Radar rainfall nowcasting and flood forecasting based on deep learning
In one sentence. U-Net variants nowcast radar rainfall one hour ahead and feed HEC-HMS flood models; TransAtt-Unet keeps simulated peak errors under 20%.
Abstract
To explore the applicability of deep learning methods to radar rainfall nowcasting and flood forecasting, U-Net, Attention-Unet and TransAtt-Unet are used to carry out rainfall nowcasting. The nowcasted rainfall results are used as inputs to the HEC-HMS hydrological model for flood forecasting. The results show that with a 1-hour lead time, Attention-Unet has the best performance in nowcasting heavy rainfall with a short duration, and the relative errors in the simulated flood peak and runoff volume by the nowcasted rainfall of TransAtt-Unet are less than 20%. Each deep learning model has a good forecasting accuracy for rainfall and flood events with large magnitudes. The rainfall intensity, rainfall totals, flood peaks and runoff volumes are significantly underestimated with a 2-hour lead time, with U-Net achieving relatively good rainfall nowcasting. The 1-hour lead time radar rainfall nowcasting and flood forecasting based on deep learning can provide a scientific reference for watershed flood prevention and mitigation.
Labels
- Topic
- Nowcasting
- Type
- Application study
- Method
- CNN / U-Net, Transformer / ViT
- Time scale
- 0–6 h
- Domain
- Regional
- Data
- Radar
- Status
- Research
- Identifiers
- DOI:10.14042/j.cnki.32.1309.2023.05.003 OpenAlex:W7158829235
- Added
- 2026-10-22
Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.