See the wind: Wind scale estimation with optical flow and VisualWind dataset

Area of Science:

• Computer Vision
• Environmental Science
• Machine Learning

Background:

• Professional weather records often lack fine-grained spatiotemporal resolution.
• Human interpretation of wind scale relies on observing environmental object dynamics, particularly trees.
• Existing methods for wind sensing are limited in scope and resolution.

Purpose of the Study:

• To develop a camera-based system for sensing wind scale using video analysis.
• To improve the spatiotemporal resolution of weather data through automated wind scale estimation.
• To create a comprehensive dataset for training and evaluating wind sensing models.

Main Methods:

• Introduction of a novel dataset with over 6000 labeled video clips of trees swaying in eleven Beaufort wind classes.
• Proposal of a dual-branch model incorporating an optical flow-based motion branch and a visual branch.
• Adaptive fusion of motion and visual information for wind scale estimation.

Main Results:

• Achieved 86.69% accuracy in estimating wind scale from video clips.
• Demonstrated superior performance compared to baseline models, including a two-stage model and a motion-only branch model.
• The dual-branch model effectively utilizes both motion dynamics and visual cues.

Conclusions:

• The proposed method effectively enables cameras to sense wind scale by analyzing tree motion.
• The developed dataset and model offer a significant advancement in high-resolution wind sensing.
• Publicly accessible dataset and code facilitate further research and application in environmental monitoring.