Figure 2 shows the scatter plot between the MI T b and the IASI T b for the whole study period and Table 2 summarizes the statistical parameters. Overall, there is excellent agreement between the two instruments' data, although the comparison results depend highly on the channel. For example, all four channels show high correlation coefficients, although the WV channel shows a slightly smaller correlation coefficient of 0.997. In terms of bias, they are all satisfactory, showing better than required NEdT, except for two cases. First of all, the WV channel shows a large negative bias value of about -0.77 K. Also, as shown in Fig. 3, the bias depends on the target temperature. For example, there is a large negative bias of about -1 K in a higher tab, which represents the clear-sky conditions. On the other hand, there is a positive bias of about 1 K in a lower tab, which is mainly obtained in cloudy conditions. Based on the work of (Wu and Yu, 2013), these kinds of temperature-dependent bias characteristics are caused mainly by the uncertainty in the SRF of the specific channel, effectively represented by the uncertainty in the center wavelength of the SRF. To derive the exact number for the SRF shift, further considerations such as the re-calibration of raw MI data along with the adjustment of derived IASI radiance are required, which are under investigation and will be reported in another paper. Another noticeable bias characteristic is shown in the daytime case of the IR1 channel; the difference between daytime and nighttime is about 0.17 K (0.06 K vs. 0.23 K), while it is about 0.06 K in the IR2 channel. On the other hand, there is no significant difference in the WV channel. The large difference in the window channel is further described later (refer Fig. 4). Finally, as for RMSD, the IR1 and IR2 channels show quite small values, less than about 0.7 K; and that of the WV channel is also close to 1 K, which is the calibration uncertainty of the MI (Kim and Ahn, 2014). However, the SWIR channel shows a much larger RMSD value of 1.39 K due mainly to the increased NEdT at the lower temperature. This demonstrates the high scatteredness and is described in more detail below.
As shown in Fig. 2, the T b difference between the MI and IASI seemingly depends on the measured T b. To characterize the difference with the measured T b, Fig. 3 shows the T b difference as a function of the MI T b. Overall, the T b difference is much smaller and stable at the higher MI T b and increases with the cooler MI T b. Bias and RMSD values of the IR1 and IR2 channels show quite a small value at the warm MI T b, begin to increase at about 250 K, and show relatively large positive bias of about 1 K at the MI T b of 200 K. This is explained by the fact that the IR channels are highly sensitive to the presence of cloud and its variability of properties such as optical depth. Thus, the sensitivity and NEdT of the IR channels increase with decreasing measured T b, which is mainly caused by the presence of high clouds. It is most significant in the SWIR channel, which shows a sudden increase of variability starting from the MI T b value of 260 K, and reaches a negative bias and large RMSD at the MI T b of 230 K. The higher variability of the difference at the lower measured T b in the SWIR channel has also been shown in other inter-comparison studies (Hewison et al., 2013; Wu et al., 2009). This can be caused by different sensitivities between T b and the radiance, i.e. the T b sensitivity against radiance increases significantly with decreasing radiance in the SWIR region. Another plausible cause is due to the wide spectral range compensation in addition to the relatively large noise of IASI radiances around the SWIR channel. This effect is exaggerated in the SWIR channel of the MI due to the increased dynamic range, as described before. On the other hand, the WV channel shows quite different characteristics. The bias shows a rather large negative value at the warm MI T b, while it slowly increases with increasing MI T b and reverses to a positive value at the coldest MI T b. The RMSD value also shows a strong MI T b dependence, being larger at the cooler MI T b, which is primarily related with the systematic bias of the MI T b.
