Hyperspectral Pixels Research Articles

Spatial-spectral graph convolutional extreme learning machine for hyperspectral image classification

ABSTRACT Hyperspectral image has excellent spectral information and abundant spatial information, and its feature quality is one of the key factors affecting the classification performance. Extreme learning machine (ELM) is widely used in hyperspectral classification problems. However, traditional ELM is usually based on regular Euclidean data, ignoring the inherent structured information between hyperspectral pixels, resulting in poor robustness. In this paper, we propose a new supervised extreme learning machine framework, termed spatial-spectral graph convolutional ELM (SSG-ELM), based on using graph convolution to extend ELM into the non-Euclidean domain. The method inherits all the advantages from ELM, and consists of a random graph convolutional layer followed by a graph convolutional regression layer, enabling it to model complex intra class variations. Otherwise, a local spectral-spatial context integration and reshaping mechanism is incorporated into the hidden layer feature representation by using a context-aware bilateral filtering procedure. Experimental results show that the proposed algorithm obtains a competitive performance and outperforms other state-of-the-art ELM-based methods.

GAUSS: Guided encoder - decoder Architecture for hyperspectral Unmixing with Spatial Smoothness

ABSTRACT This study introduces GAUSS (Guided encoder-decoder Architecture for hyperspectral Unmixing with Spatial Smoothness), a novel autoencoder-based architecture for hyperspectral unmixing (HU). GAUSS consists of an Approximation Network (AN), Unmixing Network (UN), and a Mixing Network (MN). The AN incorporates spatial context within a hyperspectral pixel’s neighborhood, while the UN utilizes a pseudo-ground truth mechanism to enhance abundance estimation. The MN provides estimated endmembers’ signatures. By incorporating UN-produced abundances, unlike the conventional AE model, GAUSS overcomes the single-layer constraint of the MN. Thereafter, a secondary training phase improves the accuracy of endmembers and abundance estimation using a reliable Signal Processing (SP) algorithm, resulting in superior HU performance. The results demonstrate the effectiveness of GAUSS on two Standard datasets and a Simulated dataset compared to the state-of-the-art SP and Deep Learning (DL) based methods. This signifies the benefit of integrating an SP algorithm in the training process, contributing to advancements in DL-based HU techniques.

MADANet: A Lightweight Hyperspectral Image Classification Network with Multiscale Feature Aggregation and a Dual Attention Mechanism

Hyperspectral remote sensing images, with their continuous, narrow, and rich spectra, hold distinct significance in the precise classification of land cover. Deep convolutional neural networks (CNNs) and their variants are increasingly utilized for hyperspectral classification, but solving the conflict between the number of model parameters, performance, and accuracy has become a pressing challenge. To alleviate this problem, we propose MADANet, a lightweight hyperspectral image classification network that combines multiscale feature aggregation and a dual attention mechanism. By employing depthwise separable convolution, multiscale features can be extracted and aggregated to capture local contextual information effectively. Simultaneously, the dual attention mechanism harnesses both channel and spatial dimensions to acquire comprehensive global semantic information. Ultimately, techniques such as global average pooling (GAP) and full connection (FC) are employed to integrate local contextual information with global semantic knowledge, thereby enabling the accurate classification of hyperspectral pixels. The results from the experiments conducted on representative hyperspectral images demonstrate that MADANet not only attains the highest classification accuracy but also maintains significantly fewer parameters compared to the other methods. Experimental results show that our proposed framework significantly reduces the number of model parameters while still achieving the highest classification accuracy. As an example, the model has only 0.16 M model parameters in the Indian Pines (IP) dataset, but the overall accuracy is as high as 98.34%. Similarly, the framework achieves an overall accuracy of 99.13%, 99.17%, and 99.08% on the University of Pavia (PU), Salinas (SA), and WHU Hi LongKou (LongKou) datasets, respectively. This result exceeds the classification accuracy of existing state-of-the-art frameworks under the same conditions.

Geometric and Radiometric Quality Assessments of UAV-Borne Multi-Sensor Systems: Can UAVs Replace Terrestrial Surveys?

