4.3.1. Radiometric Calibration Accuracy, Version History, and Caveats
 The goals for radiometric calibration accuracy developed for CRISM prior to MRO launch were (1) 10% absolute accuracy at 630 nm (to support mixture modeling using laboratory spectra); (2) 1% relative accuracy, comparing 0.45 and 0.75 μm (to measure the decrease in reflectance below 0.8 μm due to ferric iron); (3) 0.5% relative accuracy, for adjacent channels near 1 μm (to measure the 1-μm absorption due to ferrous and ferric minerals); (4) 0.25% relative accuracy, for adjacent channels near 2.3 μm (to measure absorptions due to phyllosilicates, hydrated silica and sulfates, and carbonates); and (5) 1% relative accuracy, comparing 1.6 μm versus 2.5 μm and 2.5 versus 3.1 μm (to measure spectral continuum, pyroxene absorption bands, and depth of the 3-μm H2O absorption).
 Attainment of these goals was evaluated over the VNIR wavelength range by comparing CRISM data with simultaneously acquired PANCAM measurements of large, uniform areas at the MER landing sites (e.g., “Cliff Hanger” in Figure 19 of Arvidson et al. ). Over the IR wavelength range it was evaluated by comparison of CRISM data with OMEGA measurements of the same sites. In both cases, corrections for photometric and atmospheric effects were made to simulate OMEGA or PANCAM data at CRISM geometry. The results of these comparisons and other results led to updates in the calibration; the values reported below pertain to the current version, version 2. In the comparison of CRISM's VNIR wavelength range with PANCAM data, goal 1 (absolute calibration) is met, and goal 2 (spectral slope at 0.45–0.75 μm) is close to being met. The comparison of CRISM's VNIR wavelength range at 440–1000 nm with PANCAM data suggests that absolute calibration at those wavelengths may be good typically to ±5%. Independent comparison of CRISM's VNIR wavelength range with MARCI data yields similar agreement to within ±2–3% [Wolff et al., 2009]. The relative accuracy of wavelengths near 1 μm does not meet goal 3 and is closer to 1%, but is expected to meet this goal after the upgrade to calibration version 3 described below. Systematic channel-to-channel variations appear in multiple scenes and may be attributable to errors in the radiometric model of the integrating sphere; reduction of these artifacts is planned for version 3. The relative accuracy channel-to-channel near 2.3 μm either meets or is close to goal 4, 0.25%. Goal 5 was not initially met owing to errors in continuum slope as described below, that were addressed using CRISM observations of Deimos, but this goal may now be met.
 As is typical of other imaging investigations on planetary missions, as data were acquired in-flight, previously recognized calibration issues have become better quantified and some new issues have emerged. In response, the calibration algorithm goes through periodic upgrades. Which calibration version was used to process a particular observation is coded as the last character of the root file name. In addition, the full list of calibration files applied to the data is given in the detached PDS label. The initial version of the calibration, version 0, was based on strict application of ground-derived performance (most notably the radiometric model of the integrating sphere), and using a relative simple estimate of scattered light from the grating (interpolating between values from the scattered light columns of the detector). Once Mars observations were acquired, several problems became apparent, especially (1) a field angle–dependent shape to VNIR spectra that was traced to the overly simplistic treatment of scattered light from the grating; (2) systematic, column-to-column differences in spectral shape, traced to uncorrected artifacts of noise in ground calibration data; (3) inaccurate spectral shape at VNIR wavelengths <0.5 μm, traced to random noise and systematic processing errors using the low internal integrating sphere radiance at those wavelengths; (4) spurious, sharp peaks or troughs at boundaries of detector zones, traced to light leakage at the boundaries of the order-sorting filters; (5) shallow, systematic bumps at wavelengths where effects of atmospheric H2O vapor were inaccurately removed from ground calibration data; (6) intermittently elevated values of I/F at particular detector columns and wavelengths, traced to uncorrected, transient elevated bias in particular detector elements, discussed below; and (7) a broad, shallow rise in I/F centered near 3.4 μm, resulting from incomplete implementation of a correction for leaked second-order light in zone 3 of the IR detector.
 The next version of the calibration, IR version 1 and VNIR version 2 (VNIR version 1 having been abandoned), addressed the first six of the above issues. Changes made to the processing included improved noise correction and interpolation across detector-zone boundaries in ground-based calibration data, modeling scattered light as a polynomial function of field angle, adoption of fixed values of responsivity at <0.56 μm, and introduction of a rudimentary filter to interpolate over detector elements when they exhibit transient high bias levels.
 VNIR version 2 and IR version 1 were used in the comparison with OMEGA measurements of bland, dusty areas of Mars, and a comparison of CRISM measurements of Deimos with ground-based measurements [Lynch et al., 2007]. Discrepancies drove an update in IR calibration to version 2. The comparison with OMEGA indicated systematic differences between the two data sets in continuum slope at >1.5 μm as well as a recurring, systematic “bump” near 2.55 μm. The same features appeared in the spectrum of Deimos, which is an ideal calibration target because its spectrum is believed to be smooth and it exhibits no atmospheric gas absorptions. The 2.55-μm bump was traced to incomplete removal of atmospheric water vapor from the spectra of the ground calibration standard used for the integrating sphere; the error in the continuum slope at >1.5 μm was modeled as a small pointing error at the ground calibration standard, which slightly vignetted the standard (incoming rays from the standard did not reach part of the aperture, affecting how the multizoned gratings were sampled). Corrections for both effects were derived from first principles of instrument performance. Additional changes in IR version 2 included (1) refinements to the corrections for noise and light leakage in detector-zone boundaries in ground-based calibration data; (2) increasing the threshold for filtering pixels with elevated bias, to remediate aliasing that the initial filter introduced into the data; and (3) implementation of the correction for leaked second-order light in zone 3 of the IR detector. After all these changes, a shallow “ramp” remained near 1.9 μm and it was corrected empirically.
