Study on Accuracy Improvement of Slope Failure Region Detection Using Mask R-CNN with Augmentation Method.

Area of Science:

• Geosciences
• Computer Science
• Artificial Intelligence

Background:

• Slope failure disasters necessitate efficient damage assessment.
• Current deep learning research for landslide detection is limited, facing challenges with insufficient training data from infrequent aerial imagery during disasters.
• Existing methods lack robust solutions for accurate detection in disaster scenarios.

Purpose of the Study:

• To propose an automatic detection method for slope failure regions using deep learning.
• To enhance the efficiency and accuracy of damage assessment following slope failure disasters.
• To address the challenge of limited training data through data augmentation techniques.

Main Methods:

• Utilized a semantic segmentation deep learning algorithm, Mask R-CNN, for automatic detection of slope failure regions.
• Implemented CutMix-based data augmentation to improve model performance and address data scarcity.
• Compared detection results from various augmentation patterns, including maintaining the shape of slope failure regions.

Main Results:

• Augmenting data while preserving the shape of slope failure regions significantly improved detection accuracy.
• Recall score increased by 0.186 (to 0.701) compared to non-augmented data, indicating fewer missed detections.
• F1 score improved by 0.139 (to 0.740), demonstrating a better balance between precision and recall.

Conclusions:

• The proposed Mask R-CNN-based method with CutMix augmentation is highly effective for accurate slope failure region detection.
• This approach offers a valuable tool for rapid and precise damage assessment in slope failure disaster management.
• The study demonstrates the utility of shape-preserving augmentation in enhancing deep learning models for geological hazard monitoring.