Search

This paper introduces a new hyperspectral image denoising method called Non-local Convolutional Neural Network Denoiser (NL-CNND). The technique exploits data in four bands adjacent to the target one as additional information for the restoring process, and it uses a pre-denoising step based on BM4D. All the bands paired with their pre-denoised versions in a second step feed a Convolutional Neural Network. To network generalization, one of the inputs is the noise level of the input image, allowing a single model to work with different noise levels. This restoration technique overcomes quality when compared to current eight classical and neural methods. The results show higher peak signal to noise ratio, structural similarity index, and spectral angle mapper metrics than all the other restoration methods, surpassing those achieved using Block Matching and 4D Filtering alone. Besides, the results show a higher level of detail visually while at the same time reducing over-smoothing on the input images’ features. The paper also includes an algorithm for complete image restoration, allowing for denoising full-sized hyperspectral images independent of their shape. The dataset creation used for network training is detailed, based on a small set of available hyperspectral images, encompassing data normalization, conversion, and storage.

Large-scale self-supervised pretraining of deep learning models is known to be critical in several fields, such as language processing, where its has led to significant breakthroughs. Indeed, it is often more impactful than architectural designs. However, the use of self-supervised pretraining lags behind in several domains, such as hyperspectral images, due to data scarcity. This paper addresses the challenge of data scarcity in the development of methods for spatial super-resolution of hyperspectral images (HSI-SR). We show that state-of-the-art HSI-SR methods are severely bottlenecked by the small paired datasets that are publicly available, also leading to unreliable assessment of the architectural merits of the models. We propose to capitalize on the abundance of high resolution (HR) RGB images to develop a self-supervised pretraining approach that significantly improves the quality of HSI-SR models. In particular, we leverage advances in spectral reconstruction methods to create a vast dataset with high spatial resolution and plausible spectra from RGB images, to be used for pretraining HSI-SR methods. Experimental results, conducted across multiple datasets, report large gains for state-of-the-art HSI-SR methods when pretrained according to the proposed procedure, and also highlight the unreliability of ranking methods when training on small datasets.

Multiscale Dense Network (MSDN) is a deep convolutional neural network method proposed to solve the loss of fine features in hyperspectral remote sensing images. Multiscale Dense Network (MSDN), which makes full use of the different scale information in the network, implements feature extraction of HSI in both horizontal and vertical dimensions. Since MSDN maintains all fine scales until the last layer of the network, the computational effort is large. To address this problem, this paper proposes a lightweight Inter layer Short Multiscale Dense Network (IMSDN) method for hyperspectral image classification. HSI classification is performed using network reduction and inert computation to avoid unnecessary computation, the MSDN network is divided into S blocks along the horizontal dimension, keeping the coarsest (S-i+1) scale only in the ith block and only down sampling the final output of each block in the Short Multiscale Dense Network (SMSDN) for HSI classification. On the basis of the SMSDN network, in order to maintain the classification accuracy of the network, the Lightweight Interlayer Multiscale Dense Network (IMSDN) utilizes three scales of convolution kernels in different horizontal directions to extract the image features separately, which are then fused for the classification of HSI. Experiments are conducted on two publicly available datasets and the results are compared with other methods to demonstrate that the IMSDN network has less computation and shortens the experiment time compared to the MSDN network while guaranteeing the classification accuracy.

The ever-increasing resolution puts tremendous pressure to the onboard hyperspectral imaging system. Compressed sensing technology is one of the important ways to solve this problem. Distributed compressed sensing was proposed to exploit both intra- and inter-correlation structures of hyperspectral images. However, the implementation method of distributed compressed sampling has not been reported, and the joint sparsity reconstruction algorithm cannot achieve excellent image reconstruction performance. In this paper, a distributed compressed sampling strategy inspired by distributed compressed video sensing and optical implementation model are proposed to collect compressed hyperspectral data. In the image reconstruction process, we discard the direct application of the joint sparsity constraint on the data itself. Instead, we explore the estimation method of abundance and endmember with the help of the existing spectral library. Then, the images are recovered by applying the linear mixing model of hyperspectral. The comparison experiments of various schemes show that the proposed compressed sensing scheme has an obvious advantage in reconstruction performance under the low sampling rate. The proposed compressed sensing scheme has great potential in a high-compression ratio onboard hyperspectral imaging system.

Hyperspectral unmixing has attracted considerable attentions in recent years and some promising algorithms have been developed. In this paper, collaborative representation–based unmixing (CRU) for hyperspectral images is proposed. Different from imposing the sparseness constraint on training samples in sparse representation, collaborative representation emphasizes the collaboration of training samples. Furthermore, its closed form solution greatly improves computational efficiency. In the experiments, synthetic and the real hyperspectral data are used to evaluate the effectiveness and efficiency of the proposed collaborative representation-based hyperspectral unmixing algorithm.

