Field and Research Assistants: Robert Ahl, Kerry Halligan, Carl Legleiter, Jim Rasmussen

Transcription

1 1 Application of high spatial resolution, hyperspectral imagery to riparian mapping in Yellowstone National Park Probe1 1 meter data, 8/03/99 MNF Image by W. A. Marcus W. Andrew Marcus Department of Geography University of Oregon Outline of Presentation Goal of talk: Review data types and mapping approaches that have been most successful for us in riparian mapping Concept of high spatial resolution hyperspectral (HSRH) data Field area and data collection Research approaches and findings: Data Decisions choice of appropriate spatial resolution choice of data compression approaches Classification/Mapping Algorithms spectral angle mapping for terrestrial landcover maximum likelihood for in-stream habitats regression analysis for stream depths matched filtering for downed wood Summary Collaborators on Image Acquisition, Data Collection, and Analysis Richard Aspinall, Montana State University Joe Boardman, Analytical Imaging Geophysics Robert Crabtree, Yellowstone Ecosystem Studies Don Despain, U.S. Geological Survey Earth Search Systems, Inc. Mark Fonstad, Southwest Texas State University Wayne Minshall, Idaho State University Chuck Peterson, Idaho State University Field and Research Assistants: Robert Ahl, Kerry Halligan, Carl Legleiter, Jim Rasmussen Research Objectives and Rationale Overarching goal: determine if HSRH imagery can be used to map key environmental variables in complex stream settings. Soda Butte Creek, 9ssg68, Oct 10, 1999 Funding Sources: What is Hyperspectral Imagery? The High Tech Landscapes Initiative of Yellowstone Ecosystem Studies, Bozeman, Montana 1999 Probe1 (solid lines) and TM5 spectra (symbols) for willow, cottonwood and aspen The NASA Earth Observations Commercial Appli- cations Program (EOCAP) - Hyperspectral, Stennis Space Flight Center, Mississippi The EPA bioindicators program 1999 Probe1 1 meter image, Lamar River Cottonwood Aspen Willow Lamar River, Aug 3, 1999, Image by J. Boardman

2 2 Rationale for using high spatial resolution Detection of within-stream features like riffles requires HSR imagery Landscapes: Cooke City TM (30 m) Probe1 (1 m) AVIRIS (17 m) Probe1 (8 m) Probe1 (5 m) High Spatial Resolution and Hyperspectral combined ( HSHR imagery ) provide more information on variables at the stream scale. Landscapes: Round Prairie Probe1 1 meter image Simulated true color TM image 7 bands 472 pixels 16 pixels MNF Bands 7 (invert), 6, 1 Lamar River, 8/03/99 Image by W.A. Marcus Image by W.A. Marcus pixels 128 bands 13 pixels Field Area: Northeast Yellowstone Landscapes: Lower Soda Butte Creek Soda Butte Creek 3 rd order 9msb10-4, Oct, 4, th order Lamar River 5 th order 9mul8-26, Aug 26, 1999

3 3 Landscapes: Lamar River Location of 1999 Hyperspectral Data in NE YNP Red = Hyperspectral Lines Yellow = AirSAR Data (AirSAR = Lhv, Chv, Chh) Upper Soda Butte Creek Lamar River Lower Soda Butte Creek YNP boundary Backdrop shows 1990 Landsat TM NIR Band 4. AirSAR data acquiredmay 28, Data Sets: Imagery and Field Data 1- m Imagery: Helicopter platform To collect 1-m resolution data, the Probe-1 sensor was mounted on a helicopter and flown 600 m above the ground surface.. Probe1 1 meter data, 8/03/99 MNF Image by W. A. Marcus J. Boardman, Probe1, misc6, 08/04/99 17m Landsat 7 TM 30m 4m 1m The helicopter platform created unique problems: more vibration, pitch, yaw and roll than fixed wing aircraft gyroscope overcompensation led to serious image swirl rotor blades blocked GPS signal, causing loss of coordinate data greater difficulty in coregistering field maps to imagery and GIS data b, Lamar River Aug m Aug 4, 1999, misc5

