Satellite to Radar: Sequence to Sequence Learning for precipitation nowcasting

The forecasting of rain is a complex problem with centuries of scientific work. The implications of weather for individuals and companies continue to be important. Machine Learning approaches have been shown to outperform state of the art physics based models of weather for short term predictions. Using multi-spectral satellite images as out input and radar reflectivity as the target. We investigate three different types of models: 3D U-Net, ConvLSTM and ConvLSTM with self attention. We found that ConvLSTM outperforms the other approaches for both classification and regression pixel rain intensities.