Effective Detection of Cloud Masks in Remote Sensing Images.

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Accurate cloud mask detection is crucial for weather and natural disaster studies.
• Existing deep learning models struggle with refined edge extraction and detecting small cloud targets due to unpredictable patterns.

Purpose of the Study:

• To develop an advanced deep learning model for improved cloud and cloud shadow segmentation.
• To enhance the detection of fine cloud edges and small cloud formations.

Main Methods:

• Proposes MDU-Net, a multiscale dual up-sampling segmentation network with an encoder-decoder-decoder architecture.
• Utilizes improved residual modules for multi-scale feature extraction and a dual up-sampling strategy for edge refinement and contextual information fusion.

Main Results:

• MDU-Net achieved 95.61% PA and 84.97% MIOU on a Landsat8-based cloud and cloud shadow dataset.
• Outperformed existing models in segmentation accuracy and visualization.
• Demonstrated strong generalization capabilities on the landcover.ai dataset.

Conclusions:

• MDU-Net offers superior performance in cloud and cloud shadow detection compared to other methods.
• The model effectively addresses limitations in edge refinement and small target detection.
• MDU-Net shows potential for practical applications in weather monitoring and disaster management.