Investigation of the Permafrost Table through Multi-resolution Object Oriented Fuzzy Analysis, North Slope, Alaska Justin L. Rich and Bea Csatho Department of Geology, University at Buffalo, The State University of New York, Buffalo, NY 14260 RSL Remote Sensing Laboratory Dept. of Geology, SUNY at Buffalo Data implemented for this project included an Advanced Land Imager and Landsat TM scene; as well as, a Digital Elevation Model with its derivative data. Hyperion hyperspectral data was also utilized for this project to obtain spectral characteristics of the ground surface along with field data collected with an ASD spectral radiometer. Destriping of the Hyperion image was conducted within ENVI and the Eclipse development platform utilizing the Java IDE. All images were atmospherically corrected to retrieve at-surface-reflectance values by using Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes or Empirical Line Calibration where appropriate in order to facilitate cross scene analysis of the images. Using an object oriented multi-scale segmentation approach, this study employed Definiens Professional, an image analysis application that, among other things, allowed fuzzy analysis of data and in- tegration of multiple data types within the same project. Working in conjunction with ENVI (Environment for Visualizing Images), a model based on spectral properties of the surface materials yielded more robust results than a standard pixel based classification derived from a training set. Previous studies conducted have utilized datasets that were largely moderate spatial and low spectra resolution. is study employed datasets that are also moderate spatial resolution but were reinforced with high spectral resolution data provided by Hyperion, resulting in a more accurate assessment of the surface materials and increased confidence in the model. Additionally, by first segmenting the datasets it was possible to utilize textual and contextual information that is typically lost in a pixel based classifications. is type of processing also allowed for automated processing of other datasets which facilitated an efficient temporal study and produced datasets that have undergone the same processing steps. is reduced the possibility of processing mistakes, increased confidence in the resulting surface classifications and subsequently, increased confidence in the resulting subsurface characterization of the permafrost table located, in most cases, at shallow depths below. ALI (Raw) Mosaic & Rescale Empirical Line Calibration Hyperion (Raw) Outlier Removal & Destriping Convert to Reflectance (FLAASH) Comp. with Field Spectra Field Spectra Collection Data Processing (Splice Correction) Build Spectral Library SRTM DEM Preparation Generate Slope Generate Aspect Land Surface Classification Surface Discovery Characterization Characterization Identification Identification Other Sources Import into Definiens Pro Import into Definiens Pro Build Class Definitions Defuzzification Estimate Active Layer Depth Depth Data Error Analysis Results is investigation examined the changing surface conditions of a study area near Toolik Lake, Alaska (see below) and sought to differentiate the surficial geology and geomorphology, largely influenced by glacial activity, as well as ecology of the region, in order to characterize the state of the permafrost table. e study was conducted utilizing remotely sensed images and datasets; as well as, field data, in order to conduct analysis of the landscape over multiple years. is information was subsequently used as proxy data to make observations of the state of the permafrost table which underlies this landscape. is type of study yielded continuous estimates of the ground conditions without the need for a lengthy ground campaigns that could have proved difficult in a region such as this. Advanced Land Imager 1 (ALI) Map of the Study Area (left): Band Spectral Range (μm) Resolution (m) ETM+ Pan 0.048 - 0.690 10 1 0.433 - 0.453 30 2 0.450 - 0.515 30 1 3 0.525 - 0.605 30 2 4 0.630 - 0.690 30 3 5 0.775 - 0.805 30 4 6 0.845 - 0.890 30 7 1.200 - 1.300 30 8 1.550 - 1.750 30 5 9 2.080 - 2.350 30 7 Spectral Data Cube (above) Generated from the Hyperion image near Toolik Lake, Alaska Hyperion Hyperspectral Swath 2 (leſt): Red: 32 Green: 22 Blue: 16 VNIR (bands 8 - 57) 0.427 - 0.925 μm SWIR (bands 77 - 224 0.912 - 2.395 μm Spect. Resolution ~10 nm Resolution (m) 30.38 m Scene Width 7.7 km Data Project Digital Elevation Model 5 (below): Shuttle Radar Topography Mission (SRTM) data has been used for this project with spatial resolutions of 2 arc seconds. is model has been oriented with north toward the lower leſt of the scene, also displays the Dalton Highway, Alaska Pipeline and water features for reference. (Note: this is not SRTM data and has only been used in this figure) Process Flow Spectral Profile (above) Sample spectral profile derived from data collected in the field aſter processing. Data Citation 1 - U. S. Geological Survey. August 8, 2004. EO-1 ALI Scene, Receiving Station PF1, Path 73, Row 12, Level1R. Data obtained from: USGS’s Earth Explorer at: http://edcsns17.cr.usgs.gov/EarthExplorer/. 