HTML
-
To optimize the threshold parameters for cloud screening with results from the contingency table analysis, we analyzed the variation of summary statistics and diagnostic variables independently with the OCO-2 measurements and MODIS reference state in January and July 2016. The collocated cloud screening dataset in January is composed of 71 orbits, with 311 500 soundings passing over the selected region in Europe. The collocated dataset in July is composed of 72 orbits, and a total of 304 438 soundings.
Figure 3 shows the changes of diagnostic values, throughput, agreement, and PPV, in response to altering cloud screening thresholds in a chosen range. Based on the July dataset, the trend of these changes gives a way to evaluate the influence of each threshold value, including the surface pressure difference and χ2 scale factor (SF) for the ABP method, and center value and HW of RCO2 and RH2O for the IDP method. Because the limit of χ2 is dynamically calculated for each sounding, a multiplicative SF is used to evaluate all soundings. For retrieved surface albedos, the threshold is adopted from the current OCO-2 parameter and therefore its influence is not examined in this study.
Figure 3. Changes of the throughput (left-hand column), agreement (middle column) and positive predictive value (PPV, right-hand column) for variations in the ABP surface pressure and scale factor thresholds (a–c) and the IDP RCO2 (d–f) and RH2O (g–i) thresholds based on OCO-2 and MODIS data in Europe in July 2016. The numbers in black indicate the tightened thresholds in this work, while the numbers in white indicate the original OCO-2 thresholds.
Based on the trend shown in Fig. 3, we first determined Δps and HW since the figure indicates they have a stronger influence over the changes of the outcome, and then we determined the other parameter for each pair (χ2 SF or center value, respectively), noting that the first determined parameter would be more crucial. The six major threshold parameters are adjusted back and forth, until the throughput of the combined results from the ABP and IDP methods are closely matched with the average monthly clear-sky fraction in the region from MYD08.
In general, a set of tight thresholds, i.e., lower limits of Δps, and χ2 SF, and a narrower HW range, as well as some shifts in the center value for acceptable RCO2 and RH2O, creates a more stringent cloud screening scheme, which leads to lower throughput, but a higher agreement and PPV. In other words, stringent thresholds, compared to loose ones, help to select scenes that are more “confidently clear”. The fewer scenes remaining have better agreement with the MODIS reference state and are supposed to have less influence from clouds or aerosol contamination, thus giving better quality assurance.
Similar trends are also observed in the contour plot created with the January dataset; although, compared to the July dataset, the limit for Δps in January is more than doubled to allow a reasonable throughput from the ABP method. This could be explained by the high snow cover in winter, which is known to increase errors in cloud screening and XCO2 retrievals.
A sensitivity test is performed to evaluate the rate of change of each diagnostic variable relative to different threshold values (Table 1). Five major threshold parameters are tested one at a time, while others stay the same. The χ2 SF is not tested, because significant change in the ABP results is not observed unless the SF is set to be extremely small.
Tested term* Selected value Test value Change (%) Result change (%) (for either ABP or IDP) THR AGR PPV Δps (hPa) 45 22.50 −50.00% −9.33% 3.70% 7.83% 33.75 −25.00% −3.08% 2.14% 3.12% 67.50 50.00% 4.59% −3.86% −4.61% 90.00 100.00% 8.40% −7.34% −8.17% RCO2 center 0.99 0.97 −2.02% −32.90% −22.84% 8.67% 0.98 −1.01% −11.09% −4.78% 5.46% 1.00 1.01% 7.73% −1.43% −7.07% 1.01 2.02% 14.33% −5.12% −13.73% RCO2 HW 0.04 0.030 −25.00% −11.09% −4.78% 5.46% 0.035 −12.50% −4.74% −1.22% 3.02% 0.045 12.50% 4.01% −0.29% −3.47% 0.050 25.00% 7.73% −1.43% −7.07% RH2O center 0.99 0.96 −2.04% −2.72% 0.25% 2.18% 0.97 −1.02% −1.27% 0.22% 1.08% 0.99 1.02% 1.10% −0.33% −1.04% 1.00 2.04% 2.12% −0.74% −2.06% RH2O HW 0.1 0.050 −50.00% −8.86% −0.96% 5.98% 0.075 −25.00% −3.60% 0.24% 2.83% 0.15 50.00% 5.11% −2.35% −5.20% 0.20 100.00% 9.19% −5.34% −9.66% *χ2 scale factor is not tested, because significant change in the ABP results is not observed unless the scale factor is set to be extremely small. Table 1. Sensitivity test for each threshold value.
