Further results on dissimilarity spaces for hyperspectral images RF-CBIR

Content-Based Image Retrieval (CBIR) systems are powerful search tools in image databases that have been little applied to hyperspectral images. Relevance feedback (RF) is an iterative process that uses machine learning techniques and user’s feedback to improve the CBIR systems performance. We pursued to expand previous research in hyperspectral CBIR systems built on dissimilarity functions defined either on spectral and spatial features extracted by spectral unmixing techniques, or on dictionaries extracted by dictionary-based compressors. These dissimilarity functions were not suitable for direct application in common machine learning techniques. We propose to use a RF general approach based on dissimilarity spaces which is more appropriate for the application of machine learning algorithms to the hyperspectral RF-CBIR. We validate the proposed RF method for hyperspectral CBIR systems over a real hyperspectral dataset.     

Synthesis of mixed pixel hyperspectral signatures

The general method of analysing mixed pixel spectral response is to decompose the actual spectra into several pure spectral components representing the signatures of the endmembers. This work suggests a reverse engineering of standardizing the mixed pixel spectrum for a certain spatial distribution of endmembers by synthesizing spectral signatures with varying proportions of standard spectral library data and matching them with the experimentally obtained mixed pixel signature. The idea is demonstrated with hyperspectral ultraviolet–visible–near-infrared (UV–vis–NIR) reflectance measurements on laboratory-generated model mixed pixels consisting of different endmember surfaces: concrete, soil, brick and vegetation and hyperspectral signatures derived from Hyperion satellite images consisting of concrete, soil and vegetation in different proportions. The experimental reflectance values were compared with the computationally generated spectral variations assuming linear mixing of pure spectral signatures. Good matching in the nature of spectral variation was obtained in most cases. It is hoped that using the present concept, hyperspectral signatures of mixed pixels can be synthesized from the available spectral libraries and matched with those obtained from satellite images, even with fewer bands. Thus enhancing the computational job in the laboratory can moderate the keen requirement of high accuracy of remote-sensor and band resolution, thereby reducing data volume and transmission bandwidth.     

A quantitative and comparative analysis of different preprocessing implementations of DPSO: a robust endmember extraction algorithm

Linear spectral unmixing is a very important technique in hyperspectral image analysis. It contains two main steps. First, it finds spectrally unique signatures of pure ground components (called endmembers); second, it estimates their corresponding fractional abundances in each pixel. Recently, a discrete particle swarm optimization (DPSO) algorithm was introduced to accurately extract endmembers with high optimal performance. However, because of its limited feasible solution space, DPSO necessarily needs a small amount of candidate endmembers before extraction. Consequently, how to provide a suitable candidate endmember set, which has not been analyzed yet, is a critical issue in using DPSO for unmixing problem. In this study, three representative pure pixel-based methods, pixel purity index, vertex component analysis (VCA), and N-FINDR, are quantitatively compared to provide candidate endmembers for DPSO. The experiments with synthetic and real hyperspectral images indicate that VCA is the most reliable preprocessing implementation for DPSO. Further, it can be concluded that DPSO with the proposed preprocessing implementations given in this paper is robust for endmember extraction.     

基于线性混合模型的高光谱图像端元提取

近年来,基于线性混合模型的光谱解混合技术正在越来越广泛地用在光谱数据分析和遥感地物量化中,这项技术的关键就在于确定端元(Endmember)光谱。通常,端元的荻取有两种方式：来源于光谱库以及来源于图像数据,相比之下后者得到的结果更能体现真实的地面信息。为此,从线性混合模型的特点出发,归纳了目前几种比较成熟的端元提取算法,分析了它们的主要思想和存在的优缺点,并总结了评估算法结果的依据,最后介绍了端元提取技术的发展趋势。     

A new non-negative matrix factorization method based on barycentric coordinates for endmember extraction in hyperspectral remote sensing

In this study, we present a new non-negative matrix factorization (NMF) method using the pixel's barycentric coordinates for endmember extraction, named BC-NMF. Our method applies the geometrical property of simplex in the calculation of abundance fraction. That is, for any pixel in an image, its abundance fractions are its barycentric coordinates within the endmember coordinate system. Experiments using both simulated and real hyperspectral images show that BC-NMF can generate endmembers with higher accuracy and lower computational complexity than NMF.     

