Precipitation NowcastingUsing Deep Learning with Pixel Classification

Abstract

Accurate nowcasting of heavy rains can effectively mitigate damages following meteorological disasters. Precipitation nowcasting –an important tool for forecasting rainfall intensity, has become a challenging topic of study for many meteorologic researchers. In recent years, precipitation nowcasting models based on deep learning have been receiving more attentions. In this study, an encoder-forecaster framework model is proposed for precipitation nowcasting. In this model, sufficient feature map numbers are given in every layer to effectively capture the spatiotemporal features of radar echo sequences. The mode of prediction differs traditional radar echo extrapolation methods, which only yields radar echo intensity values, while the proposed model obtains the pixel classifications. Significant improvements were attained by the proposed model for the two highest radar echo intensity levels compared to the traditional model. The extrapolation results demonstrate the effectiveness and accuracy of the proposed model for precipitation nowcasting.