<p>Geochemical data obtained from stream sediments are widely used to identify geochemical anomalies and to guide mineral exploration and prediction. However, these data commonly exhibit intricate geological features and nonlinear relationships with spatially heterogeneous distributions, presenting dual challenges for traditional linear models to effectively capture geochemical complexities and spatially dependent patterns. The integration of attention mechanisms into convolutional neural networks (CNNs) specifically addresses these challenges by enabling simultaneous enhancement of elemental sensitivity and spatial localization capabilities. The squeeze-and-excitation module (SEM) improves channel-wise feature discrimination through adaptive channel recalibration, effectively amplifying mineralization-related geochemical signatures while suppressing irrelevant elemental noise. The convolutional block attention module (CBAM) further extends this capability through sequential channel–spatial attention processing, where spatial attention weights complement channel attention by identifying spatially significant anomaly patterns. This dual attention integration resolves the key limitation of conventional CNNs in geochemical modeling, which is their inability to concurrently prioritize important elemental associations and their spatial manifestation patterns. Taking the vanadium deposits in Jiujiang City (Jiangxi Province, China) as an example, this study utilized geochemical data from stream sediments to conduct predictive research on identification anomalies related to vanadium deposits. Comparative experiments using random forest (RF), standard CNN, SEM–CNN, and CBAM + CNN were conducted to evaluate the performance of models with attention mechanisms and identify the most effective approach for recognition of geochemical anomalies related to vanadium deposits within the study area. The research found that, in the absence of attention mechanisms, the RF model demonstrated the poorest overall performance compared to all other models, while the standard CNN model outperformed the RF model. Additionally, the SEM–CNN model further optimized the prediction results with relatively higher AUC (area under the curve) values and success rates compared to those of CNN and RF models. Notably, the CBAM + CNN model, integrating channel and spatial attention mechanisms, surpassed the other models in overall performance having the highest AUC values and success rates. The performance of the CBAM + CNN model in fivefold cross-validation demonstrated its ability to adapt well to the training data while maintaining strong validation performance. Overlaying the anomaly map generated by CBAM + CNN with the outcrop outlines of the Hetang and Piyuancun formations revealed significant moderate- and strong-mineralization zones, underscoring its practical applicability for mineral exploration. This paper highlights the use of attention mechanisms in CNNs to process geochemical data for identifying geochemical anomalies.</p>