利用稳健非负矩阵分解实现无监督高光谱解混

目的 基于非负矩阵分解的高光谱图像无监督解混算法普遍存在着目标函数对噪声敏感、在低信噪比条件下端元提取和丰度估计性能不佳的缺点。因此,提出一种基于稳健非负矩阵分解的高光谱图像混合像元分解算法。方法 首先在传统基于非负矩阵分解的解混算法基础上,对目标函数加以改进,用更加稳健的L₁范数作为重建误差项,提高算法对噪声的适应能力,得到新的无监督解混目标函数。针对新目标函数的非凸特性,利用梯度下降法对端元矩阵和丰度矩阵交替迭代求解,进而完成优化求解,得到端元和丰度估计值。结果 分别利用模拟和真实高光谱数据,对算法性能进行定性和定量分析。在模拟数据集中,将本文算法与具有代表性的5种无监督解混算法进行比较,相比于对比算法中最优者,本文算法在典型信噪比20 dB下,光谱角距离（spectral angle distance,SAD）增大了10.5%,信号重构误差（signal to reconstruction error,SRE）减小了9.3%;在真实数据集中,利用光谱库中的地物光谱特征验证本文算法端元提取质量,并利用真实地物分布定性分析丰度估计结果。结论 提出的基于稳健非负矩阵分解的高光谱无监督解混算法,在低信噪比条件下,能够获得较好的端元提取和丰度估计精度,解混效果更好。     

Modified N-FINDR endmember extraction algorithm for remote-sensing imagery

The N-FINDR, developed by Winter, is one of the most widely used algorithms for endmember extraction for hyperspectral images. N-FINDR usually needs an outer loop to control the stopping rule and two inner loops for pixel replacement, so it suffers from computational inefficiency, particularly when the size of the remote-sensing image is large. Recently, geometric unmixing using a barycentric coordinate has become a popular research field in hyperspectral remote sensing. According to Cramer’s rule, a barycentric coordinate estimated by the ratios of simplex volumes is equivalent to a least-squares solution of a linear mixture model. This property implies a brand new strategy for endmember extraction. In other words, we can deduce endmembers by comparison only of abundances derived from a least-squares approach rather than a complicated volume comparison in N-FINDR. Theoretical analysis shows that the proposed method has the same performance as N-FINDR but with much lower computational complexity. In the experiment using real hyperspectral data, our method outperforms several other N-FINDR-based methods in terms of computing times.     

Spectral mixture analysis for subpixel vegetation fractions in the urban environment: How to incorporate endmember variability?

In the urban environment both quality of life and surface biophysical processes are closely related to the presence of vegetation. Spectral mixture analysis (SMA) has been frequently used to derive subpixel vegetation information from remotely sensed imagery in urban areas, where the underlying landscapes are assumed to be composed of a few fundamental components, called endmembers. A critical step in SMA is to identify the endmembers and their corresponding spectral signatures. A common practice in SMA assumes a constant spectral signature for each endmember. In fact, the spectral signatures of endmembers may vary from pixel to pixel due to changes in biophysical (e.g. leaves, stems and bark) and biochemical (e.g. chlorophyll content) composition. This study developed a Bayesian Spectral Mixture Analysis (BSMA) model to understand the impact of endmember variability on the derivation of subpixel vegetation fractions in an urban environment. BSMA incorporates endmember spectral variability in the unmixing process based on Bayes Theorem. In traditional SMA, each endmember is represented by a constant signature, while BSMA uses the endmember signature probability distribution in the analysis. BSMA has the advantage of maximally capturing the spectral variability of an image with the least number of endmembers. In this study, the BSMA model is first applied to simulated images, and then to Ikonos and Landsat ETM+ images. BSMA leads to an improved estimate of subpixel vegetation fractions, and provides uncertainty information for the estimates. The study also found that the traditional SMA using the statistical means of the signature distributions as endmember signatures produces subpixel endmember fractions with almost the same and sometimes even better accuracy than those from BSMA except without uncertainty information for the estimates. However, using the modes of signature distributions as endmembers may result in serious bias in subpixel endmember fractions derived from traditional SMA.     

A real-time unsupervised background extraction-based target detection method for hyperspectral imagery

Target detection is an important technique in hyperspectral image analysis. The high dimensionality of hyperspectral data provides the possibility of deeply mining the information hiding in spectra, and many targets that cannot be visualized by inspection can be detected. But this also brings some problems such as unknown background interferences at the same time. In this way, extracting and taking advantage of the background information in the region of interest becomes a task of great significance. In this paper, we present an unsupervised background extraction-based target detection method, which is called UBETD for short. The proposed UBETD takes advantage of the method of endmember extraction in hyperspectral unmixing, another important technique that can extract representative material signatures from the images. These endmembers represent most of the image information, so they can be reasonably seen as the combination of targets and background signatures. Since the background information is known, algorithm like target-constrained interference-minimized filter could then be introduced to detect the targets while inhibiting the interferences. To meet the rapidly rising demand of real-time processing capabilities, the proposed algorithm is further simplified in computation and implemented on a FPGA board. Experiments with synthetic and real hyperspectral images have been conducted comparing with constrained energy minimization, adaptive coherence/cosine estimator and adaptive matched filter to evaluate the detection and computational performance of our proposed method. The results indicate that UBETD and its hardware implementation RT-UBETD can achieve better performance and are particularly prominent in inhibiting interferences in the background. On the other hand, the hardware implementation of RT-UBETD can complete the target detection processing in far less time than the data acquisition time of hyperspectral sensor like HyMap, which confirms strict real-time processing capability of the proposed system.     