2 - U. S. Geological Survey. August 8, 2004. EO-1 Hyperion Scene, Receiving Station PF1, Path 73, Row 12, Level1R. Data obtained from: USGS’s Earth Explorer at: http://edcsns17.cr.usgs.gov/EarthExplorer/. 3 - NASA Landsat Program, Landsat TM scene p073r12_5t850804, SLC-Off, USGS, Sioux Falls, 08/04/1985. Source for this data set was the Global Land Cover Facility, www.landcover.org. 4 - NASA Landsat Program, 2004, Landsat ETM+ scene p073r012_7k19990702_z06, SLC-Off, USGS, Sioux Falls, 07/02/1999. Source for this data set was the Global Land Cover Facility, www.landcover.org. 5 - U.S. Geological Survey (USGS), EROS Data Center. 1999. National Elevation Dataset, Shuttle Radar Topography Mission (SRTM) 2 Arc Second Data. Data obtained from the National map seamless Server at: http://ned.usgs.gov/. 6 - Walker, D.A. and N.C. Barry. 1991. Toolik Lake permanent vegetation plots: Site factors, soil physical and chemical properties, plant species cover, photographs, and soil descriptions. Department of Energy R4D Program Data report, Joint Facility for Regional Ecosystem Analysis, Institute of Arctic and Alpine Research, Boulder, CO. Boulder, CO: National Snow and Ice Data Center. Identifier no. ARCSS018. Digital media and paper copy. Web Site RSL Home: rsl.geology.buffalo.edu RSL Data Pages: rsl.geology.buffalo.edu/data Acknowledgements Teahun Yoon: Research Technician, University at Buffalo, e State University of New York, Buffalo, NY 14260 Chien-Lu Ping: Professor of Soil Science, University of Alaska Fairbanks, Fairbanks, Alaska 99775 Erin Trochim: University of Alaska Fairbanks, Fairbanks, Alaska 99775 Land Cover 6 (above) Previously constructed land cover classification of a portion of the study area which has been included in the discovery portion of the classification. Known surfaces are compared to spectral profiles in order to catalogue the properties of materials. Field Data Collection (below): Photo of data collection using the ASD Spectral Radiometer take July 2007 near Galbraith Lake. Landsat (Raw) Convert to Reflectance (FLAASH) Change Detection Results Error Analysis Lakes & Rivers Streams Primary Vegetation Barrens Dry Acidic Dry Nonacidic Tundras Snowbeds Moist Acidic Tundras Moist Nonacidic Tundras Poor Fens Rich Fens Riparian Shrublands Shrub Tundras Legend Legend (right) is legend shows the land surface classes utilized in this project for all land surface figures shown. Photo looking south across Galbraith Lake toward Atigun Pass Derived Land Cover Classification (above) is image shows the land cover distribution of the region from an object-oriented classification of the ALI image and SRTM. Since classes are only partially defined at this point, the model had difficulty distinguishing between some of the unit types. Note that this classification is only partially completed. Landsat Processing Step (above) Landsat derived classifications are not represented on this poster; however, it will be used in the same manner as ALI. e processing step “Import into Definiens Pro” displayed in red, is the same processing step as the one above and is displayed here to demonstrate where Landsat will splice into the data flow process in place of ALI. U. S. Geological Survey. August 8, 2004. EO-1 ALI Scene, Receiving Station PF1, Path 73, Row 12, Level1R. Data obtained from: USGS’s Earth Explorer at: http://edcsns17.cr.usgs.gov/EarthExplorer/. Landsat TM 3 (leſt) True Color Composite: R3: G2: B1 Landsat ematic Mapper image of the study area captured in August of 1985 Landsat ETM+ 4 (right) True Color Composite: R3: G2: B1 Landsat Enhanced ematic Mapper image of the study area captured in July of 1999 Original Land Cover Classification 6 (above) Vegetation map featuring primary vegetation for a portion of the study region. (see Legend) Definiens (leſt) A screen capture of an ongoing surface classification within Definiens Professional 5 (Definiens AG). is classification was conducted as a test in the surface exploration