Earthquakes are among the most catastrophic natural phenomena, highlighting the need for precise magnitude prediction to support disaster risk reduction and emergency management. This research focuses on leveraging advanced deep learning techniques to classify earthquake magnitudes using a curated seismic dataset from the Los Angeles region. An extensive feature engineering process was conducted to uncover hidden geophysical patterns and enhance the quality of input data. A comparative analysis was performed using a range of models, including Random Forest, LSTM, BiLSTM, CNN-LSTM and CNN-BiLSTM architecture. The experimental evaluation demonstrates that deep learning models, particularly hybrid architectures, are significantly more effective than traditional machine learning approaches in modeling spatiotemporal dependencies. The CNN-BiLSTM model delivered the best performance and achieving an accuracy of 89.56%, underscoring its capability in accurately classifying earthquake magnitudes. These findings highlight the potential of integrating convolutional and bidirectional learning strategies for enhancing the predictive accuracy of seismic forecasting systems.