VisNet: Deep Convolutional Neural Networks for Forecasting Atmospheric Visibility.

Area of Science:

• Computer Vision
• Environmental Science
• Artificial Intelligence

Background:

• Visibility estimation is crucial for atmospheric studies and is affected by pollutants and weather.
• Current methods for visibility estimation often struggle with diverse foggy conditions.

Purpose of the Study:

• To propose VisNet, a novel deep integrated convolutional neural network for estimating visibility distances from camera imagery.
• To develop and evaluate a robust model for visibility estimation under various fog densities.

Main Methods:

• Developed VisNet, a parallel deep integrated convolutional neural network with three streams.
• Collected the largest dataset of three million outdoor images with precise visibility values.
• Applied frequency domain filtering and spectral filtering to input images for feature extraction.

Main Results:

• VisNet achieved the highest performance in classification across three diverse datasets.
• The model demonstrated superior performance compared to classical and state-of-the-art visibility estimation methods.
• Evaluated VisNet on varied fog densities using a diverse image set.

Conclusions:

• VisNet offers a significant advancement in visibility estimation accuracy, particularly in challenging foggy environments.
• The proposed deep learning approach provides a more reliable method for assessing atmospheric visibility.