High Resolution Hyperspectral Remote Sensing over ...
HIGH RESOLUTION HYPERSPECTRAL REMOTE SENSING OVER OCEANOGRAPHIC SCALES AT THE LEO 15 FIELD SITE David D. Kohler1, W. Paul Bissett1, Curtiss O. Davis2, Jeffrey Bowles2, Daniel Dye1, Robert G. Steward1, Jessica Britt1, Marcos Montes2, Oscar Schofield3, and Mark Moline4 1 Florida Environmental Research Institute, 4807 Bayshore Blvd., Suite 101, Tampa, FL 33611 2 Naval Research Laboratory, Code 7212, 4555 Overlook Ave. S. W., Washington, D. C. 20375 3 Rutgers University, 71 Dudley Road, New Brunswick, NJ 08901 4 California Polytechnic State University, San Luis Obispo, CA 93407 INTRODUCTION The complexity and variety of marine optical signals in coastal ocean areas has created a challenging environment for the development of remote sensing algorithms. To date, most of the successful oceanic algorithms have dealt with Case 1 type waters. These algorithms were developed to exploit multispectral remote sensing data streams. While these models have worked well for characterizing the open ocean waters, they have been less successful in classifying near shore environments. The efforts to derive coastal remote sensing algorithms have not only been hampered by the increasing importance of the bottom reflectance in the overall upwelling optical signal as we move towards the shore, but the relatively small spatial scale water type variability, seen near the shoreline, has also been a factor. Terrestrial influences to the coastal environment (river out flows) and the water-bottom boundary interactions (sediment resuspension) and their potential impacts on the water types, and thus their optical signals, need to be considered prior to the application of any algorithm. Additionally, care needs to be taken prior to the removal of the atmosphere from any coastal remotely sensed image; the potential for the mixture of marine and terrestrial aerosols in coastal scenes makes atmospheric correction a non-trivial matter. Promising to deliver the extra information needed to properly handle such spectrally complex scenes, hyperspectral remote sensing emerged as a collection tool nearly a decade ago. Having many more and much finer spectral bands than its multispectral predecessor, hyperspectral imagery appeared (and still does) to hold the key in tackling such problems. However, the optical signals of marine environments are relatively low and differences between various signals are subtle. Therefore, highly sensitive hyperspectral imagers are needed for such applications. During the summer of 2001, a study was set up off the coast of New Jersey in part to gather the necessary airborne hyperspectral remote sensing and ground truth data needed to undertake the challenge of classifying the coastal environment. The study was part of the ONR funded Hyperspectral Coastal Ocean Dynamics Experiment (HyCODE). This part of the experiment was centered at the Rutgers University Marine Field Station, LEO – 15, in Tuckerton, NJ. For nearly a month, twenty-three distinct research groups gathered to collect this exhaustive data set. The ground truth data collected consisted of both optical and traditional oceanographic measurements as well as atmospheric and
1
meteorological observations. The remote sensing data consisted of multispectral satellite data and airborne hyperspectral data. The gathered satellite data included SeaWiFS, FY1-C, Oceansat, AVHRR, and MODIS. The hyperspectral campaign included flights of the PHILLS I, PHILLS II, and AVIRIS sensors. The three hyperspectral instruments were flown at different altitudes, and thus, covered a wide range of spatial scales. The PHILLS I flew at 8,500 ft and gathered ~2mx2m spatial data. The PHILLS II flew at 30,000 ft and gathered ~9mx9m spatial data. And AVIRIS flew at 60,000 ft and gathered ~20mx20m spatial data. The respective spectral range and resolution of the instruments are: 400nm to 960nm at 4nm spacing, 400nm to 960nm at 4nm spacing, and 400nm to 2500nm at 10nm spacing Our particular focus in this experiment was on the calibration, deployment, and delivery of the PHILLS II data stream. The purpose of this paper is to outline the steps taken and problems addressed in the development of this data set. BACKGROUND The Portable Hyperspectral Imager for Low-Light Spectroscopy II (PHILLS II) was designed and developed by the Optical Sensing Section of the Naval Research Laboratory. The instrument is an aircraft mounted, push broom type sensor. It utilizes a two dimensional charge coupled device (CCD) camera to collect the spectral information along a single line on the ground perpendicular to the direction the aircraft is traveling. The image cube is built up as the aircraft moves forward in space. Davis et al. (2002) provides a detailed description of the sensor. What distinguishes this instrument from other hyperspectral sensors is that from conception this sensor was designed specifically for oceanic hyperspectral remote sensing. Capturing coastal optical signals from an airborne platform poses two major design challenges that are not usually considerations for terrestrial focused systems. The first challenge is signal sensitivity. Imaging over deep waters and at high altitudes, the atmosphere makes up the majority of the observed signal (~90-100%). This combined with the non linear attenuation properties of water requires a sensor have a high degree of sensitivity in order properly map the water’s subtle spectral characteristics. In the coastal environment, this is compounded by relatively bright returns from shallow areas in which the bottom albedo is visible. In order to properly capture both shallow and deep water signals effectively, the dynamic range of the sensor is crucial. Limited dynamic range sensors will need to compromise between the range of signals they can detect and the degree of sensitivity they can detect those signals. To overcome this limitation, the PHILLS II utilizes a high dynamic range camera, the 14 bit PlutoCCD PixelVision camera. The second issue is the spectral characteristics of the target itself. Water is a blue rich source. However, traditional CCD cameras are notoriously inefficient in imaging blue light. This limitation was accounted for by the employment of a Scientific Imaging Technologies’ (SITe) thinned, backside-illuminated CCD. Thinned, backside-illuminated chips are essentially normal CCD chips except the silicon wafer that the chip was constructed from was thinned and the chip is flipped upside down when it is mounted in the camera. This process lets incident photons avoid encountering the silicon nitride passivation layer and silicon dioxide and polysilicon gate structures on the front side of
2
