This work describes a novel method of estimating statistically optimum pixel sizes for classification. Historically
more resolution, smaller pixel sizes, are considered better, but having smaller pixels can cause difficulties in
classification. If the pixel size is too small, then the variation in pixels belonging to the same class could be very
large. This work studies the variance of the pixels for different pixel sizes to try and answer the question of how
small, (or how large) can the pixel size be and still have good algorithm performance. Optimum pixel size is defined
here as the size when pixels from the same class statistically come from the same distribution. The work first derives
ideal results, then compares this to real data. The real hyperspectral data comes from a SOC-700 stand mounted
hyperspectral camera. The results compare the theoretical derivations to variances calculated with real data in order
to estimate different optimal pixel sizes, and show a good correlation between real and ideal data.
NASA’s EO-1 satellite, well into it’s second decade of operation, continues to provide multispectral and hyperspectral
data to the remote sensing community. The Hyperion pushbroom type hyperspectral spectrometer aboard EO-1 can be a
rich and useful source of high temporal resolution hyperspectral data. Unfortunately the Hyperion sensor suffers from
several issues including a low signal to noise ratio in many band regions as well as imaging artifacts. One artifact is the
presence of vertical striping, which, if uncorrected, limits the value of the Hyperion imagery. The detector array reads in
all spectral bands one spatial dimension (cross-track) at a time. The second spatial dimension (in-track) arises from the
motion of the satellite. The striping is caused by calibration errors in the detector array that appear as a vertical striping
pattern in the in-track direction. Because of the layout of the sensor array each spectral band exhibits it’s own
characteristic striping pattern, each of which must be corrected independently. Many current Hyperion destriping
algorithms focus on the correction of stripes by analyzing the column means and standard deviations of each band. The
more effective algorithms utilize windowing of the column means and interband correlation of these window means. The
approach taken in this paper achieves greater accuracy and effectiveness due to not only using local windowing in the
across track dimension but also along the in‐track. This allows detection of the striping patterns in radiometrically
homogeneous areas, providing improved detection accuracy.
KEYWORDS: Hyperspectral imaging, Image classification, Algorithm development, Fuzzy logic, Sensors, Data modeling, Remote sensing, FDA class I medical device development, Image sensors, Space sensors
This study presents image complexity measures applied to hyperspectral images and their relation to the percentage of
correct classification (PCC). Specifically, it studies the relationship between these metrics and the PCC for Maximum
Likelihood and Angle Detection classifiers. First, many complexity measures were studied to determine if there was a
relation between the measure and the PCC. Results showed a correlation of above 0.7 between complexity measures
based on entropy and uncertainty and the PCC of the classifiers mentioned above. Once the relation was established,
PCC estimators based on the metrics using simple and multiple regression models were designed. This design was
performed using data from both synthetic and real images. The real images were from two hyperspectral sensors, the
space based AISA and a portable SOC 700 hyperspectral sensor, and include scenes from the Enrique Reef in La
Parguera Puerto Rico. The models were then tested with real data. Results show that confidence intervals on the PCC
can be reliably obtained for real images.
Noise reduction algorithms for improving Raman spectroscopy signals while preserving signal information were
implemented. Algorithms based on Wavelet denoising and Kalman filtering are presented in this work as alternatives
to the well-known Savitky-Golay algorithm. The Wavelet and Kalman algorithms were designed based on
the noise statistics of real signals acquired using CCD detectors in dispersive spectrometers. Experimental results
show that the random noise generated in the data acquisition is governed by sub-Poisson statistics. The proposed
algorithms have been tested using both real and synthetic data, and were compared using Mean Squared Error
(MSE) and Infinity Norm (L∞) to each other and to the standard Savitky-Golay algorithm. Results show that
denoising based on Wavelets performs better in both the MSE and (L∞) the sense.
Charge-Coupled Device (CCD) detectors are becoming more popular in spectroscopy instrumentation. In spite of
technological advances, spurious signals and noise are unavoidable in Raman spectroscopes. In general, the noise
comes from two major sources, impulsive noise caused by high energy radiation from local or extraterrestrial
sources (cosmic rays), and noise produced in Raman backscattering estimation. In this work, two algorithms
for impulsive noise removal are presented, based in spectral and spatial features of the noise. The algorithms
combine pattern recognition and classical filtering techniques to identify the impulses. Once an impulse has been
identified, it is removed and substituted with data points having similar statistical properties as the surrounding
data.
This paper presents a ground truth data collection effort along with its use in evaluating unmixing
algorithms. Unmixing algorithms are typically evaluated using synthetic data generated by selecting
endmember spectrums and adding them in different amounts and with added noise. Going from synthetic to
real data poses many problems. One of the greatest is the amount of data to be collected. Also, there will be
many unmodeled variations in real data. These include greater variation of the endmembers, additional
endmembers that are a very small percentage of the image, and nonlinear effects in the data that are not
modeled. The data collation effort produced a high resolution class map along with spectral measurements
of 153 different sampling sites to validate the map. The methodology for using this high resolution class
map for generating the ground truth data for use in the unmixing algorithms is presented. Specifically, a 1m
class map is used to generate the endmember abundances for every pixel in a 30m Hyperion image of the
Enrique Reef in Southwest Puerto Rico. The results using two unmixing algorithms, one with a sum to one
constraint and the other with a non-negative constraint are presented. The unmixing results for each
endmember are presented along with a newly developed unmixing parameter called the Correct Unmixing
Index (CUI).
