Enhanced U-Net++ for Improved Semantic Segmentation in Landslide Detection.

Area of Science:

• Geosciences and Remote Sensing
• Artificial Intelligence in Earth Observation

Background:

• Accurate landslide detection and segmentation are crucial for disaster risk management.
• Challenges include complex terrain and limited high-quality labeled datasets for training models.
• Existing methods often struggle with the intricate details of landslide-affected areas.

Purpose of the Study:

• To develop an enhanced deep learning model for precise semantic segmentation of landslides.
• To improve the accuracy and robustness of landslide segmentation in complex terrains.
• To evaluate the proposed model's performance against traditional methods using a specific dataset.

Main Methods:

• An enhanced U-Net++ architecture, named ASK-UNet++, was developed for semantic segmentation.
• The model integrates multi-scale feature extraction and attention mechanisms.
• Experiments were conducted on the CAS Landslide Dataset for the Wenchuan region.

Main Results:

• The ASK-UNet++ model achieved a mean intersection over union (mIoU) of 97.53%.
• A Dice coefficient of 98.27% and an overall accuracy of 96.04% were recorded.
• The proposed model significantly outperformed traditional semantic segmentation methods.

Conclusions:

• The enhanced U-Net++ model (ASK-UNet++) demonstrates superior performance in landslide segmentation.
• The integration of multi-scale features and attention mechanisms enhances accuracy and robustness.
• This approach holds significant potential for advancing landslide monitoring and disaster response strategies.