-
Based on trends in the January and July contour plots, we set the values of seasonal thresholds according to the average monthly clear-sky fraction in the selected area (Table 2). The current thresholds used in OCO-2 algorithms are designed to have 25%–30% throughput globally, which means 5%–10% more than the clear-sky fraction observed by MODIS (Taylor et al., 2016). In contrast, the narrowed thresholds aim to have throughput close to the observed local clear-sky fraction in each month. The reduction of inflation over the MODIS clear-sky fraction helps to minimize the chance that some cloud- or aerosol-contaminated scenes also pass the screening. It is also worth noting that cloud coverage varies greatly throughout the year. The highest clear-sky fraction occurs in summer, which is 55.1%, while the lowest occurs in winter, which is 29.5%. The spring and fall have close values, which are 39.3% and 40.0%, respectively. Therefore, custom thresholds for each season is important to fit the regional conditions.
Clear-sky fraction Δps (hPa) χ2 SF RCO2 center RCO2 +/−HW RH2O center RH2O +/−HW OCO-2(in operation) − 20 5 0.99 0.04 0.99 0.2 spring 0.30 70 3 0.99 0.04 0.97 0.1 0.28 0.31 summer 0.34 45 2 0.99 0.04 0.98 0.1 0.41 0.43 fall 0.50 40 3 0.99 0.035 0.99 0.08 0.56 0.60 winter 0.50 100 3 0.99 0.04 0.98 0.08 0.39 0.31 Table 2. Settings of the ABP and IDP cloud screening thresholds used for the seasonal OCO-2 measurements discussed in section 2.1, including the differences between the retrieved and priori surface pressure (Δps), χ2 scale factor (SF), and center and half-width (HW) range for RCO2 and RH2O.
A summary of the statistical values for the re-screened dataset in each season is given in Table 3. For scenes with a clear reference state, the TPR ranges from 0.69 to 0.84, which is lowest in spring and highest in summer. For scenes with a cloudy reference state, the TNR ranges from 0.91 to 0.94. The results suggest that, compared to the global results in winter and spring given in Taylor et al. (2016), the correctly predicted clear scenes increased about 10%, and the correctly predicted cloudy scenes increased about 5%.
Reference clear Reference cloudy Total Season NTP TPR NFN FNR NFP FPR NTN TNR Spring 49588 0.69 22633 0.31 27483 0.076 332775 0.92 Summer 188747 0.84 34628 0.16 22193 0.081 253491 0.92 Fall 173842 0.78 49533 0.22 16127 0.058 259557 0.94 Winter 52054 0.72 20167 0.28 32380 0.090 327878 0.91 ABP Season NTP TPR NFN FNR NFP FPR NTN TNR Spring 52097 0.72 20375 0.28 54808 0.12 388478 0.88 Summer 220793 0.98 4182 0.019 85834 0.29 207364 0.71 Fall 219865 0.98 5110 0.023 80070 0.27 213128 0.73 Winter 57545 0.79 14927 0.21 80848 0.18 362438 0.82 IDP Season NTP TPR NFN FNR NFP FPR NTN TNR spring 56648 0.78 15573 0.22 64347 0.18 296080 0.82 summer 189336 0.85 34504 0.15 23668 0.086 252247 0.91 fall 174424 0.78 49416 0.22 16732 0.061 259183 0.94 winter 55735 0.77 16486 0.23 57270 0.16 303157 0.84 Table 3. Contingency tables for the comparison of the OCO-2 cloud screening results to MODIS cloud mask for each season in 2016 in Europe. Results are from the combination of the ABP and IDP methods.
