Visual analysis of meteorological satellite data via model-agnostic meta-learning

Satellites detect the distribution of meteorological data worldwide. However, due to the orbital constraints, the satellite can only reach the same area again after one orbiting cycle. The interval between two detections in the same area is long, and the variation of meteorological data between the two detections is unknown. Moreover, meteorological satellite data are only located near the orbit in one cycle, while the global distribution of meteorological data is unknown. Our method allows to train a regression model with only few meteorological satellite data by taking advantage of the recent advances in deep learning. In detail, we train a model-agnostic meta-learning (MAML) model with data from ground stations instead of meteorological satellites and get the initial network parameters. Based on the initial network parameters trained by MAML, we train the regression models again for different areas. We sample the regression curves of all areas by time and get a time series of global meteorological data distribution. Through case studies conducted together with domain experts, we validate the effectiveness of our method. Graphic abstract