3D long time spatiotemporal convolution for complex transfer sequence prediction.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Spatiotemporal sequence prediction (SSP) utilizes historical observations to forecast future events.
• Deep learning models are increasingly researched for SSP, but face challenges with long-term data and capturing non-smooth changes.

Purpose of the Study:

• To develop a novel deep learning model, 3DcT-Pred, to overcome limitations in existing spatiotemporal sequence prediction methods.
• To enhance the prediction accuracy of future situations from spatiotemporal sequence data (SSD).

Main Methods:

• Proposed 3DcT-Pred model employing a two-branch 3D convolution architecture.
• Implemented a cross-structured spatio-temporal attention module to capture non-smooth local features.
• Integrated global and local features using a fusion gating module.

Main Results:

• The model mitigates the long-range forgetting problem by extracting global features.
• Enhanced capture of non-smooth local features crucial for detail reconstruction.
• Demonstrated superior performance compared to state-of-the-art models on multiple datasets, including radar echo data.

Conclusions:

• 3DcT-Pred effectively addresses key challenges in spatiotemporal sequence prediction.
• The proposed architecture improves the prediction of future spatiotemporal dynamics.
• The model shows significant potential for various SSP applications.