Long-term multi-channel satellite video prediction with attentive video predictor using residual frame

Abstract

Weather forecasting is one of the main research topics in meteorology and the satellite video can be used for weather nowcasting which refers forecasting weather within one to six hours. Therefore, predicting satellite video is very important task for weather nowcasting. In this paper, we propose a novel long-term multi-channel satellite video prediction network based on deep learning model. The proposed network consists of two parts, which are the attentive video predictor and the multiple 3D discriminators. The attentive video predictor gets multi-channel satellite videos as an input and encodes spatio-temporal features for prediction. The multiple 3D discriminators are trained to distinguish whether the input satellite video is real satellite video or predicted satellite video. As a result, the attentive video predictor can predict satellite video with realistic distribution. For the evaluation of long-term multi-channel satellite video prediction, we use satellite videos from COMS-1. Experimental results show that the proposed network predicted satellite video with remarkable quality.