the chip before being recorded by the silicon substrate. This greatly increases the quantum efficiency of the chip from ~5% to 60% at 400nm and from ~40% to 85% at 700nm. An additional benefit of the thinned, backside-illuminated design is that the high quantum yield allows for shorter sensor integration times and thus more flexibility in the deployment of the instrument. To further improve the blue light throughput, the spectrograph that was chosen for the sensor, the American Holographics (now Agilent Technologies) HyperSpec VS-15 Offner Spectrograph, was optimized to be efficient at the shorter wavelengths. No matter how thorough the development of the instrument is, the sensor is of little use if its output can not be related to physical reality. The calibration and characterization of this instrument is a multi-step process that includes radiometric calibration, spectral calibration, spatial alignment, geographical correction, and atmospheric correction. Here, we focus on the radiometric calibration and atmospheric correction procedures developed for this system. Davis et al. (2002) has details on the other procedures. RADIOMETRIC CALIBRATION Radiometric calibration relates the digital counts recorded by the sensor to the true physical units of the signal being sampled. To determine this relationship, we employed a 10 lamp integrating sphere as our known source. The lamps and sphere were “NIST” calibrated one month prior to the taking of our measurements. The tungsten-halogen lamps in the integrating sphere are red rich (Figure 1). Obviously, this is not ideal because the signals the instrument will be measuring in the field are blue rich. As can be seen in Figure 2, using the sphere measurements alone are not sufficient. Most of the ocean target spectra are outside of the calibration range of the sphere, resulting in the extrapolation of the retrieved upwelling radiance values from regions outside of the calibration series. With this extrapolation left unaccounted, the effects of small flaws and imperfections within the spectrograph and camera will be exaggerated. And these exaggerated values will get imbedded within the radiometric calibration, and thus, will propagate to the field data when the calibration is applied. Therefore, a series of filters was employed so that the calibration spectra more closely resembled the expected field spectra. The use of these filters over a range of lamp intensities allows us to provide a calibration series that covers almost the entire spectral range of expected targets (Figure 3). It should be noted that prior to utilizing the filters, the per spatial pixel angular response of the camera was used to determine the effective transmission of the filter glass at off angles (Figure 4). The procedure for this measurement is outlined in Davis et al. 2002. As already stated, the mismatch of the calibration lamp spectra and the anticipated field spectra could introduce flaws in to the final data stream by exaggerating flaws inherent to the sensor and projecting them on to the radiometric calibration. In an attempt to characterize the instrument’s inherent flaws, a set of measurements was taken in front of a sphere through several different cutoff filters. By dividing the signal measured with the filter by the signal measured without the filter, the perceived filter transmission was determined (Figure 5). As can be seen, the perceived filter does not approach the true filter transmission in the blue part of the spectrum. In an attempt explain this flaw, it was
3
hypothesized that there was some undiffracted light that was either directly, or through a reflection, corrupting the blue side of the CCD (zero order effect). To combat this effect, a mask was placed in the sensor to shield the CCD from the zero order of the diffraction. The perceived filters were then recalculated (Figure 6). While there was marked improvement after the placement of the mask, the perceived filter responses using this mask were far from ideal. Another potential source of the error is called "out-of-band response." Diffraction gratings are not 100 percent efficient; some percentage of photons will be recorded outside of their intended spectral position. This relationship can be seen here in Figure 7. Assuming that you have a measure of the “out-of-band response” probability distribution function, the calculating the sensors true response is relatively straight forward. However, directly measuring the probability function that would be needed to correct this misappropriation is very difficult. Using the above equation and the cutoff filter data gathered over several sphere lamp settings, a genetic algorithm was constructed to generate the probability matrix that would best solve the perceived filter - true filter comparison (Figure 8). The probability matrix is then used to correct all of the calibration data (as well as, all the field data). Using this corrected data, a linear regression is performed for every element of the CCD (Figure 9). It is this relationship that relates the digital counts collected by the camera to the physical units of W m-2 sr-1 µm-1. APPLICATION This calibration was applied to the data gathered by the PHILLS II at the 2001 HyCODE field study in the summer of 2001. This sensor alone gathered nearly 28,000 square kilometers of hyperspectral data. During the period of July 21st to August 2nd, there were seven flights covering the study site around Long-term Ecological Observatory-15 m (LEO-15) site off the coast of New Jersey (Figure 10). During the process of analyzing the data after the calibration was applied, it became apparent that there were small shift in the spatial dimension (Figure 11). This shift meant that the calibration created in the laboratory could not be directly applied to the field data. While exact cause of this shift remains unknown, it is suspected that in the process of mounting the sensor in the plane a small amount of pressure was applied to the unit, which resulted in a torque in the spectrograph-camera coupling. In the PHILLS II sensor, the CCD camera is not only measuring the spectral information of the target it is viewing, but it also recording the projection of the flaws and imperfections of the sensor itself. These defects could include, but are not limited to, distortions in the lens, mirrors, and diffraction grating, flaws on the slit, and inconsistent responses of different elements on the CCD. While we strive for a sensor free of these defects, by employing the assumption the imperfections remain stable over time, they can be accounted for in the radiometric calibration. It is by creating an autocorrelation of the projection of these flaws as seen in the laboratory and field data sets we can determine the shift that has taken place (Figure 12). An autocorrelation was created for each band (Figure 13). The spatial shift was then used to warp the radiometric calibration, which in turn was applied to the data (Figure 14).