phase using a basic training set and a nearest neighbor classification that took into consideration ALI bands 1-9, Elevation, Slope and Aspect. Fuzzy Classification is example shows a lake with low concentrations of suspended sediment in the water column. Note that using the current class definitions and training set the classification identified this lake correctly as a lake or pond with low sediment with a membership value of 0.950 (scale 0.000-1.000). Alternative assignments for this class include lake, river or pond with high suspended sediment in the water column with a membership value of 0.648 and Riparian Complex with a membership value of 0.493. Class Definitions For this example, class definitions and regions selected for training where not adequate in all cases to fully distinguish between all classes. In particular, more information and/ or parameters are needed to separate out units such as “Moist Acidic Tundras” and “Moist Nonacidic Tundras”. Comparison of the Original Classification and the Derived Test Classification An enlarged comparison of the original primary vegetation classification (top leſt) with the test classification of the data set consisting of the ALI image and the topographic information from the SRTM DEM. (below leſt). Points of note: 1 – While it is clear that continued work on defining units needs to be conducted, it is promising to see that with only limited effort it is possible to start pulling some surfaces cover types out of the data. Most notably in this example, Barrens and Fens. 2 – Classes need to be more defined to yield results that allow for more confidence in the model. In particular for this example, “Moist Acidic Tundras” and “Moist Nonacidic Tundras” need more work. 3 – While not all materials were identified correctly the general shape of many units has been extracted relatively well. A Source of Error ere are several factors that should be mentioned as possible error sources for this particular example. While the main goal of the classification was vegetation mapping, we haven't examined yet how much the spectral signatures of the underlying soils influenced the classification. Moreover, the elapsed time between the compilation of the original land cover map and the ALI image acquisition is great enough to cause actual surface changes. Results By using object-based classification of multisensor and multitemporal data we utilize contextual information in addition to spectral properties for identifying surface units. We expect that this approach will yield more reliable results than pixel-based classification for mapping land cover and other surface properties resulting in improved accuracy in change detection. With respect to “Result 2”, using processing steps such as those employed by Leverington and Duguay (1997) 7 or Nelson et al. (1997) 8 , who achieved average errors less then 10%, results with a high degree of accuracy will hopefully be produced on a much smaller scale. References 7 - Leverington, D. W. and Duguay, C. R. (1997). “A Neural Network Method to Determine the Presence or Absence of Permafrost near Mayo, Yukon Territory, Canada.” Permafrost and Periglacial Processes 8: 205-215. 8- Nelson, F. E., et al. (1997). “Estimating Active-Layer ickness over a Large Region: Kuparuk River Basin, Alaska, U.S.A.” Arctic and Alpine Research 29(4): 367-378 Input or Output General Processing Step Reference to Existing Processing Step More Interactive Step Indicates flow direction General Arrow Chart Key 1 2
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Investigation of the Permafrost Table through Multi-resolution Object Oriented Fuzzy Analysis, North Slope, AlaskaJustin L. Rich and Bea Csatho
Department of Geology, University at Buffalo, The State University of New York, Buffalo, NY 14260
RSL Remote Sensing LaboratoryDept. of Geology, SUNY at Bu�alo
Data implemented for this project included an Advanced Land Imager and Landsat TM scene; as well as, a Digital Elevation Model with its derivative data. Hyperion hyperspectral data was also utilized for this project to obtain spectral characteristics of the ground surface
along with �eld data collected with an ASD spectral radiometer. Destriping of the Hyperion image was conducted within ENVI and the Eclipse development platform utilizing the Java IDE. All images were atmospherically corrected to retrieve at-surface-re�ectance values by using Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes or Empirical Line
Calibration where appropriate in order to facilitate cross scene analysis of the images.