Unmanned aerial vehicles (UAVs), also known as drones, are a cost-effective alternative to traditional surveying methods, and they can be used to collect geospatial data over inaccessible or hard-to-reach locations. UAV-integrated miniaturized remote sensing sensors such as hyperspectral and LiDAR sensors, which formerly operated on airborne and spaceborne platforms, have recently been developed. Their accuracies can still be guaranteed when incorporating pieces of equipment such as ground control points (GCPs) and field spectrometers. This study conducted three experiments for geometric and radiometric accuracy assessments of simultaneously acquired RGB, hyperspectral, and LiDAR data from a single mission. Our RGB and hyperspectral data generated orthorectified images based on direct georeferencing without any GCPs. Because of this, a base station is required for the post-processed Global Navigation Satellite System/Inertial Measurement Unit (GNSS/IMU) data. First, we compared the geometric accuracy of orthorectified RGB and hyperspectral images relative to the distance of the base station to determine which base station should be used. Second, point clouds could be generated from overlapped RGB images and a LiDAR sensor. We quantitatively and qualitatively compared RGB and LiDAR point clouds in this experiment. Lastly, we evaluated the radiometric quality of hyperspectral images, which is the most critical factor of the hyperspectral sensor, using reference spectra that was simultaneously measured by a field spectrometer. Consequently, the distance of the base station for post-processing the GNSS/IMU data was found to have no significant impact on the geometric accuracy, indicating that a dedicated base station is not always necessary. Our experimental results demonstrated geometric errors of less than two hyperspectral pixels without using GCPs, achieving a level of accuracy that is comparable to survey-level standards. Regarding the comparison of RGB- and LiDAR-based point clouds, RGB point clouds exhibited noise and lacked details; however, through the cleaning process, their vertical accuracy was found to be comparable with LiDAR’s accuracy. Although photogrammetry generated denser point clouds compared with LiDAR, the overall quality for extracting the elevation data greatly relies on factors such as the original image quality, including the image’s occlusions, shadows, and tie-points, for matching. Furthermore, the image spectra derived from hyperspectral data consistently demonstrated high radiometric quality without the need for in situ field spectrum information. This finding indicates that in situ field spectra are not always required to guarantee the radiometric quality of hyperspectral data, as long as well-calibrated targets are utilized.

Hyperspectral image segmentation: a preliminary study on the Oral and Dental Spectral Image Database (ODSI-DB)

ABSTRACT Visual discrimination of clinical tissue types remains challenging, with traditional RGB imaging providing limited contrast for such tasks. Hyperspectral imaging (HSI) is a promising technology providing rich spectral information that can extend far beyond three-channel RGB imaging. Moreover, recently developed snapshot HSI cameras enable real-time imaging with significant potential for clinical applications. Despite this, the investigation into the relative performance of HSI over RGB imaging for semantic segmentation purposes has been limited, particularly in the context of medical imaging. Here we compare the performance of state-of-the-art deep learning image segmentation methods when trained on hyperspectral images, RGB images, hyperspectral pixels (minus spatial context), and RGB pixels (disregarding spatial context). To achieve this, we employ the recently released Oral and Dental Spectral Image Database (ODSI-DB), which consists of 215 manually segmented dental reflectance spectral images with 35 different classes across 30 human subjects. The recent development of snapshot HSI cameras has made real-time clinical HSI a distinct possibility, though successful application requires a comprehensive understanding of the additional information HSI offers. Our work highlights the relative importance of spectral resolution, spectral range, and spatial information to both guide the development of HSI cameras and inform future clinical HSI applications.

Semantic modeling of hyperspectral target detection with weak labels

Hyperspectral target detection is consistently restricted by the difficulty of obtaining pixel-level accurate labels. This paper proposes a semantic multiple instance neural network (Semantic MINN) with contrastive and sparse attention fusion. Semantic MINN relaxes the requirement for precise pixel-wise labels needing only patch-level labels. This network models the hyperspectral pixels as semantic signals, captures the spatial semantic information of the potential target regions with weak labels, i.e., imprecise knowledge of target presence, and adopts a sparse normalization strategy using attention mechanism to refine the prime target information. Notably, the siamese structure and a feature similarity metric constraint are utilized to promote discriminative high-level prime target representations. The proposed method is more effective on both simulated and real-field hyperspectral target detection with weak labels, compared with the classical and state-of-the-art weakly supervised techniques.