The seasonal throughput, agreement and PPV for the ABP method, the IDP method and combined outcomes are given in Fig. 4. The total throughput is 0.18 in spring, 0.42 in summer, 0.38 in fall, and 0.20 in winter. The numbers in spring and winter are close to the values from the Glint-land viewing scenario in Taylor’s work (~0.19), but lower than values from the Nadir-land (~0.26). The overall agreement with MODIS is 0.88 on average, and is relatively consistent throughout year. There is a constant improvement compared to the current OCO-2 results (~0.83). The overall PPV is 0.77, and the average PPV of spring and winter is 0.63, which is higher than the 0.58 from Taylor’s results.
Figure 4. Seasonal throughput (a), agreement (b) and positive predictive value (PPV, c) for the ABP method, IDP method, and combined outcomes.
In general, the statistics in the summer and fall seasons are much better than in winter and spring. This indicates that the remaining data might still contain influence from snow-covered surfaces. A close examination of the results from the ABP and IDP methods shows that significant improvement of the ABP method is mainly in summer and fall, wherein the FNR can be reduced to about 0.02; on the other hand, improvement of the IDP method is mainly shown in the same seasons, with the FPR reduced to less than 0.1.
-
After determining the new thresholds and re-screening the OCO-2 measurements, the remaining retrievals were compared with collocated TCCON measurements, as discussed in section 2.3. Figure 5 shows scatterplots of the seasonal daily average XCO2 from OCO-2 versus collocated TCCON observations, in the order of time (winter, spring, summer and fall). It is rather obvious from the figure that the re-screened data [(e–h) on the right-hand side], compared to the original data [(a–d) on the left-hand side], show improvement, especially for measurements deviating from the one-to-one ratio line.
Figure 5. Scatterplots of seasonal daily average XCO2 from OCO-2 versus collocated TCCON observations. The dotted line is the one-to-one ratio line, while the solid line is the regression line.
For a total of 143 days, or 143 pairs of data, 31 pairs are removed after re-screening. For the remaining 112 pairs, 97 pairs have a smaller difference compared to the original dataset. Overall, the difference between the average XCO2 from the six TCCON sites and from the OCO-2 measurements passing nearby regions reduced 34.7%, decreasing from 3.23 ppm to 2.11 ppm. The average OCO-2 XCO2 before re-screening is 398.19 ppm, the average uncertainty is 3.85 ppm, and the standard deviation is 0.76 ppm. After rescreening, the average XCO2 increases slightly to 399.71 ppm, the average uncertainty decreases to 2.52 ppm, and the standard deviation decreases to 0.71 ppm.
Among the six sites, the Garmich site shows the greatest improvement: the difference of XCO2 between the TCCON and OCO-2 measurements decreases by 59.4%. The Karlsruhe site shows the second greatest improvement with a decrease of 42.6%. Next, the difference at the Orleans site decreases by 31.0%; the difference at the Bialystok site decreases by 28.7%; and the difference at the Paris site decreases by 24.7%. The Bremen site shows the least improvement with a decrease of 15.7%. There appears to be no relation between the position of these TCCON sites and the degree of improvement they have.
Based on the similarity of trends found in this work and Taylor’s work, we believe that the same optimizing scheme can be applied to worldwide locations. We applied the same procedure to OCO-2 measurements over the land area of Japan, and compared the re-screened data with three local TCCON sites (Table 4). However, there are far fewer data collected by these TCCON sites, resulting in fewer possible comparisons during the same period. Significant improvement of agreement between the TCCON and re-screened OCO-2 XCO2 is shown in summer and winter, though the latter has a very small sample size.
Season TCCON OCO-2 Data Re-screened Data XCO2 (ppm) Count XCO2 (ppm) Diff* (ppm) R Count XCO2 (ppm) Diff (ppm) R spring 407.15 1111 402.57 4.58 0.9887 869 403.32 3.83 0.9906 summer 398.63 1056 391.37 7.26 0.9817 821 399.03 −0.40 0.9919 fall 403.54 418 401.41 2.13 0.9947 46 400.41 3.13 0.9970 winter 403.92 86 399.87 4.05 0.9899 16 403.57 0.35 0.9996 *Diff refers to the difference between average TCCON measurements and OCO-2 measurements of XCO2. Table 4. Summary of the comparisons among TCCON, OCO-2, and re-screened OCO-2 measurements of XCO2 for the Japan area during 2016.