Data Fusion Optimization and Contribution Assessment of Dual-Polarization Radar Data in Short-Term Forecasting of Convective Precipitation

EasyChair Preprint no. 11406

Abstract

China's complex geographical environment frequently leads to severe convective weather events, including thunderstorms, hail, and tornadoes, posing significant threats to the economy and public safety. Upgraded Doppler weather radar systems now provide a broader range of parameters, enhancing short-term quantitative precipitation estimation. Leveraging the Conv-LSTM method, these systems capture spatiotemporal characteristics and relationships among multiple parameters within radar echo images, enabling precise short-term forecasting of convective weather precipitation. The integration of advanced equipment and technology has yielded breakthroughs in quantitative precipitation forecasting for short-term convective events.To fully utilize Doppler weather radar parameters and select the most effective ones for precipitation estimation, an evaluation and optimization of the current convolutional neural network is essential. Our approach enhances the neural network structure by incorporating a self-attention mechanism layer to assess individual parameter contributions. This ensures that the most informative parameters receive greater importance in the forecasting process. Additionally, we introduce a dynamic allocation layer that prioritizes parameters with higher weightings for subsequent predictions.The study results reveal that within the self-attention layer, the KDP parameter exhibits the highest composite weight, underscoring its significance. When compared to the conventional ConvLSTM algorithm, our improved algorithm, which dynamically selects parameters after discerning different precipitation phases, consistently yields superior estimation performance. These findings provide a viable assessment strategy and optimization approach for the application of Doppler weather radar parameters in the estimation of precipitation during severe convective weather events.

Keyphrases: ConvLSTM, data fusion, Dual-polarization radar, evaluation model, self-attention mechanism

@Booklet{EasyChair:11406,
  author = {Guoqing Zhao and Leilei Deng and Zhiyuan Chen and Xianbiao Kang},
  title = {Data Fusion Optimization and Contribution Assessment of Dual-Polarization Radar Data in Short-Term Forecasting of Convective Precipitation},
  howpublished = {EasyChair Preprint no. 11406},

  year = {EasyChair, 2023}}