An orbital "virtual radar" from TRMM passive microwave and lightning observations

The retrieval of vertical structure from joint passive microwave and lightning observations is demonstrated. Three years of data from the TRMM (Tropical Rainfall Measuring Mission) are used as a training dataset for regression and classification neural networks; the TMI (TRMM Microwave Imager) and LIS (Lightning Imaging Sensor) provide the inputs, the PR (Precipitation Radar) provides the training targets. Both vertical reflectivity profile categorization (into 9 convective, 7 stratiform, 2 mixed and 6 anvil types) and geophysical parameters (surface rainfall, vertically integrated liquid [VIL], ice water content [IWC] and echo tops) are retrieved. Retrievals are successful over both land and ocean surfaces. The benefit of using lightning observations as inputs to these retrievals is quantitatively demonstrated; lightning essentially provides an additional convective/stratiform discriminator, and is most important for isolation of midlevel (tops in the mixed phase region) convective profile types (this is because high frequency passive microwave observations already provide good convective/stratiform discrimination for deep convective profiles). This is highly relevant as midlevel convective profiles account for an extremely large fraction of tropical rainfall, and yet are most difficult to discriminate from comparable-depth stratiform profile types using passive microwave observations alone. The retrievals proceed as follows: A principal components analysis (PCA) is performed on 33 "raw" inputs (lightning, nine passive microwave frequency/polarization brightness temperature variants, physically-based linear and nonlinear combinations of them, and metrics derved from texture analyses of them). The first 25 PCs are retained, accounting for 99.9% of the variance in the original observations. These are then used as inputs to a regression neural network (i.e., nonlinear multivariate continuous regression) for the geophysical parameters listed above, and a separate classification neural network (i.e., a nonlinear multivariate categorical regression) for the 25 different profile types (previously identified from cluster analysis of a large sample of PR data). The networks are trained with "oversampled" TMI data (i.e., at the PR pixel level), and hence incorporate TMI subpixel convective/stratiform variability effects into their predictions. The resulting classification network is shown to be unbiased with respect to profile type, and not overfitted (more than adequate training data exist for even very complex, nonlinear NNs). The preliminary predictions are then used to "cross-train" and refine the estimates (a new regression NN is trained using both the TMI/LIS PCs and the classification predictions as inputs; a new classification NN is trained using both the TMI/LIS PCs and the geophysical predictions as inputs). In addition to the predicted geophysical quantities, the profile type retrievals can be used to reconstruct a full volumetric radar reflectivity field, hence the term "virtual radar". This allows computation of reflectivity-based products other than the subset of geophysical parameters that are explicitly retrieved. The quality of retrieved fields is sufficient for data assimilation purposes, which is highly important, as data assimilation modules trained on volumetric radar data can thus be directly applied to a number of past, current and future orbital passive microwave sensors/platforms (SSM/I, TMI, AMSR-E, NPOESS, GPM, etc.). Similarly, "conventional" radar-based warning products familiar to operational decision-makers can be dervied from sensors on these platforms.