4 4 Methods: field mapping for image training and validation Morphologic units, stream depth, substrate size, and vegetation commun- ities and species were mapped directly to 1 m images to insure coregistration Reducing Dimensionality: Selecting Spectra If a feature has a unique spectral signal, selecting specific spectra may be an efficient approach for reducing dimensionality. Car 4 spectral bands shown by yellow dotted lines can be used to distinguish metal car from other features on 4-m resolution Probe1 image. J. Rasmussen, Aug, 1999 (misc12) C. Legleiter, Aug, 1999 (misc13) Data Decisions Two key considerations: Different data compression approaches Effects of spatial resolution on feature detection True color composite using 1-m Probe1 image, August 2, 1999 Reducing Data Dimensionality: PCA PC1 PC3 PC10 PC40 Probe1 1 meter data, 8/03/99 MNF Image by W. A. Marcus Hyperspectral data: reducing dimensionality Goals of reducing dimensionality: Data compression and faster processing Signal enhancement Noise reduction True color composite using 1-m Probe1 image, August 2, 1999 Reducing Data Dimensionality: MNF MNF 1 MNF3 Approaches examined: Spectral selection Principal Component Analysis (PCA) Minimum Noise Fraction (MNF) Possible future approach: Geostatistical filtering MNF10 MNF40

5 5 Reducing Data Dimensionality: PCA vs MNF Data decision: spatial resolution PC1 MNF1 Entire 4-m image Entire 1-m image PC3 MNF3 Finer scale spatial resolution better discriminates many features, but at the cost of: less area coverage for a given sensor greater data storage cost for a given area Therefore one wants to choose coarsest pixel resolution that still meets mapping needs. Reducing Data Dimensionality: PCA vs MNF PC10 MNF10 Effects of spatial resolution: human features PC40 MNF40 Findings: Reducing dimensionality Except with spectrally unique targets, spectral band extraction is generally not best approach as it discards the power of hyper- spectral imagery. Exception is stream depth (to be discussed later). PC and MNF both reduce data size and processing time by and order of magnitude, but: no a priori reason for favoring one over the other each seems to work better for certain applications: o PCA best for in-stream classification o MNF best for terrestrial land cover Further work needed to: determine why a given method works better in certain situations evaluate alternative approaches e.g., geostatistical filtering Effects of spatial resolution: Streams Scour pool 4-m pixel Spatial Eddy drop Overall resolution zones Riffles Pools Glides accuracy (m) (%) (%) (%) (%) (%) Kappa Producer's Acc uracy

6 6 Effects of spatial resolution: Downed wood 4-m pixel Downed wood Downed wood: Requires 1- m resolution for detection. True color SAM rule image Findings: Spatial resolution Appropriate spatial resolution varies with feature in question: Feature Human features In-stream habitats Stream depth Downed wood Vegetation cover Required spatial resolution (m) to 10 Accuracy at that resolution (%) N.A. 76 to to General Rule: Pixel resolution should be ¼ size of feature of i nterest. Realistic Rule: Pixel resolution should be approximately same size as feature of interest Effects of spatial resolution: Vegetation type 4-m resolution provides valley wide coverage, while still detecting individual large trees. Probe1 4-m image, Round Prairie Individual trees Hyperspectral Mapping Features of interest for riparian mapping: Terrestrial land cover (vegetation, wetland, human features) Stream habitats and depths Downed wood Classification approaches used for final maps: Spectral angle mapper (SAM) land cover Maximum likelihood in-stream habitats Matched filter (MF) downed wood Regression and hydraulic model stream depths Also evaluated mixture tuned matched filters (MTMF), pixel purity indices (PPI), and pixel unmixing. Work by Kerry Halligan, YERC % correct Conifer 100% Sedge 96.6% Decid. 100% Thistle 93.1% Average 97.4% Evaluation site for land cover mapping 30 sites visited for each category Round Prairie, YNP 4-m imagery