4
Although the spatial shift was wavelength dependant, it became clear after an initial atmospheric correction was applied to the data set that spatial correction did not fully account for the observed spectral shift (Figure 15). To map the extent of the spectral shift, the observed spectral response was compared to the known spectral position of certain atmospheric parameters. The atmospheric parameters used in this comparison were the Fraunhofer Line at 431nm, the Fraunhofer Line at 516nm, the oxygen absorption at 760nm, and the water vapor absorption at 820nm. This analysis was performed for every spatial position along the cross track of the sensor. The result was a map of the apparent spectral shift of every cell in the PHILLS II’s CCD. As can be seen in Figure 16, this procedure not only accounted for the shift of the spectral dimension of the sensor, but it also reflected the sensor’s spectral smile. Finally, this map was used to warp every frame of data collected to a common wavelength array. Prior to the reapplication of the atmospheric correction, two other issues need to be addressed regarding the data. The first is the removal of the scratch that was present in the plane’s window during the time of collection. The effects of this scratch in the data can be seen in Figures 17 and 18. A traditional flat fielding procedure can not be utilized in the removal of the scratch’s effect in the data because the off angle path radiance effects must be preserved for the atmosphere to be properly removed. Exploiting the fact the sensor’s angular response is centered in the center in the CCD and is nearly symmetrical (Figure 4), a correction was developed by comparing the side with the scratch to the side with out the scratch. This correction was then applied to data (see Figures 19 and 20). The second issue has to deal with the noise of the sensor. Before and after the collection of each flight line, a small amount of data is collected with the lens cap on. This data is then averaged and subtracted from each frame of data. This is called dark current removal. However, there is some variation around the mean of the dark current. As can be seen in Figures 21 and 22, this variation is relatively constant across the entire spectrum. However, its influence is much stronger in the longer wavelengths due to the weak response water have in this region. Unfortunately, it is these wavelengths that drive the atmospheric correction. This is further complicated by the fact that there is a temporal pattern to the dark noise. This pattern is not sufficiently removed from the raw data by the average of the dark signal (Figure 23). The magnitude of this dark current periodicity, however, is not constant across the spectrum. This indicates an instability in the CCD we are employing for the PHILLS II. As can be seen in Figure 24, it is the longer wavelengths that are most effected by this. In an attempt to minimize these noise effects prior to the correction of the atmosphere, a procedure was developed. A geographic region of interest (ROI) was created offshore. This area was in deep enough waters so that changes in bathymetry would have no impact of the remotely sensed signal. The ROI contained enough pixels to have statistical significance, but it was small enough so that the assumptions of invariant water type and atmosphere across the ROI were plausible. The ratio of the mean and standard deviation per spectral band of this ROI was compared to a threshold selected by the user. If a band’s metric did not satisfy the threshold, that band was smoothed using a 3 by 3 spatial low pass filter. After this was accomplished the metric was recomputed and retested against the threshold. This cycle continued until the threshold was satisfied.
