Modified LSTM-based Classification Model for Precipitation and Snowfall on the East Coast

The precipitation and snowfall in the eastern coastal region of the Korean Peninsula are high-risk meteorological phenomena often prone to heavy rain and heavy snow due to local geographic characteristics, making predictive research highly critical. However, most existing meteorological studies are limited to case studies, and the few predictive studies attempted are primarily restricted to classification based on given simple factors. Based on established meteorological findings that easterly winds are closely related to precipitation and snowfall in the eastern coastal region, this paper conducts a study to simultaneously predict precipitation and snowfall using meteorological factors of vertical layers tightly linked to easterly wind phenomena. In this paper, we construct a modified LSTM-based prediction model using ERA5 data that encompasses multiple vertical layers. We propose the application of vertical and horizontal data structures tailored to these specific meteorological characteristics, along with the processing of top-bottom relationships in vertical layer data by utilizing the sequential structure of the LSTM. Additionally, we propose a two-stage LSTM architecture designed to process low-dimensional features and short-term sequential dependencies in the first stage, while better learning high-dimensional long-term dependencies and non-linear patterns in the second stage. As a case study, we evaluated the performance of the proposed model at major locations in the Gangwon and Gyeongsang coastal regions, demonstrating superior performance to LSTM-FC in terms of precision, recall, and accuracy.