Abstract: There is a lack of selection bases in the geochemical anomaly identification and the reconstruction of the geochemical background conforming to the metallogenic distribution using deep learning algorithms with different network structures. Given this, based on the 1∶200 000 stream sediment data of the copper-zinc-silver metallogenic area in southwestern Fujian Province, this study extracted the combined structural characteristics, spatial distribution characteristics, and mixed characteristics of multiple elements in the samples using three unsupervised deep learning models, i.e., AE, MCAE, and FCAE. Then, these characteristics were used to reconstruct the geochemical background and simulate the metallogenic distribution. The results show that the anomaly areas delineated by the FCAE model were the most consistent with the known copper ore occurrences, followed by the MCAE and AE models. The FCAE, MCAE, and AE models had an area under the curve (AUC) score of 0.80, 0.78, and 0.61, respectively. Moreover, the FCAE and AE models were not sensitive to the change in the convolution window size. These results indicate that when deep learning algorithms are constructed for geochemical anomaly identification, the algorithms based on the extraction of spatial distribution characteristics or mixed characteristics perform well, and those based on the extraction of combined structural characteristics or mixed characteristics have a strong anti-interference ability for the noise caused by the change or inconsistency of the spatial observation scale. This study provides some effective selection bases for constructing geochemical anomaly identification models based on deep learning algorithms.