Hyperspectral unmixing addressing spectral variability remains an important challenge. In this field, unmixing methods do not exploit the possible availability of some spectral information that corresponds to known spectra of some pure materials present in an acquired scene. In this work, a hyperspectral unmixing method, which considers not only the spectral variability phenomenon but also exploits one or more available known pure material spectra, is proposed. Such a combination, initially proposed here, constitutes the originality of the conducted work that distinguishes it from other investigations in the hyperspectral unmixing topic. The proposed method, based on an informed nonnegative matrix factorization technique, employs a partial structured additively-tuned linear mixing model that deals with spectral variability. Experimental results, based on real data, show that the designed informed algorithm, which addresses spectral variability, yields very satisfactory results and outperforms tested literature approaches. Thus, such an unmixing algorithm may be used for automatically detecting and mapping, using hyperspectral data, materials of interest whose spectra are known while dealing with their spectral variability.
Spectral unmixing (SU) has been a subject of particular attention in the hyperspectral imaging literature. Most SU algorithms are based on the linear mixing model (LMM), which assumes that one pixel of the image is the linear combination of a given number of pure spectra called endmembers, weighted by their coefficients called abundances. SU is a technique to identify these endmembers and their relative abundances. We present an LMM approach based on nonnegative matrix factorization, combining the minimum volume constraint (MVC) and Kullback–Leibler (KL) divergence referred to as KL-MVC. The proposed method is evaluated using synthetic images with different noise levels and real images with different methods of initialization, and high performance has been achieved compared with the widely used LMM-based methods.
Remote sensing hyperspectral sensors, with high spectral resolution, allow precise classification of endmembers present in imaged areas. These sensors have a limited spatial resolution, which results in mixed pixels. The mixture is usually assumed to be linear and blind linear spectral unmixing (LSU) methods are used to unmix all observed pixel spectra. Most blind LSU approaches assume that each endmember is represented by a unique spectrum in all image pixels. But, in many practical applications, this assumption is not valid and more complex models are needed to describe other phenomena, e.g. when each endmember needs to be represented by slightly different spectra in all image pixels. This spectral variability must be handled by replacing the concept of endmembers by classes of endmembers, to avoid errors when processing the considered data. In this paper, a new linear mixing model is firstly introduced in order to handle the spectral variability. In the proposed model, the endmember spectra are additively tuned. Then, an algorithm, based on pixel-by-pixel nonnegative matrix factorization, is proposed for unmixing the considered data. That algorithm, which derives, for each class of endmembers, slightly different estimated spectra in all pixels, optimizes a cost function and uses additional constraints that are related to the introduced linear mixing model. Experiments, based on realistic synthetic data, are conducted to evaluate the performance of the proposed algorithm. The obtained results are compared to those of three approaches from the literature. These test results show that the proposed approach outperforms all other tested methods.
Recently, the new Geographic object-based image analysis (GEOBIA) was proposed as an alternative classification approach to pixel based ones. In GEOBIA, image segments can be depicted with various attributes such as spectral, texture, shape, deep features and context, and hence final classification can produce better land cover/use map. The presence of such a large number of features poses significant challenges to standard machine learning methods and has rendered many existing classification techniques impractical. In this work, we are interested to feature selection techniques, which are employed to reduce the dimensionality of the data while keeping the most of its expressive power. Inspired by recent works in remote sensing using Convolutional Neural Networks (CNNs), especially for hyperspectral band selection, a feature selection approach based on One-Dimensional Convolutional Neural Networks (1-D CNN) is proposed in this study. All object-based features are used to train the 1-D CNN to obtain well trained model. After testing different feature combinations, we use the well trained model to obtain their test classification accuracies, and finally we select the subset of features with the highest precision. In our experiments, we evaluate our feature selection approach on 30-cm resolution colour infrared (CIR) aerial orthoimagery. A multi-resolution segmentation is performed to segment the images into regions, which are characterized later using various spectral, textural and spatial attributes to form the final object-based feature dataset. The obtained experimental results show that the proposed method can achieve satisfactory results when compared with traditional feature selection approaches.
