2nd bimonthly progress report (August-September 2009)

More on combining independent wind information

In the first reporting period it was found that the average of model and SAR-retrieved wind speed was more accurate than model or SAR wind speed alone, compared to buoy measurements.
As a follow up to this, it is here tested how the error compared to buoy depends on the weighting coefficient, when the average wind speed is calculated by the following formula:

windspeed = w*SAR_wind_speed + (1-w)Model_wind_speed

where the weight w is between 0 and 1. The RMSE versus buoys as a function of the weight parameter is shown with a blue line on the figure below. A weight of 0 (left end of the blue line) corresponds to only NCEP wind speed, and a weight of 1 (right end of the line) corresponds to only SAR wind. The RMSE is the mean versus all buoys as used for the 1st progress report, and the SAR wind speed is calculated with the CMOD4 algorithm using wind direction from the HIRLAM model.

It is seen that the smallest error is found for a weight of 0.5 (equal weighting of SAR and NCEP wind speed), which was used on the test of the 1st progress report.

The red dashed line shows the RMSE when an error is introduced to the SAR wind speed by adding 4 m/s for all measurements. The RMSE of the SAR wind speed alone (weight = 1) increases now from about 2.7 m/s to 4.5 m/s. However, interestingly, even with this very large introduced error, the weighted average is more accurate than the pure NCEP wind speed, as long as the weight of the SAR wind speed is smaller than 0.35, with a minimum error for a relative weight of 0.2 for the SAR wind speed. The main message from this simple exercise is that even "bad" information may be useful information, it should only be given less weight. The goal of the development of the Bayesian SAR wind retrieval scheme is to determine the weights in an optimal way. However, it might also be anticipated from this exercise that the final performance will improve with the amount of (independent) wind information, and less sensitive to the method by which the information is combined/weighted. This encourages the future plan to test the performance when using triple combination of SAR, scatterometer and model wind.

CDOP function

One of the major tools of this project is the CDOP-function, which empirically relates the (C-band) Doppler Centroid Anomaly to (model) wind speed, wind direction (relative to SAR look direction) and SAR incidence angle. This function is analogous to the CMOD-function(s) which relate the radar cross section (sigma0) to the same geophysical variables. The original CDOP-function was developed by Boost Technologies (now CLS), and was provided to NERSC in IDL format, where it has been converted to Matlab language. Some first tests with this code show however disagreement with figures in published results (Johannessen et al., 2008). It has therefore been decided to repeat the exercise of tuning such a function at NERSC. Although this includes some duplication of work already done, it has some advantages: 1) NERSC will not depend on technical support for the function; 2) the work of the tuning will give more insight into both the physics and the limitations of the function, and 3) using the same model for tuning of the CDOP function as in the Bayesian scheme will improve the consistency. NERSC has established an efficient framework for using NCEP GFS model wind speed, and plan to use this for both CDOP development and for the wind retrieval scheme in general. This development has just started, and results will be shown in the next reporting period.

Calibration of the Doppler Centroid Anomaly

Although the Doppler Centroid Anomaly has proven very useful and of generally good quality, some issues remain about the calibration of this value. Some biases of unknown origin have been removed by empirical corrections, and in particular it has been a challenge to determine the level corresponding to zero range velocity. One way of determining the zero-level has been to calculate the average Doppler over land for a given scene, and then subtract this value, under assumption that the Doppler velocity should be zero over land. However, in practice large variations of the Doppler has been found over land. Some of this is attributed to artifacts since the algorighm for calculating the Doppler implicitly assumes that the roughness is constant over the ~5 km area for which the Doppler is estimated. However, the land Doppler seems also to be clearly linked to the height of the terrain. This effect has been studied together with Morten W. Hansen at NERSC, and preliminary results are shown on a separate blog: http://groups.google.com/group/nerscsat/web/surface-sar-doppler-velocity...

A possible explanation is related to a deviation of incidence angle caused by the terrain elevation (foreshortening effect). It is found that the change in incidence angle is quite small (less than 0.02 degrees), but this may still be significant when the anomaly is calculated by predicting the component of the earth rotation velocity in the SAR look direction, since this velocity is very large; about 465 m/s at equator and half of this at 60 degrees latitude. This work will continue, and will hopefully lead to improved estimation of the Doppler anomaly, of concern for both the wind and current retrieval of the INCUSAR project.

Coordination with CLS

WP1 of INCUSAR is building on work performed at CLS in France, by in partuclar Alexis Mouche, Fabrice Collard and Vincent Kerbaol. We have been in discussion with this group to avoid unnecessary duplication of work, and to aim for complementary achievements. It is suggested that we will in INCUSAR pay particular attention to the problem of existence of solutions of the Bayesian Scheme; although several examples have been shown where the algorithm performs well, there are also cases where the scheme does not converge, since there is no real solution satisfying both the CMOD and CDOP functions and other imposed restrictions. This work will also depend critically on an accurate CDOP-function, as discussed above. NERSC and CLS have submitted complementary presentations of this topic to the SeaSAR conference in ESRIN in January 2010.

WP 2 - current retrieval

WP 2 of INCUSAR was planned to start in early 2010. However, we suggest rather to start this work package gradually already now, in parallel with WP 1, to have more time to digest on critical issues which might arise. This first task of this work package is to adapt the DopRIM software written and provided by Vladimir Kudryavtsev to the programming environment at NERSC. Being a research software, it has a very low level user interface, and little documentation except for published papers on the physics. This work is now in progress at NERSC.