Learnable Optical Flow Network for Radar Echo Extrapolation
In one sentence. Fully convolutional learnable optical-flow network extrapolates six years of Guangdong radar mosaics, outperforming optical flow and other deep models for convective nowcasting.
Abstract
The impact of extreme weather on maintaining flight schedules is becoming more pronounced. Currently, radar echo extrapolation technology is widely used in the nowcasting of severe convection, in which the optical flow method is a representative example of traditional extrapolation algorithms. By training a large number of known samples to find the optimal solution, the deep extrapolation models have gradually become better than the traditional algorithms in recent years. In this study, after examining the optical flow method and other deep learning models, a learnable optical flow deep model with a fully convolutional structure is proposed. Using the convolutional deep learning of optical flow information, this new model can overcome the kernel size limitation of traditional convolutional neural networks, and it can correlate the data history further in time and space. The six-year radar mosaics of Guangdong Province, China, were used as the data set to independently train and verify the new model. The results reveal that the new model outperformed the traditional optical flow method and it is also better than other deep learning models.
Labels
- Topic
- Nowcasting
- Type
- New model or method
- Method
- CNN / U-Net
- Time scale
- 0–6 h
- Domain
- Regional
- Data
- Radar
- Status
- Research
- Identifiers
- DOI:10.1109/jstars.2020.3031244 OpenAlex:W3094401231
- Added
- 2026-10-22
Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.