Coregistration of Hyperspectral Imagery With Photogrammetry for Shallow-Water Mapping

In this paper we present coregistration of hyperspectral imagery with photogrammetry for shallow-water mapping from an Unmanned Aerial Vehicle (UAV). The coregistration is based on a methodology for georeferencing that utilizes outputs from the photogrammetry pipeline through three steps. First, we perform the photogrammetry pipeline. This gives camera poses from Structure-from-Motion (SfM) and a dense point cloud from Multi-View Stereo (MVS). Performing a refraction correction of the dense point cloud yields high-resolution bathymetry. Second, poses from SfM are fused with UAV navigation sensors in a Kalman smoother. Third, the geometric model of the hyperspectral imager is calibrated to align hyperspectral images with photogrammetry. Then, ray tracing is performed to georeference spectral measurements onto the bathymetry from MVS. Using the georeferenced hyperspectral imagery, we present spectral bathymetry estimation. The methods were demonstrated for a coastal site (depth < 6 m) in Norway, using a UAV with a camera and a hyperspectral imager. The georeferencing yielded a horizontal Mean Absolute Error (MAE) of 22 cm between hyperspectral and photogrammetry, equivalent to one hyperspectral pixel. The MVS bathymetry gave a MAE of 14 cm with respect to ground truth acoustic. The spectral bathymetry estimator was calibrated on ground truth acoustics with a MAE of 10 cm. Comparing the bathymetry from MVS with the spectral bathymetry yielded a MAE of 11 cm. The results suggest that coregistration with photogrammetry yields accurate georeferencing of hyperspectral imagery. The results also show that we can map bathymetry accurately using MVS with refraction correction or the spectral estimator.

A Fast Spatial–Spectral NMF for Hyperspectral Unmixing

This letter proposes a fast yet efficient method to solve the hyperspectral unmixing problem in the challenging unsupervised context, i.e., when the endmember spectral signatures are unknown. First, a coarse approximation of the hyperspectral image is computed by spatially averaging neighboring pixels, which significantly reduces the amount of pixels to be handled. This reduced set of hyperspectral pixels is unmixed to derive coarse solutions of the unmixing problem, i.e., coarse estimates of the endmember signatures and the corresponding low-resolution abundance maps. Then, the plain resolution abundance maps are estimated from the corresponding hyperspectral image based on the coarse endmember signatures. A sparsity promoting prior exploiting the low resolution map complements the conventional data fitting term to promote spatial smoothness while mitigating the loss of details in the edge areas. Finally, a least square optimization problem is solved to obtain the actual endmember signatures from the hyperspectral image and the abundance maps of plain resolution estimated in the previous step. Numerical experiments show that the proposed method is fast and performs well compared to state-of-the-art approaches from the literature.

S2DMSC: A Self-Supervised Deep Multilevel Subspace Clustering Approach for Large Hyperspectral Images

Subspace clustering (SC) has achieved remarkable success in hyperspectral images (HSIs) due to the powerful representation ability of handling high-dimensional complex data. However, most of the existing SC methods focus on linear subspace representation and ignore the more effective nonlinear representation. Besides, SC suffers from the bottlenecks, such as high computation load and memory capacity, due to the spectral decomposition of adjacency matrix for large HSIs. To overcome these limitations, we propose an end-to-end learnable network framework for large HSIs, called self-supervised deep multi-level subspace clustering (S <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> DMSC), which incorporates the convolutional neural network (CNN) module, multi-level subspace clustering (MSC) module, and high-quality pseudo-label-based self-supervised learning module into a unified learning framework. More concretely, the deep multi-level spatial-spectral representation from hierarchical superpixels is modeled as a sparsity-constrained self-expression module for SC to construct high-quality pseudo-labels to learn the network parameters and produce better clusters for hyperspectral pixels. Experimental results on four classical HSIs demonstrate the effectiveness of S <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> DMSC and exhibit superior clustering performance compared to the representative clustering methods.

Deep Spectral Clustering With Regularized Linear Embedding for Hyperspectral Image Clustering

The past decade has witnessed the rapid development of deep learning techniques, especially for large-scale and complex data sets. However, it is still a noteworthy problem in dealing with unsupervised hyperspectral image segmentation since inefficiency and misleading result from the absence of supervised information. Generally, spectral clustering is one of the most powerful clustering algorithms, as it often outperforms other methods for image segmentation. Unfortunately, the poor scalability and generalization severely limit the use of spectral clustering, especially for large-scale and high-dimensional hyperspectral images processing. The major motivation of this work is to solve this problem, and we designed a novel algorithm, termed Deep Spectral Clustering with Regularized Linear Embedding (DSCRLE), to benefit from both spectral graph theory and deep learning techniques. The brief procedure is first to construct a fully connected neural network to extract latent feature representations, and then normalize the feature representations by the spectral orthonormal constraint. Lastly, by introducing low-dimensional embedding, we refined the final outputs of all given unlabeled hyperspectral pixels. Extensive experiments have demonstrated that the competitiveness of the proposed method, and it outperforms state-of-art clustering approaches in the task of hyperspectral image segmentation.

