Differences

This shows you the differences between two versions of the page.

--- intercomparisons:c2_cubiccloud [2016/11/16 11:30]
claudia
+++ intercomparisons:c2_cubiccloud [2016/11/16 11:51]
claudia
@@ Line 91: / Line 91: @@
   * Stokes component Q (second row): The first column shows the mean absolute radiance (since Q can be positive or negative). The only visible difference in mean radiance is seen for case 5, here SHDOM is smaller than the Monte Carlo models, the difference is about 10%. \\ Mean standard deviations again are larger for models without variance reduction. The match fraction shows obvious systematic differences for mscart_bmc (all cases), mscart_fmc (cases 5 and 7) and SPARTA (case 6).
   * Stokes component U (third row): Largest difference for case 7, mean difference between SHDOM and MYSTIC is about 5% of mean radiance. The match fraction shows obvious systematic differences for mscart_bmc (all cases), mscart_fmc (case 7).
-  * Stokes component V (bottom): Very noisy, without variance reduction (SPARTA, 3DMCPOL) noise is much larger than mean radiance. The match fraction shows systematic differences for mscart_bmc and mscart_fmc (all cases).
+  * Stokes component V (bottom): Very noisy, without variance reduction (SPARTA, 3DMCPOL) noise is much larger than mean radiance. The match fraction shows systematic differences for mscart_bmc and mscart_fmc (all cases). These differences might be numerical because the mean radiance is very small (about 2e-6). Since the standard deviation is an order of magnitude larger than the radiance for SPARTA and 3DMCPOL, the results mostly match within 2σ, but this does not show whether the models really agree.
-These differences might be numerical because the mean radiance is very small (about 2e-6). Since the standard deviation is an order of magnitude larger than the radiance for SPARTA and 3DMCPOL, the results mostly match within 2σ, but this does not show whether the models really agree.
   * Generally the standard deviation for SPARTA is below 3DMCPOL because for SPARTA more photons were used (10e11 vs. 49e9). MYSTIC and MSCART used 49e9 photons and variance reduction techniques which decrease the standard deviation approximately by almost one order of magnitude.
 ==== Results with atmosphere ====
-The following plots show the results obtained with MYSTIC. For all other models the patterns look the same.
+The following plots show the results obtained with MYSTIC. As for the cases without atmosphere the patterns look the same for all models.
-{{:intercomparisons:phase_c:c2_cubic_cloud:c2_mystic_noatm.png|C2 without atmosphere, MYSTIC results}}
+{{:intercomparisons:phase_c:c2_cubic_cloud:c2_mystic_atm.png|C2 without atmosphere, MYSTIC results}}
-Statistics for the cases with atmosphere
-{{:intercomparisons:phase_c:c2_cubic_cloud:c2_statistics_1.png|Summary plots, statistics C2 with atmosphere}}
-{{:intercomparisons:phase_c:c2_cubic_cloud:c2_statistics_1.pdf|Summary plots, statistics C2 with atmosphere}}
+The following plots show the mean radiance (for each Stokes component), mean absolute difference with respect to MYSTIC (forward Monte Carlo), mean standard deviation, fraction of pixels which agree within 2 standard deviations (2σ), please note the non-linear scale of the colorbar.
+{{:intercomparisons:phase_c:c2_cubic_cloud:c2_statistics_2.png|Summary plots, statistics C2 with atmosphere}}
+{{:intercomparisons:phase_c:c2_cubic_cloud:c2_statistics_2.pdf|Summary plots, statistics C2 with atmosphere}}
+  * Stokes component I (upper row): Mean radiance agrees very well for all models. Mean difference is largest for SHDOM (about 5e-4), for all other models below 1e-4. The mean standard deviations are in the range from 2e-5 to 5e-5, here the difference between models with variance reduction and without is not as large as for the cases without atmosphere. The reason is that for cases with atmosphere, there are no pixels with zero radiance and all pixels are now included in the statistics. In the clearsky region the standard deviation is not reduced by variance reduction. The match fraction shows a perfect agreement between all Monte Carlo codes over the full image.
+  * Stokes component Q (second row): Mean absolute radiance agrees well for all models, again largest mean difference for SHDOM As for I the standard deviations are similar for all codes. The match fraction shows a perfect agreement between all Monte Carlo codes, except MSCART (backward mode). MYSTIC (backward mode) agrees perfectly to MYSTIC run in forward mode.
+  * Stokes component U (third row): Mean absolute radiance agrees well. Match fraction again shows perfect agreement for all Monte Carlo codes except MSCART (backward mode).
+  * Stokes component V (fourth row): Mean radiances agree within noise. Largest differences (still below 5e-6) for MSCART. Standard deviation is significantly reduced by variance reduction, because V is only non-zero in cloud (only cloud scattering causes circular polarization, molecular scattering does not). Pixels with zero V are not included in statistics, thus standard deviation is more influenced by variance reduction. Very small systematic differences (may be numerical) between MYSTIC and MSCART (forward and backward).
+  * Standard deviations are smaller in forward than in backward mode, when the same number of photons/pixel (i.e. 1e7) is used. This can be seed for MYSTIC and also for MSCART.
 ==== Detailed plots for all models, no atmosphere ====

International Polarized Radiative Transfer

User Tools

Site Tools

Differences

Page Tools