2.1. Study Area
The study area was located along the Colorado River near De Beque, Colorado (39º33’ N, -108º2’ W). Average elevation of the area is approximately 1,500 m, with the southeastern edge of the Roan Cliffs located to the northeast. The area contains substantial amounts of tamarisk that have been well-documented by a number of field survey teams. Tamarisk stands used in image classifications were located via GPS between January, 2003 and May, 2004. A total of 40 plots located within the De Beque 7.5 minute USGS quadrangle were surveyed using a purposive sample design [19
] and relevَ
e principles [20
] to maximize variation in habitat. Plots were placed in each major vegetation zone to cover the full range of environmental gradients present. The relevَ
e method quickly determines relative cover by species in a plot that represents a particular vegetation type.
The primary sampling unit was a 30 m × 30 m cluster plot which corresponded with the 30 m GSD of Hyperion and TM5 data. Within each of these 900 m2 areas, nine subplots, each 10 m × 10 m (100 m2) in dimension, allowed for analysis at the finer resolution enabled by QB data (2.5 m GSD). Within each of these plots and subplots, percent cover of tamarisk and other dominant vegetation was recorded.
2.2. Remote Sensing
TM5 and Hyperion data (USGS EROS Data Center, Sioux Falls, South Dakota, USA) acquired on July 13 and July 5, 2004, respectively, and QB data (Digital Globe, Longmont, Colorado, USA) acquired on August 16, 2004, corresponded with the approximate time of field sampling. Additionally, QB data acquired on June 8, 2005 were available for comparison with the 2004 QB data. Airborne hyperspectral data from the HyMap sensor (HyVista Corp., New South Wales, Australia) acquired on July 6, 2002, were used in calibrating the TM5 and QB data to reflectance units. While the dates of field sampling and image acquisition may differ, tamarisk populations in the study area are notably stable from year-to-year (T. Stohlgren, USGS, Fort Collins, CO, USA, personal communication). Thus, the surveyed locations of tamarisk stands were applicable for all image acquisitions.
2.3. Georectification and Reflectance Calibration
All images were registered using the nearest-neighbor resampling method to a rectified 1 m digital ortho quarter quad (DOQQ) of the De Beque area. Rectification of the Level I Hyperion and TM5 imagery was refined further to a root mean square error (RMSE) of less than 0.5 pixels. Rectification of QB standard-bundle imagery was refined to a RMSE of 0.4 and 0.5 pixels for the 2004 and 2005 acquisitions, respectively.
The HyMap imagery consisted of 126 spectral channels (bands) covering a 440 nm to 2,500 nm spectral range at a full-width-at-half-maximum (FWHM) bandwidth of approximately 15 nm. The flight altitude of 3,500 m produced a GSD of 3.7 m. The data were calibrated at-sensor to radiance units (µW/cm2 nm sr) and processed subsequently to apparent reflectance using Fast Line-of-Sight Atmospheric Analysis of Hypercubes (FLAASH) (ENVI v. 4.2, ITT Visual Information Systems, Boulder, CO, USA). FLAASH employs an atmospheric correction based on MODTRAN 4+ radiative transfer models and corrects for absorptions by atmospheric water vapor, methane, oxygen, carbon dioxide and ozone on a pixel-by-pixel basis. The FLAASH calibration to apparent reflectance was refined using an Empirical Line calibration (ENVI v. 4.2) based on the in situ reflectance of an asphalt road collected in June, 2005 (ASD FS, Analytical Spectral Devices, Boulder, CO, USA). Although field measurements of asphalt road reflectance were not made simultaneously with image acquisitions, all data were acquired when the road surface was dry. Thus, the road was acceptable as a pseudo-invariant calibration target. The Empirical Line calibration forces spectral data to match field reference data using a linear regression for each band. When only one ground target is used, as in the present case, the regression line is assumed to pass through a zero origin. HyMap bands 1 (437 nm), 31 (873 nm), 63–66 (1,405–1,447 nm), 94 (1,805 nm), 95 (1,949 nm) and 126 (2,484 nm) were deleted due to strong atmospheric interference, detector overlap or detector insensitivity. The remaining 117 bands spanned the 443 nm to 2,468 nm range.
