High-Resolution Geochemical Data Mapping With Swin Transformer-Convolution-Based Multisource Geoscience Data Fusion

Geochemical data are crucial for reflecting geological features and is extensively applied in mineral exploration, environmental impact assessment, and geological research. However, the high economic cost of geochemical data analysis hinders large-scale studies, leading to low spatial resolution, especially in remote areas. Although remote sensing data provides rich surface spectral information and shows a strong correlation with geological features, its accuracy in large-scale geochemical data inversion is insufficient. Therefore, we improve the accuracy and reliability by fusing multisource geoscience data. Vegetation information, digital elevation models, and aeromagnetic data, among other geoscience data, offer new perspectives for geochemical data analysis. This article proposes a novel multimodal spatial–spectral fusion model with swin transformer and convolutional networks for regression (MSSF-SCR). This model extracts spatial features from multisource geoscience data using a multibranch swin transformer and dynamically adjusts feature weights with the multimodal multihead convolutional attention module. The swin transformer unifies spatial features, addressing semantic disparities among diverse data sources. Spectral features from remote sensing data are then fused with spatial features through 2-D convolutional regression, producing 15 m resolution geochemical maps. Experiments conducted in the Dananhu–Tousuquan Island Arc region of East Tianshan demonstrate that MSSF-SCR achieves superior performance in terms of R-squared ($R^{2}$) score, Pearson correlation coefficient (R), mean absolute error, and root-mean-squared error indices for five elements (Al₂O₃, Fe₂O₃, K₂O, MgO, and SiO₂).