In this study, a hybrid deep learning and ensemble-based approach is introduced to predict typhoon intensity accurately by utilizing historical meteorological track data. The model is compounded by Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) to comprehensively learn spatial and temporal features of typhoon patterns. It first uses CNN to extract spatial features of typhoon track at-tributes (such as wind speed, pressure, latitude, and longitude) and time series attributes and the LSTM network captures time series features and sequential dynamics of typhoon motions. To mitigate the class imbalance problem due to the much lower number of severe typhoon instances, the Synthetic Minority Oversampling Technique (SMOTE) is utilized for balancing the dataset and for achieving better generalization of the model. In addition, a Random Forest (RF) classifier is adopted as ensemble component to further improve the robustness of classification and predictive performance. The model is tested on the Kaggle Typhoon Track Dataset, in which the typhoon intensity is divided into three ranks: Low, Moderate, and Severe. The expire-mental results show that the proposed hybrid CNN–LSTM–RF–SMOTE framework obtains an overall classification accuracy of 97.1% with F1-scores of 0.910, 0.955 and 0.992 for the Low, Moderate and Severe class, respectively. The results of experiments on real data demonstrate that the proposed HST-LSTM+ model can effectively capture the complex spatiotemporal dependencies of typhoon track data and subsequently improve the reliability of a prediction on the typhoon intensity. This prediction ability may contribute to the disaster preparedness, risk management, and early warning system for extreme weather.