Using an object oriented multi-scale segmentation approach, this study employed De�niens Professional, an image analysis application that, among other things, allowed fuzzy analysis of data and in-tegration of multiple data types within the same project. Working in conjunction with ENVI (Environment for Visualizing Images), a model based on spectral properties of the surface materials yielded more robust results than a standard pixel based classi�cation derived from a training set.
Previous studies conducted have utilized datasets that were largely moderate spatial and low spectra resolution. �is study employed datasets that are also moderate spatial resolution but were reinforced with high spectral resolution data provided by Hyperion, resulting in a more accurate assessment of the surface materials and increased con�dence in the model. Additionally, by �rst segmenting the datasets it was possible to utilize textual and contextual information that is typically lost in a pixel based classi�cations. �is type of processing also allowed for automated processing of other datasets which facilitated an e�cient temporal study and produced datasets that have undergone the same processing steps. �is reduced the possibility of processing mistakes, increased con�dence in the resulting surface classi�cations and subsequently, increased con�dence in the resulting subsurface characterization of the permafrost table located, in most cases, at shallow depths below.
ALI(Raw)
Mosaic &
Rescale
Empirical Line Calibration
Hyperion(Raw)
Outlier Removal&
Destriping
Convert to Reflectance
(FLAASH)
Comp. with Field Spectra
Field Spectra Collection
Data Processing(Splice Correction)
Build Spectral Library
SRTM
DEMPreparation
Generate Slope
Generate Aspect
Land Surface Classification
SurfaceDiscovery
Characte
rizatio
n
CharacterizationIdentification
Identif
icatio
n Other Sources
Import intoDefiniens Pro
Import intoDefiniens Pro
Build ClassDefinitions
Defuzzification
Estimate Active Layer Depth Depth Data
ErrorAnalysis
Results
�is investigation examined the changing surface conditions of a study area near Toolik Lake, Alaska (see below) and sought to di�erentiate the sur�cial geology and geomorphology, largely in�uenced by glacial activity, as well as ecology of the region, in order to characterize the state of the permafrost table. �e study was conducted utilizing remotely sensed images and datasets; as well as, �eld data, in order to conduct analysis of the landscape over multiple years. �is information was subsequently used as proxy data to make observations of the state of the permafrost table which underlies this landscape. �is type of study yielded continuous estimates of the ground conditions without the need for a lengthy ground campaigns that could have proved di�cult in a region such as this.
Advanced Land Imager1 (ALI)Map of the Study Area (left):Band Spectral Range (µm) Resolution (m) ETM+ Pan 0.048 - 0.690 10 1 0.433 - 0.453 30 2 0.450 - 0.515 30 1 3 0.525 - 0.605 30 2 4 0.630 - 0.690 30 3 5 0.775 - 0.805 30 4 6 0.845 - 0.890 30 7 1.200 - 1.300 30 8 1.550 - 1.750 30 5 9 2.080 - 2.350 30 7
Spectral Data Cube (above) Generated from the Hyperion image near Toolik Lake, Alaska
Hyperion Hyperspectral Swath2 (le�):Red: 32 Green: 22 Blue: 16 VNIR (bands 8 - 57) 0.427 - 0.925 µmSWIR (bands 77 - 224 0.912 - 2.395 µm
Spect. Resolution ~10 nmResolution (m) 30.38 mScene Width 7.7 km
Data
Project
Digital Elevation Model5 (below):Shuttle Radar Topography Mission (SRTM) data has been used for this project with spatial resolutions of 2 arc seconds. �is model has been oriented with north toward the lower le� of the scene, also displays the Dalton Highway, Alaska Pipeline and water features for reference. (Note: this is not SRTM data and has only been used in this �gure)
Process Flow
Spectral Pro�le (above)Sample spectral pro�le derived from data collected in the �eld a�er processing.
