京都大学 / 0048 / 新制・課程博士 / 博士(農学) / 甲第19771号 / 農博第2167号 / 新制||農||1040(附属図書館) / 学位論文||H28||N4987(農学部図書室) / 32807 / 京都大学大学院農学研究科地域環境科学専攻 / (主査)教授 飯田 訓久, 教授 近藤 直, 准教授 中村 公人 / 学位規則第4条第1項該当 / Doctor of Agricultural Science / Kyoto University / DFAM

The compressed sensing (CS) model of signal processing, while offering many unique advantages in terms of low-cost sensor design, poses interesting challenges for both signal acquisition and recovery, especially for signals of large size. In this work, we investigate how CS might be applied practically and efficiently in the context of natural video. We make use of a CS video acquisition approach in line with the popular single-pixel camera framework of blind, nonaptive, random sampling while proposing new approaches for the subsequent recovery of the video signal which leverage interrame redundancy to minimize recovery error. We introduce a method of approximation, which we term multihypothesis (MH) frame prediction, to create accurate frame predictions by comparing hypotheses drawn from the spatial domain of chosen reference frames to the non-overlapping, block-by-block CS measurements of subsequent frames. We accomplish this frame prediction via a novel distance-weighted Tikhonov regularization technique. We verify through our experiments that MH frame prediction via distance-weighted regularization provides state-of-the-art performance for the recovery of natural video sequences from blind CS measurements. The distance-weighted regularization we propose need not be limited to just frame prediction for CS video recovery, but may also be used in a variety of contexts where approximations must be generated from a set of hypotheses or training data. To show this, we apply our technique to supervised hyperspectral image (HSI) classification via a novel classifier we term the nearest regularized subspace (NRS) classifier. We show that the distance-weighted regularization used in the NRS method provides greater classification accuracy than state-of-the-art classifiers for supervised HSI classification tasks. We also propose two modifications to the core NRS classifier to improve its robustness to variation of input parameters and and to further increase its classification accuracy.

No description available.

Geoscience applications generate large datasets; thus, compression is necessary to facilitate the storage and transmission of geoscience data. One focus is on the coding of hyperspectral imagery and the prominent JPEG2000 standard. Certain aspects of the encoder, such as rate-allocation between bands and spectral decorrelation, are not covered by the JPEG2000 standard. This thesis investigates the performance of several JPEG2000 encoding strategies. Additionally, a relatively low-complexity 3D embedded wavelet-based coder, 3D-tarp, is proposed for the compression of geoscience data. 3D-tarp employs an explicit estimate of the probability of coefficient significance to drive a nonadaptive arithmetic coder, resulting in a simple implementation suited to vectorized hardware acceleration. Finally, an embedded wavelet-based coder is proposed for the shapeaptive coding of ocean-temperature data. 3D binary set-splitting with $k$-d trees, 3D-BISK, replaces the octree splitting structure of other shapeaptive coders with $k$-d trees, a simpler set partitioning structure that is well-suited to shapeaptive coding.

Because of the vast data volume of hyperspectral imagery, compression becomes a necessary process for hyperspectral data transmission, storage, and analysis. Three-dimensional discrete wavelet transform (DWT) based algorithms are particularly of interest due to their excellent rate-distortion performance. This thesis investigates several issues surrounding efficient compression using JPEG2000. Firstly, the rate-distortion performance is studied when Principal Component Analysis (PCA) replaces DWT for spectral decorrelation with the focus on the use of a subset of principal components (PCs) rather than all the PCs. Secondly, the algorithms are evaluated in terms of data analysis performance, such as anomaly detection and linear unmixing, which is directly related to the useful information preserved. Thirdly, the performance of compressing radiance and reflectance data with or without bad band removal is compared, and instructive suggestions are provided for practical applications. Finally, low-complexity PCA algorithms are presented to reduce the computational complexity and facilitate the future hardware design.

No description available.

