An Improved Deep Learning-Based Approach to Urban Weather Radar Echo Extrapolation

In recent years, the frequency of extreme precipitation events in urban areas has increased significantly, and the resulting disaster events have serious negative impacts on residents' daily lives. To take timely measures so as to reduce the losses caused by these events, meteorological engineers often perform extrapolation operation using maps sensed by urban weather radars, and then make now casting to estimate the intensity and extent of possible precipitation events. At present, deep learning techniques have been introduced to extrapolation tasks for their wide applicability to various conditions and their ability to mine potential patterns from large amounts of historical data. They obtained better results than traditional extrapolation methods. However, due to structural defects of these models themselves and the extremely uneven distribution of intensity values in radar echo maps, those high-intensity echoes that are closely related to potential severe convective weather events in maps predicted by existing deep models are often underestimated. To alleviate this problem, the MMST-LSTM, a structurally improved recurrent unit, is proposed in this paper, whose ability to capture and simulate the changing dynamics of high-intensity echoes is enhanced. The results of experiments that conducted on a radar echo dataset show the performance improvement of MMST-LSTM, and the alleviation of the intensity-underestimation problem in extrapolated maps compared with previous representative methods.