Landslide Displacement Prediction With Uncertainty Based on Neural Networks With Random Hidden Weights.

Area of Science:

• Geotechnical Engineering
• Artificial Intelligence
• Data Science

Background:

• Landslide displacement forecasting is crucial for hazard mitigation.
• Deterministic models often fail to capture prediction uncertainty.
• Probabilistic forecasting offers a more robust approach to quantifying uncertainty.

Purpose of the Study:

• To develop a probabilistic landslide displacement forecasting model using artificial neural networks (ANNs).
• To construct accurate prediction intervals (PIs) for landslide displacement.
• To propose a novel ANN architecture with random hidden weights for uncertainty quantification.

Main Methods:

• A new single hidden layer feedforward ANN with random hidden weights was designed for the lower-upper bound estimation (LUBE) method.
• Input weights and hidden biases were randomly initialized, requiring only output weight optimization.
• A hybrid particle swarm optimization (PSO) and gravitational search algorithm (GSA), termed PSOGSA, was employed for output weight optimization.
• A novel ANN objective function combining a modified combinational coverage width-based criterion with one-norm regularization was introduced.

Main Results:

• The proposed method successfully constructed high-quality prediction intervals (PIs) for landslide displacement.
• Experimental results on benchmark and real-world landslide datasets validated the method's capability.
• The approach effectively quantifies uncertainty in landslide displacement predictions.

Conclusions:

• The novel ANN-based probabilistic forecasting framework with random hidden weights provides a significant advancement in landslide displacement prediction.
• The integration of PSOGSA optimization and the new objective function enhances the quality of prediction intervals.
• This method offers a reliable tool for geotechnical engineers and researchers involved in landslide hazard assessment.