A multiscale Transformer-graph attention network for geochemical prospecting

Geochemical anomaly identification is a crucial task for geochemical prospecting. Recently, deep learning algorithms (DLAs) have gained increasing attention for their ability to learn spectral features (geochemical element concentrations and elemental assemblages) and spatial patterns (the spatial scale and shape of geochemical anomalies) for detecting geochemical anomalies. However, conventional DLAs for identifying geochemical anomalies face challenges in jointly modeling long-range feature dependencies and multiscale spatial patterns owing to the complex element associations and spatial heterogeneity of geochemical anomalies. In this study, a multiscale Transformer-graph attention network (MSTGAT) was designed to address the aforementioned challenges. Specifically, the Transformer was applied as a spectral branch to capture long-range contextual information in elemental sequences. A multiscale graph attention network (MSGAT) was devised as a spatial branch to extract spatial structural features across multiple scales and capture complex spatial heterogeneity. MSTGAT was applied to identify geochemical anomalies associated with mineralization in the Tianziping-Taiyangsi region in the eastern part of the Western Qinling orogenic belt of China. Quantitative evaluation showed that the proposed model achieved superior recognition performance to that of the single spectral and spatial models. Moreover, the attention weights of the Transformer and MSGAT were visualized to reveal key geochemical indicators and spatial coupling patterns, facilitating an understanding of the model operation. The identified anomaly zones can guide future mineral exploration in the study area.