A novel deep learning model for landslide mapping using cross-resolution change technique

Change detection serves as a prevalent approach for updating landslide inventories. Due to the challenges of continuously acquiring high-resolution images, practical applications often rely on bi-temporal images of varying resolutions for landslide mapping. This study introduces the Segment Anything Model (SAM) that leverages cross-resolution images for landslide mapping. Therefore, expanding the utilization of diverse data sources to enhance the temporal frequency of landslide inventory updates. Three unique modules are developed to enable SAM for landslide mapping with automatic prompt generation capability based on cross-resolution images. The cross-scale feature fusion module is designed to align features from low-resolution images with those from high-resolution images through cross-correlation. The multi-scale feature extraction module enhances the model’s capacity to identify landslides of all sizes, especially smaller ones. An Auto-Prompt module is introduced to transform the model into an end-to-end system that autonomously generates prompts for change detection with high generalization capability. Three experiments were carried out to evaluate the model’s performance across three datasets. The first experiment involved testing on a dataset within the same domain as the training dataset, while the second explored change detection on datasets from regions not included in the training dataset. These two experiments were conducted at resolution ratios of 1:2, 1:4, and 1:8. The third experiment assessed the model’s performance by substituting pre-event images with different image sources. Results demonstrate that the proposed model outperforms existing state-of-the-art methods in all three experiments. The average F1 scores of the proposed model at all three resolution ratios in the first experiment and the second experiment are 83.8 and 78.9, surpassing the worst performance model by 24.0 and 48.4, respectively. In the third experiment, the F1 score for the proposed model is 86.1, which is significantly higher than the worst-performing model (52.4). These findings highlight the model’s ability for cross-resolution landslide change detection with high generalization capabilities across diverse regions and data sources.