695 Dual Application of Convolutional Neural Networks: Forecasts of Radar Precipitation Intensity and Offshore Radar-Like Mosaics

Tuesday, 9 January 2018
Exhibit Hall 3 (ACC) (Austin, Texas)
Christopher J. Mattioli, MIT Lincoln Laboratory, Lexington, MA; and M. S. Veillette and H. Iskenderian

Convolutional Neural Networks (CNNs) have proven to be extremely effective in image classification and regression tasks. In weather applications, CNNs have high potential in the areas of weather forecasting and data fusion. In this work, we leverage CNNs to create a global, probabilistic forecast of a radar-derived precipitation quantity with the ultimate goal of improving a data fusion capability. Specifically, our forecasted quantity is Vertically Integrated Liquid (VIL) – a radar-derived measure of precipitation intensity often used by Air Traffic Management. A CNN is trained using several fields from the Global Forecast System (GFS). The regression target is VIL derived from reflectivity measured by the Dual-frequency Precipitation Radar aboard the Global Precipitation Measurement Mission (GPM) satellite. The result is a CNN pipeline that uses GFS forecast fields as inputs and outputs a precipitation-intensity range of highest probability. The results of training this model to produce a probabilistic VIL forecast are promising and will be described.

This probabilistic VIL forecast is then used in a data fusion capability known as the Offshore Precipitation Capability (OPC). The OPC creates a radar-like mosaic of VIL in areas devoid of weather radar data by fusing three disparate, non-radar datasets together in a CNN pipeline. The three datasets used in OPC are geostationary satellite imagery, global lightning, and numerical weather model forecasts. Currently, the numerical model forecast features of OPC are the lowest contributors of skill when compared to the satellite and lightning features. The proposed CNN VIL forecast model described above will be used as a pre-processing step for these numerical model forecast fields in the pipeline. Instead of being trained with features from the unaltered numerical model forecast fields, the OPC will be trained using this new probabilistic VIL forecast. The strength of this approach is that OPC will be trained on data close to its regression target of VIL. We will demonstrate the value added to OPC by incorporating this CNN-derived VIL forecast feature into the OPC CNN pipeline.

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner