Wonkook Kim

Adaptive Classification for Hyperspectral Image Data Using Manifold Regularization Kernel Machines

Localized training data typically utilized to develop a classifier may not be fully representative of class signatures over large areas but could potentially provide useful information which can be updated to reflect local conditions in other areas. An adaptive classification framework is proposed for this purpose, whereby a kernel machine is first trained with labeled data and then iteratively adapted to new data using manifold regularization. Assuming that no class labels are available for the data for which spectral drift may have occurred, resemblance associated with the clustering condition on the data manifold is used to bridge the change in spectra between the two data sets. Experiments are conducted using spatially disjoint data in EO-1 Hyperion images, and the results of the proposed framework are compared to semisupervised kernel machines.

Exploring Nonlinear Manifold Learning for Classification of Hyperspectral Data

Increased availability of hyperspectral data and greater access to advanced computing have motivated development of more advanced methods for exploitation of nonlinear characteristics of these data. Advances in manifold learning developed within the machine learning community are now being adapted for analysis of hyperspectral data. This chapter investigates the performance of popular global (Isomap and KPCA) and local manifold nonlinear learning methods (LLE, LTSA, LE) for dimensionality reduction in the context of classification. Experiments were conducted on hyperspectral data acquired by multiple sensors at various spatial resolutions over different types of land cover. Nonlinear dimensionality reduction methods often outperformed linear extraction methods and rivaled or were superior to those obtained using the full dimensional data.

Multiresolution manifold learning for classification of hyperspectral data

Nonlinear manifold learning algorithms assume that the original high dimensional data actually lie on a low dimensional manifold defined by local geometric distances between samples. Most of the traditional methods have focused only on the spectral distances in calculating the local dissimilarity of samples, whereas in the case of image data, the spatial distribution and localized contextual information of image samples could provide useful information. As a framework for integrating spatial and spectral information associated with image samples, a hierarchical spatial-spectral segmentation method is investigated for constructing the manifold structure. The new approach, which develops the manifold for the purpose of classification, incorporates an updating scheme whereby the spatial information and class labels are transferred through the segmentation hierarchy. It is applied to hyperspectral data collected by the Hyperion sensor on the EO-1 satellite over the Okavango Delta of Botswana. Classification accuracies and generalization capability are compared to those achieved by the best basis binary hierarchical classifier, the hierarchical support vector machine classifier, and the shortest path k-nearest neighbor classifier.

Spatially Adapted Manifold Learning for Classification of Hyperspectral Imagery with Insufficient Labeled Data

A classifier derived from labeled samples acquired over an extended area may not perform well for a specific sub-region if the spectral signatures of classes vary across the image. However, characterizing the local effects are an ill-posed problem, particularly for hyperspectral data, since an adequate number of labeled samples is not typically available for every location. This problem is addressed using semi-supervised learning and manifold learning, which both exploit the information provided by unlabeled samples in the image. A spatially adaptive classification method that uses Laplacian regularization is proposed, with the updating scheme using a combination of labeled and unlabeled samples.

Retrieval of Substrate Bearing Strength from Hyperspectral Imagery during the Virginia Coast Reserve (VCR’07) Multi-Sensor Campaign

Hyperspectral imagery (HSI) derived from remote sensing can delineate surface properties of substrates such as type, moisture, and grain size. These are critical parameters that determine the substrate bearing strength. Although HSI only sees the surface layer, statistics can be derived that relate surface properties to the likely bearing strength of soils in particular regions. This information can be used to provide an initial map estimate on large scales of potential bearing strength. We describe an initial validation study at the Virginia Coast Reserve relating airborne HSI to in situ spectral and geotechnical measurements through a spectral-geotechnical lookup table (LUT).

A novel adaptive classification method for hyperspectral data using manifold regularization kernel machines

Remote sensing data sets are often difficult to compare directly due to environmental changes between acquisitions of two data sets. This paper proposes an adaptive framework for robust classification when no reference data are available in a new area or time period. Labels of test data are recovered during iterative applications of kernel machines by reflecting geometry of unlabeled samples via the manifold regularization term, so that the labeled/unlabeled samples form clusters on the data manifold. A one-against-one scheme is used for the extension of the binary classifier to multiclass problems, where semi-labels are used for iterative training of classifier. The proposed method is applied to a series of data pair of Hyperion and AVIRIS hyperspectral data and compared to other non-adaptive classification methods.

