Multi-element geochemical anomaly recognition applying geologically-constrained convolutional deep learning algorithm with Butterworth filtering of frequency domain information

Deep learning networks have extensively been used to detect multi-element geochemical anomalies linked to different mineral deposits in recent decades. Efficient detection of multi-element geochemical anomalies is a significant issue that should be done through capable processing methods. The two-dimensional (2D) convolutional neural networks (CNN) are a powerful subset of machine learning algorithms which can present magnificent conclusions. Because they have the ability to extract high-level features of complex inputs. However, there are several shortcomings in applying 2D images as input of the CNNs for multi-element geochemical anomaly mapping. This challenge can be solved by designing one-dimensional (1D) CNN to convolve geochemical data table to decrease uncertainties. On the other hand, great feature extraction ability of the CNNs can be enhanced through (i) defining geological constraints and (ii) training their frameworks with frequency domain data which is contained superlative information. In this regard, a novel geologically-constrained convolutional deep learning (GCDL) algorithm was developed to classify multi-element geochemical data table. Accordingly, the GCDL could demonstrate impressive results applying frequency domain training data of the Robat Sefid district, NE Iran. Based on success-rate curves, frequency domain geochemical map has indicated 86.22% Cr occurrences via 28% of the study area while spatial domain geochemical map has indicated 79.91% Cr occurrences via same study area.