• Proposed a geologically fused multiscale convolutional autoencoder (GMCAE). • Integrated lithological and structural data to guide feature learning. • Integrated multiscale convolution and squeeze-and-excitation (SE) attention to capture complex geochemical anomalies. Geochemical anomaly detection is a critical step in mineral exploration but remains challenging due to the high dimensionality, spatial heterogeneity, and limited labeled data in most geochemical surveys. In this paper, a multiscale convolutional autoencoder for fusing geochemical element concentrations and key geological information (GMCAE) is designed for unsupervised detection of geochemical anomalies, with application to uranium prospectivity mapping in the Yuhuashan district, Jiangxi Province, China. The model integrates dual-branch multiscale convolutional encoders and a squeeze-and-excitation (SE) attention mechanism to enhance feature extraction across multiple spatial resolutions, while key geological information such as fault structures and lithological units are incorporated to improve spatial coherence and interpretability. A composite loss function combining mean absolute error (L1 Loss) and structural similarity index (SSIM) further strengthens anomaly sensitivity. Experimental results demonstrate that the model GMCAE achieves improved predictive performance compared to single-scale and unconstrained multiscale convolutional autoencoder, effectively delineating anomaly zones that closely align with known uranium occurrences and ore-controlling faults. The proposed framework provides a scalable and interpretable tool for data-driven mineral exploration in complex geological settings
Liu et al. (Sun,) studied this question.