National Technical University of Athens (NTUA)
Methods and algorithms for hyperspectral data processing
Abstract
dc:descriptionHyperspectral remote sensing and spectroscopy collect and process information from across theelectromagnetic spectrum with the purpose of finding objects, identifying materials, or detectingprocesses. Hyperspectral imaging systems, due to their low spatial resolution, obtain pixels that may contain several distinct materials. This is the problem of spectral mixing for which linear and nonlinear models have been proposed in order to spectrally unmix the data. For both models, butespecially for the linear one, many methods that implement the required stages of spectral unmixing have been developed.This thesis addresses several key issues on hyperspectral remote sensing and spectroscopyconcerning spectral unmixing, spectral matching, spectral discrimination, quantification ofparameters using spectral data and relative radiometric normalization of hyperspectral data. Theseissues are related to real world hyperspectral remote sensing applications, which require bothmeasurements and processing of field data, as well as hyperspectral acquisitions and processing ofthe same area in many different time periods. The main contributions of the thesis include theintroduction of new innovative concepts and ideas, and exploitation of existing ones. A new novelunmixing concept and six new methods have been developed within the thesis: estimation of thenumber of endmembers and extraction, abundance fraction estimation, spectral matching, spectraldiscrimination, spectroscopy, and relative radiometric normalization. The first three perform the main spectral unmixing steps in hyperspectral imaging, the next two address key issues of spectroscopy on the distinction of species/varieties and estimation of the object parameters, while the sixth allows the simultaneous processing of remote sensing data of different chronologies / sensors, etc.The first method called Fractional Distance (FD) introduces a novel distance-based method, forestimating the number of endmembers and extracts them. The method adopts the network conceptaccording to which, every pixel of the image creates a network. The endmembers and the pixel arethe nodes of the network while the fractional distance parameter quantify the network links. Whenthe pixel is replaced by an endmember, then an endmember network is created. The set of theendmember networks formulates the endmember fractional distance matrix. The Fractional Distance (FD) method, exploits the mathematical properties of the endmember fractional distance matrix and provides the number and the signatures of the endmembers with low computational cost. Fractional distance matrices satisfactorily reveal the linear independence of the endmember vectors, as well as, the maximization of the volume of the simplex formed by the endmembers. FD was evaluated using simulated and real data, and it was also compared with the existing random counterparts of state of the art algorithms. Evaluation of the method showed that the method yields robust and satisfactory results, and outperforms its competitors.The second method called Network Based Method (NBM) is a new non-parametric algorithm forcalculating the abundance fraction values for given endmembers which relies on the network concept. Based on the fractional distances which imply nonlinearity, the method quantifies interactions between pixels and endmembers. Relationships between fractional distances and abundance fractions are built through the networks. Because an equal spectral mixture of endmembers is likely impossible, the proposed method, due to its mathematical concept, reveals the presence of unknown endmembers.Three versions of the method have been developed: the non-constrained (U-NBM), the sum-to-one,(S-NBM) and the fully constrained (F-NBM) versions. Evaluation of the method using synthetic andreal data showed that the method is robust with clear and interpretable results and provides reliable abundance fractions, particularly the sum-to-one and the fully constrained versions of the method.The new unmixing method has also been compared with the fully constrained least squares method. As part of the matching/labeling, the third method is a new methodology for automaticallylabeling extracted endmembers from hyperspectral images, using a reference Spectral Library (SL).The SL is structured with three main layers: main consisting material, semantic meaning, and illumination conditions in order to classify the endmembers. The labeling process is based on twosimilarity metrics: the Spectral Angle Mapper and the Cross Correlation. These two metrics areinitially introduced in spectral inequalities which characterize each endmember of the reference SL.During the labeling procedure, inequalities should be respected for labeling the endmembers,otherwise the values of metrics are used. Updating of the SL refers to incorporation of the not labeled endmembers in the