Research on tropical cyclone intensity prediction in the North Indian Ocean based on neural networks and multisource data analysis

To address the spatiotemporal modeling challenges in predicting tropical cyclone peak intensity, this study proposes a multimodal feature fusion deep learning framework. By jointly analyzing the spatial distribution of meteorological grid data and the temporal evolution patterns of tropical cyclone trajectories, the model achieves end-to-end intensity prediction from the formative to mature stages of tropical cyclones. The framework employs a two-dimensional convolutional neural network (2D CNN) to extract spatial structural features of the pressure field in tropical cyclone regions, combined with a bidirectional long short-term memory network (BiLSTM) to encode temporal dynamics of path parameters. Feature concatenation and nonlinear dimensionality reduction strategies are designed to generate spatiotemporal fusion representations. Experimental results demonstrate that extending the input time window effectively captures long-range dependencies in tropical cyclone intensity evolution, while the multimodal fusion strategy significantly improves prediction accuracy compared to single-modality inputs. Furthermore, independent encoding architectures exhibit differentiated advantages in predicting distinct physical parameters. During training, the Huber loss function and dynamic learning rate scheduling mechanism synergistically optimize model robustness and convergence efficiency. This study validates the modeling potential of deep learning in capturing abrupt tropical cyclone intensity changes, providing a novel technical pathway for early warning systems and disaster prevention decision-making.