The accurate identification of traffic congestion plays a pivotal role in optimizing traffic flow management and distribution, thereby enhancing the operational efficiency of transportation systems. To address the limitations in the precision of existing traffic congestion identification methods, this study proposes a novel approach based on an enhanced EfficientNetB0 model. The methodology comprises three main phases: Firstly, a proprietary traffic congestion dataset is subjected to data augmentation techniques, including fuzzy processing and noise injection, to generate an enriched dataset, which is subsequently partitioned into training and testing subsets. Secondly, an advanced attention mechanism, the Efficient Spatial Attention Module (ESPAM), is integrated into the conventional EfficientNetB0 architecture to get the improved EfficientNetB0-ESPAM model. Finally, the performance of the proposed EfficientNetB0-ESPAM model is rigorously evaluated against other prevalent neural network models using the aforementioned dataset. Experimental results demonstrate that the EfficientNetB0-ESPAM model achieves significant improvements across multiple performance metrics, with enhancements of 0.68% in accuracy, 0.67% in precision, 0.74% in recall, and 0.70% in F1-score compared to the baseline EfficientNetB0 model. Furthermore, the proposed model exhibits superior performance across various evaluation indices when compared to other classical network architectures. This study contributes to the field by providing a novel and effective reference framework for traffic congestion identification, with potential implications for intelligent transportation system optimization.