Diagnosing Convective Instability from GOES-8 Radiances

Abstract

Statistical algorithms are developed to diagnose the vertical change in equivalent potential temperature (?Θe) between 920 and 620 hPa from GOES-8 radiance data. The models are prepared using a training dataset of radiosonde releases from 10 United States cities. Simulated GOES-8 channel brightness temperatures are calculated from these soundings. The training data are stratified into several subsets (depending on time and location). Models trained only on 0000 or 1200 UTC data explain approximately 7% more of the variance in observed ?Θe than those trained on both 0000 and 1200 UTC. Values of R2 from models using training data from only one are superior to those trained on multiple stations. Inclusion of the imager channels adds little information to the algorithms. These models then are applied to data from the Limited Area Mesoscale Prediction System model to see which performs consistently better over diurnally varying conditions. Models trained only with 0000 UTC data give the best results, explaining between 63% and 81% of the variance in the independent data. The model that performed best is studied further. Biases are present when this model is applied to times other than 0000 UTC. These biases are caused by temperature differences between the 0000 UTC training data and those at the times being examined. Strong regional biases also occur when a model trained on only one location is applied to a large area. A second model is incorporated into the procedure to reduce this bias. The two-model algorithm explains more variance than the initial one-model version (93% vs 77%), and the area of strong regional bias is greatly reduced. This statistical procedure for ?Θe is then tested on observed GOES-8 data. A new statistical model is formed using the observed GOES-8 brightness temperatures and ?Θe?s calculated from collocated radiosonde observations. The new model yields an R2 of 67% when applied to an independent dataset. This value is smaller than those from the models using simulated data, most likely due to several additional sources of discrepancy. Finally, simulated GOES-7 Visible Infrared Spin Scan Radiometer (VISSR) Atmospheric Sounder (VAS) radiances are used to prepare a ?Θe algorithm. The VAS model explains approximately 5% less variance than its GOES-8 counterpart, due to the reduced vertical resolution available on VAS. These analyses show that regionally trained regression models can accurately diagnose convective instability while using relatively little computational time.