A novel landslide susceptibility mapping method based on a backpropagation neural network algorithm with optimized non-landslide samples and hyperparameters

Landslide susceptibility mapping is essential for early warning and hazard mitigation. To address the uncertainty in non-landslide sample selection and optimize model hyperparameters, this study introduces a multi-sample label learning (MSLL) approach. Additionally, a meta-heuristic optimization algorithm, Gradient-based optimizer (GBO), is used to optimize the Back Propagation Neural Network (BPNN) model. A spatial database is constructed using 167 historical landslide events and 12 evaluation factors. Non-landslide samples are selected with buffer control sampling (BCS) and MSLL methods. The GBO algorithm is then applied to optimize the hyperparameters of the MSLL-BPNN model. Model performance is evaluated using Area Under the Curve (AUC), Confusion Matrix, Accuracy, and Precision. Results show that the MSLL method improves the AUC by approximately 3 % for both training and testing samples, compared to BCS, indicating a more reliable selection of non-landslide samples. GBO optimization further increases the AUC of the BPNN model by 4 % for training and 3 % for testing, demonstrating the importance of hyperparameter optimization in improving model accuracy. Overall, the MSLL method effectively reduces uncertainty in non-landslide sample selection, while GBO optimization significantly enhances the BPNN model's performance, both are crucial for accurate landslide susceptibility mapping.

Authors

• Shandong Liu
• Xingyuan Jiang
• Xudong He
• Haijun Qiu
• Yi Yang
• Rongqian Zhao
• Hongsen Hu

Keywords

• Landslide susceptibility mapping
• Back propagation neural network
• Non-landslide sample
• Multi-sample label learning
• Gradient-based optimizer