The CIMSS Satellite Consensus (SATCON) product blends tropical cyclone intensity estimates derived from multiple objective algorithms to produce an ensemble estimate of intensity for current tropical cyclones worldwide. The algorithm uses individual ADT, CIMSS AMSU, and CIRA AMSU intensity estimates utilizing a statistically-derived weighting scheme which maximizes/minimizes the strength/weaknesses of each technique to produce a consensus estimate of the current tropical cyclone intensity.

 
A statistical analysis of each of the 3 member algorithms was used to determine the individual member performance in a variety of TC structures. Each algorithm has strengths and weaknesses that are a function of the algorithm limitations, scanning geometry, instrument resolution or a combination of these factors. For example the ADT algorithm assigns a scene type to each IR image of a TC. ADT performance is strongly dependent on scene type with the best performance for scenes when a clear eye is present and decreased algorithm performance for other scenes. Because of this dependence the ADT is weighted according to scene type.
Examples of ADT Scene Types


Both AMSU algorithms are sensitive to TC core size as compared to the AMSU-A instrument resolution. The CIMSS AMSU algorithm uses an estimate of the TC Radius of Maximum Winds (RMW) obtained from either IR (an algorithm within the ADT estimates the RMW) or from the ATCF working Best Track files (RMW is input as part of the ATCF by the responsible warning agency). The RMW is used to estimate TC inner core size. If the inner core is sub-sampled by the AMSU-A instrument a correction is applied. In addition because of the relatively coarse instrument resolution it is possible that the true TC center may not be co-located with the instrument Field of View (FOV) used for the estimate. This source of sub-sampling is addressed using information from the AMSU-B moisture sounder. CIMSS AMSU performance is best when the TC eye is sufficiently large compared to the AMSU-A resolution therefore the CIMSS AMSU is weighted according to whether or not the TC inner core is resolved.

The CIRA AMSU algorithm performance is also a function of how well the TC inner core is resolved. An estimate of Cloud Liquid Water (CLW) and TC size are used by the algorithm as predictors to help address sub-sampling. Like the CIMSS AMSU algorithm however if the TC eye is small and the TC center is not co-located with the actual AMSU-A Field of View (FOV) center then addition sub-sampling can occur due to position offset (bracketing). In other words the TC may be positioned in between adjacent FOV and therefore only partially sampled by those FOV.

Below are examples of 2 corrections applied to CIMSS AMSU estimates. The first correction (left) is applied when the TC center is determined to be significantly offset from the AMSU-A position (bracketed). Other scan position/TC position geometries may result in corrections being applied. The correction on the right represents sub-sampling due specifically to TC eye size. When the TC eye is smaller than the AMSU-A FOV the TC warm core will be sub-sampled.


The goal of SATCON is to produce an estimate of TC intensity that is superior to both the individual members and a simple average of the members. 460 Cases of coincident estimates for all 3 algorithms from 1999-2009 were used to evaluate SATCON. Initial work involved using a simple weighting scheme to weight each member and then produce a SATCON estimate. Each member was weighted based on TC structure or instrument resolution characteristics. The ADT was weighted based on scene type, CIMSS AMSU was weighted based on if the TC inner core was resolved (based on the RMW) and CIRA was weighted based on the amount of eyewall convection within the AMSU-A FOV as determined by AMSU-B 89 Ghz. The results of this revealed performance that was better than the individual members and better than a simple average of the members for both MSLP and MSW (see tables below).

Information Sharing
Each algorithm uses (or creates) a variety of parameters in the process of producing an intensity estimate. Becasue these parameters contain information related to TC structure, possible instrument deficiencies or TC/instrument geometry they can be used to apply corrective input to the other algorithm members. Once the member algorithms have been adjusted then a consensus of the adjusted members can be preformed. A listing of the current corrections used in the SATCON algorithm follows.

ADT to AMSU: The ADT produces an estimate of the RMW. The ADT RMW estimate is used to estimate TC inner core size that can then be compared to the AMSU-A FOV resolution in order to make corrections to the CIMSS AMSU estimates of MSLP.

CIMSS AMSU to CIRA AMSU: As mentioned earlier the AMSU-B 89 Ghz channel is used by the CIMSS ASMU algorithm to address sub-sampling due to bracketing. There is also a strong relationship between this term and the CIRA estimates therefore the 89 Ghz signal is used to adjust both the MSLP and MSW estimates from the CIRA algorithm. The AMSU-B 89 Ghz channel with a nadir resolution of 16 km has three times the resolution of the AMSU-A sounder channels. The 89 Ghz channel is used as a measure of convective vigor with colder brightness temperatures (Tb) corresponding to stronger convection. Therefore if a portion, or all, of the TC eyewall is located within an AMSU-A FOV the AMSU-B 89 Ghz signal within that AMSU-A FOV can be used to adjust the AMSU estimates. Even after applying the 89 GHz correction the RMSE for the CIRA estimates increases as the amount of 89 GHz correction increases. In other words there is decreased confidence in the CIRA estimates when large corrections are required. Therefore the 89 Ghz signal is used to weight the CIRA estimates after corrections are applied with more weight being given to CIRA estimates when only small 89 GHz corrections are needed.
Examples of the relationship between the AMSU-A FOV location, TC center and the AMSU-B 89 Ghz signal. The yellow circle represents the AMSU-A FOV scan spot. Panel A indicates the possible scenario where the TC eye is large compared to the AMSU-A FOV. Only warm 89 Ghz Tb's are located within the FOV and thus there is little to no correction applied to the estimate. Panel B indicates a scenario where the AMSU-A FOV position is offset from the true TC center placing the FOV within the eyewall. Cold 89 GHz Tb's within the FOV indicate this bracketed geometry and a significant correction is applied. Panel C shows a scenario where the AMSU-A FOV location is coincident with the true TC center however the eye is small compared to the AMSU-A FOV resulting in a cold 89 Ghz Tb signal and required correction. Panel D represents the worst possible geometry. In this scenario (Iris 2001) the TC core is so small that even the higher resolution AMSU-B can not resolve the intense convection near the core resulting in a relatively warm Tb signal. In addition the AMSU-A FOV is not collocated with the true TC center and is instead located within the moat region resulting in significant sub-sampling that cannot be corrected sufficiently using the 89 Ghz signal.


Additional Corrections: The ADT was trained on MSLP estimates not delta_P therefore estimates may be weak/strong when located in regions of anomalously high/low environmental pressure. ATCF working Best Track files contain an estimate of environmental pressure and this term is used to adjust the ADT MSLP estimates accordingly.

Both the ADT and the CIRA AMSU algorithm were developed using MSW from Best Track data. This Best Track data contains storms moving at a variety of different speeds with an average of about 11 knots. As such the wind component imparted by the storm motion is not explicitly accounted for. Therefore an estimate of the storm motion is added to the ADT and CIRA MSW estimates (the CIMSS AMSU algorithm was trained using TC-relative MSW data). Currently 50% of the deviation from 11 knots is used. For example a storm moving at 35 knots would result in 12 knots being added to both the ADT and CIRA MSW estimates. This may not be optimum and is likely more dynamic than a simple static percentage for all cases.

After all corrections are applied to the individual members a weighted consensus is then performed. This approach results in improved performance over the previous method of weighting the un-corrected members as seen in the tables below.


 

Go to SatCon Real-Time analysis page.
Valid HTML 4.01 Transitional