The effects of spectral dimensionality reduction on hyperspectral pixel classification: A case study.

This paper presents a systematic study of the effects of hyperspectral pixel dimensionality reduction on the pixel classification task. We use five dimensionality reduction methods-PCA, KPCA, ICA, AE, and DAE-to compress 301-dimensional hyperspectral pixels. Compressed pixels are subsequently used to perform pixel classifications. Pixel classification accuracies together with compression method, compression rates, and reconstruction errors provide a new lens to study the suitability of a compression method for the task of pixel classification. We use three high-resolution hyperspectral image datasets, representing three common landscape types (i.e. urban, transitional suburban, and forests) collected by the Remote Sensing and Spatial Ecosystem Modeling laboratory of the University of Toronto. We found that PCA, KPCA, and ICA post greater signal reconstruction capability; however, when compression rates are more than 90% these methods show lower classification scores. AE and DAE methods post better classification accuracy at 95% compression rate, however their performance drops as compression rate approaches 97%. Our results suggest that both the compression method and the compression rate are important considerations when designing a hyperspectral pixel classification pipeline.

Non-linear hyperspectral unmixing with 3D convolutional encoders

ABSTRACT Deep learning-based methods are accepted as a viable alternative to conventional statistical and geometrical methods for hyperspectral unmixing in recent years. These methods are however mainly based on linear mixture assumption on the hyperspectral data. The vast majority of presented algorithms process individual hyperspectral pixels while neglecting the spatial relationships between pixels. In order to address these two missing aspects, we propose a convolutional autoencoder-based hyperspectral unmixing method in this paper. The proposed structure incorporates the spatial neighbourhood relation with its convolutional layers in the first stage and possible non-linearities in the observed data with the included non-linear layer in the final stage. The experiments have first revealed that Adam optimizer have the best performance among different optimization methods for the proposed network. Second, the proposed method has indicated about 20–40% accuracy improvement in terms of mean squared error (MSE) metric compared to traditional hyperspectral abundance estimation methods. Third, the contribution of the non-linear layer is verified by comparing the proposed network with the conventional LMM-based autoencoder structure without the non-linear layer. Finally, the accuracy improvement for the proposed network with non-linear layer compared to the state-of-the-art deep learning-based methods using linear mixture assumption is evaluated in terms of MSE and reported as about 10% and 20% for synthetic and real data, respectively.

Классификация гиперспектральных данных дистанционного зондирования Земли с использованием комбинированных 3D--2D сверточных нейронных сетей

Hyperspectral image classification is used for analyzing remote Earth sensing data. Convolutional neural network is one of the most commonly used methods for processing visual data based on deep learning. The article considers the proposed hybrid 3D--2D spectral convolutional neural network for hyperspectral image classification. At the initial stage, a simple combined trained deep learning model was proposed, which was constructed by combining 2D and 3D convolutional neural networks to extract deeper spatial-spectral features with fewer 3D--2D convolutions. The 3D network facilitates the joint spatial-spectral representation of objects from a stack of spectral bands. Functions of 3D--2D convolutional neural networks were used for classifying hyperspectral images. The algorithm of the method of principal components is applied to reduce the dimension. Hyperspectral image classification experiments were performed on Indian Pines, University of Pavia and Salinas Scene remote sensing datasets. The first layer of the feature map is used as input for subsequent layers in predicting final labels for each hyperspectral pixel. The proposed method not only includes the benefits of advanced feature extraction from convolutional neural networks, but also makes full use of spectral and spatial information. The effectiveness of the proposed method was tested on three reference data sets. The results show that a multifunctional learning system based on such networks significantly improves classification accuracy (more than 99 %)

A Bipartite Graph Partition-Based Coclustering Approach With Graph Nonnegative Matrix Factorization for Large Hyperspectral Images

