A hybrid generative adversarial network for weakly-supervised cloud detection in multispectral images.

Area of Science:

• Earth Observation
• Remote Sensing
• Computer Vision

Background:

• Cloud detection is vital for optical satellite image analysis, but traditional methods are sensor-specific and deep learning approaches require extensive labeled data.
• Clouds obscure ground features, limiting the utility of satellite imagery for land applications.

Purpose of the Study:

• To develop a novel, weakly-supervised cloud detection method that overcomes limitations of existing techniques.
• To enable accurate cloud masking with reduced reliance on pixel-level annotations.

Main Methods:

• A hybrid approach combining Generative Adversarial Networks (GAN) with a physics-based Cloud Distortion Model (CDM).
• Weakly-supervised training utilizing patch-level labels on a new Landsat 8 dataset (WHUL8-CDb).
• Pixel-level cloud mask generation during both training and testing phases.

Main Results:

• The proposed GAN-CDM method significantly outperforms baseline deep learning methods on Landsat 8 images (90.20% vs 72.09%).
• Demonstrates strong transferability, achieving high accuracy on Sentinel-2 images (92.54% vs 77.00%).
• Achieves superior cloud detection accuracy compared to existing methods.

Conclusions:

• The GAN-CDM method offers an effective and adaptable solution for accurate cloud detection in optical satellite imagery.
• The approach shows excellent performance and transferability across different satellite sensors.
• Enables efficient cloud masking with reduced annotation effort.