Deep Learning-Based Feature Importance for Rainfall Nowcast Driven by GNSS PWV and CAPE

Accurate rainfall nowcasting remains one the most challenging tasks in weather forecasting. Previous studies mainly developed rainfall forecast models by combining GNSS derived precipitable water vapor (PWV) with shallow neural networks or simple machine learning algorithms. However, these shallow models suffer from capturing time-dependent phenomena, difficulties in achieving stable solutions, filtering insignificant inputs, and overfitting. Additionally, the quantitative contributions from predictors for rainfall forecasting are still unclear. Therefore, a deep learning rainfall nowcast (DLRN) model based on a Long Short-Term Memory network and Decision Tree Regression, driven by GNSS PWV, convective available potential energy (CAPE), and multiple meteorological data, is proposed in this study. Hourly PWV, CAPE, temperature, relative humidity, pressure, wind speed, wind direction, and rainfall data from 20 GNSS stations in Taiwan Province, recorded over a period of five years, were selected to evaluate the performance of the proposed DLRN model. The root mean square error (RMSE), mean absolute error (MAE), and correlation coefficients for the DLRN model are 1.25 mm, 0.37 mm, and 0.75, respectively. Compared to existing quantitative rainfall forecast studies, the DLRN model achieved a 34% improvement in RMSE, and a 43% improvement in MAE. In addition, a wavelet coherence analysis indicated that most predictors contribute to rainfall occurrence. Feature importance experiments illustrated that CAPE and PWV are the top two contributing factors to rainfall nowcasting. Experimental results confirmed satisfactory performance of the proposed DLRN model, and that it possesses the capability of deployment in practical scenarios for high-precision rainfall nowcasting.