The optical path length to the instrument onboard the satellite changes greatly with the observation geometry, thus affecting the measured radiances. Therefore, it is interesting to check whether the simple threshold test applied for the viewing geometry results in any significant residual effects in the comparision results. For this, the T b differences as a function of the MI zenith angle are shown in Fig. 4. At first glance, there seems to be no significant variation in the MI zenith angle for all four channels. However, closer inspection reveals that bias and STD are worst near the nadir and improve with increasing zenith angle. For example, in the SWIR channel where the variation is most prominent, bias (STD) is -0.35 (3.5) K at the MI zenigh angle of 10°, while it is 0.12 (2.7) K and 0.00 (2.4) K at 25° and 35°, respectively. The slight improvement in the bias and STD toward the higher MI zenith angle is shown in all IR channels except the WV channel, which shows little variation. The zenith angle dependence is thought to be caused by the fact that the viewing angle difference is the largest at nadir and smallest at the scan edge. However, the overall variation of statistical parameters with MI zenith angle is much smaller than the calibration uncertainty of the MI; therefore, it is concluded that the simple threshold check does not introduce a significant residual.
Based on the monitoring results of legacy instruments, such as the imagers onboard the GOES (Geostationary Operational Environmental Satellite) series, the calibration coefficient and its stability show diurnal and seasonal variation (Weinreb and Han, 2003), due mainly to the relative geometry between the sun and satellite, which introduces the temperature variation of the instrument's optics. Thus, it is also imperative to check if there is any temporal variability in the measured T b. For this, the inter-comparison results are given as a time series of monthly mean bias and RMSD in Figs. 5 and 6, respectively. There are several characteristics to be noted in the monthly mean values. First, it is clear that there is no significant long-term drift in both bias and RMSD for all channels. According to the RMSD, the overall performance is slightly better during the recent months compared to the early operation period.
Second, there is seasonal dependence in both bias and RMSD (weak in bias and rather prominent in RMSD) in all four channels. For example, RMSD is maximum during the summer, around July, and minimum around Feburary. The variation is largest in the SWIR channel, where it is most sensitive to the noise signal at lower T b. This kind of seasonal dependence has also been shown in comparison results based on similar instruments onboard GOES-11 and MTSAT-2, although it is much weaker in the case of GOES-12 (Hewison et al., 2013, Figs. 9-11). The seasonal dependence is closely related to the variation of solar illumination angles, which results in an increase of the detector patch temperature and stronger optics variation, including internal blackbody, scan mirror etc. (Weinreb and Han, 2003).
The large day and night difference in bias during the summer months shown in Figs. 5 and 6 is closely related with that of the IR1 and IR2 channels in Table 2. During the early period of operation, say September 2011, the day and night difference in bias is as large as 0.8 K in the IR1 channel and about 0.4 K in the IR2 channel, while it is insignificant in the WV channel. One plausible explanation for the channel and seasonal dependency of the day and night difference is the increased skin temperature over the land mass during the daytime of the summer season. Then, any inhomogeneity or difference between the MI targets and IASI pixels could be amplified. However, the bias and RMSD derived from the ocean-only dataset shows only a slight improvement, 0.20 K vs 0.23 K, for the daytime in the IR1 channel, while there are no changes in the WV channel. Thus, the exact causes of the rather large day and night difference, albeit which has gradually decreased since the commissioning of the satellite, remain to be solved.
Finally, the monthly mean bias for the WV channel shows a clear negative bias and its absolute magnitude is much larger than that of the other three channels. However, its seasonal variation is much weaker. This is also true for the RMSD, which shows a smooth variation but with comparatively smaller absolute values, near 1 K for all months. This kind of large negative mean bias shown in the MI is not apparent in the comparison results from either MTSAT-2 or the GOES series. MTSAT-2 shows a stable negative bias with much lower value (around -0.3 K), while the GOES series show positive and negative values, within 0.3 to 0.4 K (Hewison et al., 2013). Based on the stable RMSD and consistent bias trend with the comparison results from other satellites, we consider this characteristic is largely due to the uncertainty in the spectral response function, although it has not been solidly demonstrated (Wu and Yu, 2013).
