High Performance Computing in Remote Sensing - Chapter 6

Chapter 6 Computing for Analysis and Modeling of Hyperspectral Imagery Gregory P. Asner, Carnegie Institution of Washington Robert S. Haxo, Carnegie Institution of Washington David E. Knapp, Carnegie Institution of Washington Contents 6.1 Introduction ........................................................... 110 6.2 Hyperspectral Imagery and Analysis .................................... 112 6.2.1 Typical Imaging Systems ....................................... 112 6.2.2 Typical Analysis Steps .......................................... 114 6.2.3 Computing Challenges in Hyperspectral Analysis ................ 117 6.3 Meeting the Computing Demands ...................................... 120 6.3.1 High-Performance Computing Jobs ............................. 121 6.3.1.1 Job Class 1 – Independent Calculations on Different Data Sets ................................. 121 6.3.1.2 Job Class 2 – Independent Calculations on a Single Data Set ................................... 121 6.3.1.3 Job Class 3 – Dependent Calculations on a Single Data Set ................................... 122 6.3.2 Computing Architectures ....................................... 122 6.3.2.1 Cluster Compute Nodes ................................ 122 6.3.2.2 Cluster Front-End Computer ........................... 122 6.3.2.3 Networking ............................................ 123 6.3.2.4 Data Storage .......................................... 124 6.4 Future Possibilities of Hyperspectral Computing ........................ 125 6.4.1 Onboard Processing and Data Compression ...................... 125 6.4.2 Distributed Internet Computing ................................. 126 6.5 Acknowledgments ..................................................... 127 References .................................................................. 127 109 © 2008 by Taylor & Francis Group, LLC 110 High-Performance Computing in Remote Sensing Hyperspectral remote sensing is increasingly used for Earth observation and anal-ysis, but the large data volumes and complex analytical techniques associated with imagingspectroscopyrequirehigh-performancecomputingapproaches.Inthischap-ter, we highlight several analytical methods employed in vegetation and ecosystem studies using airborne and space-based imaging spectroscopy. We then summarize the most common high-performance computing approaches used to meet these ana-lytical demands, and provide examples from our own work with computing clusters. Finally,wediscussseveralemergingareasofhigh-performancecomputing,including data processing onboard aircraft and spacecraft and distributed Internet computing, that will change the way we carry out computations with high spatial and spectral resolution observations of ecosystems. 6.1 Introduction There is an increasing demand for high spatial and spectral resolution remote sensing data for environmental studies ranging from ecological dynamics of terrestrial and aquaticsystemstourbandevelopment.Agoodexampleishyperspectralremotesens-ing, also called imaging spectroscopy, which is rapidly advancing from the remote sensing research arena to a mapping and analysis science in support of conservation, management, and policy development. Hyperspectral remote sensing is the measure-ment,innarrowcontiguouswavelengthbands,ofsolarradiationreflectedbymaterials in the environment (Figure 6.1). These measurements express the chemical compo-sition and structural properties of the materials of interest. In industrial operations, spectroscopyisusedformaterialidentification,manufacturing,andqualityassurance. In Earth observation, imaging spectroscopy is used to estimate chemical concentra-tions and structures in vegetation, phytoplankton, soils, rocks, and a wide range of synthetic materials [1, 2]. The advent of hyperspectral remote sensing technology represents a progression from basic panchromatic and multispectral camera-like imaging of the past to a more data-rich and physically-based imaging and analysis arena for 21st century science. Field, airborne, and even space-based hyperspectral sensors are available today to government, commercial, and private organizations, yet the collection and analysis of imaging spectrometer data continue to be a challenge. Both the data volume and the processing techniques currently require a level of technological and scientific investment that is beyond the reach of many agencies and organizations. Continued effort is thus needed to advance the science of imaging spectroscopy from an esoteric specialty area to a mainstream set of applied methods for earth science. This is particularlytruetodayastheearthsciencecommunityischallengedtodemonstratethe societalbenefitofitsobservationsandstudies,especiallyfromexpensiveinvestments such as remote sensing. In this chapter, we summarize the major processing challenges and steps involved in hyperspectral image data collection and analysis, and we provide examples of © 2008 by Taylor & Francis Group, LLC Computing for Analysis and Modeling of Hyperspectral Imagery 111 4000 3000 Green Vegetation Dead Vegetation Bare Soil 2000 Water Body (lake) 1000 0 400 700 1000 1300 1600 1900 2200 2500 Wavelength (nm) Figure 6.1 Imaging spectrometers collect hyperspectral data such that each pixel contains a spectral radiance signature comprised of contiguous, narrow wavelength bands spanning a broad wavelength range (e.g., 400–2500 nm). Top shows a typical hyperspectralimagecube;eachpixelcontainsadetailedhyperspectralsignaturesuch as those shown at the bottom. © 2008 by Taylor & Francis Group, LLC 112 High-Performance Computing in Remote Sensing how high-performance computing is used to meet these challenges. We highlight why specific types of high-performance computing approaches are matched to the demands of different types of scientific algorithms employed for hyperspectral data analysis. Whereas some analytical methods require true parallel processing, others benefit from the strategic use of distributed computing techniques. We also look into thefuturebyoutliningaframeworkforprocessinghyperspectraldataonboardaircraft and spacecraft. Near-real-time processing of data is the next frontier in bringing hyperspectral imaging from a specialty to a mainstream science for environmental research and monitoring. 