Landslide Susceptibility Prediction Modeling Based on Remote Sensing and a Novel Deep Learning Algorithm of a...

Area of Science:

• Geosciences
• Artificial Intelligence
• Remote Sensing

Background:

• Landslide susceptibility prediction (LSP) is crucial but challenged by complex landslide feature correlations.
• Conventional machine learning models show limited performance due to the nonlinear nature of landslide factors.
• Remote sensing (RS) images and geographic information systems (GIS) are vital for analyzing landslide-related environmental factors.

Purpose of the Study:

• To propose a novel deep learning model, cascade-parallel LSTM-CRF, for enhanced LSP.
• To leverage RS images and GIS data for accurate landslide susceptibility mapping.
• To overcome the limitations of traditional machine learning algorithms in LSP.

Main Methods:

• Developed a cascade-parallel Long Short-Term Memory (LSTM) and Conditional Random Field (CRF) model.
• Utilized frequency ratio values of environmental factors as input layers.
• Employed cascade-parallel LSTM for feature extraction and CRF for landslide/non-landslide state modeling and optimization.

Main Results:

• The cascade-parallel LSTM-CRF model demonstrated superior performance compared to traditional methods.
• Achieved a higher landslide prediction rate: positive predictive rate of 72.44%, negative predictive rate of 80%, and total predictive rate of 75.67%.
• Successfully applied the model to Shicheng County, China, validating its effectiveness in a real-world scenario.

Conclusions:

• The proposed cascade-parallel LSTM-CRF is an effective data-driven deep learning approach for LSP.
• This novel model overcomes the limitations of conventional machine learning algorithms.
• The study confirms the model's potential for achieving promising and accurate landslide susceptibility predictions.