Geotechnical “Facial Recognition” Challenge

The site recognition challenge is to quantify site uniqueness, directly or indirectly, so that sparse site-specific data can be supplemented by big indirect data (BID) to produce a quasi-site-specific model. Engineers appreciate that every site is unique to some extent, but they do not have the ability to quantify this uniqueness. In all likelihood, the difference between two sites depends on the feature of interest (stratification, spatial variability, properties, anomaly, etc.). All sites in a region share common general features as a result of their geologic origin, but they are different at a detailed level when investigated using site investigation tools. This is broadly similar to the facial recognition challenge. However, the training set for facial recognition (photographs) is much bigger and more complete than the available training set for the site recognition challenge. In fact, there is no 3D photograph of a ground volume for any feature of engineering interest. Nonetheless, for the case of soil properties, this paper shows that it is possible to construct a quasi-site-specific (or quasilocal) transformation model from site specific small data and big indirect data. The strategies include the hierarchical Bayesian method, tailored clustering method, outlier detection, dimension reduction, and dictionary learning. These Bayesian machine learning methods can be combined to produce inferences that are more accurate and less biased than those from the classical probabilistic multiple regression (PMR) method. The computational cost of PMR is the lowest, but its disadvantages are significant. One example is that PMR cannot handle multivariate, uncertain and unique, sparse, incomplete, and potentially corrupted (MUSIC) which is a common set of data attributes for real site data. This paper reviews the state-of-the-art research in the site recognition challenge in the presence of MUSIC. Spatial variability and geologic uncertainty are not included in this review. The performance of each method is illustrated using an actual soil database. The comparison between methods using the same database is conducted for the first time in this paper.