Discussion and conclusions

A total of 5 multichannel microwave radiometer algorithms have been compared on 7 areas including combinations of open water (W), multi-year ice (MY) and first-year ice (FY).
In order to calibrate all algorithms, mean brightness temperatures from a two-month period have been calculated from W, MY and FY areas. These brightness temperatures are properly transformed to tiepoints for each individual algorithm by employing e.g. a given atmosphere model to the satellite measured brightness temperatures.
The COMFREQ and NORSEX algorithms have been shown theoreticly to produce the same results when the tiepoints to the NORSEX algorithm are transformed according to its atmosphere model. This has been demonstrated by this comparison.

All algorithms have been used to calculate the total mean ice concentration in a two-month winter period, and thereby it has been shown, that the algorithms always produces results with an absolute deviation of less than 2 %. All algorithms are unbiased, i.e. in the tie-point areas of W, MY and FY, they all produce ice concentrations of 0 %, 100 % and 100 %, respectively. Plots of the ice concentration through the period shows, that all algorithms produces similar results. Within the tie-point areas, the curves are very similar, and the calculated changes in ice concentration is believed to be created by the same effects in all algorithms. However, the COMPOL algorithm has a worse performance than the other algorithms in the open water area.
The standard deviation on the daily measurements through the period ranges from 1.2 % to 11 %. A significant contribution comes from changing ice conditions in the period, and the figures are only suitable for this intercomparison. The small difference between the algorithms shows, that the geophysical influence have almost the same effect on all algorithms.

The algorithms ability to divide the ice types in multi-year ice and first-year ice have been investigated through the multi-year ice fraction. Again in the tie-point areas, all algorithms produces average values very close to the expected figures. In the tie-point areas, the curves from the different algorithms are very similar. There are rather big variations through the period; almost equal in all algorithms. This is probably caused by changing ice- or geophysical conditions in the period, and these effects are having almost the same influence in all algorithms. In the other areas, the average multi-year ice fractions differs less than 6 %. In a W/FY area, the multi-year ice fraction ranges between -8.4 % and -9.9 %, which indicates, that it is underestimated. Also in a W/MY area, the multi-year ice fraction is underestimated.
The standard deviation on the daily multi-year ice fractions ranges between 7.6 % and 17.5 %, and again it is primarily caused by changing ice conditions in the period.

The sensitivity of the algorithms to changes in the physical ice temperature have been analyzed. Hereby it has been shown, that the NORSEX/COMFREQ algorithms are most sensitive, with changes in the total ice concentration up to 1 %/K. The COMPOL algorithm is less sensitive than the AES, NORSEX and COMFREQ algorithms. The NASA algorithm is insensitive over consolidated ice, which is caused by its use of the polarization ratio and spectral gradient ratio. The impact of changing temperatures on the multi-year ice fraction is investigated, and here the NASA algorithm is totally insensitive, while the other four algorithms shows a comparable sensitivity of up to 2 %/K.
The sensitivity of the algorithms to measurement noise have been investigated by adding Gaussian distributed noise to the input signals. This has shown, that the COMPOL algorithm is most sensitive to measurement noise, and that the standard deviation of the ice concentration is more than 2.5 % for a 0.6K standard deviation on the input signal. The least sensitive to measurement noise is the AES algorithm, which produces a standard deviation close to 1 % for a 0.6K standard deviation on the input. The influence on the multi-year ice fraction is most severe in the NASA algorithm, while the other four are about equal sensitive.

The only clear conclusion from this comparison is, that the COMPOL algorithm always has the biggest standard deviations on its ice estimates. It only uses 37GHz channels, and it therefore gives the best spatial resolution. The 37GHz channels are more sensitive to geophysical influence, and the price to be paid for a better spatial resolution is a degradation in radiometric resolution.
A major result is, that all algorithms on average produces very similar results. This is caused by the careful tie-point selection, where tiepoints are transformed according the the algorithm atmosphere model. The important issues are therefore not to select a prober algorithm, but to adjust the tiepoints to the selected algorithm. Further, plots of ice concentrations and multi-year ice fractions through a two month winter period have shown, the the variation between the algorithms normally are of little significance compared to the total variations. The total variations are dominated by geophysical variations, and an effort in incorporating these parameters in the algorithms might be the key to more precise algorithms in future.
For precise short term estimates, measurement- and geophysical noise can be of significance, and a prober algorithm for reducing the principal noise components has to be chosen. The noise analysis has shown, that a big retrieval triangle reduces the sensitivity to measurement noise, and that a certain algorithm in a certain area can be insensitive to changing ice temperatures with respect to either the ice concentration or the multi-year ice fraction.

Acknowledgment. I would like to thank National Snow and Ice Data Center for delivering measured SSM/I brightness temperatures, and also to Leif Toudal Pedersen for discussions over the obtained results.