The research on landslide detection in remote sensing images based on improved DeepLabv3+ method

Area of Science:

• Earth Science
• Computer Science
• Artificial Intelligence

Background:

• Classical semantic segmentation models struggle with accurate landslide edge extraction in high-resolution images.
• Existing models often have a large number of parameters and long training times, limiting their practical application.
• There is a need for efficient and accurate landslide detection methods for disaster management.

Purpose of the Study:

• To propose a lightweight and efficient landslide detection model, Landslide Detection Network (LDNet).
• To improve the accuracy and speed of landslide detection compared to existing semantic segmentation models.
• To reduce model parameters and training time for real-time landslide monitoring.

Main Methods:

• Developed Landslide Detection Network (LDNet) based on DeepLabv3+ architecture.
• Replaced the Xception backbone with the lightweight MobileNetv2 network.
• Integrated a dual attention mechanism using the Convolutional Block Attention Module for enhanced feature detection.

Main Results:

• LDNet achieved high performance metrics: 93.37% precision, 91.93% recall, 92.64% F1-score, 86.30% IoU, 89.79% mIoU, and 95.28% overall accuracy.
• Demonstrated significant improvements over the original DeepLabv3+ model (13-14% increase across metrics).
• Outperformed other models like UNet, PSPNet, HRNet, and Swin Transformer in accuracy, parameter count, and training speed.

Conclusions:

• The proposed LDNet model offers a significant advancement in lightweight and accurate landslide detection.
• LDNet's efficiency and accuracy make it suitable for real-time landslide monitoring and disaster response.
• The model shows strong generalization capabilities and potential for widespread application in geological hazard assessment.