A Generative Adversarial Model for Weather Radar Echo Extrapolation based on Cycle Generation Structure

Precipitation forecasting is crucial given that extreme precipitation occurrences have a significant influence on human civilization. Radar echo extrapolation is a method for forecasting echo maps in the future based on observations of past echo maps, and it can forecast approaching short-term weather events. With moderate effectiveness, deep learning-based models have recently been used for radar echo extrapolation. However, high intensity radar echo values fail to be accurately predicted. In this paper, a novel generative adversarial model for weather radar echo extrapolation is proposed, which is composed of dual criss-cross transposed attention mechanism and cycle generation units. The findings show that the suggested model commonly performs better when in comparison to existing deep learning models.