Rainfall Forecasting with LSTM by Combining Cloud Image Feature Extraction with CNN and Weather Information

Abstract

Concern about rainfall increase due to climate change and other factors is growing, inexpensive and easy-to-use rainfall forecasting methods are required. Therefore, this study developed a rainfall forecasting model using a neural network that uses readily available weather information such as cloud images, precipitation, and humidity. The proposed model achieved 89% accuracy for 24-hour-ahead classification, exceeding the 85% accuracy of the Japan Meteorological Agency.(JMA) In addition, by focusing on the seasonality of weather and introducing time information into the forecast model, the stability of the forecast was improved. Finally, a rainfall forecast model was developed and simulated by applying AdaBelief to EfficientNetV2+Bi-LSTM. Consequently, the accuracy of both 2-hour and 24-hour-forecasts exceeded the forecast precision of the previous study and the JMA. In particular, the 24-hour-ahead rainfall forecast precision was improved by more than 10% compared to the previous research, indicating a significant improvement in precision.