Brain-Computer Interfaces (BCIs) are advancing rapidly, enabling direct communication between the brain and external systems through non-invasive EEG signals. However, accurately classifying motor imagery (MI) remains a key challenge due to signal non-stationarity, low signal-to-noise ratio, and inter-subject variability. This paper presents AMCL-Net: an Attention-based Multi-scale CNN-LSTM Ensemble Network designed to address these challenges by combining spatial-temporal attention mechanisms with multi-scale convolutional and recurrent processing. The architecture integrates fine, medium, and coarse scale CNN branches to capture diverse temporal features, followed by BiLSTM layers to model sequential dependencies. Dual attention modules selectively emphasize discriminative spatial and temporal patterns, while EEG-specific data augmentation and k-fold ensemble learning enhance robustness and generalization. Evaluated on the BCI Competition IV Dataset 2a, AMCL-Net achieves an average accuracy of 85.32% and a kappa score of 0.804, outperforming existing state of the art methods.