A spatial-temporal frequency feature fusion framework (SSF-ConvLSTM) was proposed, in order to address the issue that auditory attention detection methods ignore the band-specific contribution of EEG signals. This framework systematically evaluated the differentiated contributions of the $\delta $(1~4 Hz), $\theta $(4~8 Hz), $\alpha $(8~13 Hz), $\beta $(13~30 Hz), and $\gamma $(30~50 Hz) frequency bands, thereby achieving quantitative screening and dynamic coupling modeling of key frequency bands. Firstly, the spatial weight distribution of neural activities in different frequency bands was revealed through brain topographic maps, thereby screening those closely related to target speech encoding. Secondly, the SSF-ConvLSTM model was constructed. The spatial features of the frequency bands were extracted through convolutional layers, and the modeling ability of the Long Short-Term Memory (LSTM) network for the time-varying dynamics of attention was integrated, thereby enabling the joint decoding of spatial-temporal dynamic features across frequency bands. The algorithm was verified on the public KUL and DTU datasets. The results showed that as the frequency continuously increased, the weights of the frontal and temporal lobes related to auditory attention decoding reached their peak in the $\alpha $ band and then gradually decreased in the $\gamma $ band. Through model analysis on the KUL dataset, the $\alpha $ low-frequency band had the optimal decoding accuracy of 93.38% in the 5-second decision window, which was 9.78 percentage points higher than that of the baseline model. On the DTU dataset, the decoding accuracy of the α band was significantly improved by 5.5 percentage points compared with the baseline model. This study confirmed the key role of band-specific features in AAD decoding, thereby providing a theoretical basis for the development of a new type of band-spatial-temporal coupled brain-computer interface based on feature optimization.