Interpretability-Weighted and Phase-Constrained Autoencoder-Based Extraction of Cu Mineralization Anomalies in the Eastern Tianshan

To address the challenge of identifying copper (Cu) mineralization anomalies masked by high backgrounds in the complex tectonic-magmatic overlapping area of Eastern Tianshan, where traditional smoothing and thresholding methods are often ineffective, this study proposes an integrated anomaly extraction workflow based on ″interpretable weighting and phase-constrained deep reconstruction.″ By utilizing typical deposits as positive samples and background areas as negative samples, SHAP-XGBoost is employed to quantify multi-element feature contributions and automatically identify key indicator factors, thereby enhancing the geological significance of the resulting anomalies. A Phase-Constrained Convolutional Autoencoder (PC-CAE) is introduced, utilizing unsupervised convolutional autoencoding to learn high-dimensional background distributions while phase constraints are employed to capture geological topological structures. This approach enhances the sensitivity to discontinuous signals, ultimately identifying mineralization anomalies masked by high backgrounds through reconstruction residuals. ROC evaluation using known deposits indicates that the Area Under the Curve (AUC) for the integrated anomaly reaches 0.903, outperforming traditional Kriging interpolation and the S-A fractal method. The results demonstrate that the method integrating interpretable weighting and phase-constrained autoencoder extraction exhibits robust performance, providing a reliable geochemical basis and technical support for copper exploration and deployment in Eastern Tianshan.