Next, regions-of-interest (ROI) representing specific pseudo-invariant targets of lake water and bare gravel were created using the HyMap data. ROI spectra of these surfaces were extracted to a spectral library (ENVI v. 4.2). The water served as a low-reflectance target while reflectance of the bare gravel was substantially greater. HyMap spectral reflectances of these dark-to-bright targets were re-sampled to TM5 or QB spectral bands (ENVI v. 4.2) and used in an Empirical Line calibration of the TM5 and QB data. This corrected the data for atmospheric interference and yielded units of percentage reflectance. QB data were acquired in four broad spectral bands centered at 485, 560, 660 and 830 nm (Table 1
) at 2.5 m GSD and 11-bit radiometric resolution. TM5 data included these bands as well as mid-infrared bands at 1,650 and 2,215 nm (Table 1
) and were acquired at 30 m GSD with 8-bit radiometric resolution.
Hyperion data include 242 spectral bands ranging from 356 nm to 2,577 nm at a FWHM bandwidth of 10 nm and 12-bit radiometric resolution. Because Hyperion bandwidths are narrower than the ca. 15 nm bandwidths in HyMap data, the latter could not be used to calibrate the Hyperion data to reflectance units. Instead, Hyperion data were spectrally subset to remove bands 1–8 (357–417 nm) and 225-242 (2,406–2,577 nm), owing to data noise, and bands 58–70 (925–1,068 nm) and 71–77 (852–912 nm) were eliminated due to detector overlap. The remaining 196 Hyperion bands covered the 426 nm to 2,396 nm spectral range and were calibrated to apparent reflectance using FLAASH. The field spectroradiometric data could not be used to refine the Hyperion calibration because the gravel road was not clearly visible in the 30 m GSD Hyperion image. Inspection of the reflectance-calibrated Hyperion data revealed substantial noise or extreme image striping in the 1,356–1,457 nm, 1,820–1,992 nm, 2,022–2,042 nm, 2,062–2,082 nm and 2,284–2,396 nm regions. These spectral regions were deleted, leaving 148 bands for potential use in tamarisk mapping.
2.4. Image Classification
Initially, an unsupervised classification algorithm (isodata, ENVI v. 4.2) was applied to the TM5, QB and Hyperion data. This allowed a preliminary assessment of tamarisk discrimination in the De Beque area. Additionally, principal components analysis (PCA) [21
] and the Minimum Noise Fraction (MNF) procedure [21
] were applied to the Hyperion data (ENVI v. 4.2). However, band correlation analysis (ENVI v. 4.2) applied to each reflectance-calibrated data set indicated substantial redundancy among spectral bands in the reflectance of vegetated terrain (see Results). Based on these initial assessments, data dimensionality was reduced by removing redundant bands. This facilitated use of the Maximum Likelihood (ML) algorithm and normalized-difference indices [21
] in subsequent comparisons among sensor data in tamarisk delineation. ML was selected because it is widely accepted and generally provides the greatest accuracy among various supervised classification procedures [21
]. It computes the probability that a certain pixel belongs to one of a pre-defined number of classes, taking into account the variability in each ROI and assuming that training data statistics in each band for each class are normally distributed. The pixel is then assigned to the class to which it most likely belongs. Inspection of training data for the bands selected from each sensor indicated general normality with slight skewing in some bands, but no bi-modal distributions. Nevertheless, as a parametric method, ML is robust to violations of training-data normality and performs well when training data are limited [23
Training data for supervised classifications of TM5 or Hyperion images were based on 30 m plots which contained 80% or greater coverage by tamarisk. For QB classifications, 10 m subplots containing 100% tamarisk cover were used. The same 30 m plots and 10 m subplots were used in determining value ranges of remote sensing indices that were representative of tamarisk stands. Indices applied were the NDVI [23
] and a similar index which incorporated green-band rather than red-band reflectance (Green NDVI or GNDVI). Additionally, data from these plots and subplots were used to assess image classification accuracy. For data from each sensor, 40% of plot data were used in classification training. The remaining 60% were reserved for post-classification accuracy assessment. All image classifications were based on sample plot (training) areas of at least 0.3 ha.
Classification accuracy was determined by error matrix and the Khat
coefficient of agreement [21
]. This produced values for errors of omission (percentage of tamarisk pixels that were not classed as tamarisk) and errors of commission (percentage of non-tamarisk pixels that were classed as tamarisk). Khat
represents the extent to which a given classification procedure improved classification accuracy relative to a random classifier [22
]. Thus, for example, Khat
= 0.33 would indicate a 33% improvement in accuracy relative to classification by chance.