Prediction of Rain intensity classes from near-real-time geOsynchronous Satellites and PrECipitation imagery Time-series (PROSPECT)

Extreme precipitation events are becoming more frequent and in areas where previously never existed. Accurate/prior prediction of light, medium and intense rain events is important on a time-scales of hours to allow adequate warnings and mitigation planning, especially in remote and developing countries. Indeed, such predictions are available (albeit not optimal yet) by numerical weather prediction (NWP) models or by ground-based radar systems.

However, NWP models are very costly in terms of processing time, memory, and expert-knowledge needs, while ground-based radars are not widespread enough, especially not in developing countries or in remote regions (e.g. desert areas or over the ocean). 

Here, we propose to utilize available near-real-time wide-coverage high temporal resolution geosynchronous satellite imagery and sparse precipitation data to construct a machine learning model for predicting rain intensity classes for short-term (0-6 hours) applications. We propose a tailor-made machine-learning architecture that combines both spatial (e.g. CNN) and time-series (ESN, LSTM) machine-learning approaches that will allow the ingestion and fusion of satellite-data on different times and spatial scales.

We are looking to apply our expertise in cloud detection and classification and retrieval approaches, with the support from the TAD center experts to derive a fast and simple yet useful enough rain predictor that can be used globally and can be easily embedded and implemented on a publicly available web-platform.