Manifold Learning for Multi-classifier Systems via Ensembles

Statistical classification of hyperspectral data is challenging because the inputs are high in dimension, while the quantity of labeled data is typically limited. The resulting classifiers are often unstable and have poor generalization. Nonlinear manifold learning algorithms assume that the original high dimensional data actually lie on a low dimensional manifold defined by local geometric differences between samples. Recent research has demonstrated the potential of these approaches for nonlinear dimension reduction and representation of high dimensional observations. Nonlinear scattering phenomena associated with processes observed in remote sensing data suggest that these may be useful for analysis of hyperspectral data. However, computational requirements limit their applicability for classification of remotely sensed data. Multi-classifier systems potentially provide a means to exploit the advantages of manifold learning through decomposition frameworks, while providing improved generalization. This paper reports preliminary results obtained from an ensemble implementation of Landmark Isomap in conjunction with a kNN classifier. The goal is to achieve improved generalization of the classifier in analysis of hyperspectral data in a dynamic environment with limited training data. The new method is implemented and applied to Hyperion hyperspectral data collected over the Okavango Delta of Botswana.

Correction of Stray-Light-Driven Interslot Radiometric Discrepancy (ISRD) Present in Radiometric Products of Geostationary Ocean Color Imager (GOCI)

The radiometric calibration of satellite data is critical in many environmental studies and applications that are based on remote sensing data. The Geostationary Ocean Color Imager (GOCI) has suffered from what is called an interslot radiometric discrepancy (ISRD), which creates clear inconsistency between the adjacent slots in GOCI Level 1B (L1B) radiometric products, the largest source of which is currently believed to be the stray light generated in the sensor instrument. Difficulties in removing the stray-light-driven anomalies are that the intensity and the spatial extent vary with time and location, depending on the reflectance of nearby bright targets, such as cloud and land. This paper proposes an image-based correction method that removes the stray-light-driven radiometric inflation without involving an independent reference so that the method can be used for GOCI operational data processing. First, the radiometric inflation pattern is characterized by independent sources, such as Moderate Resolution Imaging Spectrometer (MODIS) data, and the inflation pattern is modeled by the minimum noise fraction transform of the input data. The modeled inflation patterns in individual slots are then adjusted across the slots in such a way that the overall ISRD in all slot boundaries is minimized. The analysis shows that the stray-light-driven radiometric anomalies can be up to 20% of the normal signals in Bands 6 (680 nm) and 8 (865 nm) of the uncorrected L1B images, and the proposed correction method reduces it to less than 2% in most of the cases, recovering the spatial continuity of natural variability across the slots.

Remote Sensing, Public Health & Disaster Mitigation

The authors review advances in applications for geotechnologies, specifically earth-observing satellite remote sensing, geo-positioning (i.e. USA’s Global Positioning System (GPS), Russia’s Global’naya Navigatsionnaya Sputnikovaya Sistema (GLONASS), Europe’s Galileo and China’s Beidou/Compass) and selected geo-spatial modeling software for public health and disaster management applications, with an emphasis on environmental health and environmental sustainability. Specific applications addressed include the use of remote sensing for infectious disease vector habitat identification and ecologically sustainable disease vector population mitigation, as well as the integration of GPS into mobile CD4 testing devices for HIV/AIDS. Public domain software models described include the Spatio-Temporal Epidemiological Modeler (STEM) and the Hydrologic Engineering River Analysis System (HEC-RAS) for flood modeling. Examples of regional, national and global real-time data acquisition and near-real-time data product development and distribution for time-critical events are offered, specifically through the Purdue Terrestrial Observatory (PTO), the United States Geological Survey (USGS) supported AmericaView and the International Charter – Space & Major Disasters.

Remote quantification of Cochlodinium polykrikoides blooms occurring in the East Sea using geostationary ocean color imager (GOCI)

Accurate and timely quantification of widespread harmful algal bloom (HAB) distribution is crucial to respond to the natural disaster, minimize the damage, and assess the environmental impact of the event. Although various remote sensing-based quantification approaches have been proposed for HAB since the advent of the ocean color satellite sensor, there have been no algorithms that were validated with in-situ quantitative measurements for the red tide occurring in the Korean seas. Furthermore, since the geostationary ocean color imager (GOCI) became available in June 2010, an algorithm that exploits its unprecedented observation frequency (every hour during the daytime) has been highly demanded to better track the changes in spatial distribution of red tide. This study developed a novel red tide quantification algorithm for GOCI that can estimate hourly chlorophyll-a (Chl a) concentration of Cochlodinium (Margalefidinium) polykrikoides, one of the major red tide species around Korean seas. The developed algorithm has been validated using in-situ Chl a measurements collected from a cruise campaign conducted in August 2013, when a massive C. polykrikoides bloom devastated Korean coasts. The proposed algorithm produced a high correlation (R2=0.92) with in-situ Chl a measurements with robust performance also for high Chl a concentration (300mg/m3) in East Sea areas that typically have a relatively low total suspended particle concentration (<0.5mg/m3).