<p>Multisource geoscientific data provide critical constraints for geological mapping. While conventional machine learning approaches typically establish models on training datasets and validate their performance on test sets, this study investigates the application of multisource data fusion in geological mapping and evaluates diverse data combinations across various study areas. We propose a two-stage approach for automatic-transfer geological mapping that integrates multisource data fusion from original geology, remote sensing, and geophysical inputs. The two-stage mapping strategy is designed to first perform coarse segmentation and then refine subdivisions, thereby addressing the challenge of simultaneously classifying all geological units via convolutional neural networks. During the coarse segmentation stage, the prior geological knowledge and total radioactive gamma spectrum are fed into the input layer of a UNet network to facilitate automatic-transfer geological mapping, initially partitioning the region into five units. Subsequently, in the subdivision stage, multisource datasets are concatenated and fed into the network input layer to further divide these five units into 22 detailed units, thereby refining the initial coarse output to complete the automatic-transfer geological mapping. To assess the performance of the automated mapping, the predicted boundaries are benchmarked against manually interpreted geological mapping and field-observed attitude data. Validation against field measurements indicates that an 80% match rate (12/15) is identified for attitude data during the coarse stage, improving to 71.1% (32/45) as more units are added in the subdivision stage. Overall classification accuracy ranges from 94.95% to 81.58% across the two stages, reflecting the inherent trade-off between spatial resolution and classification certainty by the network.</p> Graphical Abstract <p></p>