Introducing a novel multi-layer perceptron network based on stochastic gradient descent optimized by a meta-heuristic...

Area of Science:

• Geosciences
• Artificial Intelligence
• Environmental Modeling

Background:

• Landslide susceptibility mapping is crucial for hazard mitigation and land-use planning.
• Existing methods often require complex variable selection and model optimization.
• Developing accurate and efficient landslide prediction models is an ongoing challenge.

Purpose of the Study:

• To present a novel methodological approach for landslide susceptibility mapping.
• To integrate Information Theory, a Neural Network (NN) with Stochastic Gradient Descent (SGD), and a Genetic Algorithm (GA).
• To optimize NN parameters and classify landslide-related variables for improved prediction.

Main Methods:

• Shannon's entropy index was used for variable class determination, maximizing the information coefficient.
• The Certainty Factor method was employed for weighting landslide-related variables.
• A NN-SGD-GA model was developed and validated against Logistic Regression (LR) and Random Forest (RF) models using spatial data from Yanshan County, China.

Main Results:

• The NN-SGD-GA model achieved the highest prediction accuracy (88.10%), outperforming RF (86.26%) and LR (85.82%).
• The proposed model demonstrated superior performance in Area Under Curve (0.8212) and relative landslide density (65.09) validation metrics.
• The study successfully identified and classified prognostic variables and optimized model complexity.

Conclusions:

• The integrated NN-SGD-GA approach offers a robust and accurate method for landslide susceptibility mapping.
• This novel methodology provides an intelligent and efficient alternative for spatial investigation and hazard assessment.
• The model's ability to reduce complexity and maintain generalization capacity makes it a valuable tool for geoscientists.