The threshold values used for the preparation of the matchup dataset are similar but slightly different from the previous GSICS activities (Tahara, 2008; Hewison et al., 2013). Although most of the participating organizations utilize the approaches and recommended threshold values, the specific process and threshold values could be modified to accommodate their own characteristics (Wu and Yu, 2011; Hewison et al., 2013). Thus, to perform a simple sensitivity test, we check the variability of the comparison results with the different sets of thresholds used for each criteria test. For simplicity, we vary the threshold values by halving and doubling them for the time and uniformity tests. For the viewing geometry, we apply an extremely small difference of 0.003 and a moderately small difference of 0.01, which is the value recommended for all conditions if the number of data is not a significant issue (Wu and Yu, 2011). A total of eight different cases are selected and the combinations of the different threshold values are summarized in Table 3.
Table 4 summarizes the variation of the statistical parameters with the different sets of the threshold values applied to all-day data (the variability characteristics for day and night are almost the same; thus, we show only the results from all-day data), and Fig. 7 shows the scatter plot of the MI T b and the IASI T b of the IR1 channel for the different cases. First, the utilization of a 1 minute temporal threshold value (Case I) reduces the number of matchup data by about 80%, with a slight improvement in bias and RMSD. When the threshold value is relaxed to 15 minutes (Case II), the data points increase more than tenfold, with a slight degradation of bias and RMSD. Thus, unless there is a strong requirement for an increase of the matchup dataset, the 5 minute threshold looks a reasonable choice. For the viewing geometry, if the threshold value is reduced to 0.01, the available data points decrease to about one-third, with slight improvements in both bias and RMSD, especially in the SWIR channel. Even with the threshold value of 0.003, the improvements in bias and RMSD are not that impressive, although the number of data points is reduced to about one-tenth. Thus, it turns out that the error statistics are not very sensitive to the threshold values for the temporal and viewing geometry and the current threshold values provide a reasonable number of data points and error statistics.
On the other hand, when we tighten the uniformity constraint by twofold (Case V), the bias and RMSD show significant variation, with data reduction of about 40%. In the IR1 channel, both bias and RMSD improve by 0.03 K and 0.22 K, respectively. The IR2 channel shows similar improvement. However, the SWIR channel shows an opposite result, i.e. bias and RMSD degrade by about 0.17 K and 0.09 K, respectively. In the case of the WV channel, bias is degraded slightly by 0.05 K, although the RMSD improves significantly, by 0.14 K. This might explain some of the MI T b dependence of the RMSD of the WV channel shown in Fig. 4. Thus, with a tightened threshold value, the statistical parameters improve significantly, with the exception of the SWIR channel, which shows a significant degradation of bias and RMSD. On the other hand, when the constraint is relaxed by twofold (Case VI), then the statistical parameters, especially the RMSD, show clear degradation in all channels, by as much as 0.43 K for RMSD in the IR2 channel. Thus, the comparison results are sensitive to the threshold value used for the uniformity test and the responses are different for different channels and error parameters.
Finally, the threshold for the radiance difference between the target pixel and the environment pixel is varied by halving (Case VII) and doubling (Case VIII) the value. Overall, the number of data points varies linearly with the threshold value, although the bias and RMSD do not show any significant variation. The largest difference is seen for bias in the SWIR channel with the relaxed constraint, with the degradation being about 0.04 K. Thus, as found for other threshold values that are not changed, changing the threshold value for the radiance difference does not produce significant variation in the comparison results.
From the simple sensitivity test, we can draw several conclusions. First, the comparison results are significantly sensitive to the changes in the threshold values for the environment uniformity test, especially a large change in RMSD. However, the sensitivity also depends on the channel, which suggests it is wise to use different sets of threshold values for different channels, if more detailed analysis for a certain channel is required. As a result, the changes in threshold values for other tests do not introduce significant variation in the comparison results and the baseline sets give quite representative numbers, which are comparable with the expected instrument performance (Kim and Ahn, 2014), and a similar inter-satellite comparison study (Kim et al., 2014).