The spatial structural features and compositional relationships of multivariate geochemicals are influenced by complex geological processes (e.g., diagenesis and mineralization), and can help identify geochemical anomalies and provide key references for mineral resource exploration. However, previous machine-learning-based models often treat spatial structural features or compositional relationships separately. Based on the multitask stack autoencoder structure, this study proposes a feature fusion convolutional autoencoder (FCAE) to extract and fuse the spatial structural features and compositional relationships of multivariate geochemicals for identifying geochemical anomalies. In addition, a three-stage training (3ST) strategy combining greedy layerwise pretraining and overall fine-tuning is adopted to calibrate the FCAE. To assess the performance, the proposed FCAE was used to identify the anomalies related to the Cu ore in the southwest area of the Wuyishan polymetallic metallogenic belt in China. The results showed that fusing both spatial structural features and compositional relationships effectively improved the accuracy of the anomaly identification. The FCAE outperformed several existing models by achieving an AUC of 0.863, a recall of 0.909, and the highest intersection point of the P-A plot in the experiments. In addition, the FCAE is less sensitive to the size of the convolution window, which makes the method more applicable and reliable for mineral resource exploration.