Estimation of precipitation extremes from space at scales relevant for hazards: Diagnostics and recent advances in merging of multi-satellite observations via Deep Learning

In this talk we will present our efforts on multi-scale space-time decomposition of state-of the-art satellite precipitation estimates to diagnose the nature of the error and assess the ability of the current global precipitation products to capture extremes at scales of interest to hydrologic applications and hazards. Diagnosing and quantifying systematic and random errors allows also to characterize uncertainty at any time and location and across a range of scales. For hydrologic applications, getting the timing of storms right is also important and our event-based error decomposition allows to detect and diagnose systematic biases in the start, peak and end time of storms as well as event intensity. We apply the knowledge gained from our spectral decomposition analysis into constructing appropriate loss functions for training the new class of Machine Learning (ML) retrieval algorithms and show significant improvements in preserving extremes and the space-time structure of precipitation across multiple scales, compared to the usual mean squared error (MSE) loss function. We ask the question as to whether the retrieval bias depends on the age of the clouds, and quantify this dependence by tracking clouds from geostationary infrared (GEO IR) imagery. To harness this dependence for improving retrievals, we propose a new conditional generative deep neural diffusion model that combines the information from the instantaneous Passive Microwave (PMW) snapshots taken by Low Earth Orbit (LEO) satellites with the dynamical temporal information provided by GEO IR before and after the time of the LEO overpass, and show a considerable improvement as well as the ability of the conditional generative diffusion model to capture extremes and provide uncertainty estimates.