6.2 Hyperspectral Imagery and Analysis The term hyperspectral is used in a wide variety of ways by remote sensing practi-tioners, and there is some confusion in the literature as to what hyperspectral really means. Some call any imaging system with more than about 5–10 channels ‘hyper-spectral.’However,ahyperspectralsensorisasystemthatcollectsimageswithpixels containing a series of contiguous, narrowband spectral channels covering a particular region of the spectrum (e.g., 400–1050 nm or 400–2500 nm). The spectrum in each pixel provides information on absorption and scattering features of materials in that pixel. Because the data are collected in image format with each spectrum spatially lo-cated, the measurements are organized as 3-D ‘cubes’ that allow analysis of remotely sensedmaterialsinageographiccontext(Figure6.1).Inthissection,wediscusssome of the most common hyperspectral imaging systems and analytical techniques that subsequently demand high-performance computing techniques. 6.2.1 Typical Imaging Systems Imaging spectrometers are used to collect hyperspectral data from field, airborne, or space-based vantage points. Imaging spectrometers vary in design, but the two most commonsystemsemployeitherscanning‘whiskbroom’sensorsorpushbroomarrays (Table 6.1). The data vary in spatial resolution (ground instantaneous field-of-view; TABLE 6.1 The Basic Characteristics of Several Well-Known Imaging Spectrometers Sensor Airborne Visible/Infrared Imaging Spectrometer Compact Airborne Spectrographic Imager-1500 Digital Airborne Imaging Spectrometer Earth Observing-1 Hyperion HyMap Imaging Spectrometer PROBE-1 Bands 220 288 79 220 128 100–200 Range (nm) 360–2510 400–1050 400–12,000 400–2500 400–2500 400–2400 © 2008 by Taylor & Francis Group, LLC Computing for Analysis and Modeling of Hyperspectral Imagery 113 GIFOV)basedonflyingaltitude,aircraftspeed,scananddatarate,anddesiredsignal-to-noise(SNR)propertiesoftheimagery.SmallerGIFOVmeasurementsfromaircraft (e.g. < 5 m) require slower speed over ground, high sensor SNR, and fast data rates. Many applications require these high spatial resolution measurements, but few airborne and no space-based spectrometers can deliver the information. Probably the most well-known airborne sensor to do so is the NASA Airborne Visible and Infrared Imaging Spectrometer (AVIRIS), which can fly sufficiently low and slow to acquire data at about 1.8 m resolution; otherwise, this sensor is most often flown at higher altitudes to obtain data in the 4–20 m resolution range. AVIRIS is a very high fidelity spectrometer that collects spectra in 9.6 nm wide bands (full width at half maximum; FWHM) spanning the 360–2510 nm wavelength range and with a cross-track swath of614pixels[3,4].AVIRIShasbeencontinuouslyimprovedoverthepast15ormore yearsandisnowinitsfifthmajorversionforuseinEarthobservationandanalysis[4]. Anotherimagingspectrometerthathasundergonecontinualimprovementsoverthe years is the Compact Airborne Spectral Imager (CASI) [5]. CASI collects spectral data at 2.4 nm FWHM sampling across a wavelength range of about 400–1050 nm. The SNR and overall fidelity of CASI are now roughly similar to those of AVIRIS, but CASI stands as a unique instrument because of its programmability and very high spectralresolution.ItspushbroomarrayallowstheCASItobeoperatedatlowaltitude to achieve spatial resolutions of less than 1 m. The most recent version of CASI has 1500 cross-track pixels. Other spectrometers such as HyMap and PROBE-1 provide a near-contiguous spectral sampling of the 400–2500 nm wavelength range, but at differing operational spatial resolutions, SNRs, and swath widths. There are very few spaceborne imaging spectrometers available for scientific use; the space-based technology has not been made operational for the earth sciences. The earth Observing-1 (EO-1) satellite does carry the Hyperion imaging spectrom-eter (Table 6.1), which was placed in low Earth orbit in December 1999. EO-1 is a technology demonstration and thus does not provide large-scale coverage of the Earth’s surface; however, Hyperion data can be requested from the U.S. Geological Survey. The Hyperion imagery has a spatial resolution (GIFOV) of 30 m, and the spectra cover the 400–2500 nm wavelength region in 220 channels [6]. Hyperion is a pushbroom imager with relatively low signal-to-noise and image uniformity as compared to systems such as AVIRIS and HyMap, but it does provide a chance to test imaging spectroscopy concepts and analysis methods just about anywhere in the world. The data volumes associated with airborne hyperspectral data are significantly larger than typical multispectral data for a given spatial resolution and coverage. For a given length of distance flown, the width of the sensor scan can also affect the data volume. AVIRIS collects 640 pixels per scan, whereas the CASI-1500 collects 1500 pixels across its linear array. In the spectral domain, the AVIRIS and the CASI-1500 sensors contain 220 and 288 spectral bands per pixel, respectively, while the Landsat Thematic Mapper (TM) and Enhanced Thematic Mapper Plus (ETM+) sensors con-tain only seven multispectral bands. Furthermore, many multispectral sensors only have an 8-bit dynamic range of intensity per pixel per band, whereas AVIRIS and CASI data have a greater dynamic range and are stored as 14- and 16-bit values per © 2008 by Taylor & Francis Group, LLC ... - tailieumienphi.vn