 Six known issues remaining in VNIR and IR version 2 are being investigated, and will be addressed in an upgrade to “version 3” or later versions. The first issue is systematic channel-to-channel variations, discussed above, with will be addressed by removing the corresponding systematic error in the integrating sphere model. The second issue is the temporal drift in wavelength calibration, that is, the mapping of detector rows to wavelengths. Seasonal variations in optics temperatures introduced a time-dependent variation in CRISM's wavelength calibration about 1 nm in magnitude (0.15 detector rows) [Smith et al., 2009]. The inaccuracy in wavelength calibration results in errors in removal of CO2 gas absorptions and elevation-dependent errors in the 2-μm wavelength region in atmospherically corrected data. The resulting artifact introduces error into 1.9- to 2.1-μm absorptions owing to ices and bound water. A time-dependent wavelength calibration has subsequently been developed that has with an accuracy of about 0.13 nm, as has a time-segmented volcano-scan atmospheric correction. Either the usage of the time-segmented volcano-scan correction or the incorporation of a time-dependent wavelength calibration into the Lambert albedo correction largely removes this effect. The third issue is high time-frequency variations in detector bias, or “noisy pixels,” which affect up to a couple percent of the IR detector elements. In spatial images at some wavelength, this appears as streaks, and the effect is worse and affects more detector elements at higher detector temperatures. The effects of these bias variations are being addressed using a kernel filter that identifies outlying values from various detector elements statistically, and then interpolates over them from the surrounding data. The fourth issue is a spurious peak or trough near 3.18 μm that appears in parts of the field of view, intermittently. It has been traced to an artifact of the zone 1-zone 2 boundary in the correction for leaked second-order light in zone 3 of the IR detector, and is corrected in version 3. The fifth issue is time variations in spectral shape at <1.5 μm which result from slight irreproducibility in shutter position viewing the internal integrating sphere, that change the way integrating sphere radiance is sampled by the two zones of the IR grating. This is being corrected using the known effect of the phenomenon, which can be scaled from observation to observation by its signature at the boundary of the two VNIR detector zones. This may be remedied by shifting the wavelength of the correction. The final issue is high spectral frequency oscillations at >3300. These are due to a Fabry-Perot effect in the zone of the detector-mounted filter that blocks out-of-order light at >2.7 μm. Small variations in that filter's temperature (thickness) between observations of Mars and of the internal integrating sphere cause ∼5% oscillations in system response as a function of wavelength not to cancel out. This will be addressed by using weighted averages of temporally adjacent integrating sphere measurements, instead of using the closest one in time as is the current practice.
 Two other radiometric calibration issues are so integral to the data that further corrections are not likely. The first of these latter issues is incomplete removal of grating scatter at <0.44 μm. At these wavelengths, for most Mars scenes, radiance is low and strongly affected by residuals from the correction for grating scatter. The second issue is leakage of light though boundaries of different zones of the detector-mounted filters. This leakage is scene-dependent and presently cannot be accurately estimated and removed.
 Given the various issues outlined above, some guidelines can be prescribed for reliability of the radiometric calibration at different wavelengths. The following channels at the ends of detector zones can be routinely excluded: (1) VNIR wavelengths less than 410 nm, between 644 and 684 nm, and greater than 1023 nm. (2) IR wavelengths less than 1021 nm, 2694 and 2701 nm, and greater than 3924 nm.
 The following channels may be degraded in their accuracy in some observations, but intrascene variations appear to be valid. In other words, information from the following channels can be recovered by ratioing to some spectrally bland part of the same scene, preferably in the same column(s) of an image: (1) VNIR wavelengths 410–442 nm (due to artifacts from the correction for grating scatter, in very contrasty scenes). (2) VNIR wavelengths 970–1023 nm and IR wavelengths 1021–1047 nm (the radiances can misalign between detectors; the reason is speculated to be uncorrected effects of beam splitter temperature). (3) IR wavelengths near 1650 nm surrounding the zone 1-zone 2 boundary. (4) IR wavelengths 2660–2800 nm (the reason is uncertain but may be due to problems with correction of water vapor in measurements of the ground calibration standard). (5) IR wavelengths greater than 3700 nm (there is a scene-dependent turndown in radiances beyond 3700 nm due to unknown causes).
 In terms of scientific interpretation of the data, the implications of calibration artifacts depend on preprocessing and the scene. Where it is possible to ratio a region of interest to a spectrally bland region in or near the same detector column, narrow absorptions due to phyllosilicates, carbonates, silica, or sulfates can be measured in the relative spectra to an accuracy typically of much better than 1%. The information lost in the ratioing process is the visible wavelength red slope and the exact shape of the 1-μm absorption due to ferrous minerals. In unratioed spectra, the primary cautions on interpretations are that (1) uncorrected temporal variations in shape of IR spectra at <1.5 μm can lead to errors in models of the relative abundances of olivine and pyroxene; (2) the 3.18-μm artifact can obscure the 3.4-μm carbonate absorption, where carbonate is present; and (3) errors in correction for atmospheric CO2 at times of extreme instrument temperatures lead to a superimposed “ripple” near 2 μm that can partially obscure the 1.9-μm absorption owing to molecular H2O-bearing phases.