Advanced prediction of radar reflectivity using U-Net with visual and infrared data from Himawari-8

Precise prediction of radar composite reflectivity (CREF) is essential in meteorology for monitoring and forecasting extreme weather events. Himawari-8 satellite imagery provides valuable data for this purpose, but the main challenge lies in effectively integrating multiple input channels from these satellite images to achieve precise prediction. Therefore, this study aimed to evaluate and develop a U-Net model using visible (VIS) and infrared (IR) channels from Himawari-8 imagery for radar reflectivity prediction. The model was compared with Attention U-Net, which had been used in similar studies. The results showed that U-Net model with VIS channels performed better than other configurations. This model demonstrated lower Mean Squared Error (MSE) and Mean Absolute Error (MAE) compared to models using IR channels. Despite the superior performance of VIS model, there are challenges in handling IR inputs, indicating the need for further modifications. The study concluded that selecting optimal input channels and modifying the model architecture could significantly improve the accuracy and efficiency of radar reflectivity prediction. These improvements had the potential to contribute substantially to the advancement of remote sensing-based weather forecasting systems.