Death, Taxes and Aerosol

Abstract:

In this world, nothing can be said to be certain, except death and taxes (Benjamin Franklin, possibly paraphrasing Daniel Defoe).

Observations and model predictions of aerosol come with uncertainties. How do we estimate those uncertainties, what are their causes and what are their consequences? In the context of the evaluation of global models with remote sensing data, this talk will attempt to give a whirlwind overview of possible answers to these questions.

Uncertainties in observations (e.g. satellite AOD) can be characterised in several ways: by comparison with a truth reference dataset (e.g. AERONET), by intercomparison with multiple other datasets (e.g. other satellite AOD), or by assumptions on the underlying causes (e.g. calibration uncertainties) and forward modelling. Each of these approaches has strengths and limitations that will be discussed. E.g. the reference dataset may be sparse, the other satellite AOD products are based on similar retrieval schemes or measurements, and the forward modelling may miss important error sources. In essence, there will always be uncertainty in estimated observational uncertainties.

In addition, two important aspects of uncertainty (the distinction between bias and random error; and the impact of spatio-temporal scale) will be discussed. I will argue that the usual way to represent satellite AOD uncertainty is poorly suited to inform modellers of the limitations in observational datasets.

Assessing models with observations introduces a further uncertainty in that the observations may not be fully representative of what the model predicts. This is usually a matter of scales, that is the different granularity with which observations and models attempt to resolve nature (e.g. spatial resolution, or microphysical detail). Estimating this representativeness is a relatively new field but the associated uncertainty is often similar or larger than observational uncertainty. I will argue that more work is urgently needed, not just on the representativeness of individual observations but on that of entire datasets.

The evaluation of models with observations then starts with a proper understanding of observational and representational uncertainty. Using satellite AAOD observations, I will show how despite significant biases in diverse products, a meaningful model evaluation is possible. Ultimately this is because in this case model errors are larger than observational and representational uncertainties. Whether this is reason for joy or sadness is left up to the individual researcher.

Throughout the talk, I will give examples from my own work on model evaluation and data assimilation. If time permits, I will also talk about the quantification of uncertainties in models and the interpretation of the cause of errors.