Heavy rainfall prediction from satellite and radar observations based on Deep Anomaly Detection

Abstract:

Heavy rainfall prediction is essential to prevent the losses of life and property from a severe natural disaster such as landslides and flash floods. However, it poses challenges related to its inherent nonlinearity of the precipitation process and its irregular spatial and temporal distributions. Recently, deep learning techniques have shown the great possibility of predicting the weather for nowcasting with enormous amounts of data. However, extreme weather events, including heavy rainfall, are considered anomalous behavior in the climate and weather system. Thus, when applying deep learning techniques to predict heavy rainfall, it faces the imbalance of data amount due to its low occurrence. In this study, we applied deep anomaly detection techniques to predict heavy rainfall and introduced ExRainNet using satellite and radar observations. ExRainNet uses convolutional autoencoder architecture to train the normal condition with light to moderate precipitation and then calculate anomalous scores to predict heavy rainfall. The autoencoder has been known for its ability to detect anomaly by reconstructing the input with its key features. Moreover, we also consider spatial and temporal features of a normal conditioned rainfall system with convolutional LSTM layers to train the networks to predict heavy rainfall. We have applied ExRainNet to three different types of dataset, including Integrated Multi-satellitE Retrievals for GPM (IMERG) products, Himawari-8, and radar observations over the Korea peninsula. The preliminary result shows that the averaged anomalous scores of the abnormal conditions are 50 % greater than those of normal conditions for the IMERG dataset. We will discuss the prediction results from each dataset and their effectiveness for nowcasting.