JP1.1
Producing satellite retrievals for NWP model initialization using artificial neural networks
Robert J. Kuligowski, Penn State Univ., University Park, PA; and A. P. Barros
Proper specification of initial conditions is crucial for optimizing the forecast skill of numerical weather prediction (NWP) models. This is especially true in the case of quantitative precipitation forecasting (QPF), since the production of precipitation is strongly influenced by very fine-scale processes. While traditional observation techniques do not provide data at sufficiently fine spatial resolution to depict these processes, the high resolution of multispectral radiance fields from satellites provide a significant opportunity to obtain temperature and moisture data at a resolution that will be much more appropriate for this task. However, a number of challenges must be addressed, including the mixing of satellite data at multiple spatial resolutions and the relation of satellite radiances to temperature and moisture profiles in the atmosphere--an inverse problem which is ill-posed and does not provide unique solutions.
This work presents an application of artificial neural networks and fractal scaling techniques to retrieve temperature and moisture profiles using data from the NOAA-15 polar orbiting satellite. Fractal scaling techniques are used to make the satellite data compatible with one another in terms of scale. (The spatial resolution of the satellite data range from 1 to 50 km in scale.)
Artificial neural networks that have been trained using collocated radiosonde data and satellite radiances are used to retrieve first-guess profiles of temperature and moisture from these rescaled satellite fields. A radiative transfer model then solves the inverse problem to generate temperature and moisture profiles using the neural network-retrieved profiles as a first guess. Finally, these temperature and moisture profiles are used along with wind fields generated by the Penn State/NCAR Fifth-Generation Mesoscale Model (MM5) to initialize a high-resolution (1-km) nested numerical weather prediction model that focuses on precipitation processes. The performance of this technique is evaluated by comparing the retrievals to available observations and by evaluating their effect on QPF from a nested orographic precipitation model over the Pocono Mountains of northeastern Pennsylvania.
Joint Poster Session 1, (Joint with 10th Conference on Satellite Meteorology and Oceanography and Second Conference on Artificial Intelligence)
Tuesday, 11 January 2000, 4:30 PM-5:45 PM
Next paper