一种端元可变的混合像元分解方法

混合像元线性分解是高光谱影像处理的常用方法，它使用相同的端元矩阵对像元进行分解，其结果是分解精度不高。为此提出了一种端元可变的混合像元分解方法，在确定端元矩阵时，首先考察混合像元与端元的光谱相似性，结合地物空间分布特点，实现了可变端元的混合像元分解。试验结果表明，该分解方法分解精度优于传统线性模型，符合实际情况。     

Spatial–spectral preprocessing for spectral unmixing

Most techniques available in the endmember extraction rely on exploiting the spectral information of the data alone. In this paper, we improve the utilization of data information by dividing a pixel into four subpixels which are redefined by the scalar factor related to the spatial–spectral similarity. The spatial information is integrated into the spectral information in a certain spatial neighbourhood domain, which can make extracted endmembers more precisely, because the effect of noise and outliers can be suppressed with preprocessing (PP). Meanwhile, the accuracy of spectral unmixing will be improved without modification to the conventional methods applied to spectral-based endmember extraction. Experimental results with both synthetic and real hyperspectral images demonstrate the unmixing accuracy is better than that without PP.     

A new volume formula for a simplex and its application to endmember extraction for hyperspectral image analysis

Based on the geometric properties of a simplex, endmembers can be extracted automatically from a hyperspectral image. To avoid the shortcomings of the N-FINDR algorithm, which requires the dimensions of the data to be one less than the number of endmembers needed, a new volume formula for the simplex without the requirement of dimension reduction is presented here. We demonstrate that the N-FINDR algorithm is a special case of the new method. Moreover, whether the null vector is included as an endmember has an important effect on the final result of the endmember extraction. Finally, we compare the new method with previous methods for endmember extraction of hyperspectral data collected by the Advanced Visible and Infrared Imaging Spectrometer (AVIRIS) over Cuprite, Nevada.     

基于拉格朗日的高光谱解混算法研究*

针对混合像元分解误差问题,提出一种基于拉格朗日算法的高光谱解混算法。通过变分增广拉格朗日算法提取出部分端元,由于端元组中存在相似端元影响解混精度,利用基于梯度的光谱信息散度算法进行光谱区分,除去相似端元。通过对得到的端元进行排序,依次增加端元进行光谱解混,将满足条件的端元增加进端元组,最终得到优选端元。该方法不仅有效去除了相似端元的干扰,而且不需要不断搜索端元的组合,根据每个端元对于混合像元的重要性做出相应次数的非限制性最小二乘法计算,得到更精确高光谱端元的子集,该方法对高光谱混合像元解混的效率以及可靠性均有所提高。     

基于数据场的端元提取算法研究

传统端元提取算法一般需要人工指定端元数目,易导致多选或漏选端元。利用数据场自然拓扑聚类、可视化的特性,提出了基于数据场的端元提取方法。首先对图像进行分区处理,然后应用数据场思想计算各区域数据点的势能,并分别选择一定数量的特征点,将所有特征点集合成特征图像,再计算特征图像的数据场;最后根据数据场形成的拓扑聚类结构,可视化地提取端元,获得最佳端元的数目和位置。利用Cuprite矿区的AVIRIS数据进行端元提取实验,结果表明:该方法是合理有效的,能够应用于高光谱图像的端元提取中。     

非监督正交子空间投影的高光谱混合像元自动分解

利用混合像元线性分解技术处理高光谱影像，以获取研究区域中同一像元的不同组份是遥感应用的主要目的之一。近年来，研究者们发展了一种正交子空间投影技术(0SP)，用来探测感兴趣目标，进一步可以用来分解混合像元，然而应用这种方法分解混合像元的缺陷是需要有研究区域的先验信息，这就制约了它在这方面的应用。为此针对这种不足，提出一种非监督的正交子空间投影(UOSP)技术，用来自动获取影像端元光谱，同时进行混合像元分解。并用成像光谱数据(PHI)实例测试了这个方法，结果表明该方法自动获取的端元比较合理，且分解混合像元精度较高。     

Surface Emissivity Image Simulationfor Atmospheric Absorption Bands based on Spectral Mixing