The Hyperspectral Image Analysis Toolbox (HIAT) is a collection of algorithms that extend the capability of
the MATLAB numerical computing environment for the processing of hyperspectral and multispectral imagery.
The purpose the Toolbox is to provide a suite of information extraction algorithms to users of hyperspectral
and multispectral imagery. HIAT has been developed as part of the NSF Center for Subsurface Sensing and
Imaging (CenSSIS) Solutionware that seeks to develop a repository of reliable and reusable software tools that can be shared by researchers across research domains. HIAT provides easy access to feature extraction/selection,
supervised and unsupervised classification algorithms, unmixing and visualization developed at Laboratory of
Remote Sensing and Image Processing (LARSIP). This paper presents an overview of the tools, application
available in HIAT using as example an AVIRIS image. In addition, we present the new HIAT developments,
unmixing, new oversampling algorithm, true color visualization, crop tool and GUI enhancement.
This work studies the end-to-end performance of hyperspectral classification and unmixing systems. Specifically, it compares widely used current state of the art algorithms with those developed at the University of Puerto Rico. These include algorithms for image enhancement, band subset selection, feature extraction, supervised and unsupervised classification, and constrained and unconstrained abundance estimation. The end to end performance for different combinations of algorithms is evaluated. The classification algorithms are compared in terms of percent correct classification. This method, however, cannot be applied to abundance estimation, as the binary evaluation used for supervised and unsupervised classification is not directly applicable to unmixing performance analysis. A procedure to evaluate unmixing performance is described in this paper and tested using coregistered data acquired by various sensors at different spatial resolutions. Performance results are generally specific to the image used. In an effort to try and generalize the results, a formal description of the complexity of the images used for the evaluations is required. Techniques for image complexity analysis currently available for automatic target recognizers are included and adapted to quantify the performance of the classifiers for different image classes.
Remote sensing is increasingly being used as a tool to quantitatively assess the location, distribution and relative health of coral reefs and other shallow aquatic ecosystems. As the use of this technology continues to grow and the analysis products become more sophisticated, there is an increasing need for comprehensive ground truth data as a means to assess the algorithms being developed. The University of Puerto Rico at Mayaguez (UPRM), one of the core partners in the NSF sponsored Center for Subsurface Sensing and Imaging Systems (CenSSIS), is addressing this need through the development of a fully-characterized field test environment on Enrique Reef in southwestern Puerto Rico. This reef area contains a mixture of benthic habitats, including areas of seagrass, sand, algae and coral, and a range of water depths, from a shallow reef flat to a steeply sloping forereef. The objective behind the test environment is to collect multiple levels of image, field and laboratory data with which to validate physical models, inversion algorithms, feature extraction tools and classification methods for subsurface aquatic sensing. Data collected from Enrique Reef currently includes airborne, satellite and field-level hyperspectral and multispectral images, in situ spectral signatures, water bio-optical properties and information on habitat composition and benthic cover. We present a summary of the latest results from Enrique Reef, discuss our concept of an open testbed for the remote sensing community and solicit other users to utilize the data and participate in ongoing system development.
Benthic habitats are the different bottom environments as defined by distinct physical, geochemical, and biological
characteristics. Remote sensing is increasingly being used to map and monitor the complex dynamics associated with
estuarine and nearshore benthic habitats. Advantages of remote sensing technology include both the qualitative benefits
derived from a visual overview, and more importantly, the quantitative abilities for systematic assessment and
monitoring. Advancements in instrument capabilities and analysis methods are continuing to expand the accuracy and
level of effectiveness of the resulting data products. Hyperspectral sensors in particular are rapidly emerging as a more
complete solution, especially for the analysis of subsurface shallow aquatic systems. The spectral detail offered by
hyperspectral instruments facilitates significant improvements in the capacity to differentiate and classify benthic
habitats. This paper reviews two techniques for mapping shallow coastal ecosystems that both combine the retrieval of
water optical properties with a linear unmixing model to obtain classifications of the seafloor. Example output using
AVIRIS hyperspectral imagery of Kaneohe Bay, Hawaii is employed to demonstrate the application potential of the two
approaches and compare their respective results.
