7A.2 Radar Quantitative Precipitation Estimate Results Using a Convolution Neural Network

Wednesday, 15 January 2020: 8:45 AM
156BC (Boston Convention and Exhibition Center)
Micheal Simpson, NOAA/NSSL, Norman, OK; and J. Zhang and K. W. Howard

Complex terrain interacting with abundant moisture can lead to complex microphysical processes that produce heavy local rainfalls and flash flooding over very small spatial and temporal scales. High-resolution and high-accuracy quantitative precipitation estimation (QPE) in mountainous areas is critical for the monitoring and prediction of flash floods and mudslides. However, obtaining such QPEs has been a challenge due to non-linear interactions of many contributing factors in the complex microphysical processes. Machine learning has become a tool within the past several decades to show great promise in solving complex problems. Furthermore, a subset of machine learning, deep learning, has been recently utilized within the meteorological community to help forecast severe weather processes based on spatially gridded datasets. This study utilizes a convolutional neural network, a specific type of deep learning model, to quantitatively estimate precipitation amounts from weather radar observations. The initials results showed a superior performance over traditional physically-based radar QPE models, with a high computational efficiency and low I/O usage that was feasible for potential operational applications. Several examples will be shown from different domains to demonstrate the ability of the convolutional neural network to estimate precipitation accurately and efficiently.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner