Fuan Tsai

Derivative Analysis of Hyperspectral Data

Abstract With the goal of applying derivative spectral analysis to analyze high-resolution, spectrally continuous remote sensing data, several smoothing and derivative computation algorithms have been reviewed and modified to develop a set of cross-platform spectral analysis tools. Emphasis was placed on exploring different smoothing and derivative algorithms to extract spectral details from spectral data sets. A modular program was created to perform interactive derivative analysis. This module calculated derivatives using either a convolution (Savitzky–Golay) or finite divided difference approximation algorithm. Spectra were smoothed using one of the three built-in smoothing algorithms (Savitzky–Golay smoothing, Kawata–Minami smoothing, and mean-filter smoothing) prior to the derivative computation procedures. Laboratory spectral data were used to test the performance of the implemented derivative analysis module. An algorithm for detecting the absorption band positions was executed on synthetic spectra and a soybean fluorescence spectrum to demonstrate the usage of the implemented modules in extracting spectral features. Issues related to smoothing and spectral deviation caused by the smoothing or derivative computation algorithms were also observed and are discussed. A scaling effect, resulting from the migration of band separations when using the finite divided difference approximation derivative algorithm, can be used to enhance spectral features at the scale of specified sampling interval and remove noise or features smaller than the sampling interval.

A derivative-aided hyperspectral image analysis system for land-cover classification

The large number of spectral bands in hyperspectral data seriously complicates their use for classification. Selection of a useful subset of bands or derived features (spectral ratios, differences, derivatives) is always desirable, strongly affects the accuracy of the classification, and is often a practical necessity to keep the processing speed and memory requirements under control. This paper examines one possible procedure for selecting spectral derivatives to improve supervised classification of hyperspectral images. The procedure is designed to identify derivative features that are more effective at separating target classes and then add them to a base subset of features for classification. The goal is to create the smallest set of features that will result in the best classification result. A key issue in this process is the interplay of the number of features and the size of the training data sets since classification accuracy declines if the dimensionality of the feature space is too large relative to the number of training samples.

Spectrally segmented principal component analysis of hyperspectral imagery for mapping invasive plant species

Principal component analysis (PCA) is one of the most commonly adopted feature reduction techniques in remote sensing image analysis. However, it may overlook subtle but useful information if applied directly to the analysis of hyperspectral data, especially for discriminating between different vegetation types. In order to accurately map an invasive plant species (horse tamarind, Leucaena leucocephala) in southern Taiwan using Hyperion hyperspectral imagery, this study developed a spectrally segmented PCA based on the spectral characteristics of vegetation over different wavelength regions. The developed algorithm can not only reduce the dimensionality of hyperspectral imagery but also extracts helpful information for differentiating more effectively the target plant species from other vegetation types. Experiments conducted in this study demonstrated that the developed algorithm performs better than correlation‐based segmented principal component transformation (SPCT) and conventional PCA (overall accuracy: 86%, 76%, 66%; kappa value: 0.81, 0.69, 0.57) in detecting the target plant species, as well as mapping other vegetation covers.

Striping Noise Detection and Correction of Remote Sensing Images

This paper presents an image destriping system for correcting striping noise of remote-sensing images. The developed system identifies stripe positions based on edge-detection and line-tracing algorithms. Pixels not affected by striping are used as control points to construct cubic spline functions describing spatial gray level distributions of an image. Detected stripes are corrected by replacing the pixels with more reasonable gray values computed from constructed spline functions. Gray values of clean pixels not affected by stripes are not altered to preserve data genuineness. Several experimental results demonstrate that the developed system can correctly detect stripes in remote-sensing images and effectively repair them. Evaluations of the results based on an quantitative image quality index indicate that the image quality has been improved significantly after destriping. The destriped images are not only visually more plausible but also can provide better interpretability and are more suitable for computerized analysis.

Feature Extraction of Hyperspectral Image Cubes Using Three-Dimensional Gray-Level Cooccurrence

This paper presents a novel approach for the feature extraction of hyperspectral image cubes. In this paper, hyperspectral image cubes are treated as volumetric data sets. Features that are most helpful in separating different targets are effectively extracted from the hyperspectral image cubes using a newly developed high-order texture analysis method. The traditional texture measure of the gray-level cooccurrence matrix is extended to a 3-D tensor field to explore the complicated volumetric data more effectively and to extract discriminant features for better classification. As the kernel size is one of the most important parameters in statistics-based texture analysis, a semivariance analysis and a spectral separability measure are used to determine the most appropriate kernel size in the spatial and spectral domains, respectively, for computing 3-D gray-level cooccurrence. In addition, a few statistical indexes are also extended to third-order forms in order to calculate quantitative texture properties of the generated cooccurrence tensor field. An airborne hyperspectral data set and an EO-1 Hyperion image are used to test the performance of the developed algorithms. Experimental results indicate that the developed 3-D texture analysis outperforms conventional second-order texture descriptors and the support vector machine-based classifier in supervised classifications of both hyperspectral data sets.