Clustering large hyperspectral images (HSIs) is a very challenging problem because large HSIs have high dimensionality, large spectral variability, and large computational and memory consumption. Recently, sparse subspace clustering (SSC) has achieved remarkable success in HSI clustering. However, most SSC-based methods suffer from the following bottlenecks for large HSIs: 1) high computational consumption and memory space during the construction of the similarity matrix and decomposition of the graph Laplacian matrix and 2) failure to capture the relationships among dictionary atoms, sparse coefficients, and hyperspectral pixels. To address these challenges, we propose a novel algorithm that extends SSC to cocluster large HSIs, called bipartite graph partition with graph nonnegative matrix factorization (BGP-GNMF). Specifically, to fully explore the characteristics of the spectral and spatial contexts in HSIs, we propose a novel superpixel and pixel coclustering framework with bipartite graph partitioning in the joint sparse representation domain, where superpixel-based dictionary atoms are defined as disjoint vertex sets of the bipartite graph and the joint sparsity representation is mapped into the adjacency matrix of the undirected bipartite graph. To overcome the challenges of high computational consumption and large memory space for large HSIs, the bipartite graph partition with orthonormal constrained nonnegative matrix factorization is proposed to simultaneously cluster the structured dictionary atoms and hyperspectral pixels with an indicator matrix. Finally, to exploit the intrinsic geometry of HSIs, we incorporate manifold regularization into the bipartite graph partition to improve final clustering accuracy. The effectiveness and efficiency of the proposed method are verified on three classical HSIs, and the experimental results illustrate the superiority of the proposed method compared with other state-of-the-art HSI clustering methods.

Graph Convolutional Sparse Subspace Coclustering With Nonnegative Orthogonal Factorization for Large Hyperspectral Images

Sparse subspace clustering (SSC) is a representative data clustering paradigm that has been broadly applied in the unsupervised classification of hyperspectral images (HSIs). Existing SSC methods usually produce a subspace affinity matrix between representations of hyperspectral pixels first, followed by spectral clustering for the affinity matrix. To this end, the separated framework fails to exploit the dualities contained in both features and pixels or higher order entities at the same time, and thus, it is difficult to compute coclusters simultaneously. In addition, SSC methods often require expensive computational consumption and memory capacity to approximate the spectral decomposition of the affinity matrix, thus hindering the applicability of SSC methods for large HSIs. To overcome these limitations, we propose a novel graph convolutional sparse subspace coclustering (GCSSC) model with nonnegative orthogonal factorization for large HSIs in which affinity matrix learning and spectral coclustering are integrated into a unified optimizing model to obtain the optimal clustering results. Specifically, to form a more compact self-representation, the superpixel-based adaptive dictionary construction strategy is proposed instead of the global dictionary to precisely represent the pixels. To explore the spatial–contextual and spectral neighboring characteristics between dictionary atoms, graph convolution is incorporated into the dictionary atoms to aggregate the local neighborhood information, and the affinity matrix in the proposed coclustering framework is constructed under a joint sparsity constrained representation model. To reduce high computational consumption and memory capacity, a nonnegative orthogonal factorization constraint is proposed to offer an alternative spectral clustering for hyperspectral pixels and dictionary atoms simultaneously. The clustering performance of the proposed method is evaluated for three classical HSIs, and the experimental results illustrate that the proposed method is memory and computationally efficient and outperforms the state-of-the-art HSI clustering methods.

Mapping Coastal Wetlands Using Transformer in Transformer Deep Network on China ZY1-02D Hyperspectral Satellite Images

Coastal wetlands mapping is a big challenge in remote sensing fields because of similar spectrum of different ground objects and their severe fragmentation and spatial heterogeneity. In this paper, we propose a Hyperspectral Image Transformer iN Transformer (HSI-TNT) method for mapping coastal wetlands on ZY1-02D hyperspectral images, which uses two transformer deep networks to fuse local and global features. First, we put forward the idea that each hyperspectral pixel can be considered as a superpixel in spectral dimension, and subsequent position encodings are employed aiming to retain spatial information. After that, in each HSI-TNT block, the local information between pixels is extracted by Inner T-Block, and added to the patch space by linear transformation to extract the global information by Outer T-Block. Finally, the stacked HSI-TNT block, also known as HSI-TNT framework, is used for classification and mapping. Experimental results show that HSI-TNT achieves the best results on both Yancheng and Yellow River Delta wetlands data, with OCA of 95.57% and 93.69% respectively. The HSI-TNT combined with ZY1-02D satellite hyperspectral data has huge potentials in mapping coastal wetlands.

