Validation of Large-Scale Geophysical Estimates Using In Situ Measurements with Representativeness Error

Abstract:

Validating large-scale estimates of geophysical variables by comparing them to in situ measurements neglects the fact that these in situ measurements are not generally representative of the larger area. That is, in situ measurements contain some `representativeness error’. They also have their own sensor errors. The naïve approach of characterizing the errors of a remote sensing or modeling dataset by comparison to in situ measurements thus leads to error estimates that are spuriously inflated by the representativeness and other errors in the in situ measurements. Nevertheless, this naïve approach is still very common in the literature. In this work, we introduce an alternative estimator of the large-scale dataset error that explicitly takes into account the fact that the in situ measurements have some unknown error. The performance of the two estimators is then compared in the context of soil moisture datasets under different conditions for the true soil moisture climatology and dataset biases. The new estimator is shown to lead to a more accurate characterization of the dataset errors under the most common conditions. If a third dataset is available, the principles of the triple collocation method can be used to determine the errors of both the large-scale estimates and in situ measurements. However, triple collocation requires that the errors in all datasets are uncorrelated with each other and with the truth. We show that even when the assumptions of triple collocation are violated, a triple collocation-based validation approach may still be more accurate than a naïve comparison to in situ measurements that neglects representativeness errors.