Search

Deep neural networks are currently trained under data-parallel setups on high-performance computing (HPC) platforms, so that a replica of the full model is charged to each computational resource using non-overlapped subsets known as batches. Replicas combine the computed gradients to update their local copies at the end of each batch. However, differences in performance of resources assigned to replicas in current heterogeneous platforms induce waiting times when synchronously combining gradients, leading to an overall performance degradation. Albeit asynchronous communication of gradients has been proposed as an alternative, it suffers from the so-called staleness problem. This is due to the fact that the training in each replica is computed using a stale version of the parameters, which negatively impacts the accuracy of the resulting model. In this work, we study the application of well-known HPC static load balancing techniques to the distributed training of deep models. Our approach is assigning a different batch size to each replica, proportional to its relative computing capacity, hence minimizing the staleness problem. Our experimental results (obtained in the context of a remotely sensed hyperspectral image processing application) show that, while the classification accuracy is kept constant, the training time substantially decreases with respect to unbalanced training. This is illustrated using heterogeneous computing platforms, made up of CPUs and GPUs with different performance. [ABSTRACT FROM AUTHOR]

Convolutional neural networks (CNNs) exhibit good performance in image processing tasks, pointing themselves as the current state-of-the-art of deep learning methods. However, the intrinsic complexity of remotely sensed hyperspectral images still limits the performance of many CNN models. The high dimensionality of the HSI data, together with the underlying redundancy and noise, often makes the standard CNN approaches unable to generalize discriminative spectral–spatial features. Moreover, deeper CNN architectures also find challenges when additional layers are added, which hampers the network convergence and produces low classification accuracies. In order to mitigate these issues, this paper presents a new deep CNN architecture specially designed for the HSI data. Our new model pursues to improve the spectral–spatial features uncovered by the convolutional filters of the network. Specifically, the proposed residual-based approach gradually increases the feature map dimension at all convolutional layers, grouped in pyramidal bottleneck residual blocks, in order to involve more locations as the network depth increases while balancing the workload among all units, preserving the time complexity per layer. It can be seen as a pyramid, where the deeper the blocks, the more feature maps can be extracted. Therefore, the diversity of high-level spectral–spatial attributes can be gradually increased across layers to enhance the performance of the proposed network with the HSI data. Our experiments, conducted using four well-known HSI data sets and 10 different classification techniques, reveal that our newly developed HSI pyramidal residual model is able to provide competitive advantages (in terms of both classification accuracy and computational time) over the state-of-the-art HSI classification methods [ABSTRACT FROM AUTHOR]

Abstract: Latest generation remote sensing instruments (called hyperspectral imagers) are now able to generate hundreds of images, corresponding to different wavelength channels, for the same area on the surface of the Earth. In previous work, we have reported that the scalability of parallel processing algorithms dealing with these high-dimensional data volumes is affected by the amount of data to be exchanged through the communication network of the system. However, large messages are common in hyperspectral imaging applications since processing algorithms are pixel-based, and each pixel vector to be exchanged through the communication network is made up of hundreds of spectral values. Thus, decreasing the amount of data to be exchanged could improve the scalability and parallel performance. In this paper, we propose a new framework based on intelligent utilization of wavelet-based data compression techniques for improving the scalability of a standard hyperspectral image processing chain on heterogeneous networks of workstations. This type of parallel platform is quickly becoming a standard in hyperspectral image processing due to the distributed nature of collected hyperspectral data as well as its flexibility and low cost. Our experimental results indicate that adaptive lossy compression can lead to improvements in the scalability of the hyperspectral processing chain without sacrificing analysis accuracy, even at sub-pixel precision levels. [Copyright &y& Elsevier]

The purpose of content-based image retrieval (CBIR) is to retrieve, from real data stored in a database, information that is relevant to a query. In remote sensing applications, the wealth of spectral information provided by latest-generation (hyperspectral) instruments has quickly introduced the need for parallel CBIR systems able to effectively retrieve features of interest from ever-growing data archives. To address this need, this paper develops a new parallel CBIR system that has been specifically designed to be run on heterogeneous networks of computers (HNOCs). These platforms have soon become a standard computing architecture in remote sensing missions due to the distributed nature of data repositories. The proposed heterogeneous system first extracts an image feature vector able to characterize image content with sub-pixel precision using spectral mixture analysis concepts, and then uses the obtained feature as a search reference. The system is validated using a complex hyperspectral image database, and implemented on several networks of workstations and a Beowulf cluster at NASA's Goddard Space Flight Center. Our experimental results indicate that the proposed parallel system can efficiently retrieve hyperspectral images from complex image databases by efficiently adapting to the underlying parallel platform on which it is run, regardless of the heterogeneity in the compute nodes and communication links that form such parallel platform. Copyright © 2009 John Wiley & Sons, Ltd. [ABSTRACT FROM AUTHOR]

