Airborne electromagnetic data interpretation with deep learning-based stochastic inversion and posterior distribution clustering with application to salinization detection

Stochastic inversion of AEM data poses a significant computational challenge. In this study, we propose an innovative framework for AEM data interpretation. This framework integrates a deep learning-based stochastic inversion operator with posterior distribution clustering. We first develop an invertible neural network (INN) for AEM data to predict the posterior distribution of subsurface conductivity structures and quantitatively evaluate the model uncertainty. We train the INN using a synthetic data set and validate its adaptability to field AEM data acquired from the lower Murray Basin of South Australia for salinization study. The efficiency of stochastic inversion is significantly improved. We further classify the derived posterior distributions into different groups using K-means clustering to identify the spatial patterns of salinization. The clustering results exhibit a pronounced correlation with the geological environment and previous studies. The proposed AEM interpretation framework can support fast uncertainty evaluation of underground structures and inform conceptual hydrogeological model development and refinement.