DASUNet: a deeply supervised change detection network integrating full-scale features.

Area of Science:

• Remote Sensing
• Computer Vision
• Geospatial Analysis

Background:

• Change detection (CD) technology is crucial for interpreting land surface alterations.
• Deep learning (DL) methods offer high accuracy and broad applicability in CD.
• Existing DL-based CD methods often fail to fuse features at full scale and rely on transfer learning.

Purpose of the Study:

• To propose a novel deep learning network, DASUNet, for improved change detection.
• To address limitations in feature fusion and reliance on transfer learning in current DL-based CD methods.
• To enable end-to-end training and exploit multi-scale feature information effectively.

Main Methods:

• Developed a deeply supervised (DS) change detection network (DASUNet) with a Siamese architecture.
• Implemented an atrous spatial pyramid pooling (ASPP) module in the encoding stage for enhanced feature extraction.
• Utilized a DS module in the decoding stage to leverage feature information across all scales for prediction.

Main Results:

• The proposed DASUNet demonstrates state-of-the-art performance on benchmark datasets.
• Achieved an F1 score of 94.32% on the CDD dataset.
• Achieved an F1 score of 90.37% on the WHU-CD dataset.

Conclusions:

• DASUNet effectively fuses full-scale features for superior change detection.
• The network's architecture, incorporating ASPP and DS modules, enhances feature utilization.
• The proposed method represents a significant advancement in deep learning-based land surface change detection.