7 7 Land cover mapping: Spectral Angle Mapper Land cover mapping: SAM rule images Dark Point 0Band 2 Reflectance Material 2 Material 1 0 Band 1 Reflectance Spectral angle Sedge (marsh) rule image Map of sedge in Round Prairie Similarity between pixels is defined by the angle between their vectors. The advantage of this technique is that illumination differencesacross large landscapes (e.g. different aspects) do not create false differences between pixels of the same composition. Land cover mapping: Data Processing Data transformations prior to spectral angle mapping: None use raw untransformed imagery Principal components (PCA) images Minimum noise fraction (MNF) images SAM terrain map of Round Prairie, WY, based on MNF bands 1-20, Probe1 4-m imagery The best classifications occurred with MNF transformed data. It is not immediately apparent why this is the case, but it seems likely that MNF does a better job of: Removing image-wide noise prior to data analysis Isolating signal related to subtle variations in vegetation cover CRITICAL to using MNF using an homogenous dark patch or dark bands for defining noise statistics. asphalt: gray water: blue gravel: white bare soil: red sparse veg: orange grass: yellow marsh: medium green conifer: dark green deciduous: light green Spectral Angle Mapping of Algae SAM rule images by W.A. Marcus Probe1 1 m data, Aug 03, 1999 SAM requires that the user define the threshold at which the feature is found. The user should therefore have knowledge about the landscape for feature ID to work. Hyperspectral mapping: In-stream habitats High gradient riffle (white water) Low gradient riffle (no white water) Scour pool The algae training site Eddy drop zone (fine sediment accumulation) Glide Round Prairie, 9srp104, October, 1999

8 8 In-stream habitats: Data processing Raw reflectance values, Probe1, Lamar River, 1999 Data transformations evaluated prior to mapping: None use raw untransformed imagery Principal components (PCA) images Minimum noise fraction (MNF) images The best classifications occurred with PCA transformed data. It appears the low reflectance from water causes MNF transforms to classify signal as noise, thereby reducing the spectral information available for analysis. CRITICAL to using PCA in streams calculate the covariance statistics using the water portion only of the stream (i.e., mask out remainder of image when calculating statistics). Reflectance at 707 nm Eddy drop zones (red) Pools (dark blue) Riffles (cyan) Glides (green) Reflectance at 555 nm Take raw image In-stream habitats: PCA images Mask to calculate covar- iance statistics for water Create pc data that highlight within-stream variations Glides (in green) are shallow with gravel substrates Rapids (in white) Turbulence Category Riffle/rapid Glide Pool Eddy drop zone Accuracy (%) Pools (in blue) Image Fine grained sediment (in yellow) Mask of mapped units PC bands 8,4,1 In-stream habitats: Maximum likelihood approach Calculates the distance of a given pixel from the average based on statistical distribution of training sites in spectral space. Results of 4 Unit Stream Classification Field map 1 meter classification Riffles Glides Pools Eddy drop zones Example of maximum likelihood classification, (Lillesand and Kiefer, 1994, pages 590 &596)

9 9 Hyperspectral mapping: stream depths Steps to reach this goal: Extract pc scores for depth pixels Conduct multiple regression analysis on depth vs. pc scores Error analysis Create depth image using regression results Soda Butte Creek, 9ssg53, Oct 10, 1999 Stream mapping: Depths in the Lamar River Multiple regression of pc scores (x) vs measured depths (y) In-stream habitat Eddy drop zone High gradient riffle Low gradient riffle Pool Rough water runs Runs Glides Number of sites Number of pc bands in regression Adjusted R2 (%) Multiple Regression Results, Lamar River, WY Summary statistics for multiple regression of depth (y) as a function of principal component scores (x, x 1 2 x n ) Depth Range All depths <= 60 cm Number of sites Number of pc's included in regression 8 4 Goal: determine if 1 m imagery can map contin- uous variations in depth. Signifi- cant at R2 adjusted 59% 65% R2 predicted 53% 62% Cutoff of relation between depth and reflectance data at > 60 cm consistent with Winterbottom and Gilvear, 1998, but... Stream depth mapping: The HAB-1 model Basis for model (developed by Mark Fonstad, SWTU): Manning equation Jarret s roughness Q, S and W are measured Key assumptions: roughness estimate is reasonable triangle approximation: D max = 2*D avg avg depth a avg reflectance signal (implies depth driving all variations in reflectance). Reflectance is a function of: depth and turbulence and substrate Stratifying by in-stream habitats should therefore help to normalize the relation of depth to reflectance. Stream depth mapping: Input data Applying the HAB-1 Model: acquire Q and S from remote sources estimate D avg and D max at multiple cross sections determine avg, max, min reflectances at cross sections regress reflectance (x) vs depth (y) apply regression to remainder of remote sensing image Rough water transects (yellow) Soda Butte Crk, July, 1999, Untitled 1-2 Smooth water transects (green)

