Abstract
Two Pattern recognition procedures are developed to provide improvements to first-guess fields for satellite temperature retrievals. The first is a technique whereby a radiometer measurement may be used to select one or more historical radiosonde temperature profiles as analog estimates of ambient thermal structure. The vertical scales of the analogs are those of radiosondes—the vertical resolving power of the satellite radiometer being relevant only to a decision process. The analog selection process is shown to be much more effective if implemented in an orthogonalized space of measurement information. The second procedure is one which partitions a priori dependent data into shape-coherent pattern libraries using structure information inherent in the data itself. This is an alternative to traditional partitioning schemes whereby proxy classifiers such as season, location and surface type are used.
These pattern recognition techniques are shown to be capable of reducing first-guess profile errors by nearly 50%, in an independent test of about 800 diverse retrievals. The impact of pattern recognition on temperature retrieval error is assessed using regression and physical-iterative retrieval algorithms. The influence of improved first-guess fields is markedly different on these two types of algorithms. Pattern recognition is shown to have a strong, positive impact on the physical-iterative method but little significant impact on regression when evaluated in an overall batch sense. A case study suggests that a small number of very poor retrievals may particularly mask the potential benefits of pattern recognition on both methods.