5
With spectral shift and smile accounted for, the window’s scratch removed, and the dark current’s noise better accounted for, atmospheric correction can be applied. Atmospheric correction was applied using the tabular TAFKAA model (Montes et al., 2001, Gao et. al., 2000). TAFKAA was generated using many elements from Gao’s ATREM model, including most of the geometry, solar spectra, water vapor subroutines (Gao, et. al. 2000). To speed up and automate some of the atmospheric correction processing, TAFKAA uses lookup tables that have been prepared for a variety of solar and observational geometries as well as for five aerosol types (each at five humidities, three wind speeds, and ten optical depths). TAFKAA can run in two modes. One is an optimization mode which uses the “black pixel” assumption (Gordon and Morel, 1983) where the atmospheric variables are adjusted to minimize Lw at a specified near infrared (NIR) wavelength for each pixel. The other mode is by fixing the inputs to known field measurements. As to be expected due to the high NIR scattering in the coastal New Jersey waters, the optimization mode performed poorly. The aerosol types and optical depths that were selected by the optimization clearly mimicked the in water optical features and tended to overcorrect the PHILLS II data. Therefore, it was decided to use fixed inputs over the entire scene. These inputs were derived from coincident field measurements with the exception of wind speed. Of the three wind speed selections (2, 6, 10 m/s), 10 would be the most appropriate for the conditions. However, any speed over 2 m/s generates excessive atmospheric removal and overcorrects the PHILLS II data. A possible explanation for this is how the wind field correction is implemented in TAFKAA and whether the winds are for the surface or a 10m sensor above the surface. FUTURE WORK As can be seen in Figure 25, in the green and blue parts of the spectrum there is good agreement in the initial comparison between the atmospherically corrected data and ground truth spectra with regards to the spectra’s shape. The slight offset in magnitude is likely due to an inaccuracy in the wind speed setting for the atmospheric removal (as outlined above). The disagreement in the red part of the spectrum is likely due to the fact that the ground truth data used a “black pixel” assumption in its correction. However, this is inappropriate due to the high turbidity of this water. Although the comparison showed some promise, the spectral shift procedure outlined here did not completely account for spikiness in the PHILLS II data. Improving this process is clearly a priority. Once this condition has been satisfied, an exhaustive comparison between the PHILLS II atmospherically corrected data and the coincident ground truth data is warranted. Also needed are some improvements to the atmospheric removal procedure itself. To save processing time, TAFKAA was designed to use a single solar geometry per scene and geographically bin the across track field of view. Since each PHILLS II LEO flight line has been treated as an image, the 7 minute flight time and 65 kilometer flight length allows for significant changes in the solar position. The cross track geographic binning produced by TAFKAA results in slight vertical banding in the atmospherically corrected image output. The banding is most noticeable at nadir and at the outer edges of the scene in the cross track dimension. Work is underway to include variable solar
6
geometry and higher resolution spatial binning in TAFKAA, perhaps with a pixel by pixel solution. REFERENCES C. O. Davis, J. Bowles, R. A. Leathers, D. Korwan, T. V. Downes, W. A. Snyder, W. J. Rhea, W. Chen, J. Fisher, W. P. Bissett, and R. A. Reisse (2002), Ocean PHILLS hyperspectral imager: design, characterization, and calibration, Optics Express 10:4, 210--221 H. R. Gordon and A. Y. Morel, Remote Assessment of Ocean Color for Interpretation of Satellite Visible Imagery: A Review Springer-Verlag, New York, 1983.
7
Figure 1: The radiance output of the Naval Research Laboratory’s NIST calibrated integrating sphere for different lamp settings.
Figure 2: The radiometric range the integrating sphere’s lamp cover. Superimposed on the graph are typical sand and water spectra without the atmosphere removed.
8
Figure 3: The radiometric range the integrating sphere’s lamp cover with the use of various filters. Superimposed on the graph are typical sand and water spectra without the atmosphere removed.
Figure 4: The angular resolution viewed per spatial pixel for the PHILLS II sensor.
9
Figure 5: The original perceived filter and true filter responses for three cutoff filters and a blue balancing filter.
Figure 6: The post mask perceived filter and true filter responses for three cutoff filters and a blue balancing filter. 10
True1[1 − ( P12 + P13 + L)] + (True2 P21 + True3 P31 + L) = Measured1 True2 [1 − ( P21 + P23 + L)] + (True1 P12 + True3 P32 + L) = Measured2 True3[1 − ( P31 + P32 + L)] + (True1 P13 + True2 P23 + L) = Measured3 M Truem [1 − ( Pm1 + Pm 2 + L)] + (True1 P1n + True2 P2 n + L) = Measuredm Figure 7: The equations governing the out-of-band response phenomenon. P is the probability that characterizes the photons that were found in band n but belong to band m.
Figure 8: The post out-of-band response perceived filter and true filter responses for three cutoff filters and a blue balancing filter.
11
Figure 9: The regression lines for four different spectra at a single spatial position.
Figure 10: A three color mosaic of the PHILLS II data taken on July 31st, 2001 at the LEO-15 field site.
12
Figure 11: A subsection of a flight line after the radiometric calibration was applied. The vertical banding present is an indication of a mismatch between the field and laboratory calibration.
Figure 12: A horizontal profile of the lab and field data at ~ 800nm. The displacement physical characteristics of each data set can be compared to determine the direction and degree of the spatial shift that occurred.
13
Figure 13: The degree of spatial shift needed per spectral band.
Figure 14: The PHILLS II data after the shifted radiometric calibration was applied. This image covers the same geographic bounds as the image in Figure 11 does. 14
Figure 15: A PHILLS II atmospherically corrected water signal. The spikiness of the spectrum is likely contributed to a spectral offset between the Fraunhofer Lines and absorption bands used by TAFKAA and their position as viewed by PHILLS II.
Figure 16: This is a map of the spectral shift and spectral smile of the PHILLS II sensor for the CCD. The vertical dimension is the spectral dimension and the horizontal dimension is the spatial dimension. Spectrally shorter wavelengths are at the top of the image.
15
Figure 17: An image highlighting the effect of the scratch in the plane’s optical flat. The scratch can be seen running vertically slightly to the right side of the center of the image.
Figure 18: The horizontal profile of the image in Figure 17. The scratch starts around spatial pixel 400 and can be most easily seen in the blue profile.
