Principal Component Analysis of Arctic Solar Irradiance Spectra
Research Staff: Maura Rabbette and Peter Pilewskie
During the FIRE (First ISCPP Regional Experiment) Arctic Cloud Experiment and coincident SHEBA (Surface Heat Budget of the Arctic Ocean) campaign in Alaska, detailed moderate resolution solar spectral measurements were made to study the radiative energy budget of the coupled Arctic Ocean -Atmosphere system. The NASA Ames Solar Spectral Flux Radiometers (SSFRs) were deployed on the NASA ER-2 and at the SHEBA ice camp. Using the SSFRs we acquired continuous solar spectral irradiance (380-2200 nm) throughout the atmospheric column.
Principal Component Analysis (PCA) was used to characterize the several tens of thousands of retrieved SSFR spectra and to determine the number of independent pieces of information that exist in the visible to near-infrared solar irradiance spectra. It was found in both the upwelling and downwelling cases that almost 100% of the spectral information (irradiance retrieved from 1820 wavelength channels) was contained in the first six extracted principal components. The majority of the variability in the Arctic downwelling solar irradiance spectra was explained by a few fundamental components including infrared absorption, scattering, water vapor and ozone. PCA analysis of the SSFR upwelling Arctic irradiance spectra successfully separated surface ice and snow reflection from overlying clouds into distinct components.
An image of an entire upwelling solar irradiance spectral data array (1821 x 11260), acquired by the airborne SSFR on the high altitude NASA ER-2 as it flew over the Arctic terrain during the FIREIII/SHEBA experiment is shown in Figure 1a. This is an example of a typical SSFR data matrix. The 1821 columns represent the time series of upwelling solar irradiance for wavelengths between 380-2200 nm and the 11260 rows are individual solar irradiance spectra acquired during partial sampling between May 22 and June 6, 1998. The relatively darker rows indicate cloudy conditions. The bright regions depict the mid-visible maximum irradiance at solar noon. The distinct dark columns on the right side of the image are water vapor bands at 940, 1400 and 1860 nm. The dark columns due to the oxygen A and B bands (762 nm and 690 nm) are also evident. The average spectrum of all 11260 upwelling visible to near-infrared (380-2200 nm) solar irradiance spectra is shown in Figure 1b. The above mentioned water vapor and oxygen absorption bands are the most prominent spectral features. At shorter wavelengths (less than 500 nm), some features due to the solar Fraunhofer lines can be seen.
Figure 2a shows the first four rotated principal components extracted from the ER-2 SSFR upwelling solar irradiance spectral array. As can be seen from Figure 2b the first six principal components account for 97.4% of the spectral variability of more than eleven thousand upwelling solar irradiance spectra retrieved over varying albedos including tundra, open water, ice/snow, and cloud cover. PC1 at 52% variance has high weightings in the visible region and a comparison between its time series and the ER-2 latitude position time series suggests that this component is actually linked to surface albedo. PC2 at 20% is mainly an infrared component and its time series appears to be associated with cloud cover. When cloud is introduced over a surface it reduces the near-infrared atmosphere/surface absorption. The extraction of PC1 and PC2 is an important and useful result for the separation of a highly reflective surface from an overlying reflective cloud layer. This type of analysis could be applied to satellite remote sensing to resolve problems associated with distinguishing arctic ice from cloud. PC3 at 10% variance has amplitude maxima in the water vapor bands. PC4 has broad band loadings at infrared wavelengths between 1010 - 1100 nm and 2000 - 2200 nm. The physical significance of PC4, is not understood. The remaining 2.6% of the variance is comprised mainly of random noise.
Principal component analysis when applied to several thousand continuous irradiance spectra gives an overall statistical picture that helps us understand and appreciate the bigger ensemble. With PCA it is possible to quantify the variability in the spectral flux and we find that in all cases only a few principal components are needed to account for almost 100% of the spectral variance. PCA is also an effective instrument diagnostic tool. The use of PCA on ground and airborne SSFR spectra is sufficiently mature to give us confidence in applying similar methods to satellite spectra. One goal is to automate the PCA tools and to further develop retrieval algorithms and on-board data processing capabilities that will be necessary for future "smart" sensors. These developments will be very important for future remote sensing experiments and onboard reduction of large volumes of satellite hyperspectra, for performing instrument diagnostics, instrument validation, for relating measured spectral variability to physical causes, and ultimately constraining climate models.
Collaborators: Sean Twomey
Point of Contact: Maura Rabbette, 650/604-0128, firstname.lastname@example.org
Figure 1:(a) Image of entire 11260 upwelling solar irradiance spectra acquired with the airborne ER-2 SSFR from May 22 - June 6 1998. Rows (11260): One row = One spectrum. The relatively darker rows indicate cloudy conditions. The bright regions depict the mid-visible maximum irradiance at solar noon. Columns (1820): SSFR channels (wavelengths) 380-2200 nm. The distinct dark columns on the right side of the image are water vapor bands at 1400 and 1860 nm. (b) Mean of all 11260 upwelling solar irradiance spectra acquired with the ER-2 SSFR. (The gap centered around the 1000 nm channel is due to poor signal/noise in that region.)
Figure 2:(a) The first four rotated principal components extracted from the ER-2 SSFR upwelling solar irradiance visible-infrared (380-2200 nm) array. (b) The percentage of variance accounted for by the first 21 vis-nir components.