Field determination of optimal dates for the discrimination of invasive wetland plant species using derivative spectral analysis

Mapping invasive plant species in aquatic and terrestrial ecosystems helps to understand the causes of their progression, manage some of their negative consequences, and control them. In recent years, a variety of new remote-sensing techniques, like Derivative Spectral Analysis (DSA) of hyperspectral data, have been developed to facilitate this mapping. A number of questions related to these techniques remain to be addressed. This article attempts to answer one of these questions: Is the application of DSA optimal at certain times of the year? Field radiometric data gathered weekly during the summer of 1999 at selected field sites in upstate New York, populated with purple loosestrife (Lythrum salicaria L.), common reed (Phragmites australis (Cav.)) and cattail (Typha L.) are analyzed using DSA to differentiate among plant community types. First, second and higher-order derivatives of the reflectance spectra of nine field plots, varying in plant composition, are calculated and analyzed in detail to identify spectral ranges in which one or more community types have distinguishing features. On the basis of the occurrence and extent of these spectral ranges, experimental observations suggest that a satisfactory differentiation among community types was feasible on 30 August, when plants experienced characteristic phenological changes (transition from flowers to seed heads). Generally, dates in August appear optimal from the point of view of species differentiability and could be selected for image acquisitions. This observation, as well as the methodology adopted in this article, should provide a firm basis for the acquisition of hyperspectral imagery and for mapping the targeted species over a broad range of spatial scales.

3D computation of gray level co-occurrence in hyperspectral image cubes

This study extended the computation of GLCM (gray level co-occurrence matrix) to a three-dimensional form. The objective was to treat hyperspectral image cubes as volumetric data sets and use the developed 3D GLCM computation algorithm to extract discriminant volumetric texture features for classification. As the kernel size of the moving box is the most important factor for the computation of GLCM-based texture descriptors, a three-dimensional semi-variance analysis algorithm was also developed to determine appropriate moving box sizes for 3D computation of GLCM from different data sets. The developed algorithms were applied to a series of classifications of two remote sensing hyperspectral image cubes and comparing their performance with conventional GLCM textural classifications. Evaluations of the classification results indicated that the developed semi-variance analysis was effective in determining the best kernel size for computing GLCM. It was also demonstrated that textures derived from 3D computation of GLCM produced better classification results than 2D textures.

Texture augmented analysis of high resolution satellite imagery in detecting invasive plant species

Abstract During recent decades, a considerable number of alien species have been brought into Taiwan and have caused significant impacts to local ecosystems and biodiversity. High resolution satellite imagery can provide detailed spatial characteristics over a large area and has a great potential for accurate vegetation mapping. However, most traditional multispectral image classification techniques focus on spectral discrimination of ground objects and may overlook useful spatial information provided by high resolution images. To achieve the best result, analysis of high resolution imagery should also incorporate spatial variations of the data. Therefore, this paper has looked into using a texture augmented procedure to analyze a high resolution satellite (QuickBird) image in order to detect an invasive plant species (Leucaena leucocephala) in southern Taiwan. Samples of primary vegetation covers were selected from the image to determine suitable texture analysis parameters for extracting texture features helpful for classification. Validation with ground truth data showed that the analysis produced high accuracies in detecting the target plant species and overall classification for primary vegetation types within the study site.

Polygon-based texture mapping for cyber city 3D building models

The three‐dimensional building model is one of the most important components in a cyber city implementation. Currently, however, most building models do not have sufficient and accurate texture information. The lack of texture not only makes 3D building models less realistic in visualization, it may also fail to provide needed information in intricate applications. This study developed a polygon‐based texture mapping system to produce near photo‐realistic texture mappings for 3D building models. Textures of building exteriors were generated from mosaics of close‐range photographs acquired with commodity digital cameras. The developed system integrated multiple digital photographs to create texture mosaics that were continuous in geometric outlines and smooth in colour shadings, and correctly mapped them onto corresponding building model façades. A test example demonstrated that the resultant building model had more complete and accurate texture features as well as a near‐photo‐realistic appearance.

LOD generation for 3d polyhedral building model

This paper proposes an algorithm for the automatic generation of Levels-of-Detail (LODs) for 3D polyhedral building models. In this study a group of connected polyhedrons is considered as “one building”, after which the generalization is applied to each building consecutively. The most detailed building models used is the polyhedral building model which allows for an elaborate roof structure, vertical walls and a polygonal ground plan. In the work the term “Pseudo-Continuous LODs” is described. The maximum distinguishable “feature resolution” can be estimated from the viewer distance to a building and is used to simplify the building structure by the polyhedron merging and wall collapsing with regularization processes. Experimental results demonstrate that the number of triangles can be reduced as a function of the feature resolution logarithm. Some case studies will be presented to illustrate the capability and feasibility of the proposed method including both regular and irregular shape of buildings.