16
Figure 19: PHILLS II data after the scratch removal was applied This image covers the same geographic bounds as the image in Figure 17 does..
Figure 20: The horizontal profile of the image in Figure 19.
17
Figure 21: The blue line is typical deep water spectra taken by the PHILLS II sensor. The red line represents the typical dark current and the black lines represent a single deviation around the dark current.
Figure 22: This is an enlargement of Figure 21, focused on the longer wavelengths. It is clear that the variation in the dark signal can be very influential in the overall magnitude of the true signal.
18
Figure 23: This is a single band image, at approximately 900nm. The red and blue lines indicate direction and position of the wave field. The effects of the dark current’s periodicity can be seen clearly in the horizontal line pattern highlighted by the green box.
Figure 24: This is a variance map of 1024 frames of dark current as seen by the CCD. The vertical dimension is the spectral dimension and the horizontal dimension is the spatial dimension. Spectrally shorter wavelengths are at the top of the image. While right half of the CCD has far more variation that the left half, the “gull wing” pattern seen in the lower two quadrants is more troubling. This pattern is so evident in this image because cyclical changes in this area over the collection of the dark frames.
19
Figure 25: A comparison between PHILLS II data and ground truth data. These measurements were taken in the Great Bay area near LEO-15. The bay’s waters have a high sediment load. All three are not direct measurements of remote sensing reflectance. Different assumptions were needed to be made for each case, which may be while the agreement between the observations is not better.
20
1
meteorological observations. The remote sensing data consisted of multispectral satellite data and airborne hyperspectral data. The gathered satellite data included SeaWiFS, FY1-C, Oceansat, AVHRR, and MODIS. The hyperspectral campaign included flights of the PHILLS I, PHILLS II, and AVIRIS sensors. The three hyperspectral instruments were flown at different altitudes, and thus, covered a wide range of spatial scales. The PHILLS I flew at 8,500 ft and gathered ~2mx2m spatial data. The PHILLS II flew at 30,000 ft and gathered ~9mx9m spatial data. And AVIRIS flew at 60,000 ft and gathered ~20mx20m spatial data. The respective spectral range and resolution of the instruments are: 400nm to 960nm at 4nm spacing, 400nm to 960nm at 4nm spacing, and 400nm to 2500nm at 10nm spacing Our particular focus in this experiment was on the calibration, deployment, and delivery of the PHILLS II data stream. The purpose of this paper is to outline the steps taken and problems addressed in the development of this data set. BACKGROUND The Portable Hyperspectral Imager for Low-Light Spectroscopy II (PHILLS II) was designed and developed by the Optical Sensing Section of the Naval Research Laboratory. The instrument is an aircraft mounted, push broom type sensor. It utilizes a two dimensional charge coupled device (CCD) camera to collect the spectral information along a single line on the ground perpendicular to the direction the aircraft is traveling. The image cube is built up as the aircraft moves forward in space. Davis et al. (2002) provides a detailed description of the sensor. What distinguishes this instrument from other hyperspectral sensors is that from conception this sensor was designed specifically for oceanic hyperspectral remote sensing. Capturing coastal optical signals from an airborne platform poses two major design challenges that are not usually considerations for terrestrial focused systems. The first challenge is signal sensitivity. Imaging over deep waters and at high altitudes, the atmosphere makes up the majority of the observed signal (~90-100%). This combined with the non linear attenuation properties of water requires a sensor have a high degree of sensitivity in order properly map the water’s subtle spectral characteristics. In the coastal environment, this is compounded by relatively bright returns from shallow areas in which the bottom albedo is visible. In order to properly capture both shallow and deep water signals effectively, the dynamic range of the sensor is crucial. Limited dynamic range sensors will need to compromise between the range of signals they can detect and the degree of sensitivity they can detect those signals. To overcome this limitation, the PHILLS II utilizes a high dynamic range camera, the 14 bit PlutoCCD PixelVision camera. The second issue is the spectral characteristics of the target itself. Water is a blue rich source. However, traditional CCD cameras are notoriously inefficient in imaging blue light. This limitation was accounted for by the employment of a Scientific Imaging Technologies’ (SITe) thinned, backside-illuminated CCD. Thinned, backside-illuminated chips are essentially normal CCD chips except the silicon wafer that the chip was constructed from was thinned and the chip is flipped upside down when it is mounted in the camera. This process lets incident photons avoid encountering the silicon nitride passivation layer and silicon dioxide and polysilicon gate structures on the front side of
2
