SAASNets: Shared attention aggregation Siamese networks for building change detection in multispectral remote...

Area of Science:

• Remote Sensing
• Computer Vision
• Geospatial Analysis

Background:

• Traditional Convolutional Neural Network (CNN) methods for remote sensing building change detection are limited by receptive fields, hindering detailed change acquisition.
• Redundant information reuse in CNN encoding stages degrades feature representation and detection performance.

Purpose of the Study:

• To design a Siamese network with shared attention aggregation (SAASNet) for improved multispectral remote sensing building change detection.
• To enhance the learning of detailed building semantics and overcome limitations of existing methods.

Main Methods:

• Introduced a special attention embedding module to promote multi-scale feature interaction and global feature representation.
• Incorporated a channel and position multi-head attention module for encoding positional details and sharing channel information.
• Utilized a feature aggregation module with a residual strategy to fuse multi-stage Siamese network features for detecting varied building scales and irregular objects.

Main Results:

• The proposed SAASNet demonstrated superior accuracy and robustness in building change detection.
• Experimental results on LEVIR-CD and CDD datasets validated the effectiveness of the designed network.

Conclusions:

• The developed Siamese network effectively addresses limitations in traditional CNN-based change detection.
• SAASNet offers a promising approach for accurate and robust detection of detailed building changes in multispectral remote sensing images.