3.2. AC Overall Analysis
The number of match-ups depends on the selected AC: 18, 16, 19, 35, and 17 for BlAC, C2R-CC, C2R-CCAltNets, Polymer, and NASA, respectively, out of 37 valid match-ups. Figure 3
presents the scatterplots between the in situ
(x-axis) and the estimated
for all AC (y-axis) for
between 400 and 754 nm. A visual inspection of the scatterplots shows that the accuracy of the retrievals was wavelength-dependent. For
between 400 and 443 nm, the scattering of the retrievals was quite high for all ACs except C2R-CCAltNets. This is particularly true for
at 400 nm. For
between 490 and 674 nm, the scattering tended to decrease, and the AC performances seemed to increase, getting close to each other with increasing wavelength. For
between 681 and 754 nm, the accuracy decreased, in particular for NASA at 681 nm.
The scatterplots analysis is in accordance with the spectral variation of the statistical parameters (Figure 4
). The slope of the regression line,
, increased with the wavelength until 560 nm and flattened for
greater than 560 nm. It was especially pronounced for NASA where a clear inflection was observed at 443 nm. C2R-CCAltNets presented the closest
values to one at 400 and 510 nm. Between 560 and 681 nm, it was C2R-CC, except at 620 nm (BlAC). For the longest wavelengths (≥709 nm),
was the closest to one for NASA. For most of the ACs,
increased from 400 nm–490 nm, except for C2R-CCAltNets, for which
was always close to one and did not seem to be wavelength-dependent.
The variations of R
(bottom right of Figure 4
) were similar to those of
, but more pronounced, i.e., steeper increased values from 400–665 nm. All ACs had an R
value ≤90% between 400 and 560 nm, C2R-CCAltNets showing the highest values for this spectral range. From 560 nm, NASA presented the highest values, close to 100%.
The magnitude of the bias was wavelength-dependent, while its sign was algorithm-dependent. C2R-CCAltNets showed the lowest values at 400, 560, 620, 665, 674, 681, and 709 nm, whereas Polymer showed the lowest values at 412 nm. For 443 and 510 nm, C2R-CC was the least biased algorithm. BlAC was always negatively biased, while C2R-CCAltNets was always positively biased. For the other ACs, the sign of the bias depended on the wavelength.
RE showed the standard “smiley” shape with high values in the blue and in the red and low values in the green. Both Polymer and C2R-CC showed the lowest RE for all wavelengths, and their spectral variation were less pronounced compared to the other algorithms.
The spectral variation of RMSE was different from the other statistical parameters as its values continuously decreased with wavelength. For between 400 and 510 nm, C2R-CC showed the lowest values of RMSE, while the highest values were obtained for BlAC. For between 560 and 754 nm, C2R-CC and C2R-CCAltNets showed the lowest values, while the highest values were obtained for Polymer. It is worth noting that retrievals with NASA were affected by smile effect correction failure at 681 nm with peak of values for , bias, RE, and RMSE.
BlAC and NASA performed better in French Guiana than in the Eastern English Channel waters. The other ACs performed equally well in both regions.
gives the values of QAS,
, SAM, and
for each AC for a time window of ±2-h. C2R-CCAltNets had the maximum value of QAS (0.98), while BlAC and NASA had the lowest value (0.82). C2R-CCAltNets and Polymer had the minimum values of
(2.28 and 2.00, respectively), while BlAC had the highest (3.30). When the accuracy of the ACs decreased,
increased due to the increase of the relative squared errors of the spectral bands normalized at 560 nm. Polymer presented the lowest value of SAM (7.29) and NASA the maximum value (14.60). SAM had to be 0
when the AC spectra were retrieved accurately, leading to a null angle to superpose the AC spectrum on the in situ-measured spectrum. For
, C2R-CCAltNets was the best-performing AC and had the maximum score of 56.46. In second rank, C2R-CC had a value of 51.99. BlAC showed the lowest value of 36.01.
3.3. AC Concomitant Analysis
The preceding section presented the general evaluation of the AC for all match-ups. Here, we present the same analysis, but considering only the common match-ups to all AC. The number of match-ups decreased from 37 to 14; 10 in the Eastern English Channel and four in French Guiana. Figure 5
and Figure 6
are similar to Figure 3
and Figure 4
. The trends observed in the previous section were similar when taking only the common match-ups. The retrievals were not accurate for
between 400 and 443 nm and between 665 and 754 nm, while they were more accurate for
between 490 and 620 nm.
Furthermore, for the spectral statistics variation, the ACs spectral shapes of the AC common match-ups were similar for , , , and RMSE. Moreover, C2R-CC and especially C2R-CCAltNets statistical performances decreased in comparison to the initial performance analyses, while the accuracy for the other ACs (NASA, BlAC, and Polymer) showed a relative increase. Here, Polymer became the most efficient AC when taking only the AC concomitant match-ups.
is similar to Table 2
. Taking into account only the common match-ups and the ±2-h time window gives quite a similar analysis in comparison to the other
t. C2R-CCAltNets had the highest QAS value (0.99), while BlAC had the lowest value (0.82). For
, NASA had the lowest value (2.18), while BlAC had the highest value (4.19).
In terms of SAM, the minimum value was obtained for Polymer (6.80) and the highest for NASA (15.01). For , Polymer had the maximum value (65.33) followed closely by C2R-CCAltNets (58.05), while BlAC presented the lowest score (36.11). On the common match-ups, the results were slightly different, with Polymer being the most accurate AC on our dataset, closely followed by C2R-CCAltNets. In all cases, the least accurate AC was BlAC.