A novel landslide susceptibility prediction framework based on contrastive loss

Abstract

Recently, the positive unlabeled (PU) learning algorithms have proven highly effective in generating accurate landslide susceptibility maps. The algorithms categorize samples exclusively into positive samples (landslides) and unlabeled samples for training, eliminating random or subjective selection of non-landslide samples. However, existing PU learning algorithms face limitations in capturing correct negative samples in multi-genesis landslide areas, leading to lower prediction accuracy. To address this issue, the PU-pullbaggingDT algorithm was proposed in this study. This integrated method combines the strengths of two techniques: the superior ranking performance of the PU-baggingDT algorithm and the low bias of the contrastive learning approach. The contrastive learning introduces a new contrastive loss of PU learning (PUpull loss), which pulls the distance of similar landslide samples in the projection space closer based on the validation accuracy threshold, without the need for data augmentation and class probability. The PUpull loss relaxes the tightness toward unlabeled samples, reducing the impact of incorrectly defined non-landslide samples on prediction results in multi-genesis landslide areas. The proposed algorithm outperforms existing PU-learning and machine learning methods (support vector machine, decision tree, logistic regression, AdaBoost, and XGBoost) with random selection of negative samples for predicting landslide susceptibility in China’s Zigui County, as demonstrated by comprehensive evaluation metrics. The landslide susceptibility mapping utilizes equal interval division and ranking, achieving approximately a 90% landslide percentage in areas with very high and high susceptibility in Zigui County. This demonstrates the capability of the proposed algorithm to effectively predict landslide susceptibility in complex geological settings.

Keywords

• Landslide susceptibility prediction
• contrastive loss
• positive unlabeled learning
• remote sensing