Cloud Detection in Remote Sensing Images Based on a Novel Adaptive Feature Aggregation Method

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Cloud detection is crucial for remote sensing data preprocessing.
• Accurate identification of cloud boundaries and thin clouds remains challenging, especially in complex scenarios.

Purpose of the Study:

• To design and evaluate NFCNet, a novel network model for enhanced cloud detection.
• To improve the accuracy of cloud boundary segmentation and thin cloud localization.

Main Methods:

• NFCNet incorporates three key submodules: Hybrid Convolutional Attention Module (HCAM), Spatial Pyramid Fusion Attention (SPFA), and Dual-Stream Convolutional Aggregation (DCA).
• HCAM extracts multi-scale features and prioritizes critical information.
• SPFA adaptively fuses features to recover lost details and reinforce important information during upsampling.
• DCA integrates high-level and low-level features to maintain sensitivity to fine details.

Main Results:

• NFCNet demonstrated superior performance on the HRC_WHU, CHLandsat8, and 95-Cloud datasets.
• The proposed algorithm achieved finer segmentation of cloud boundaries compared to existing optimal methods.
• NFCNet provided more precise localization of subtle thin clouds.

Conclusions:

• NFCNet effectively addresses the challenges in cloud boundary detection and thin cloud identification.
• The network's architecture enables more accurate and detailed cloud segmentation in remote sensing imagery.