AbstractDeep autoencoder (AE) networks show a powerful ability for geochemical anomaly identification. Because of little contribution to the AE network, small probability samples (again, please check this) having comparatively high reconstructed errors can be recognized by the trained model as anomalous samples. However, different autoencoder networks have different abilities for anomaly identification. To test these methods for geochemical anomaly identification, we based our study on stream sediment data of the Cu‐Zn‐Ag metallogenic area in southwest Fujian province as samples. Three unsupervised deep learning models: the autoencoder (AE), multi‐convolutional autoencoder (MCAE), and fusion convolutional autoencoder (FCAE), were used to extract the combined structural, spatial distribution, and mixed features of multiple‐elements. The results showed that the anomalous area delineated by the FCAE model had the best consistency with the known copper mineral occurrences, followed by the MCAE and AE models, with area under the curve values (AUC) of 0.80, 0.78, and 0.61, respectively. FCAE and AE were insensitive to changes in convolution window size, while MCAE extracted more spatial distribution or mixed features. Overall, FCAE focused more on structural distribution or mixed features, combining the advantages of both MCAE and AE. Therefore, FCAE performed best among the three deep learning methods. This study provides a practical basis for selecting and constructing geochemical anomaly recognition models based on deep learning algorithms.