the chip before being recorded by the silicon substrate. This greatly increases the quantum efficiency of the chip from ~5% to 60% at 400nm and from ~40% to 85% at 700nm. An additional benefit of the thinned, backside-illuminated design is that the high quantum yield allows for shorter sensor integration times and thus more flexibility in the deployment of the instrument. To further improve the blue light throughput, the spectrograph that was chosen for the sensor, the American Holographics (now Agilent Technologies) HyperSpec VS-15 Offner Spectrograph, was optimized to be efficient at the shorter wavelengths. No matter how thorough the development of the instrument is, the sensor is of little use if its output can not be related to physical reality. The calibration and characterization of this instrument is a multi-step process that includes radiometric calibration, spectral calibration, spatial alignment, geographical correction, and atmospheric correction. Here, we focus on the radiometric calibration and atmospheric correction procedures developed for this system. Davis et al. (2002) has details on the other procedures. RADIOMETRIC CALIBRATION Radiometric calibration relates the digital counts recorded by the sensor to the true physical units of the signal being sampled. To determine this relationship, we employed a 10 lamp integrating sphere as our known source. The lamps and sphere were “NIST” calibrated one month prior to the taking of our measurements. The tungsten-halogen lamps in the integrating sphere are red rich (Figure 1). Obviously, this is not ideal because the signals the instrument will be measuring in the field are blue rich. As can be seen in Figure 2, using the sphere measurements alone are not sufficient. Most of the ocean target spectra are outside of the calibration range of the sphere, resulting in the extrapolation of the retrieved upwelling radiance values from regions outside of the calibration series. With this extrapolation left unaccounted, the effects of small flaws and imperfections within the spectrograph and camera will be exaggerated. And these exaggerated values will get imbedded within the radiometric calibration, and thus, will propagate to the field data when the calibration is applied. Therefore, a series of filters was employed so that the calibration spectra more closely resembled the expected field spectra. The use of these filters over a range of lamp intensities allows us to provide a calibration series that covers almost the entire spectral range of expected targets (Figure 3). It should be noted that prior to utilizing the filters, the per spatial pixel angular response of the camera was used to determine the effective transmission of the filter glass at off angles (Figure 4). The procedure for this measurement is outlined in Davis et al. 2002. As already stated, the mismatch of the calibration lamp spectra and the anticipated field spectra could introduce flaws in to the final data stream by exaggerating flaws inherent to the sensor and projecting them on to the radiometric calibration. In an attempt to characterize the instrument’s inherent flaws, a set of measurements was taken in front of a sphere through several different cutoff filters. By dividing the signal measured with the filter by the signal measured without the filter, the perceived filter transmission was determined (Figure 5). As can be seen, the perceived filter does not approach the true filter transmission in the blue part of the spectrum. In an attempt explain this flaw, it was
3
hypothesized that there was some undiffracted light that was either directly, or through a reflection, corrupting the blue side of the CCD (zero order effect). To combat this effect, a mask was placed in the sensor to shield the CCD from the zero order of the diffraction. The perceived filters were then recalculated (Figure 6). While there was marked improvement after the placement of the mask, the perceived filter responses using this mask were far from ideal. Another potential source of the error is called "out-of-band response." Diffraction gratings are not 100 percent efficient; some percentage of photons will be recorded outside of their intended spectral position. This relationship can be seen here in Figure 7. Assuming that you have a measure of the “out-of-band response” probability distribution function, the calculating the sensors true response is relatively straight forward. However, directly measuring the probability function that would be needed to correct this misappropriation is very difficult. Using the above equation and the cutoff filter data gathered over several sphere lamp settings, a genetic algorithm was constructed to generate the probability matrix that would best solve the perceived filter - true filter comparison (Figure 8). The probability matrix is then used to correct all of the calibration data (as well as, all the field data). Using this corrected data, a linear regression is performed for every element of the CCD (Figure 9). It is this relationship that relates the digital counts collected by the camera to the physical units of W m-2 sr-1 µm-1. APPLICATION This calibration was applied to the data gathered by the PHILLS II at the 2001 HyCODE field study in the summer of 2001. This sensor alone gathered nearly 28,000 square kilometers of hyperspectral data. During the period of July 21st to August 2nd, there were seven flights covering the study site around Long-term Ecological Observatory-15 m (LEO-15) site off the coast of New Jersey (Figure 10). During the process of analyzing the data after the calibration was applied, it became apparent that there were small shift in the spatial dimension (Figure 11). This shift meant that the calibration created in the laboratory could not be directly applied to the field data. While exact cause of this shift remains unknown, it is suspected that in the process of mounting the sensor in the plane a small amount of pressure was applied to the unit, which resulted in a torque in the spectrograph-camera coupling. In the PHILLS II sensor, the CCD camera is not only measuring the spectral information of the target it is viewing, but it also recording the projection of the flaws and imperfections of the sensor itself. These defects could include, but are not limited to, distortions in the lens, mirrors, and diffraction grating, flaws on the slit, and inconsistent responses of different elements on the CCD. While we strive for a sensor free of these defects, by employing the assumption the imperfections remain stable over time, they can be accounted for in the radiometric calibration. It is by creating an autocorrelation of the projection of these flaws as seen in the laboratory and field data sets we can determine the shift that has taken place (Figure 12). An autocorrelation was created for each band (Figure 13). The spatial shift was then used to warp the radiometric calibration, which in turn was applied to the data (Figure 14).
