Towards assimilation methods that adequately reflect distinct sources of error: A Bayesian perspective [INVITED]

Environmental predictions are affected by a multitude of uncertainties, which arise from sampling errors in the data, from epistemic structural uncertainties in the current understanding of large-scale (catchment and above) environmental dynamics, and from numerical approximations in the model implementation. The first two sources are particularly challenging from a methodological perspective, yet must be dealt with if one is to obtain quantitatively meaningful probabilistic predictions and forecasts. Accommodating recursive assimilation of (potentially real-time) observed data, including that on internal model states, is also a key interest, especially in operational contexts. This talk discusses the treatment of these problems using the Bayesian Total Error Analysis (BATEA) framework. We discus several distinct data-driven strategies for deriving increasingly informative probabilistic descriptions of data uncertainty, including rainfall error analysis using gauge and radar networks and a novel technique for describing rating curve runoff errors. When integrated into a self-consistent inference-assimilation-prediction system that exposes its assumptions and hypotheses to posterior scrutiny and improvement, these advances pave the way for a better characterization of structural model errors. We also discus the formulation of BATEA in recursive form suitable for data assimilation, and the selection of Monte Carlo techniques for their implementation. Pilot applications to hydrological forecasting with the Australian Bureau of Meteorology are overviewed.