10 10 Results: Smooth and rough water Reflectance variations from turbulence Up valley wind creating waves on a glide False turbulence Downstream hydraulic riffle Discussion: Reasons for inaccuracies with in-stream mappings Reflectance variations due to viewing angle Looking towards sun Reflectance from streams may not be directly proportional to depth due to: variations in substrate turbulence and false turbulence viewing angle relative to sun variations in background (mixed pixels) turbidity Soda Butte Crk,Silver Gate, Oct 1999 (9ssg54) Looking down at same time and site (arrows show same transect) Soda Butte Crk,Silver Gate, Oct 1999 (9ssg55) Reflectance variations due to substrate Reflectance variations due to mixed pixels Soda Butte Crk, Silver Gate, Oct, 1999 (9ssg57) One pixel Soda Butte Crk, Round Prairie, Oct, 1999 (9srp96) Soda Butte Crk, Oct 1999 (9srp94)

11 11 Reflectance variations due to turbidity Depth at both sites along Soda Butte Creek is ~1.5 m Hyperspectral mapping: Downed wood Data transformations prior to matched filter mapping: None use raw untransformed imagery Principal components (PCA) images Minimum noise fraction (MNF) images The best classifications occurred with MNF transformed data. Classification approaches used for final maps matched filter (MF) Also evaluated mixture tuned matched filters (MTMF), pixel purity indices (PPI), pixel unmixing, spectral angle mapper, maximum likelihood classification. Discussion: Are the depths accurate? r 2 = 0.30, seems low, but general trend captured (slope@ 1, y 0) cross sections seem real hydraulics are correct Downed wood: Matched filter mapping Create pc image Choose training sites, create matched filter images and assess accuracy at different thresholds. mf > 0.18 mf > 0.50 mf > 1.00 Discussion: The power of visualization Field map Downed Wood: Matched filter mapping Field reality MF classes One pixel Matched filter image by J. Boardman Proportion of wood in each pixel, red = high, green - medium, blue = low

12 12 Summary: high definition measurement Overall Accuracies: In-stream habitats: 76% 91% Depths: 67% 99% Woody debris: 82% Algae: 75% Wetlands: 100% Vegetation: 93% - 100% Classification map high gradient riffle low gradient riffle glide pool run eddy drop zone Summary There is great potential for HSRH mapping of riparian environments. Probe1 1 meter data, 8/03/99 MNF Image by W. A. Marcus But further work is needed: Rationale for choosing specific data compression and classification algorithms for different applications Evaluation of physical factors driving the signals Alternative modes of assessing accuracy Search for spectrally driven descriptors of streams High precision coregistration for change detection Incorporation of approaches into automated software sequences Ethics of remote measurement Summary: Sequence of steps & decisions (1) Data acquisition decisions Location and time Spatial scale Human features: 1 m General land cover: 2 to 10 m Streams: 1 m Downed wood: 1 m Spectral properties: Optical for cover estimates and stream depths Radar for biophysical measurements & moisture detection (not assessed in this study) (2) Data preprocessing Atmospheric (not done for these studies) Geometric (especially critical if coregistration to be done) Data compression/separation of signal from noise: Spectral extraction (biomass, unique spectral signatures) PCA for water features (develop covariance stats for features of interest only mask out other features) MNF for terrestrial landcover features (develop noise stats over dark homogenous features) Summary: Sequence of steps (3) Application of Classification/Mapping Algorithms Terrestrial land cover over broad areas: SAM on MNF transformed data Downed wood: MF on MNF transformed data Streams: In-stream habitats: Maximum likelihood on PCA data Depths: Multiple regression on PCA if ground data available HAB-1 model on raw spectral bands if no depth data from ground (4) Map production Assignment of data to classes (e.g. SAM rule images) Data overlays to generate comprehensive maps