4
Although the spatial shift was wavelength dependant, it became clear after an initial atmospheric correction was applied to the data set that spatial correction did not fully account for the observed spectral shift (Figure 15). To map the extent of the spectral shift, the observed spectral response was compared to the known spectral position of certain atmospheric parameters. The atmospheric parameters used in this comparison were the Fraunhofer Line at 431nm, the Fraunhofer Line at 516nm, the oxygen absorption at 760nm, and the water vapor absorption at 820nm. This analysis was performed for every spatial position along the cross track of the sensor. The result was a map of the apparent spectral shift of every cell in the PHILLS II’s CCD. As can be seen in Figure 16, this procedure not only accounted for the shift of the spectral dimension of the sensor, but it also reflected the sensor’s spectral smile. Finally, this map was used to warp every frame of data collected to a common wavelength array. Prior to the reapplication of the atmospheric correction, two other issues need to be addressed regarding the data. The first is the removal of the scratch that was present in the plane’s window during the time of collection. The effects of this scratch in the data can be seen in Figures 17 and 18. A traditional flat fielding procedure can not be utilized in the removal of the scratch’s effect in the data because the off angle path radiance effects must be preserved for the atmosphere to be properly removed. Exploiting the fact the sensor’s angular response is centered in the center in the CCD and is nearly symmetrical (Figure 4), a correction was developed by comparing the side with the scratch to the side with out the scratch. This correction was then applied to data (see Figures 19 and 20). The second issue has to deal with the noise of the sensor. Before and after the collection of each flight line, a small amount of data is collected with the lens cap on. This data is then averaged and subtracted from each frame of data. This is called dark current removal. However, there is some variation around the mean of the dark current. As can be seen in Figures 21 and 22, this variation is relatively constant across the entire spectrum. However, its influence is much stronger in the longer wavelengths due to the weak response water have in this region. Unfortunately, it is these wavelengths that drive the atmospheric correction. This is further complicated by the fact that there is a temporal pattern to the dark noise. This pattern is not sufficiently removed from the raw data by the average of the dark signal (Figure 23). The magnitude of this dark current periodicity, however, is not constant across the spectrum. This indicates an instability in the CCD we are employing for the PHILLS II. As can be seen in Figure 24, it is the longer wavelengths that are most effected by this. In an attempt to minimize these noise effects prior to the correction of the atmosphere, a procedure was developed. A geographic region of interest (ROI) was created offshore. This area was in deep enough waters so that changes in bathymetry would have no impact of the remotely sensed signal. The ROI contained enough pixels to have statistical significance, but it was small enough so that the assumptions of invariant water type and atmosphere across the ROI were plausible. The ratio of the mean and standard deviation per spectral band of this ROI was compared to a threshold selected by the user. If a band’s metric did not satisfy the threshold, that band was smoothed using a 3 by 3 spatial low pass filter. After this was accomplished the metric was recomputed and retested against the threshold. This cycle continued until the threshold was satisfied.
5
With spectral shift and smile accounted for, the window’s scratch removed, and the dark current’s noise better accounted for, atmospheric correction can be applied. Atmospheric correction was applied using the tabular TAFKAA model (Montes et al., 2001, Gao et. al., 2000). TAFKAA was generated using many elements from Gao’s ATREM model, including most of the geometry, solar spectra, water vapor subroutines (Gao, et. al. 2000). To speed up and automate some of the atmospheric correction processing, TAFKAA uses lookup tables that have been prepared for a variety of solar and observational geometries as well as for five aerosol types (each at five humidities, three wind speeds, and ten optical depths). TAFKAA can run in two modes. One is an optimization mode which uses the “black pixel” assumption (Gordon and Morel, 1983) where the atmospheric variables are adjusted to minimize Lw at a specified near infrared (NIR) wavelength for each pixel. The other mode is by fixing the inputs to known field measurements. As to be expected due to the high NIR scattering in the coastal New Jersey waters, the optimization mode performed poorly. The aerosol types and optical depths that were selected by the optimization clearly mimicked the in water optical features and tended to overcorrect the PHILLS II data. Therefore, it was decided to use fixed inputs over the entire scene. These inputs were derived from coincident field measurements with the exception of wind speed. Of the three wind speed selections (2, 6, 10 m/s), 10 would be the most appropriate for the conditions. However, any speed over 2 m/s generates excessive atmospheric removal and overcorrects the PHILLS II data. A possible explanation for this is how the wind field correction is implemented in TAFKAA and whether the winds are for the surface or a 10m sensor above the surface. FUTURE WORK As can be seen in Figure 25, in the green and blue parts of the spectrum there is good agreement in the initial comparison between the atmospherically corrected data and ground truth spectra with regards to the spectra’s shape. The slight offset in magnitude is likely due to an inaccuracy in the wind speed setting for the atmospheric removal (as outlined above). The disagreement in the red part of the spectrum is likely due to the fact that the ground truth data used a “black pixel” assumption in its correction. However, this is inappropriate due to the high turbidity of this water. Although the comparison showed some promise, the spectral shift procedure outlined here did not completely account for spikiness in the PHILLS II data. Improving this process is clearly a priority. Once this condition has been satisfied, an exhaustive comparison between the PHILLS II atmospherically corrected data and the coincident ground truth data is warranted. Also needed are some improvements to the atmospheric removal procedure itself. To save processing time, TAFKAA was designed to use a single solar geometry per scene and geographically bin the across track field of view. Since each PHILLS II LEO flight line has been treated as an image, the 7 minute flight time and 65 kilometer flight length allows for significant changes in the solar position. The cross track geographic binning produced by TAFKAA results in slight vertical banding in the atmospherically corrected image output. The banding is most noticeable at nadir and at the outer edges of the scene in the cross track dimension. Work is underway to include variable solar
6
geometry and higher resolution spatial binning in TAFKAA, perhaps with a pixel by pixel solution. REFERENCES C. O. Davis, J. Bowles, R. A. Leathers, D. Korwan, T. V. Downes, W. A. Snyder, W. J. Rhea, W. Chen, J. Fisher, W. P. Bissett, and R. A. Reisse (2002), Ocean PHILLS hyperspectral imager: design, characterization, and calibration, Optics Express 10:4, 210--221 H. R. Gordon and A. Y. Morel, Remote Assessment of Ocean Color for Interpretation of Satellite Visible Imagery: A Review Springer-Verlag, New York, 1983.
