Weather Classification by Utilizing Synthetic Data.

Area of Science:

• Computer Vision
• Machine Learning
• Meteorology

Background:

• Weather prediction using real-world images is challenging due to data variance.
• Existing datasets often lack diversity in weather conditions and locations.
• Bias in vision-based datasets hinders accurate classification.

Purpose of the Study:

• To explore the capabilities of a custom driver simulator for generating diverse weather conditions.
• To assess the performance of a synthetic dataset created by the simulator.
• To improve the training efficiency of Convolutional Neural Networks (CNNs) for weather prediction.

Main Methods:

• Development and utilization of a custom driver simulator to create synthetic weather images.
• Generation of a novel synthetic dataset covering a wide range of weather conditions.
• Training and evaluation of CNN models using combined real-world and synthetic datasets.

Main Results:

• The custom driver simulator effectively generated a diverse range of weather conditions.
• Synthetic datasets significantly improved CNN training efficiency.
• A performance increase of up to 74% in training efficiency was observed when using combined datasets.

Conclusions:

• Synthetic datasets are valuable for augmenting real-world data in weather prediction tasks.
• Combining synthetic and real-world data effectively addresses dataset bias and variance.
• This approach offers a promising solution for improving the robustness of vision-based weather classification systems.