This work aims to develop a deep learning model for land cover quantification through hyperspectral unmixing using an unsupervised autoencoder. Land cover identification and classification is instrumental in urban planning, environmental monitoring and land management. With the technological advancements in remote sensing, hyperspectral imagery which captures high resolution images of the earth's surface across hundreds of wavelength bands, is becoming increasingly popular. The high spectral information in these images can be analyzed to identify the various target materials present in the image scene based on their unique reflectance patterns. An autoencoder is a deep learning model that can perform spectral unmixing by decomposing the complex image spectra into its constituent materials and estimating their abundance compositions. The advantage of using this technique for land cover quantification is that it is completely unsupervised and eliminates the need for labelled data which generally requires years of field survey and formulation of detailed maps. We evaluate the performance of the autoencoder on various synthetic and real hyperspectral images consisting of different land covers using similarity metrics and abundance maps. The scalability of the technique with respect to landscapes is assessed by evaluating its performance on hyperspectral images spanning across 100m x 100m, 200m x 200m, 1000m x 1000m, 4000m x 4000m and 5000m x 5000m regions. Finally, we analyze the performance of this technique by comparing it to several supervised learning methods like Support Vector Machine (SVM), Random Forest (RF) and multilayer perceptron using F1-score, Precision and Recall metrics and other unsupervised techniques like K-Means, N-Findr, and VCA using cosine similarity, mean square error and estimated abundances. The land cover classification obtained using this technique is compared to the existing United States National Land Cover Database (NLCD) classification standard. / Master of Science / This work aims to develop an automated deep learning model for identifying and estimating the composition of the different land covers in a region using hyperspectral remote sensing imagery. With the technological advancements in remote sensing, hyperspectral imagery which captures high resolution images of the earth's surface across hundreds of wavelength bands, is becoming increasingly popular. As every surface has a unique reflectance pattern, the high spectral information contained in these images can be analyzed to identify the various target materials present in the image scene. An autoencoder is a deep learning model that can perform spectral unmixing by decomposing the complex image spectra into its constituent materials and estimate their percent compositions. The advantage of this method in land cover quantification is that it is an unsupervised technique which does not require labelled data which generally requires years of field survey and formulation of detailed maps. The performance of this technique is evaluated on various synthetic and real hyperspectral datasets consisting of different land covers. We assess the scalability of the model by evaluating its performance on images of different sizes spanning over a few hundred square meters to thousands of square meters. Finally, we compare the performance of the autoencoder based approach with other supervised and unsupervised deep learning techniques and with the current land cover classification standard.

Approved for public release; distribution in unlimited. / PC and MNF transforms are two widely used methods that are utilized for various applications such as dimensionality reduction, data compression and noise reduction. In this thesis, an in-depth study of these two methods is conducted in order to estimate their performance in hyperspectral imagery. First the PCA and MNF methods are examined for their effectiveness in image enhancement. Also, the various methods are studied to evaluate their ability to determine the intrinsic dimension of the data. Results indicate that, in most cases, the scree test gives the best measure of the number of retained components, as compared to the cumulative variance, the Kaiser, and the CSD methods. Then, the applicability of PCA and MNF for image restoration are considered using two types of noise, Gaussian and periodic. Hyperspectral images are corrupted by noise using a combination of ENVI and MATLAB software, while the performance metrics used for evaluation of the retrieval algorithms are visual interpretation, rms correlation coefficient spectral comparison, and classification. In Gaussian noise, the retrieved images using inverse transforms indicate that the basic PC and MNF transform perform comparably. In periodic noise, the MNF transform shows less sensitivity to variations in the number of lines and the gain factor. / Lieutenant, Hellenic Navy

One of the most important parts of engineered and biological systems is the ability to acquire and interpret information from the surrounding world accurately and in time-scales relevant to the tasks critical to system performance. This classical concept of efficient signal acquisition has been a cornerstone of signal processing research, spawning traditional sampling theorems (e.g. Shannon-Nyquist sampling), efficient filter designs (e.g. the Parks-McClellan algorithm), novel VLSI chipsets for embedded systems, and optimal tracking algorithms (e.g. Kalman filtering). Traditional techniques have made minimal assumptions on the actual signals that were being measured and interpreted, essentially only assuming a limited bandwidth. While these assumptions have provided the foundational works in signal processing, recently the ability to collect and analyze large datasets have allowed researchers to see that many important signal classes have much more regularity than having finite bandwidth. One of the major advances of modern signal processing is to greatly improve on classical signal processing results by leveraging more specific signal statistics. By assuming even very broad classes of signals, signal acquisition and recovery can be greatly improved in regimes where classical techniques are extremely pessimistic. One of the most successful signal assumptions that has gained popularity in recet hears is notion of sparsity. Under the sparsity assumption, the signal is assumed to be composed of a small number of atomic signals from a potentially large dictionary. This limit in the underlying degrees of freedom (the number of atoms used) as opposed to the ambient dimension of the signal has allowed for improved signal acquisition, in particular when the number of measurements is severely limited. While techniques for leveraging sparsity have been explored extensively in many contexts, typically works in this regime concentrate on exploring static measurement systems which result in static measurements of static signals. Many systems, however, have non-trivial dynamic components, either in the measurement system's operation or in the nature of the signal being observed. Due to the promising prior work leveraging sparsity for signal acquisition and the large number of dynamical systems and signals in many important applications, it is critical to understand whether sparsity assumptions are compatible with dynamical systems. Therefore, this work seeks to understand how dynamics and sparsity can be used jointly in various aspects of signal measurement and inference. Specifically, this work looks at three different ways that dynamical systems and sparsity assumptions can interact. In terms of measurement systems, we analyze a dynamical neural network that accumulates signal information over time. We prove a series of bounds on the length of the input signal that drives the network that can be recovered from the values at the network nodes~[1--9]. We also analyze sparse signals that are generated via a dynamical system (i.e. a series of correlated, temporally ordered, sparse signals). For this class of signals, we present a series of inference algorithms that leverage both dynamics and sparsity information, improving the potential for signal recovery in a host of applications~[10--19]. As an extension of dynamical filtering, we show how these dynamic filtering ideas can be expanded to the broader class of spatially correlated signals. Specifically, explore how sparsity and spatial correlations can improve inference of material distributions and spectral super-resolution in hyperspectral imagery~[20--25]. Finally, we analyze dynamical systems that perform optimization routines for sparsity-based inference. We analyze a networked system driven by a continuous-time differential equation and show that such a system is capable of recovering a large variety of different sparse signal classes~[26--30].