Abstract: In this work, neural network-based models involved in hyperspectral image spectra separation are considered. Focus is on how to select the most highly informative samples for effectively training the neural architecture. This issue is addressed here by several new algorithms for intelligent selection of training samples: (1) a border-training algorithm (BTA) which selects training samples located in the vicinity of the hyperplanes that can optimally separate the classes; (2) a mixed-signature algorithm (MSA) which selects the most spectrally mixed pixels in the hyperspectral data as training samples; and (3) a morphological-erosion algorithm (MEA) which incorporates spatial information (via mathematical morphology concepts) to select spectrally mixed training samples located in spatially homogeneous regions. These algorithms, along with other standard techniques based on orthogonal projections and a simple Maximin-distance algorithm, are used to train a multi-layer perceptron (MLP), selected in this work as a representative neural architecture for spectral mixture analysis. Experimental results are provided using both a database of nonlinear mixed spectra with absolute ground truth and a set of real hyperspectral images, collected at different altitudes by the digital airborne imaging spectrometer (DAIS 7915) and reflective optics system imaging spectrometer (ROSIS) operating simultaneously at multiple spatial resolutions. [Copyright &y& Elsevier]

The article discusses the study which aims to develop an advance heterogeneous parallel algorithm for hyperspectral image processing through HeteroMPI, an extension of message passing interface (MPI) for programming high-performance computations on heterogeneous networks of computers (HNOCs). It describes hyperspectralimaging algorithm, which performs joint spatial and spectral analysis of the data in combined fashion, and further develops HeteroMPI-based parallel implementations of the algorithm. The study concludes through citing the effectiveness of the HeteroMPI-based implementation.

The main objective of this paper is to describe a realistic framework to understand parallel performance of high-dimensional image processing algorithms in the context of heterogeneous networks of workstations (NOWs). As a case study, this paper explores techniques for mapping hyperspectral image analysis techniques onto fully heterogeneous NOWs. Hyperspectral imaging is a new technique in remote sensing that has gained tremendous popularity in many research areas, including satellite imaging and aerial reconnaissance. The automation of techniques able to transform massive amounts of hyperspectral data into scientific understanding in valid response times is critical for space-based Earth science and planetary exploration. Using an evaluation strategy which is based on comparing the efficiency achieved by an heterogeneous algorithm on a fully heterogeneous NOW with that evidenced by its homogeneous version on a homogeneous NOW with the same aggregate performance as the heterogeneous one, we develop a detailed analysis of parallel algorithms that integrate the spatial and spectral information in the image data through mathematical morphology concepts. For comparative purposes, performance data for the tested algorithms on Thunderhead (a large-scale Beowulf cluster at NASA’s Goddard Space Flight Center) are also provided. Our detailed investigation of the parallel properties of the proposed morphological algorithms provides several intriguing findings that may help image analysts in selection of parallel techniques and strategies for specific applications. [ABSTRACT FROM AUTHOR]

Spectral mixture analysis provides an efficient mechanism for the interpretation and classification of remotely sensed multidimensional imagery. It aims to identify a set of reference signatures (also known as endmembers) that can be used to model the reflectance spectrum at each pixel of the original image. Thus, the modeling is carried out as a linear combination of a finite number of ground components. Although spectral mixture models have proved to be appropriate for the purpose of large hyperspectral dataset subpixel analysis, few methods are available in the literature for the extraction of appropriate endmembers in spectral unmixing. Most approaches have been designed from a spectroscopic viewpoint and, thus, tend to neglect the existing spatial correlation between pixels. This paper presents a new automated method that performs unsupervised pixel purity determination and endmember extraction from multidimensional datasets; this is achieved by using both spatial and spectral information in a combined manner. The method is based on mathematical morphology, a classic image processing technique that can be applied to the spectral domain while being able to keep its spatial characteristics. The proposed methodology is evaluated through a specifically designed framework that uses both simulated and real hyperspectral data. [ABSTRACT FROM AUTHOR]

One of the most important tasks in hyperspectral imaging is the classification of the pixels in the scene in order to produce thematic maps. This problem can be typically solved through machine learning techniques. In particular, deep learning algorithms have emerged in recent years as a suitable methodology to classify hyperspectral data. Moreover, the high dimensionality of hyperspectral data, together with the increasing availability of unlabeled samples, makes deep learning an appealing approach to process and interpret those data. However, the limited number of labeled samples often complicates the exploitation of supervised techniques. Indeed, in order to guarantee a suitable precision, a large number of labeled samples is normally required. This hurdle can be overcome by resorting to unsupervised classification algorithms. In particular, autoencoders can be used to analyze a hyperspectral image using only unlabeled data. However, the high data dimensionality leads to prohibitive training times. In this regard, it is important to realize that the operations involved in autoencoders training are intrinsically parallel. Therefore, in this paper we present an approach that exploits multi-core and many-core devices in order to achieve efficient autoencoders training in hyperspectral imaging applications. Specifically, in this paper, we present new OpenMP and CUDA frameworks for autoencoder training. The obtained results show that the CUDA framework provides a speed-up of about two orders of magnitudes as compared to an optimized serial processing chain. [ABSTRACT FROM AUTHOR]

Remotely sensed hyperspectral imaging instruments are capable of collecting hundreds of images corresponding to different wavelength channels for the same area on the surface of the Earth. For instance, NASA is continuously gathering high-dimensional image data with instruments such as the Jet Propulsion Laboratory's Airborne Visible-Infrared Imaging Spectrometer (AVIRIS). This advanced sensor for Earth observation records the visible and near-infrared spectrum of the reflected light using more than 200 spectral bands, thus producing a stack of images in which each pixel (vector) is represented by a spectral signal that uniquely characterizes the underlying objects. [ABSTRACT FROM PUBLISHER]