A particular challenge in hyperspectral remote sensing of benthic habitats is that the signal exiting from the water is a small component of the overall signal received at the satellite or airborne sensor. Therefore, in order to be able to discriminate different ecological areas in benthic habitats, it is important to have a high signal to noise ratio (SNR). The SNR can be improved by building better sensors; SNR improvements however, we believe, are also achievable by means of signal processing and by taking advantage of the unique characteristics of hyperspectral sensors. One approach for SNR improvement is based on signal oversampling. Another approach for SNR improvement is Reduced Rank Filtering (RRF) where the small Singular Values of the image are discarded and then reconstruct a lower rank approximation to the original image. This paper presents a comparison in the use of oversampling filtering (OF) versus RRF as SNR enhancement methods in terms of classification accuracy and class separability when used as a pre-processing step in a classification system. Overall results show that OF does a better job improving the classification accuracy than RRF and at much lower computational cost, making it an attractive technique for Hyperspectral Image Processing.
This paper presents a comparison between classification results when hyperspectral imagery is pre-processed by spectral resolution enhancement algorithms and/or atmospheric correction algorithms. Different combinations of pre-processing options were investigated. Overall, resolution enhancement does improve classification accuracy with and without atmospheric correction. Furthermore, classification accuracy using radiance data enhanced by resolution enhancement techniques was higher than accuracies obtained by atmospherically corrected data even when it was enhanced. AVIRIS data from the Indian Pines test site in NW Indiana was used to illustrate the different concepts.
The spectrum of most objects in a hyperspectral image is oversampled in the spectral dimension due to the images having many closely spaced spectral samples. This oversampling implies that there is redundant information in the image which can be exploited to reduce the noise, and so increase the correct classification percentage. Oversampling techniques have been shown to be useful in the classification of hyperspectral imagery. These previous techniques consist of a lowpass filter in the spectral dimension whose characteristics are chosen based on the average spectral density of many objects to be classified. A better way of selecting the characteristics of the filter is to calculate the spectral density and oversampling of each object, and use that to determine the filter. The algorithm proposed here exploits the fact that the system is supervised to determine the oversampling rate, using the training samples for this purpose. The oversampling rate is used to determine the cutoff frequency for each class, and the highest of these is used to filter the whole image. Two pass approaches, where each class in the image is filtered with its own filter, were studied, but the increase in performance did not justify the increase in computational load. The results of applying these techniques to data to be classified are presented. The results, using AVIRIS imagery, show a significant improvement in classification performance.
This paper investigates if and how oversampling techniques can be applied in a useful manner to hyperspectral images. Oversampling occurs when the signal is sampled higher than the Nyquist frequency. If this occurs, the higher sampling rate can be traded for precision. Specifically, one bit of precision can be gained if the signal has been oversampled by a factor of four. This paper first investigates if spectral oversampling actually occurs in hyperspectral images, then looks at its usefulness in classification. Simulations were done with synthetic and real images. The results indicate that oversampling does occur for many real objects, so a knowledge of what is being searched for is crucial for knowing if oversampling techniques can be used. The classification results indicate that it takes a relatively large amount of noise for these techniques to have a significant impact on classification with synthetic images. For real images however, an improvement in classification for both supervised and unsupervised algorithms was observed for all simulations.
An important aspect of hyperspectral pattern recognition is selecting a subset of bands to perform the classification. This is generally necessary because the statistical algorithms on which classification is based need probabilistic estimates to work. The great number of spectral bands in hyperspectral images means that there is not enough data to accurately perform these estimates. In typical hyperspectral pattern recognition, the band selection and classification stages are done separately. This paper presents research done with an iterative system that integrates the band selection and classification. The objective is to choose an optimal subgroup of bands by maximizing the distance between the centroids of the classified data. The results of the study show that: (1) the algorithm correctly chooses the best bands based on centroid separability with synthetic data, (2) the system converges, and (3) the percentage of samples classified correctly using the iterative system is greater than the percentage using all the bands.
Lossless compression algorithms typically do not use spectral prediction, and typical algorithms that do, use only one adjacent band. Using one adjacent band has the disadvantage that if the last band compressed is needed, all previous bands must be decompressed. One way to avoid this is to use a few selected bands to predict the others. Exhaustive searches for band selection have a combinatorial problem, and are therefore not possible except in the simplest cases. To counter this, the use of a fast approximate method for band selection is proposed. The bands selected by this algorithm are a reasonable approximation to the principal components. Results are presented for exhaustive studies using entropy measures, sum of squared errors, and compared to the fast algorithm for simple cases. Also, it was found that using six bands selected by the fast algorithm produces comparable performance to one adjacent band.
Nonlinear predictors based on feedforward artificial neural networks are investigated for use in lossless compression of AVHRR Imagery. The FNN predictors are designed and compared to the optimum nonlinear Mean Square Error predictor, and to the linear predictor. The predictors are compared based on the first order entropy of the predictor error, on run time, and memory requirements. The FNN predictors can be designed to have a wide range of performance with a trade off between first order entropy error, and memory and run time. There is little difference in prediction errors between the linear and FNN predictors for large sample sizes, when the image is segmented into large areas. The difference can be greater for smaller sample sizes, when the image is segmented into smaller areas such as the typical 8 X 8 pixel size. The results indicate there is no advantage to using nonlinear predictors when compression and run time requirements are taken into account.
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.