Air mass pre-classification to improve performance of physical retrievals for CMIS

A method has been developed for pre-classification according to air mass for the forthcoming Conical-scanning Microwave Imager/Sounder (CMIS) on NPOESS. This presentation covers the performance of the pre-classification method, its integration with the CMIS physical inversion algorithm for atmospheric sounding, and its impact on CMIS retrievals. The reference standard we used for classifying air masses was the Thermodynamic Initial Guess Retrieval (TIGR-3) dataset developed at the Laboratoire de Météorologie Dynamique. TIGR-3 consists of 2311 profiles of temperature and water vapor that have been separated into five classes defined with regard to profile structure, rather than according to geographic boundaries. With pre-classification, we aim to select the correct air mass class at any location on the basis of the CMIS brightness temperature measurements. The physical inversion can then use a class-specific a priori (background) estimate and error covariance. The method we used for classification is a probabilistic neural network (PNN). The inputs are CMIS channels in the oxygen absorption band whose response is maximal near the tropopause and the stratosphere, and which have virtually no sensitivity to the surface under any meteorological conditions. Use of channels affected by the surface would prevent the classification from being robust, because of the high variability of surface emissivity. To train the PNN, we computed brightness temperatures for each of the TIGR-3 profiles, using a radiative transfer model and including simulated measurement noise. The first level of validation of the PNN was to compute the rate of correct classification of the training (dependent) profiles. The PNN was able to correctly classify the dependent TIGR-3 profiles with no errors when provided three CMIS brightness temperatures as inputs. For independent validation, we used a set of 7137 profiles from the NOAA-88 database. We estimated the “true” class of each NOAA-88 profile by applying the PNN in a configuration where the inputs were the profile level temperatures, rather than brightness temperatures, training on TIGR-3 and applying to NOAA-88. This method works the same as when classifying on brightness temperatures, but removes the effects of radiometric smoothing and measurement noise. We compared the classification of NOAA-88 profiles based on brightness temperature to the “true” classification of NOAA-88 profiles and found that a large majority (83%) of the profiles were assigned to the same class by either method, and there were no instances of gross inconsistency, where tropical and polar profiles were cross-classified. Retrieval experiments were performed to assess the impact of pre-classifying cases according to air mass on retrieval results. Retrievals made with a global background (a priori) were compared with retrievals made with the classified background. When we perform retrievals for test scenes and select a background on the basis of brightness-temperature classification, there will be some misclassified cases that would be outliers in relation to the selected background and thus would have heavily degraded retrieval performance. The retrievals were made more robust by computing the background estimate and error covariance from profiles that were segregated according to classification by brightness temperatures, including cloud, emissivity, and noise effects. In this manner, the background classes were defined consistently with the classifications of retrieval scenes, including misclassifications. In the retrieval tests, the primary benefit of the classification was to reduce the temperature profile retrieval errors near the tropopause by about 0.3 K, but the classification also benefited profile retrieval performance near the surface, including for water vapor. With pre-classification, there is a risk that fields of retrievals will have abrupt, artificial changes between neighboring regions were different classes were selected. In the CMIS approach, where the backgrounds are derived from data classified by brightness temperature, the borders between backgrounds are “fuzzy”, because noise in the brightness temperatures causes profiles that are near the border between classes to be mixed among the classes. With this approach, the retrieval impact of changing from one class to a neighboring class should be diminished, and abrupt changes in fields of retrievals should be avoided. We tested the algorithms performance in this regard by focusing on the impact of noise on retrievals. The impact of noise at the classification stage of retrieval was considerably less than the impact at the inversion stage, even when limiting the analysis to the border profiles for which a change in noise realization caused a change in classification. We did an additional test for abrupt class-related boundaries by performing retrievals over continuous swaths of simulated CMIS data derived from National Centers for Environmental Prediction (NCEP) Global Forecast System (GFS) analysis data. Retrieved temperature and water vapor fields showed no signs of boundaries coincident with changes in selected atmospheric class.