Multi-sensor rainfall data assimilation using ensemble approaches

Abstract

Rainfall is a major process transferring water mass and energy from the atmosphere to the surface. Rainfall data is needed over large scales for improved understanding of the Earth climate system. Although there are many instruments for measuring rainfall, none of them can provide continuous global coverage at fine spatial and temporal resolutions. This thesis proposes an efficient methodology for obtaining a probabilistic characterization of rainfall over an extended time period and spatial domain. The characterization takes the form of an ensemble of rainfall replicates, each conditioned on multiple measurement sources. The conditional replicates are obtained from ensemble data assimilation algorithms (Kalman filters and smoothers) based on a recursive cluster rainfall model. Satellite measurements of cloud-top temperatures are used to identify areas where rainfall can possibly occur. A variational field alignment algorithm is used to estimate rainfall advective velocity field from successive cloud-top temperature images. A stable pseudo-inverse improves the stability of the algorithms when the ensemble size is small. The ensemble data assimilation is implemented over the United States Great Plains during the summer of 2004.

(cont.) It combines surface rain-gauge data with three satellite-based instruments. The ensemble output is then validated with ground-based radar precipitation product. The recursive rainfall model is simple, fast and reliable. In addition, the ensemble Kalman filter and smoother are practical for a very large-scale data assimilation problem with a limited ensemble size. Finally, this thesis describes a multi-scale recursive algorithm for estimating scaling parameters for popular multiplicative cascade rainfall models. In addition, this algorithm can be used to merge static rainfall data from multiple sources.