Multi‑feature geological hazard susceptibility assessment by integrating improved ResNet and transfer learning: A case study of the Loess Plateau in Northern Shaanxi
This study addressed the issues of insufficient feature utilization, poor spatial coherence, and weak generalization in conventional models for Geological hazard susceptibility assessment (GHSA) on the Loess Plateau in Northern Shaanxi. Thirteen conditioning factors were quantitatively evaluated and...
Saved in:
| Main Authors: | , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-09-01
|
| Series: | Results in Engineering |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2590123025020183 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | This study addressed the issues of insufficient feature utilization, poor spatial coherence, and weak generalization in conventional models for Geological hazard susceptibility assessment (GHSA) on the Loess Plateau in Northern Shaanxi. Thirteen conditioning factors were quantitatively evaluated and their underlying mechanisms analyzed using the Information quantity model and point density statistics. A lightweight deep network framework was then developed by simplifying the ResNet-18 backbone and embedding a Self-Attention mechanism and a convolutional block attention module. Transfer learning was applied by fine-tuning pretrained ResNet-18 parameters on the target region. Results showed that the transfer learned SFResNet-CBAM achieved an accuracy of 0.826, precision of 0.831, recall of 0.825, F1 of 0.825, and AUC of 0.918 on the test set, representing improvements of 3.0 %, 3.2 %, 2.9 %, 2.6 %, and 5.0 % over the ResNet-18 baseline; AUC increased by 4.3 % compared to the non-attention SFResNet. The model also achieved breakthroughs in high-risk zone boundary delineation and sensitivity to micro-deformation, with the generated GHSA maps closely matching field surveys. This study innovatively combined multi-source data fusion, network lightweight design, dual-attention mechanisms, and a transfer learning strategy. The proposed framework reduced model parameters to 24.1 % of the original ResNet-18 while ensuring reliable predictions under small-sample conditions. It significantly enhances early warning capability for regional geological hazards and offers a feasible theoretical and technical foundation for real-time updates and cross-regional, multi-hazard GHSA on grassroots disaster-prevention platforms. |
|---|---|
| ISSN: | 2590-1230 |