Data Citation
1 - U. S. Geological Survey. August 8, 2004. EO-1 ALI Scene, Receiving Station PF1, Path 73, Row 12, Level1R. Data obtained from: USGS’s Earth Explorer at: http://edcsns17.cr.usgs.gov/EarthExplorer/.2 - U. S. Geological Survey. August 8, 2004. EO-1 Hyperion Scene, Receiving Station PF1, Path 73, Row 12, Level1R. Data obtained from: USGS’s Earth Explorer at: http://edcsns17.cr.usgs.gov/EarthExplorer/.3 - NASA Landsat Program, Landsat TM scene p073r12_5t850804, SLC-O�, USGS, Sioux Falls, 08/04/1985. Source for this data set was the Global Land Cover Facility, www.landcover.org.4 - NASA Landsat Program, 2004, Landsat ETM+ scene p073r012_7k19990702_z06, SLC-O�, USGS, Sioux Falls, 07/02/1999. Source for this data set was the Global Land Cover Facility, www.landcover.org.5 - U.S. Geological Survey (USGS), EROS Data Center. 1999. National Elevation Dataset, Shuttle Radar Topography Mission (SRTM) 2 Arc Second Data. Data obtained from the National map seamless Server at: http://ned.usgs.gov/.6 - Walker, D.A. and N.C. Barry. 1991. Toolik Lake permanent vegetation plots: Site factors, soil physical and chemical properties, plant species cover, photographs, and soil descriptions. Department of Energy R4D Program Data report, Joint Facility for Regional Ecosystem Analysis, Institute of Arctic and Alpine Research, Boulder, CO. Boulder, CO: National Snow and Ice Data Center. Identi�er no. ARCSS018. Digital media and paper copy.
Web SiteRSL Home: rsl.geology.bu�alo.edu
RSL Data Pages: rsl.geology.bu�alo.edu/data
Acknowledgements
Teahun Yoon: Research Technician, University at Bu�alo, �e State University of New York, Bu�alo, NY 14260Chien-Lu Ping: Professor of Soil Science, University of Alaska Fairbanks, Fairbanks, Alaska 99775
Erin Trochim: University of Alaska Fairbanks, Fairbanks, Alaska 99775
Land Cover6 (above)Previously constructed land cover classi�cation of a portion of the study area which has been included in the discovery portion of the classi�cation. Known surfaces are compared to spectral pro�les in order to catalogue the properties of materials.
Field Data Collection (below):Photo of data collection using the ASD Spectral Radiometer take July 2007 near Galbraith Lake.
Landsat(Raw)
Convert to Reflectance
(FLAASH)
Change Detection
Results
ErrorAnalysis
Lakes & RiversStreams
Primary VegetationBarrensDry AcidicDry Nonacidic TundrasSnowbedsMoist Acidic TundrasMoist Nonacidic TundrasPoor FensRich FensRiparian ShrublandsShrub Tundras
Legend
Legend (right)�is legend shows the land surface classes utilized in this project for all land surface �gures shown.
Photo looking south across Galbraith Lake toward Atigun Pass
Derived Land Cover Classi�cation (above)�is image shows the land cover distribution of the region from an object-oriented classi�cation of the ALI image and SRTM. Since classes are only partially de�ned at this point, the model had di�culty distinguishing between some of the unit types. Note that this classi�cation is only partially completed.
Landsat Processing Step (above)Landsat derived classi�cations are not represented on this poster; however, it will be used in the same manner as ALI. �e processing step “Import into De�niens Pro” displayed in red, is the same processing step as the one above and is displayed here to demonstrate where Landsat will splice into the data �ow process in place of ALI.