This paper proposes three multisharpening approaches to enhance the spatial resolution of urban hyperspectral remote sensing images. These approaches, related to linear-quadratic spectral unmixing techniques, use a linear-quadratic nonnegative matrix factorization (NMF) multiplicative algorithm. These methods begin by unmixing the observable high-spectral/low-spatial resolution hyperspectral and high-spatial/low-spectral resolution multispectral images. The obtained high-spectral/high-spatial resolution features are then recombined, according to the linear-quadratic mixing model, to obtain an unobservable multisharpened high-spectral/high-spatial resolution hyperspectral image. In the first designed approach, hyperspectral and multispectral variables are independently optimized, once they have been coherently initialized. These variables are alternately updated in the second designed approach. In the third approach, the considered hyperspectral and multispectral variables are jointly updated. Experiments, using synthetic and real data, are conducted to assess the efficiency, in spatial and spectral domains, of the designed approaches and of linear NMF-based approaches from the literature. Experimental results show that the designed methods globally yield very satisfactory spectral and spatial fidelities for the multisharpened hyperspectral data. They also prove that these methods significantly outperform the used literature approaches.
In this paper, a new pansharpening method, which uses nonnegative matrix factorization, is proposed to enhance the spatial resolution of remote sensing multispectral images. This method, based on the linear spectral unmixing concept and called joint spatial-spectral variables nonnegative matrix factorization, optimizes, by new iterative and multiplicative update rules, a joint-variables criterion that exploits spatial and spectral degradation models between the considered images. This criterion considers only two unknown high spatial-spectral resolutions variables. The proposed method is tested on synthetic and real datasets and its effectiveness, in spatial and spectral domains, is evaluated with established performance criteria. Results show the good performances of the proposed approach in comparison with other standard literature ones.
In this paper, a new Spectral-Unmixing-based approach, using Nonnegative Matrix Factorization (NMF), is proposed to locally multi-sharpen hyperspectral data by integrating a Digital Surface Model (DSM) obtained from LIDAR data. In this new approach, the nature of the local mixing model is detected by using the local variance of the object elevations. The hyper/multispectral images are explored using small zones. In each zone, the variance of the object elevations is calculated from the DSM data in this zone. This variance is compared to a threshold value and the adequate linear/linearquadratic spectral unmixing technique is used in the considered zone to independently unmix hyperspectral and multispectral data, using an adequate linear/linear-quadratic NMF-based approach. The obtained spectral and spatial information thus respectively extracted from the hyper/multispectral images are then recombined in the considered zone, according to the selected mixing model. Experiments based on synthetic hyper/multispectral data are carried out to evaluate the performance of the proposed multi-sharpening approach and literature linear/linear-quadratic approaches used on the whole hyper/multispectral data. In these experiments, real DSM data are used to generate synthetic data containing linear and linear-quadratic mixed pixel zones. The DSM data are also used for locally detecting the nature of the mixing model in the proposed approach. Globally, the proposed approach yields good spatial and spectral fidelities for the multi-sharpened data and significantly outperforms the used literature methods.
This paper presents a new fusion approach for pan-sharpening multispectral remote sensing images. This approach,
related to Linear Spectral Unmixing (LSU) techniques, includes Extended Nonnegative Matrix Factorization (ExNMF)
for combining low spatial resolution multispectral and high spatial resolution panchromatic data. ExNMF is applied to
different real multispectral and panchromatic data sets with different spatial resolutions and different number of spectral
bands. The quality of pan-sharpened multispectral images is evaluated by the jointly spectral and spatial Quality with No
Reference (QNR) index. Obtained results show that our proposed method outperforms the Principal Component Analysis
(PCA) and Gram-Schmidt (GS)-based standard literature methods.
Remote sensing has become an unavoidable tool for better managing our environment, generally by realizing maps of
land cover using classification techniques. The classification process requires some pre-processing, especially for data
size reduction. The most usual technique is Principal Component Analysis. Another approach consists in regarding each
pixel of the multispectral image as a mixture of pure elements contained in the observed area. Using Blind Source
Separation (BSS) methods, one can hope to unmix each pixel and to perform the recognition of the classes constituting
the observed scene. Our contribution consists in using Non-negative Matrix Factorization (NMF) combined with sparse
coding as a solution to BSS, in order to generate new images (which are at least partly separated images) using HRV
SPOT images from Oran area (Algeria). These images are then used as inputs of a supervised classifier integrating
textural information. The results of classifications of these "separated" images show a clear improvement (correct pixel
classification rate improved by more than 20%) compared to classification of initial (i.e. non separated) images. These
results show the contribution of NMF as an attractive pre-processing for classification of multispectral remote sensing
imagery.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.