A simulation method based on spectral mixing is proposed for surface emissivity image generation in atmospheric absorption bands,in order to provide surface input data for the corresponding end to end image simulation.First,endmember selection is conducted on data source to acquire image endmember spectra.Then,substitute endmembers are selected from surface measured spectra by spectral matching with image endmembers,and used for abundance inversion.Finally,using emissivity spectra of substitute endmember in the absorption bands and abundance maps,emissivity images are simulated based on linear spectral mixing model.In the simulation experiment,Landsat8 OLI images were used as data source,and JHU and USGS spectral library data were assumed to be ground spectra of the test case.Since actual emissivity images in absorption bands are unavailable,accuracy analysis is conducted by comparing OLI reflectance images with its simulations generated by the proposed method.Total RSMEs of simulated OLI images are 0.045 and 0.049,respectively;which shows the image simulation method is feasible and can produce images with high accuracy.     

Integration of spatial-spectral information for the improved extraction of endmembers

Spectral-based image endmember extraction methods hinge on the ability to discriminate between pixels based on spectral characteristics alone. Endmembers with distinct spectral features (high spectral contrast) are easy to select, whereas those with minimal unique spectral information (low spectral contrast) are more problematic. Spectral contrast, however, is dependent on the endmember assemblage, such that as the assemblage changes so does the “relative” spectral contrast of each endmember to all other endmembers. It is then possible for an endmember to have low spectral contrast with respect to the full image, but have high spectral contrast within a subset of the image. The spatial-spectral endmember extraction tool (SSEE) works by analyzing a scene in parts (subsets), such that we increase the spectral contrast of low contrast endmembers, thus improving the potential for these endmembers to be selected. The SSEE method comprises three main steps: 1) application of singular value decomposition (SVD) to determine a set of basis vectors that describe most of the spectral variance for subsets of the image; 2) projection of the full image data set onto the locally defined basis vectors to determine a set of candidate endmember pixels; and, 3) imposing spatial constraints for averaging spectrally similar endmembers, allowing for separation of endmembers that are spectrally similar, but spatially independent. The SSEE method is applied to two real hyperspectral data sets to demonstrate the effects of imposing spatial constraints on the selection of endmembers. The results show that the SSEE method is an effective approach to extracting image endmembers. Specific improvements include the extraction of physically meaningful, low contrast endmembers that occupy unique image regions.     

Modelling land‐cover types using multiple endmember spectral mixture analysis in a desert city

Spectral mixture analysis is probably the most commonly used approach among sub‐pixel analysis techniques. This method models pixel spectra as a linear combination of spectral signatures from two or more ground components. However, spectral mixture analysis does not account for the absence of one of the surface features or spectral variation within pure materials since it utilizes an invariable set of surface features. Multiple endmember spectral mixture analysis (MESMA), which addresses these issues by allowing endmembers to vary on a per pixel basis, was employed in this study to model Landsat ETM+ reflectance in the Phoenix metropolitan area. Image endmember spectra of vegetation, soils, and impervious surfaces were collected with the use of a fine resolution Quickbird image and the pixel purity index. This study employed 204 (3×17×4) total four‐endmember models for the urban subset and 96 (6×6×2×4) total five‐endmember models for the non‐urban subset to identify fractions of soil, impervious surface, vegetation and shade. The Pearson correlation between the fraction outputs from MESMA and reference data from Quickbird 60 cm resolution data for soil, impervious, and vegetation were 0.8030, 0.8632, and 0.8496 respectively. Results from this study suggest that the MESMA approach is effective in mapping urban land covers in desert cities at sub‐pixel level.     

On Endmember Identification in Hyperspectral Images Without Pure Pixels: A Comparison of Algorithms

Hyperspectral imaging is an active area of research in Earth and planetary observation. One of the most important techniques for analyzing hyperspectral images is spectral unmixing, in which mixed pixels (resulting from insufficient spatial resolution of the imaging sensor) are decomposed into a collection of spectrally pure constituent spectra, called endmembers weighted by their correspondent fractions, or abundances. Over the last years, several algorithms have been developed for automatic endmember extraction. Many of them assume that the images contain at least one pure spectral signature for each distinct material. However, this assumption is usually not valid due to spatial resolution, mixing phenomena, and other considerations. A?recent trend in the hyperspectral imaging community is to design endmember identification algorithms which do not assume the presence of pure pixels. Despite the proliferation of this kind of algorithms, many of which are based on minimum enclosing simplex concepts, a rigorous quantitative and comparative assessment is not yet available. In this paper, we provide a comparative analysis of endmember extraction algorithms without the pure pixel assumption. In our experiments we use synthetic hyperspectral data sets (constructed using fractals) and real hyperspectral scenes collected by NASA’s Jet Propulsion Laboratory.