Explainable artificial intelligence (XAI) for interpreting predictive models and key variables in flood susceptibility
The black-box nature of machine learning limits its explainability and practical application. This study highlights the importance of enhancing interpretability in flood modeling and prediction by investigating the interactions between flood-related explanatory variables and their contributions to m...
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/S2590123025020481 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The black-box nature of machine learning limits its explainability and practical application. This study highlights the importance of enhancing interpretability in flood modeling and prediction by investigating the interactions between flood-related explanatory variables and their contributions to model performance using Explainable Artificial Intelligence (XAI) with the SHapley Additive exPlanations (SHAP) method. We selected a watershed in northwest Iran with available historical flood events and flood-related variables. Various machine learning models were tested, with XGBoost achieving the best performance for flood susceptibility prediction (RMSE = 0.333, AUC = 0.890). The SHAP-based interpretability analysis using the XGBoost’s outputs showed that distance to streams (DTS) was the most influential variable, followed by the topographic wetness index (TWI) and elevation. Other significant contributors included stream power index (SPI), precipitation, and slope. Variables like land use, normalized difference vegetation index (NDVI), aspect, lithology, and curvature had less impact, while soil order showed minimal influence. Short DTS increased flood susceptibility, as did higher TWI, SPI, precipitation, and lower elevations. Urban land use heightened flood risk, while higher NDVI and permeable lithology reduced it. Aspect, curvature, and soil order had marginal effects. The analysis also identified interactions, such as DTS strongly interacting with NDVI at low values, TWI and precipitation showing mid-range interactions, elevation influencing predictions nonlinearly with aspect, and SPI interacting with NDVI at low values but diminishing beyond a threshold. By integrating XAI techniques, this study enhanced flood prediction interpretability, providing clearer insights into variable interactions and a more transparent approach to flood susceptibility modeling. |
|---|---|
| ISSN: | 2590-1230 |