A semblance measure for model comparison

Abstract

Algorithmic and computational advances have made it possible that geophysical survey and earth model design can be aided by many systematic trial inverse-modelling runs with synthetic data. Such may, for example, come up in machine-learning approaches. Automated image appraisal pertaining to such applications will involve common statistical tests for goodness-of-data fit as a primary evaluation method. However, solution non-uniqueness may render multiple images equivalent in terms of their data fit, requiring secondary categorizers. A logical choice for classifying synthetic-imaging results quantifies the goodness of model fit where a known reference model replaces the observational input. The task of model intercomparison in terms of measuring the resemblance to the reference model poses challenges to common distance-based metrics like root mean square error and mean absolute error. First, distance-based metrics can introduce spurious contributions when smooth models with fuzzy target contours are to be compared against a sharp reference. Second, large differences due to parameter-estimation overshoots can dominate distance metrics. The remedy proposed here is referred to as semblance and is based on the idea of logistic functions, where a binary-dependent variable adds non-zero or zero accumulation terms for the, respectively, passing or failing of preset target thresholds. This classifying approach is amenable to an objective where model feature recognition is primary. Numerical comparisons to distance-based metrics provide evidence for the advantages of the semblance in view of this objective. Geophysical imaging in conjunction with machine-learning is seen as a benefitting upcoming application area.