Satellite Image Prediction Relying on GAN and LSTM Neural Networks

Abstract

Satellite image is an important resource for weather forecast. It can indicate the evolution of weather systems and is beneficial in terms of guiding people to make accurate weather forecasting. However, the use of satellite images is encountered with the dilemma of such as small data volume and of poor real-time performance. Hence it is important to make accurate prediction for satellite images. The goal of satellite image prediction is to predict the next few images of the image sequence. Essentially, it is a a spatiotemporal sequence prediction problem, where the prediction of satellite images is difficult due to its large-scale observation area. In this paper, we propose a generative adversarial networks-long short-term memory (GAN-LSTM) model for the satellite image prediction by combining the generating ability of the GAN with the forecasting ability of the LSTM network. For evaluation, we conduct our experiments on the FY-2E satellite cloud maps. In addition, we use a score correct rate (CR) to measure the degree of similarity between predictions and ground truth. Experiment results show that the proposed GAN-LSTM network is capable of efficiently capturing the evolution rules of weather systems, which outperforms the traditional autoencoder-LSTM.