U. S. Geological Survey. August 8, 2004. EO-1 ALI Scene, Receiving Station PF1, Path 73, Row 12, Level1R. Data obtained from: USGS’s Earth Explorer at: http://edcsns17.cr.usgs.gov/EarthExplorer/.
Landsat TM3 (le�)True Color Composite: R3: G2: B1Landsat �ematic Mapper image of the study area captured in August of 1985
Landsat ETM+4 (right)True Color Composite: R3: G2: B1Landsat Enhanced �ematic Mapper image of the study area captured in July of 1999
Original Land Cover Classification6 (above)Vegetation map featuring primary vegetation for a portion of the study region. (see Legend)
De�niens (le�)A screen capture of an ongoing surface classi�cation within De�niens Professional 5 (De�niens AG). �is classi�cation was conducted as a test in the surface exploration phase using a basic training set and a nearest neighbor classi�cation that took into consideration ALI bands 1-9, Elevation, Slope and Aspect.
Fuzzy Classi�cation�is example shows a lake with low concentrations of suspended sediment in the water column. Note that using the current class de�nitions and training set the classi�cation identi�ed this lake correctly as a lake or pond with low sediment with a membership value of 0.950 (scale 0.000-1.000). Alternative assignments for this class include lake, river or pond with high suspended sediment in the water column with a membership value of 0.648 and Riparian Complex with a membership value of 0.493.
Class De�nitionsFor this example, class de�nitions and regions selected for training where not adequate in all cases to fully distinguish between all classes. In particular, more information and/ or parameters are needed to separate out units such as “Moist Acidic Tundras” and “Moist Nonacidic Tundras”.
Comparison of the Original Classi�cation and the Derived Test Classi�cationAn enlarged comparison of the original primary vegetation classi�cation (top le�) with the test classi�cation of the data set consisting of the ALI image and the topographic information from the SRTM DEM. (below le�). Points of note:
1 – While it is clear that continued work on de�ning units needs to be conducted, it is promising to see that with only limited e�ort it is possible to start pulling some surfaces cover types out of the data. Most notably in this example, Barrens and Fens.
2 – Classes need to be more de�ned to yield results that allow for more con�dence in the model. In particular for this example, “Moist Acidic Tundras” and “Moist Nonacidic Tundras” need more work.
3 – While not all materials were identi�ed correctly the general shape of many units has been extracted relatively well.
A Source of Error�ere are several factors that should be mentioned as possible error sources for this particular example. While the main goal of the classi�cation was vegetation mapping, we haven't examined yet how much the spectral signatures of the underlying soils in�uenced the classi�cation. Moreover, the elapsed time between the compilation of the original land cover map and the ALI image acquisition is great enough to cause actual surface changes.
ResultsBy using object-based classi�cation of multisensor and multitemporal data we utilize contextual information in addition to spectral properties for identifying surface units. We expect that this approach will yield more reliable results than pixel-based classi�cation for mapping land cover and
other surface properties resulting in improved accuracy in change detection. With respect to “Result 2”, using processing steps such as those employed by Leverington and Duguay (1997)7 or Nelson et al. (1997)8, who achieved average errors less then 10%, results with a high degree of accuracy will hopefully be produced on a much smaller scale.
References
7 - Leverington, D. W. and Duguay, C. R. (1997). “A Neural Network Method to Determine the Presence or Absence of Permafrost near Mayo, Yukon Territory, Canada.” Permafrost and Periglacial Processes 8: 205-215.8- Nelson, F. E., et al. (1997). “Estimating Active-Layer �ickness over a Large Region: Kuparuk River Basin, Alaska, U.S.A.” Arctic and Alpine Research 29(4): 367-378
Input or Output
General Processing Step
Reference to Existing Processing
Step
More Interactive
Step
Indicates �ow direction
General Arrow
Chart Key
1
2
Related Documents
