Uncertainty Analysis of Deep Learning-Based Geochemical Models Using a New Approach (Southeast Jiroft, Kerman)

Identifying geochemical anomalies is key in mineral exploration and requires the application of advanced analytical methods due to the complexity of spatial data and the presence of multiple sources of uncertainty. In this study, the performance of five algorithms, including two classical methods (SOM and BSOM) and three deep learning algorithms (Embedded Autoencoder, Deep Embedded Clustering, and LSTM) was evaluated in detecting geochemical anomalies within a portion of the Urumieh-Dokhtar magmatic belt. The results indicated that each algorithm exhibits its unique performance and demonstrates relative efficiency under different conditions. Subsequently, to leverage the relative advantages of each model and reduce uncertainty, a hybrid model based on the Maximum Likelihood Estimation (MLE) method was designed and implemented. The hybrid model achieved superior performance compared to any individual algorithm by integrating the most frequent labels among the models. The validity of the obtained results was confirmed through field visits and microscopic studies of the collected samples, which revealed that valuable elements such as gold, copper, and arsenic were significantly enriched in the areas proposed by the models. Overall, the findings of this research highlight the superiority of deep learning methods, particularly the AESOM algorithm, in anomaly detection. It is recommended to use hybrid models as an effective strategy for managing exploration uncertainties.