Rainfall Nowcasting: A Convolutional LSTM Approach with Various Grayscale Representations of Weather Radar Images

Accurate rainfall forecasts empower authorities and localities to make informed decisions and preempt potential disasters. In line with this, rainfall nowcasting presents a promising solution, enabling precise forecasts of rainfall patterns in the short term. In this context, this study adopts the ConvLSTM approach to develop prediction models for rainfall nowcasting. To achieve accurate predictions, the study explores three types of grayscale representations of satellite and radar images as inputs for model training. This is to identify the best approach to obtaining the most accurate forecast results. Overall, the research activity includes data collection of rainfall image data, systematic refinement of data through image preprocessing, conducting comprehensive ConvLSTM model training with various grayscale datasets, and evaluating predictive visualization and accuracy metrics. Notably, when comparing the result metrics among the three types of grayscale representations, namely raw image grayscale, map lines and cloud-only grayscale, as well as cloudonly grayscale, it becomes evident that cloud-only grayscale consistently outperforms the other two grayscale representations. The study's findings have noteworthy implications, specifically for improving the performance of rainfall prediction models. Moreover, in general, these findings can be applied to various rainfall nowcasting applications, including disaster management, agriculture, and infrastructure planning.