Testing ensemble-based estimates of potential observation impact

While ensemble methods to estimate the impact of assimilated observations on forecast error have been widely used (known as EFSO), similar methods to estimate the benefit of potential observations not assimilated have received less attention. For this presentation we use a toymodel to illustrate these methods and highlight their strengths and weaknesses. We show that these methods work well over a wide range of lead times and for different types of observations, but only when the localization used in ensemble data assimilation to mitigate sampling errors is accounted for. While previous studies struggled to achieve quantitative results because they treated the localization inconsistently, we found three methods to overcome this limitation.