Pixel-Wise Sky-Obstacle Segmentation in Fisheye Imagery Using Deep Learning and Gradient Boosting.

Area of Science:

• Computer Vision
• Remote Sensing
• Renewable Energy Technologies

Background:

• Accurate sky-obstacle segmentation is vital for solar energy forecasting and environmental monitoring.
• Current methods are costly, require specific training data, and produce imprecise boundaries, often neglecting fisheye lens optics.

Purpose of the Study:

• To develop a low-cost, accurate hemispherical sky segmentation framework for fisheye imagery.
• To improve robustness and generalizability across different imaging devices and environments.

Main Methods:

• Generation of synthetic fisheye training images from street-view panoramas.
• Lens-aware data augmentation to simulate fisheye projection and photometric effects.
• A hybrid deep-learning pipeline combining Convolutional Neural Networks (CNN) and Gradient-Boosted Decision Trees (GBDT) for boundary refinement.

Main Results:

• Achieved high accuracy on real fisheye images (IoU: 96.63%, F1: 98.29%) using smartphones and low-cost lenses.
• Demonstrated strong cross-dataset generalization on an external panoramic dataset.
• The hybrid CNN-GBDT approach effectively sharpened sky-obstacle boundaries.

Conclusions:

• The proposed framework offers an accurate, cost-effective solution for hemispherical sky segmentation.
• Enables practical applications in solar energy and environmental monitoring using widely deployable imaging technology.