Data Fusion Approach for Precipitation Nowcasting with ConvLSTM

Heavy rainfall events can lead to significant destructive consequences and impacts in urban centers and socially vulnerable regions.To address this critical issue, we propose a modeling approach rooted in Convolutional Structures within the framework of the Long Short-Term Memory Model (ConvLSTM). This neural network architecture specializes in spatio-temporal prediction, in this case will be applied for short-term forecasts spanning up to 6-hour length. The model provides probabilistic outputs indicating different categories of a threshold of accumulated precipitation.

Leveraging input data from multiple sources, including the infrared channels of the GOES satellite, orographic features derived from MODIS product, and insights from the Geostationary Lightning Mapper (GLM), our approach incorporates ConvLSTM architecture. The target dataset encompasses Global Precipitation Measurement (GPM IMERGE, version 'final run') as a training target. Subsequently testing and validation were conducted against precipitation events occurring in the central region of Brazil outside the training period to validate the ability of the model generalization. Preliminary results of cases study, indicate categorical fields of model responses very similar to those observed by the GPM IMERGE, despite some discrepancies between rain gauges, tests performed with a 6-hour prediction window showed a qualitative ability to estimate categories in a dataset outside of training data, indicating a relative capacity of generalization of the model for the domain.