Recognizing spatial geochemical anomaly patterns using deformable convolutional networks guided with geological knowledge

This study tackles the limited quantification of irregular spatial geochemical anomaly patterns and weak interpretability in deep learning models in geochemical anomaly recognition. We propose a hybrid approach that integrates geological knowledge (GK) into deformable convolutional networks (DCN), creating a model termed GK-DCN, with the aim of enhancing both the performance and transparency of geochemical anomaly recognition. This model introduces learnable parameters that allow the convolution kernel to adaptively adjust its shape according to the characteristics of its sampling position, enabling them to more accurately capture complex and irregular geochemical anomaly patterns caused by mineralization. To enhance geological consistency, ore-controlling faults are incorporated as geological knowledge constraint, guiding the network to prioritize spatial correlations between deposits and faults. Experimental results in southern Tianshan Au-Cu polymetallic ore district demonstrate that the GK-DCN, verified across multiple evaluation metrics, significantly enhances the accuracy and reliability of geochemical anomaly recognition, Besides, it produces more distinct spatial anomalous patterns and higher consistency with known mineral deposits by adaptively adjusting the receptive field. Visualization of the kernel offsets revealed the model's superior adaptive spatial sampling mechanism. Furthermore, feature significance heatmaps generated by Grad-CAM (Gradient-weighted Class Activation Mapping) highlighted the key features that the model focused during geochemical anomaly recognition. These visualizations significantly improve the interpretability and prove the effectiveness in capturing complex geochemical patterns. This work provides an effective intelligent method for spatial geochemical pattern recognition and offers a reference for interpretable deep learning in geochemical exploration through multi-angle visualization.