Hyperspectral images (HSIs) are usually corrupted by various noises during the image acquisition process, e.g., Gaussian noise, impulse noise, stripes, deadlines and many others. Such complex noise severely degrades the data quality, reduces the interpretation accuracy of HSIs, and restricts the subsequent HSI applications. In this paper, a spatial non-local and local rank-constrained low-rank regularized Plug-and-Play (NLRPnP) model is presented for mixed noise removal in HSIs. Specifically, we first divide HSIs into local overlapping patches. Local rank-constrained low-rank matrix recovery is adopted to effectively separate the low-rank clean HSI patches from the sparse noise and a part of Gaussian noise, and to significantly preserve local structure and detail information in HSIs. Then the spatial non-local based denoiser is introduced to promote the non-local self-similarity and obviously depress the Gaussian noise. Without increasing the difficulty of solving optimization problems, we combine the local and non-local based methods into the Plug-and-Play framework, and develop an efficient algorithm for solving the proposed NLRPnP model by using the alternating direction method of multipliers method. Finally, several experiments are conducted in both simulated and real data conditions to illustrate the better performance of the proposed NLRPnP model than the existing state-of-the-art denoising models.

Hyperspectral data processing is a computationally intensive task that is usually performed in high-performance computing clusters. However, in remote sensing scenarios, where communications are expensive, a compression stage is required at the edge of data acquisition before transmitting information to ground stations for further processing. Moreover, hyperspectral image compressors need to meet minimum performance and energy-efficiency levels to cope with the real-time requirements imposed by the sensors and the available power budget. Hence, they are usually implemented as dedicated hardware accelerators in expensive space-grade electronic devices. In recent years though, these devices have started to coexist with low-cost commercial alternatives in which unconventional techniques, such as run-time hardware reconfiguration are evaluated within research-oriented space missions (e.g., CubeSats). In this paper, a run-time reconfigurable implementation of a low-complexity lossless hyperspectral compressor (i.e., CCSDS 123) on a commercial off-the-shelf device is presented. The proposed approach leverages an FPGA-based on-board processing architecture with a data-parallel execution model to transparently manage a configurable number of resource-efficient hardware cores, dynamically adapting both throughput and energy efficiency. The experimental results show that this solution is competitive when compared with the current state-of-the-art hyperspectral compressors and that the impact of the parallelization scheme on the compression rate is acceptable when considering the improvements in terms of performance and energy consumption. Moreover, scalability tests prove that run-time adaptation of the compression throughput and energy efficiency can be achieved by modifying the number of hardware accelerators, a feature that can be useful in space scenarios, where requirements change over time (e.g., communication bandwidth or power budget).

Spectral clustering is one of the most popular clustering approaches and has been applied in Hyperspectral Image (HSI) clustering well. However, most of these methods are not suitable for large scale HSI. In this paper, based on anchor graph and spatial information, we propose a novel method, called fast spectral clustering based on spatial information (FSCS), which could deal with large scale HSI and have better performance in user's accuracy, average accuracy, overall accuracy and so on. Firstly, based on the physical characteristic of HSI, FSCS algorithm combines the spatial information with spectral information, and uses the spatial nearest points to reconstructs the center point and reveal the local spatial structure. As a result, the correlation of pixels is strengthened and the clustering accuracy is improved. Secondly, the new adjacency matrix is constructed based on anchor graph and thus computational complexity is reduced significantly. Finally, in order to avoid tuning the heat-kernel parameter, the parameter-free strategy is adopted in FSCS. Experiments demonstrate the efficiency and effectiveness of the proposed FSCS algorithm.

Dimension reduction (DR) is an essential preprocessing for hyperspectral image (HSI) classification. Recently, nonnegative matrix factorization (NMF) has been shown as an effective tool for the DR of hyperspectral data given the fact that it provides interpretable results. However, the basic NMF ignores the geometric structure information of the HSI data, thus limiting its performance. To this end, a novel regularized NMF method, termed NMF with adaptive graph regularizer (NMFAGR), is proposed for the spectral-spatial dimension reduction of hyperspectral data in this paper. Specifically, to enhance the preservation ability of the geometric structure information, the NMFAGR performs the dimension reduction and graph learning simultaneously. Regarding the mutual correlation between these two tasks, a graph regularizer is added as an interaction. Moreover, to effectively utilize complementary information among spectral-spatial features, the NMFAGR allocates feature weight factors automatically without requiring any additional parameters. An efficient algorithm is utilized to solve the optimization problem. The effectiveness of the proposed method is demonstrated on three benchmark hyperspectral data sets through experimentation.