<p>Typhoons are natural disasters with immense destructive potential, and accurate prediction of their tracks is crucial for mitigating disaster losses. However, traditional typhoon track prediction methods typically rely solely on image-based or data-based forecasts, failing to fully capture the dynamic changes of typhoons. In contrast, existing methods that integrate data into images have shown improvements in accuracy. To better capture the continuous feature changes in time-series imagery, especially changes in the position of key points, this study introduces a novel network structure, Convolutional Module-Modified Bidirectional Long Short-Term Memory with Self-Attention (C-MBLSTM-SA), for predicting typhoon tracks. This approach employs a multiscale adaptive method for optimized marking of typhoon satellite images, coupled with image enhancement techniques to further improve feature recognizability. The constructed convolution module extracts image features, and the bidirectional LSTM network models the time series, with the addition of a self-attention mechanism to enhance the model's handling of sequential information. Experimental results demonstrate that this model exhibits high accuracy and stability in typhoon track prediction tasks, with a significant performance improvement compared to many other typhoon track prediction models.</p>