Charles G. Brown and Dr. David G. Long, Electrical and Computer Engineering
BACKGROUND
Since near-surface ocean winds affect global climate, several space-borne instruments have been deployed to measure ocean winds. The most widely known of these instruments are the satellite scatterometers, such as the Seasat Scatterometer (SASS), the NASA Scatterometer (NSCAT), and the Active Microwave Instruments (AMI) aboard the European Remote Sensing satellites ERS-1 and ERS-2. All of these devices measure near-surface ocean wind by transmitting microwave radiation to the ocean’s surface and then measuring the power level of the radiation scattered back to the satellite. The recorded power measurements provide information on the roughness of the ocean’s surface. Since wind alters the shape of the surface in a predictable fashion, the power measurements also allow engineers to infer the speed and direction of the near-surface ocean wind.
Traditional methods of estimating ocean wind from scatterometer measurements yield several possible wind vectors for each scatterometer resolution element (usually about 50 km square). The presence of multiple solutions necessitates an additional data processing step, called ambiguity removal, which selects a unique wind vector for each resolution element. Ambiguity removal following traditional wind estimation is notoriously error-prone, so alternatives for wind estimation are being explored.
One of the most promising alternatives to traditional wind estimation is field-wise estimation. The field-wise method estimates ocean wind for many resolution elements at once using a wind field model. The wind field model describes a vector field as a function of several model parameters, and the task of the field-wise technique is to determine optimal sets of model parameters. As in traditional wind estimation, field-wise estimation produces several possible solutions. However, since the solutions are for sets of resolution elements instead of single elements, the solutions from successive regions can be made to overlap. Overlapping solutions may ameliorate ambiguity removal since selecting a unique field-wise solution for each region is like fitting together a piece of a jigsaw puzzle. Also, it is expected that ambiguity removal errors will be easier to locate and correct than in the traditional scheme.
RESEARCH OVERVIEW
While the field-wise method may overcome some of the problems of the traditional technique, it also is computationally more intensive. In order to determine the optimal sets of model parameters, and hopefully the best set, it is necessary to employ non-linear global optimization methods. Although computing resources are usually adequate for such purposes, field-wise estimation is still time consuming, and methods to simplify the process are being investigated.
One such simplifying technique is to reduce the model complexity where possible. I have observed that simple wind field models, those requiring only a few model parameters, suffice to model a large percentage of wind fields. The remaining wind fields require more complex models with larger numbers of parameters. I have developed an algorithm to detect which wind fields require complex models. A point-wise data statistic (PDS) is computed based on the ambiguities of the traditional wind estimation process. If the PDS is above a user-defined threshold, the region is classified as a complex wind field; otherwise, it is classified as a simple wind field. For more details about the PDS and the calculation of the threshold, see (3) and (4).
In order to test the algorithm, I ran it on over 3500 wind fields obtained from NSCAT. For one particular threshold value, I observed that the probability of correctly detecting a complex wind field is 93% and the probability of incorrectly classifying a simple wind field is 26%. The probability of correctly detecting a complex wind field is more critical than the probability of incorrectly classifying a simple wind field, since the former reflects the chance that the algorithm uses a simple model to model a complex wind field. Large modeling inaccuracies may result from such an error. In contrast, the probability of incorrectly classifying a simple wind field is the chance that the algorithm uses a complex model where it really should use a simple one. The latter kind of classification error does not increase modeling error, but it does decrease the computational savings offered by the algorithm.
CONCLUSIONS
These results indicate that it is possible to divide wind fields into those requiring complex models and those needing only simple models. Appropriate models for each class can then be used in field-wise estimation. Since simple models are used on the simple fields, instead of complex models being used all of the time, the computational burden of field-wise estimation can be decreased without significantly increasing the error in the estimates. Decreasing the computational load of field-wise wind estimation will result in faster data processing rates, which will make field-wise estimation an even more attractive alternative to traditional wind estimation.
The results from this study, including work on global optimization techniques for field-wise estimation, were presented at two technical conferences. Refs. (1) and (3) were presented at SPIE, which was held in San Diego, CA. Refs. (2) and (4) were presented at the International Geoscience and Remote Sensing Symposium (IGARSS) in Singapore.
REFERENCES
- C. G. Brown and D. G. Long, “Algorithms for Field-Wise Scatterometer Wind Estimation,” SPIE, San Diego, CA, 27 July-1 August, 1997.
- C. G. Brown and D. G. Long, “Global Optimization Algorithms for Field-Wise Scatterometer Wind Estimation,” IGARSS Proceedings, Singapore, pp. 353-355, 1997.
- C. G. Brown et al., “Wind Field Models and Classification,” SPIE, San Diego, CA, 27 July-1 August, 1997.
- C. G. Brown et al., “Wind Field Models and Model Order Selection for Wind Estimation,”
IGARSS Proceedings, Singapore, pp. 1847-1849, 1997.