7
Figure 1: The radiance output of the Naval Research Laboratory’s NIST calibrated integrating sphere for different lamp settings.
Figure 2: The radiometric range the integrating sphere’s lamp cover. Superimposed on the graph are typical sand and water spectra without the atmosphere removed.
8
Figure 3: The radiometric range the integrating sphere’s lamp cover with the use of various filters. Superimposed on the graph are typical sand and water spectra without the atmosphere removed.
Figure 4: The angular resolution viewed per spatial pixel for the PHILLS II sensor.
9
Figure 5: The original perceived filter and true filter responses for three cutoff filters and a blue balancing filter.
Figure 6: The post mask perceived filter and true filter responses for three cutoff filters and a blue balancing filter. 10
True1[1 − ( P12 + P13 + L)] + (True2 P21 + True3 P31 + L) = Measured1 True2 [1 − ( P21 + P23 + L)] + (True1 P12 + True3 P32 + L) = Measured2 True3[1 − ( P31 + P32 + L)] + (True1 P13 + True2 P23 + L) = Measured3 M Truem [1 − ( Pm1 + Pm 2 + L)] + (True1 P1n + True2 P2 n + L) = Measuredm Figure 7: The equations governing the out-of-band response phenomenon. P is the probability that characterizes the photons that were found in band n but belong to band m.
Figure 8: The post out-of-band response perceived filter and true filter responses for three cutoff filters and a blue balancing filter.
11
Figure 9: The regression lines for four different spectra at a single spatial position.
Figure 10: A three color mosaic of the PHILLS II data taken on July 31st, 2001 at the LEO-15 field site.
12
Figure 11: A subsection of a flight line after the radiometric calibration was applied. The vertical banding present is an indication of a mismatch between the field and laboratory calibration.
Figure 12: A horizontal profile of the lab and field data at ~ 800nm. The displacement physical characteristics of each data set can be compared to determine the direction and degree of the spatial shift that occurred.
13
Figure 13: The degree of spatial shift needed per spectral band.
Figure 14: The PHILLS II data after the shifted radiometric calibration was applied. This image covers the same geographic bounds as the image in Figure 11 does. 14
Figure 15: A PHILLS II atmospherically corrected water signal. The spikiness of the spectrum is likely contributed to a spectral offset between the Fraunhofer Lines and absorption bands used by TAFKAA and their position as viewed by PHILLS II.
Figure 16: This is a map of the spectral shift and spectral smile of the PHILLS II sensor for the CCD. The vertical dimension is the spectral dimension and the horizontal dimension is the spatial dimension. Spectrally shorter wavelengths are at the top of the image.
15
Figure 17: An image highlighting the effect of the scratch in the plane’s optical flat. The scratch can be seen running vertically slightly to the right side of the center of the image.
Figure 18: The horizontal profile of the image in Figure 17. The scratch starts around spatial pixel 400 and can be most easily seen in the blue profile.
16
Figure 19: PHILLS II data after the scratch removal was applied This image covers the same geographic bounds as the image in Figure 17 does..
Figure 20: The horizontal profile of the image in Figure 19.
17
Figure 21: The blue line is typical deep water spectra taken by the PHILLS II sensor. The red line represents the typical dark current and the black lines represent a single deviation around the dark current.
Figure 22: This is an enlargement of Figure 21, focused on the longer wavelengths. It is clear that the variation in the dark signal can be very influential in the overall magnitude of the true signal.
18
Figure 23: This is a single band image, at approximately 900nm. The red and blue lines indicate direction and position of the wave field. The effects of the dark current’s periodicity can be seen clearly in the horizontal line pattern highlighted by the green box.
Figure 24: This is a variance map of 1024 frames of dark current as seen by the CCD. The vertical dimension is the spectral dimension and the horizontal dimension is the spatial dimension. Spectrally shorter wavelengths are at the top of the image. While right half of the CCD has far more variation that the left half, the “gull wing” pattern seen in the lower two quadrants is more troubling. This pattern is so evident in this image because cyclical changes in this area over the collection of the dark frames.
19
Figure 25: A comparison between PHILLS II data and ground truth data. These measurements were taken in the Great Bay area near LEO-15. The bay’s waters have a high sediment load. All three are not direct measurements of remote sensing reflectance. Different assumptions were needed to be made for each case, which may be while the agreement between the observations is not better.
20
