How to derive regional/global averages from sparsely sampled data

Aerosol observations always suffer from limited sampling, be they in-situ or remote sensing measurements. This introduces representation errors when using these observations to derive regional estimates, or when using them to evaluate models. In the latter case, one can collocate the model data with the observations to alleviate the problem. The downside is that only part of the model data (the collocated part) can be compared to observations).

We present a technique to homogenize observations (i.e. "remove" their limited sampling). In essence, this is possible because aerosol exhibits spatial and temporal correlations. In practice, we use models to provide this information. Here the technique is applied to satellite data.

First we use synthetic observations to show that remaining representation errors due to this homogenization technique are below 10%. Next, we show that after applying this homogenization technique, estimates of regional AOD (or AAOD) from 14 (or 5) different observational datasets are far better in agreement than without homogenziation. Lastly, we present evidence that remaining differences in homogenzied (A)AOD in these daatsets is dominated by retrieval error.

We also discuss the evaluation of AEROCOM models with these homogenized data. In particular, we highlight existing biases in modelled AAOD.

Although we have not tried it yet, the technique should also be applicable to in-situ data.