Identifying social vulnerability profiles for coastal flood using supervised and unsupervised machine learning: A case study of Lekki Peninsula, Lagos, Nigeria
Coastal flooding disproportionately impacts households based on pre-existing vulnerability characteristics. Identifying these vulnerabilities is critical for effective flood risk reduction. Despite its significance, there is a paucity of techniques for identifying suitable Social Vulnerability Indic...
Saved in:
| Published in: | International journal of disaster risk reduction Vol. 127; p. 105693 |
|---|---|
| Main Authors: | , , , |
| Format: | Journal Article |
| Language: | English |
| Published: |
Elsevier Ltd
01.09.2025
|
| Subjects: | |
| ISSN: | 2212-4209, 2212-4209 |
| Online Access: | Get full text |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Coastal flooding disproportionately impacts households based on pre-existing vulnerability characteristics. Identifying these vulnerabilities is critical for effective flood risk reduction. Despite its significance, there is a paucity of techniques for identifying suitable Social Vulnerability Indicators at a local scale. This study investigates an evidence-based indicator approach to rank factors contributing to social vulnerability to coastal flooding using a purposive sample of 1334 flood-affected households in Lekki Peninsula, Nigeria. By integrating the Expectation Maximization Algorithm with Support Vector Regression (EM-SVR), and employing permutation feature importance, we identified distinct social vulnerability clusters and their associated indicator profiles. The findings reveal that a substantial (over 60 %) of the case study had moderate level of vulnerability, with clusters of similar rankings exhibiting variations in indicator profiles. Also, significant differences within the wards were observed across all areas, especially in Ajiran/Osapa and Maroko/Okun Alfa. The EM-SVR models were evaluated using various metrics, which revealed that the EM-SVR achieved a high R-squared accuracy across the seven clusters, ranging from 88.8 % to 95.7 % for the training set and 90.2 %–96.1 % for the testing set. Furthermore, the models demonstrated a low Mean Absolute Error, ranging from 0.051 to 0.075 for training and 0.051 to 0.077 for testing. Financial instability, poor social networks, lack of insurance, and pre-existing health conditions consistently emerged as the most influential indicators across clusters. These findings offer actionable insight for decision-makers by providing a well-structured and targeted approach to identifying vulnerable households and enhancing mitigation strategies.
[Display omitted]
•We proposed a machine learning framework for assessing social vulnerability indicators.•Permutation features importance was used in the extraction of indicators significant to EM-SVR.•Financial stability and social networks were the most influential predictors of social vulnerability.•Some wards with moderate overall vulnerability still have pockets of extreme social vulnerability. |
|---|---|
| ISSN: | 2212-4209 2212-4209 |
| DOI: | 10.1016/j.ijdrr.2025.105693 |