Landslide susceptibility assessment using deep learning considering unbalanced samples distribution.

Area of Science:

• Geosciences
• Artificial Intelligence
• Geological Disaster Management

Background:

• Landslide susceptibility assessment (LSA) is crucial for mitigating geological disasters.
• Traditional LSA models struggle with unbalanced datasets, where landslide occurrences are less frequent than non-landslide areas.
• Deep learning offers potential but requires methods to address data imbalance.

Purpose of the Study:

• To develop and evaluate a novel deep learning approach (DNN-MSFM) for enhanced LSA.
• To address the challenge of unbalanced sample distribution in LSA datasets.
• To improve the accuracy and reliability of landslide susceptibility predictions.

Main Methods:

• A deep neural network (DNN) model was combined with a mean squared false misclassification loss function (MSFM).
• The DNN-MSFM approach was designed to algorithmically handle unbalanced samples.
• Model performance was evaluated using statistical metrics (Overall Accuracy, AUC) and compared against DNN and Support Vector Machine (SVM) models on a real-world dataset.

Main Results:

• The DNN-MSFM model achieved a high Overall Accuracy of 0.889 and an AUC of 0.84.
• Significant performance enhancement was observed compared to baseline DNN and SVM models.
• The model demonstrated effectiveness in learning landslide susceptibility features and providing improved predictions on unbalanced data.

Conclusions:

• The DNN-MSFM approach is effective for LSA, particularly in scenarios with unbalanced landslide sample data.
• The study highlights the importance of balanced loss functions in training deep neural networks for LSA.
• This research contributes to advancing deep learning applications in geological disaster assessment and management.