SL. The method can also enable the creation of SLs that are related to specific sensors and their characteristics. The proposed method is evaluated using CASI-550 radiance images.The methodology produced satisfactory results. The fourth methodology is a new methodology for the discrimination of plant species and their varieties using hyperspectral data. The concept lies on the combination of spectral pre-processing algorithms (SPPA) that enhance spectral discrimination between species and their varieties. SPPA use as input a single spectral signature and transform it according to the SPPA function. A k-step combination of SPPA uses k pre-processing algorithms serially. Initially each spectral signature is used as input to the first SPPA. The result of this SPPA is used as input to the second SPPA, and so on until the desi red pre-processed signatures are reached. These signatures are then discriminated by applying spectral matching algorithms. The performance of the combination is evaluated based on the number of correctly matched signatures. In this work a k-step combination of SPPA has been set up, with k ranging from 1 to 3. Sixteen SPPAs have been investigated in total. There is a very large number of possible combinations of the aforementioned SPPAs, thus a Simple Genetic Algorithm has been used for finding optimum combinations. The evaluation set included the spectral signatures of 9 varieties of vetches and 9 varieties of lentils, measured by the GER1500 spectroradiometer. For all the samples, the spectral signatures were measured at two slightly different times in the growing season. The results showed that several combinations exist which can successfully discriminate and label the spectral signatures in terms of variety, and they are independent from the time of the spectral signature measurement.Continuing on the k-step methodology, the fifth method presents a new algorithm for automatically determining the optimum regression model, which enables the estimation of a parameter. The combination of SPPAs are used as input to three different regression methods: the Normalized band Difference Regression (NDR), the Multiple Linear Regression (MLR) and the Partial Least Squares Regression (PLSR). Three Simple Genetic Algorithms (SGAs) are used, one for each regression method, for the selection of the optimum combination of k SPPAs. The performance of the SGAs is evaluated based on the RMS error of the regression models. The evaluation not only indicates the selection of the optimum SPPA combination but also the regression method that produces the optimum prediction model. The proposed method was applied on soil spectral measurements in order to predict Soil Organic Matter (SOM). In this study, the maximum value assigned to k was 3. PLSR yielded the highest accuracy while NDR’s accuracy was satisfactory compared to its complexity. MLR method showed several drawbacks due to the presence of noise in terms of collinearity at the spectral bands. Most of the regression methods required a 3-step combination of SPPAs for achieving the highest performance. The selected preprocessing algorithms were different for each regression method since each regression method handles with a different way the explanatory variables.Last but not least, the method called Normalized Proximity and Similarity Methodology (NPSM)is proposed. NPSM is a new automatic methodology for radiometric normalization of multi-temporalhyperspectral images. The concept of a relative radiometric normalization (RRN) for multitemporalmultispectral and hyperspectral images is investigated. The NPS methodology introduces twospectral metrics, the absolute normalized difference (AND) and the absolute normalized ratiodifference (ANRD), to select Invariant Pixels (IPs). A new metric is also introduced for the evaluation of the proposed methodology: the root mean square error (RMSE) between the values of the initial subject image and the relevant reconstructed subject image. The NPSM is validated using 9 ASTER and 14 Hyperion images. Results demonstrate the high performance of the methodology on both hyperspectral and multispectral data.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2016
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Sykas, Dimitrios
- Συκάς, Δημήτριος
Subjects
dc:subject × 21- Τηλεπισκόπηση
- Επεξεργασία υπερφασματικής δεδομένων
- Φασαμτικός διαχωρισμός
- Διαχωρισμός
- Φασματοσκοπία
- Σχετική ραδιομετρική κανονικοποίηση
- Φασματική ταύτιση
- Remote sensing
- Hyperspectral processing
- Spectral unmixing
- Spectroscopy
- Relative radiometric normalization
- Spectral matching
- Φυσικές Επιστήμες
- Γεωεπιστήμες και Επιστήμες Περιβάλλοντος
- Επιστήμες Μηχανικού και Τεχνολογία
- Άλλες Επιστήμες Μηχανικού και Τεχνολογίες
- Natural Sciences
- Earth and Related Environmental Sciences
- Engineering and Technology
- Other Engineering and Technologies
Rights
dc:rights- Statement dc:rights
-
- BY_NC_ND
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/44469