Combined application of numerical simulation and machine learning in debris flow hazard mapping
Abstract Debris flow hazard mapping (DFHM) played an important role in reducing the threat of debris flows. Conventional DFHM usually requires numerical simulations to obtain debris flow intensity, which is usually quite time-consuming. This paper is to introduce a combined application framework of...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Nature Portfolio
2025-08-01
|
| Series: | Scientific Reports |
| Subjects: | |
| Online Access: | https://doi.org/10.1038/s41598-025-15744-9 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract Debris flow hazard mapping (DFHM) played an important role in reducing the threat of debris flows. Conventional DFHM usually requires numerical simulations to obtain debris flow intensity, which is usually quite time-consuming. This paper is to introduce a combined application framework of numerical simulation and machine learning to improve the efficiency of DFHM. The FLO-2D model was employed to simulate debris flows with different recurrence intervals of 20 years, 50 years and 100 years, respectively. Corresponding maximum accumulation depth and flow velocity were collected as dependent variables. Rainfall with corresponding recurrence intervals, altitude, slope, plane curvature, profile curvature, topographic humidity index, normalized difference vegetation index and Manning’s coefficients were collected as independent variables. Then the 20-year, 50-year and 100-year datasets were prepared by combining the independent and dependent variables. The 20-year and 50-year datasets were used for training a machine learning model called gradient boosted decision tree (GBDT). The 100-year dataset was used for validation. The results showed that the predicted results are quite close to the simulated results, which verified the validity and rationality of the proposed method. In addition, the training and prediction process of machine learning models is more than 10 times faster than the running process of numerical model. This study proposed the potential use of machine learning models as alternatives to hydraulic simulations, which could provide a more efficient solution for large-scale DFHM. |
|---|---|
| ISSN: | 2045-2322 |