Tiny-Scene Embedding Network for Coastal Wetland Mapping Using Zhuhai-1 Hyperspectral Images

The fine mapping of coastal wetlands is a major challenge due to the spectral aliasing of vegetation. In this letter, we selected Zhuhai-1 hyperspectral images (HSIs) for coastal wetland mapping and proposed a tiny-scene embedding network (TSE-Net) based on scene representation and attention mechanism. In TSE-Net, the tiny-scene representation associated with each hyperspectral pixel was extracted and used to enhance the spectral discrimination of ground objects. DenseNet was chosen as the backbone network, and the attention mechanism was introduced into the dense blocks to extract remarkable features. Experiments on the Yellow River estuary coastal wetland showed that the results of TSE-Net had a significant improvement in accuracy compared to other models, especially for the coastal wetland vegetation with confusing spectra, such as Spartina alterniflora, Suaeda salsa, Phragmites australis, and Tamarix.

Cascade Residual Capsule Network for Hyperspectral Image Classification

Theconvolution neural network (CNN) has recently shown the good performance in hyperspectral image (HSI) classification tasks. Many CNN-based methods crop image patches from original HSI as inputs. However, the input HSI cubes usually contain background and many hyperspectral pixels with different land-cover labels. Therefore, the spatial context information on objects of the same category is diverse in HSI cubes, which will weaken the discrimination of spectral–spatial features. In addition, CNN-based methods still face challenges in dealing with the spectral similarity between HSI cubes of spatially adjacent categories, which will limit the classification accuracy. To address the aforementioned issues, we propose a cascade residual capsule network (CRCN) for HSI classification. First, a residual module is designed to learn high-level spectral features of input HSI cubes in the spectral dimension. The residual module is employed to solve the problem of the spectral similarity between HSI cubes of spatially adjacent categories. And then two 3-D convolution layers are exploited to extract high-level spatial–spectral features. Finally, a capsule structure is developed to characterize spatial context orientation representations of objects, which can effectively deal with the diverse spatial context information on objects of the same category in HSI cubes. The capsule module is composed of two 3-D convolution layers and the capsule structure, which is connected to the residual module in series to construct the proposed CRCN. Experimental results on four public HSI datasets demonstrate the superiority of the proposed CRCN over six state-of-the-art models.

SC-EADNet: A Self-Supervised Contrastive Efficient Asymmetric Dilated Network for Hyperspectral Image Classification

Unsupervised and semisupervised feature learning has recently emerged as an effective way to reduce the reliance on expensive data collection and annotation for hyperspectral image (HSI) analysis. Existing unsupervised and semisupervised convolutional neural network (CNN)-based HSI classification works still face two challenges: underutilization of pixel-wise multiscale contextual information for feature learning and expensive computational cost, for example, large floating-point operations per seconds (FLOPs), due to the lack of lightweight design. To utilize the unlabeled pixels in the HSIs more efficiently, we propose a self-supervised contrastive efficient asymmetric dilated network (SC-EADNet) for HSI classification. There are two novelties in the SC-EADNet. First, a self-supervised multiscale pixel-wise contextual feature learning model is proposed, which generates multiple patches around each hyperspectral pixel and develops a contrastive learning framework to learn from these patches for HSI classification. Second, a lightweight feature extraction network EADNet, composed of multiple plug-and-play efficient asymmetric dilated convolution (EADC) blocks, is designed and inserted into the contrastive learning framework. The EADC block adopts different dilation rates to capture the spatial information of objects with varying shapes and sizes. Compared with other unsupervised, semisupervised, and supervised learning methods, our SC-EADNet provides competitive classification performance on four hyperspectral datasets, including Indian Pines, Pavia University, Salinas, and Houston 2013, but few FLOPs and fast computational speed.

Spatial-Aware Hyperspectral Nonlinear Unmixing Autoencoder With Endmember Number Estimation

In this article, we develop a novel fully unsupervised autoencoder-based scheme for nonlinear hyperspectral pixel unmixing. A unique approach is derived where high noise and unresponsive pixels are accounted for, by a unique averaging approach based on spatially aware filters built using radial basis function (RBF) kernels. A novel technique is implemented via calculating rank-equivalent kernel covariance matrices in order to estimate the unknown number of endmembers contributing to the data. Utilization of spatial information is done via RBF-based weighted averaging, which is then followed by endmember estimation via K-means clustering. The RBF distances from the cluster centers are determined to measure the position of the mixed pixels in relation to the centers, which is utilized as a preliminary estimation of the abundances. The proposed framework is robust in the presence of unresponsive pixels, while highly versatile working with different nonlinear unmixing models. Extensive numerical tests establish the superiority of the novel approach with respect to state-of-the-art methods.