Deep learning-based landslide susceptibility mapping.

Area of Science:

• Geosciences
• Artificial Intelligence
• Natural Hazard Assessment

Background:

• Landslides pose a significant threat in Iran, causing substantial damage and loss of life.
• Effective landslide susceptibility mapping is crucial for mitigation and planning in hazardous areas.

Purpose of the Study:

• To develop and evaluate a deep convolutional neural network (CNN-DNN) for landslide susceptibility mapping in the Isfahan province, Iran.
• To assess the performance of the CNN-DNN model against benchmark machine learning techniques.

Main Methods:

• A deep convolutional neural network (CNN-DNN) was developed and trained using historical landslide data, remote sensing imagery, and various geomorphological, geological, environmental, and human activity factors.
• Model accuracy was evaluated using confusion matrix statistics and receiver operating characteristic (ROC) curve error indices.
• The CNN-DNN model was compared against Support Vector Machine (SVM), Logistic Regression (LR), Gaussian Naïve Bayes (GNB), Multilayer Perceptron (MLP), Bernoulli Naïve Bayes (BNB), and Decision Tree (DT) classifiers.

Main Results:

• The CNN-DNN model demonstrated superior prediction accuracy compared to benchmark algorithms, achieving an Area Under the Curve (AUC) of 90.9%.
• Performance metrics included Image Retrieval (IRs) of 84.8%, Mean Squared Error (MSE) of 0.17, Root Mean Squared Error (RMSE) of 0.40, and Mean Absolute Percentage Error (MAPE) of 0.42.
• The susceptibility map identified high-risk zones in the west and southwest of Isfahan province, correlating with the Zagros trend.

Conclusions:

• The CNN-DNN model offers a highly accurate and effective tool for landslide susceptibility mapping.
• The findings provide valuable insights for landslide risk management and land use planning in the Isfahan province.