Abstract: Accurately decoding Motor Imagery Electroencephalogram (MI-EEG) signals is the core prerequisite for building high-performance Brain-Computer Interfaces (BCIs). However, due to the extremely low signal-to-noise ratio and significant inter-subject variability inherent in MI-EEG signals, traditional decoding models—which often rely on manual frequency band preprocessing or are limited to single- or dual-dimension features—frequently fail to capture crucial neural patterns. To break through the bottleneck of current decoding accuracy, it is imperative to comprehensively and deeply synergize the multidimensional complementary information embedded in the signals, including temporal evolution, spatial topology, and frequency rhythms.To address these challenges, we propose an Adaptive Temporal-Spatial-Frequency Network (ATSFNet) designed for multidimensional feature collaborative representation. This network innovatively constructs three targeted parallel feature streams to achieve a comprehensive analysis of complex EEG data: the temporal branch utilizes multi-scale temporal convolutions to finely capture transient neural dynamics across different receptive fields; the spatial branch introduces dynamic Chebyshev graph convolutions to deeply model the non-Euclidean spatial topological correlations between EEG channels; and the frequency branch integrates an Efficient Channel Attention (ECA) mechanism to accurately extract specific frequency band rhythms. Furthermore, to integrate the complementary advantages of temporal, spatial, and frequency features, we propose a decision-level fusion strategy based on deep supervision and learnable parameter weighting.Extensive experimental validation of ATSFNet was conducted on two public datasets, BCI Competition IV-2a and OpenBMI, achieving state-of-the-art classification accuracies of 83.10% and 80.30%, respectively. Quantitative visualization analyses, combining t-SNE dimensionality reduction and Fisher score evaluations, further confirm that this multi-branch fusion architecture effectively enhances the inter-class separability of deep features and sharpens decision boundaries. The results demonstrate that the joint temporal-spatial-frequency modeling paradigm established by ATSFNet successfully overcomes the representational limitations of single-perspective approaches, providing robust technical support for the algorithmic design of highly reliable motor imagery BCIs.