Lightweight remote sensing change detection with progressive multi scale difference aggregation.

Area of Science:

• Remote Sensing
• Geospatial Analysis
• Computer Vision

Background:

• Change detection (CD) in remote sensing analyzes ground surface changes using multi-temporal images.
• Deep learning models offer advanced feature representation but often require substantial computational resources.
• Existing lightweight CD methods may overlook crucial shallow features.

Purpose of the Study:

• To develop a computationally efficient and lightweight deep learning network for remote sensing change detection.
• To address the limitations of high parameter counts and computational demands in current neural network-based CD approaches.
• To enhance the representativeness of detected changes by fusing shallow and deep features.

Main Methods:

• Proposed a novel lightweight network combining MobileNetV2 as an encoder and a modified UNet as a decoder.
• Utilized MobileNetV2 for efficient feature extraction from bi-temporal remote sensing images.
• Implemented layer-by-layer fusion of difference images within the UNet decoder for improved change representation.

Main Results:

• The proposed Mobile-CDNet achieved the lowest computational cost (2.38G) and fewest parameters (2.95M) among compared lightweight networks.
• Demonstrated superior performance on three public datasets (SYSU-CD, BCDD, LEVIR-CD) with F1 scores of 82.84%, 94.51%, and 90.89%, respectively.
• Validated the effectiveness of the proposed architecture in accurately identifying ground surface changes.

Conclusions:

• The developed Mobile-CDNet offers a practical and efficient solution for remote sensing change detection.
• The fusion strategy effectively enhances the detection of subtle changes by leveraging multi-level features.
• The method provides a valuable alternative for applications with limited computational resources.