Ground Truth for Calibrating Optical Imagery to Reflectance

Transcription

1 Visual Information Solutions Ground Truth for Calibrating Optical Imagery to Reflectance The by: Thomas Harris Whitepaper

2 Introduction: Atmospheric Effects on Optical Imagery Remote sensing of the Earth s surface with satellite or high-altitude aircraft-mounted instruments is impacted by the Earth s atmosphere: as sunlight passes through the atmosphere, atmospheric gasses interact with solar radiation, scattering, absorbing and transmitting light photons. Raw optical imagery includes information about the Earth surface and atmospheric gasses present at time of image acquisition. Figure 1 shows atmospheric transmission across the optical portion of the electromagnetic spectrum from 400 to 2500 nm and shows specific frequencies of the spectrum where specific atmospheric gasses strongly absorb electromagnetic radiation. Atmospheric calibration of optical imagery seeks to minimize the effect of atmosphere on remotely sensed signatures and is essential for many remote sensing applications seeking to detect specific materials of interest based on known reflectance signatures. Figure 1: Atmospheric transmittance from 400 to 2500 nm. pg. 1 of 7

3 Many techniques are available for calibration to reflectance, from techniques based on simple assumptions about the way light interacts with the atmosphere, to advanced techniques that seek to remove atmospheric effects by modeling the interaction of light with the atmosphere. The most accurate technique for removing atmospheric effects on optical imagery is the empirical line calibration (ELC), a method that forces image spectra to match reflectance spectra collected in the field. This paper describes the practical application of ELC to a Quickbird multispectral image (Digital Globe, Longmont, CO) collected over Socompa Volcano on the Argentinean- Chilean border in the Atacama Desert, South America (Figure 2). Figure 2: Field site location. Socompa Volcano, Argentina pg. 2 of 7

4 Empirical Line Calibration ELC is used to force spectral data to match selected field reflectance spectra, and is dependent on reference reflectance signatures collected from known field locations that can be readily identified in imagery. Because of the dependence on field-collected reference signatures, ELC is often impractical if access to field spectroscopy instrumentation is not available, or if field sites are inaccessible. However, for applications where the field site can be accessed with a spectrometer, ELC will provide the best possible atmospheric correction. Field-collected reflectance signatures have less atmospheric distortions than satellite or aircraft-mounted sensors because of the close proximity of the field spectrometer to the surface; radiation reflected from the land surface and entering the spectrometer passes through significantly less atmosphere. ELC is applied by defining a linear regression for each band equating DN to reflectance, thereby removing the solar irradiance curve, atmospheric gas absorptions and path radiance. The following equation shows how the empirical line gain and offset values, derived from the linear regression, are used, in this case, to go from radiance to reflectance. Reflectance (field spectrum) = gain * radiance (input data) + offset ELC in ENVI requires at least one field, laboratory, or other reference spectrum that can be matched with a signature extracted from a pixel or group of pixels in an image. If more than one spectrum is used, the regression for each band will be calculated by fitting the regression line through all of the spectra. Best results are achieved by visiting dark and bright regions in the field and collecting spectra which can be matched against pixels extracted from the image. This provides a more accurate linear regression. If only one spectrum is used, then the regression line will be assumed to pass through the origin (zero reflectance equals zero DN). Methods As part of a study investigating the potential for using high-resolution commercial multispectral imagery to detect and map high-altitude vegetation on Socompa Volcano, Argentina, a QuickBird multispectral image from April 13, 2005 was acquired from the DigitalGlobe (DG) image archive (Figure 2). QuickBird has four multispectral bands with bands centered in the blue, green, red and near infrared portions of the electromagnetic spectrum. DigitalGlobe s QuickBird image data is typically distributed in relative radiance. Therefore, the ENVI QuickBird Radiance utility was used to convert the relative radiance into absolute radiance in units of mw / (cm2 nm sr) pg. 3 of 7

5 Spectral signatures were acquired in the field during a visit to Socompa Volcano in February, An Analytical Spectral Devices (ASD, Boulder, CO) FieldSpec Handheld (FSHH) spectrometer, covering a wavelength range from 325 to 1075 nm with a sampling interval of 0.6 nm, was used to measure solar reflectance from gravel and rock surfaces around a volcanic vent near 5800 meters above sea level. In order to determine the reflectance of surface materials at the study site, two measurements were required: the spectral response of a Spectralon (Labsphere, North Sutton, NH) reference sample and that of the target material. The reflectance was then computed by dividing the spectral response of the target material by that of the reference sample. Eight individual signatures of rock and gravel were compared, and based on their spectral similarity, averaged together to create a mean field-measured rock and gravel reflectance signature representative of the study site (Figure 3). ELC results are often improved with the inclusion of multiple spectral pairs. If only one field spectrum is available, as in this exercise, it is possible to improve results by matching image pixels with reflectance signatures of fundamental materials such as snow, ice, or quartzite-derived sand where reflectance signatures are relatively well-characterized. This technique would likely not be desirable with hyperspectral data. In this case, correcting coarse spectral resolution data, spectral signatures are relatively generalized and absolute accuracy is not as important. A snow signature (Coarse Granular Snow) was selected from the Johns Hopkins University Spectral Library provided with the ENVI software (Figure 3, Lab-measured Snow Reflectance). The QuickBird imagery contained several prominent snow fields; 10 pixels were selected as representative of snow radiance for pairing with the spectral library snow signature in the ENVI ELC. Locations of snow pixels are visible in Figure 2. Figure 3: Field and lab-measured reflectance signatures for snow and rock/gravel pg. 4 of 7

6 A Trimble (Sunnyvale, CA) GeoXT global positioning system (GPS) was used to record the location of the study area on Socompa. Differential correction post-processing was applied to the GPS positions using base station data acquired from Salta, Argentina, resulting in average horizontal precisions of 0.8 meters. The accuracy of locating field spectra sampling stations in the imagery was approximate because, as part of Digital- Globe s Standard Imagery creation process, a coarse normalization for topographic relief was applied with a digital elevation model (DEM). The degree of topographic normalization was generally small, so while this product had terrain corrections, it could not be considered orthorectified. The Standard Imagery product has published accuracies within 23 meters, excluding any topographic displacement and off-nadir viewing angle. For these reasons, and also because of the relative homogeneity of the volcanic landscape around the field site, the location of the field site could not be precisely determined in the imagery and was approximated based on field experience. We selected five pixels from the image as representative of rock and gravel, and ten pixels were selected as representative of snow. The mean radiance signatures from each of these groups of pixels are displayed in Figure 4. The ENVI ELC was applied to the QuickBird image by inputting two spectral pairs: Mean Field-measured Rock & Gravel Reflectance : Rock & Gravel Radiance Lab-measured Snow Reflectance : Snow Radiance Figure 4: Mean QuickBird radiance signatures for two selected calibration targets pg. 5 of 7

7 Results Six calibrated QuickBird pixel signatures from bare rock and gravel areas were manually selected and are shown in Figure 5 for comparison with the mean field-measured rock and gravel reflectance. Similarly, Figure 6 shows six selected QuickBird snow signatures compared against lab-measured snow reflectance from the Johns Hopkins University Spectral Library. All selected image reflectance spectra show good agreement with field and lab-measured signatures and indicate that the calibrated QuickBird image is representative of true surface reflectance. Figure 5: Field-measured rock and gravel reflectance compared with QuickBird reflectance signatures from select pixels containing rock and gravel pg. 6 of 7

8 Figure 6: Lab-measured snow reflectance compared with QuickBird reflectance signatures from select pixels containing snow pg. 7 of 7