Satellite cloud image segmentation based on lightweight convolutional neural network.

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Over 50% of optical satellite imagery is cloud-covered, limiting its utility.
• Accurate cloud segmentation is crucial for satellite image analysis, but challenging due to image variability and snow cover.
• Existing methods struggle with precise boundary identification and efficient processing.

Purpose of the Study:

• To develop a fast and accurate cloud segmentation method for optical satellite images.
• To improve the extraction of semantic information and handling of small-scale features.
• To address edge feature loss caused by noise and training artifacts.

Main Methods:

• A lightweight convolutional neural network (CNN) architecture is proposed.
• Channel attention mechanisms optimize feature maps for enhanced semantic extraction.
• Fusion of high and low-dimensional feature maps improves small-scale feature detection.
• A feature aggregation module adaptively weights multi-level features.
• Fully connected conditional random fields refine segmentation boundaries.

Main Results:

• The proposed method achieved an overall accuracy of 0.9695 and a recall of 0.8218.
• Demonstrated higher segmentation accuracy compared to state-of-the-art methods.
• Achieved the shortest processing time among compared methods.

Conclusions:

• The lightweight CNN effectively segments clouds in optical satellite images.
• The integration of attention, feature fusion, and CRF enhances segmentation performance.
• The method offers a promising solution for efficient and accurate cloud removal in satellite data.