propose a simple unsupervised framework to effectively select and combine spectral information and spatial features for Support Vector Machine (SVM)-based classification when spatial features are the widely used morphological profiles (MPs). To overcome the difficulty of high dimensionality of resulting features, it is a common practice that MPs are extracted from principal components (PCs). In this paper, we investigate another technique on spectral feature selection, which is unsupervised band selection (BS). We find out that using selected bands as spectral features can improve classification performance because they contain more critical characteristics for classification; in particular, using the selected bands, combined with the MPs extracted from PCs, can yield the highest accuracy, due to the fact that major PCs contain less noise for extracting more reliable MPs. The overall unsupervised nature of feature selection provides the flexibility of implementation. We believe that such finding is instructive to feature selection and extraction for spectral/spatial-based hyperspectral image classification. [ABSTRACT FROM PUBLISHER]

This paper presents a fuzzy local double neighborhood information ${\text{c}}$ -means (FLDNICM) clustering algorithm for remotely sensed imagery classification, which incorporates flexible and accurate local spatial and spectral information. First, a trade-off weighted fuzzy factor is established based on a pixel spatial attraction model that considers spatial distance and class membership differences between the central pixel and its neighbor simultaneously. This factor can adaptively and accurately estimate the spatial constraints from neighboring pixels. To further enhance robustness to noise and outliers, another fuzzy prior probability function is also defined, which integrates the mutual dependence information from a pixel and its neighbor in a fuzzy logical way for obtaining accurate spatial contextual information. The FLDNICM enhances the conventional fuzzy ${\text{c}}$ -means algorithm by producing homogeneous segmentation while reducing the edge blurring artifacts. The new trade-off weighted fuzzy factor and prior probability function are both parameter free and fully adaptive to the image content. Experimental results demonstrate the superiority of FLDNICM over competing methodologies, considering a series of synthetic and real-world images classification applications. [ABSTRACT FROM AUTHOR]

Airborne light detection and ranging (lidar) data are widely used for high-resolution land cover mapping. The lidar elevation data are typically used as complementary information to passive multispectral or hyperspectral imagery to enable higher land cover classification accuracy. In this paper, we examine the capabilities of a recently developed multispectral airborne laser scanner, manufactured by Teledyne Optech, for direct classification of multispectral point clouds into ten land cover classes including grass, trees, two classes of soil, four classes of pavement, and two classes of buildings. The scanner, Titan MW, collects point clouds at three different laser wavelengths simultaneously, opening the door to new possibilities in land cover classification using only lidar data. We show that the recorded intensities of laser returns together with spatial metrics calculated from the three-dimensional (3D) locations of laser returns are sufficient for classifying the point cloud into ten distinct land cover classes. Our classification methods achieved an overall accuracy of 94.7% with a kappa coefficient of 0.94 using the support vector machine (SVM) method to classify single-return points and an overall accuracy of 79.7% and kappa coefficient of 0.77 using a rule-based classifier on multireturn points. A land cover map is then generated from the classified point cloud. We show that our results outperform the common approach of rasterizing the point cloud prior to classification by ∼4% in overall accuracy, 0.04 in kappa coefficient, and by up to 16% in commission and omission errors. This improvement however comes at the price of increased complexity and computational burden. [ABSTRACT FROM AUTHOR]

The classification of urban areas in polarimetric synthetic aperture radar (PolSAR) data is a challenging task. Moreover, urban structures oriented away from the radar line of sight pose an additional complexity in the classification process. The characterization of such areas is important for disaster relief and urban sprawl monitoring applications. In this paper, a novel technique based on deep learning is proposed, which leverages a synthetic target database for data augmentation. The PolSAR dataset is rotated by uniform steps and collated to form a reference database. A stacked autoencoder network is used to transform the information in the augmented dataset into a compact representation. This significantly improves the generalization capabilities of the network. Finally, the classification is performed by a multilayer perceptron network. The modular architecture allows for easy optimization of the hyperparameters. The synthetic target database is created and the classification performance is evaluated on an L-band airborne UAVSAR dataset and L-band space-borne ALOS-2 dataset acquired over San Francisco, USA. The proposed technique shows an overall accuracy of 91.3\% . An improvement over state-of-the-art techniques is achieved, especially in urban areas rotated away from the radar line of sight. [ABSTRACT FROM PUBLISHER]