La première partie de ce travail présente un état de l'art des principaux critères non supervisés, non paramétriques, d'évaluation d'une partition, des méthodes d'estimation préliminaires du nombre de classes, et enfin des méthodes de classification supervisées, semi-supervisées et non supervisées. Une analyse des avantages et des inconvénients de ces critères et méthodes est menée. L'analyse des performances des méthodes de classification et des critères d'évaluation a été également conduite via l'application visée dans cette thèse. Une approche de partitionnement non supervisée, non paramétrique et hiérarchique s'avère la plus adaptée au problème posé. En effet, ce type d'approche et plus particulièrement la classification descendante donne un partitionnement à plusieurs niveaux et met en évidence des informations plus détaillées d'un niveau à l'autre, ce qui permet une meilleure interprétation de la richesse d'information apportée par l'imagerie hyperspectrale et ainsi conduire à une meilleure décision. Dans ce sens, la deuxième partie de cette thèse présente, tout d'abord l'approche de classification descendante hiérarchique non supervisée (CDHNS) développée. Cette approche non paramétrique, permet l'obtention de résultats stables et objectifs indépendamment des utilisateurs finaux. Le second développement conduit, porte sur la sélection de bandes spectrales parmi celles qui composent l'image hyperspectrale originale afin de réduire la quantité d'information à traiter avant le processus de classification. Cette méthode est également non supervisée et non paramétrique. L'approche de classification et la méthode de réduction ont été expérimentées et validées sur une image hyperspectrale synthétique construite à partir des images réelles puis sur des images réelles dont l'application porte sur l'identification des différentes classes algales. Les résultats de partitionnement obtenus sans réduction montrent d'une part, la stabilité des résultats et, d'autre part, la discrimination des classes principales (végétation, substrat et eau) dès les premiers niveaux. Les résultats de la sélection des bandes spectrales font apparaître leur bonne répartition sur toute la gamme spectrale du capteur (visible et proche-infrarouge). Les résultats montrent aussi que le partitionnement avec et sans réduction sont globalement similaires. De plus, le temps de calcul est fortement réduit. / The upstream location of the different algal species causing clogging in the EDF nuclear power plants cooling systems along the Channel coastline, by analyzing hyperspectral aerial image is today the most appropriate means. Indeed, hyperspectral imaging allows, through its spatial resolution and its broad spectral range covering the areas of visible and near infrared, the objective discrimination of plant species on the foreshore, necessarily yielding accurate maps on large coastal areas. To provide a solution to this problem and achieve the objectives, the work conducted within the framework of this thesis lies in the development of unsupervised partitioning approaches to data with large spectral and spatial dimensions. The first part of this work presents a state of the art of main unsupervised criteria, and nonparametric, for partitioning evaluation, the preliminary methods for estimating the number of classes, and finally, supervised, semi-supervised and unsupervised classification methods. An analysis of the advantages and drawbacks of these methods and criteria is conducted. The analysis of the performances of these classification methods and evaluation criteria was also conducted through the application targeted in this thesis. An unsupervised, nonparametric, hierarchical partitioning approach appears best suited to the problem. Indeed, this type of approach, and particularly the descending classification, gives a partitioning at several levels and highlights more detailed information from one level to another, allowing a better interpretation of the wealth of information provided by hyperspectral imaging and therefore leading to a better decision. In this sense, the second part of this thesis presents, firstly the unsupervised hierarchical descending classification (UHDC) approach developed. This nonparametric approach allows obtaining stable and objective results regardless of end users. The second development proposed concerns the selection of spectral bands from those that make up the original hyperspectral image, in order to reduce the amount of information to be processed before the classification process. This method is also unsupervised and nonparametric. The classification approach and the reduction method have been tested and validated on a synthetic hyperspectral image constructed from real images, and then on real images, with application to the identification of different algal classes. The partitioning results obtained without reduction show firstly, the stability of the results and, secondly, the discrimination of the main classes (vegetation, substrate and water) from the first levels. The results of the spectral bands selection method show that the retained bands are well distributed over the entire spectral range of the sensor (visible and near-infrared). The results also show that partitioning results with and without reduction are broadly similar. Moreover, the computation time is greatly reduced.