Machine Learning Enhances Flood Susceptibility Mapping in Kenya’s Nyando Watershed

Global: Machine Learning Enhances Flood Susceptibility Mapping in Kenya’s Nyando WatershedIn May 2024, a research team employed machine learning techniques to assess flood risk in the River Nyando watershed of western Kenya, aiming to improve disaster preparedness and land-use planning.

Data Collection

The investigators processed dual‑polarization Sentinel‑1 Synthetic Aperture Radar (SAR) imagery from the May 2024 flood event to generate a binary flood inventory, which served as the primary training dataset. Complementary environmental and hydrological variables—including slope, elevation, aspect, land‑use/land‑cover, soil type, and distance from streams—were compiled to characterize the study area.

Machine Learning Approach

Four supervised classifiers—Logistic Regression (LR), Classification and Regression Trees (CART), Support Vector Machines (SVM), and Random Forest (RF)—were trained using the SAR‑derived flood inventory combined with the six conditioning factors. The models were evaluated through accuracy, Cohen’s Kappa, and Receiver Operating Characteristic (ROC) analysis.

Model Evaluation

Random Forest delivered the highest predictive performance, achieving an accuracy of 0.762 and a Kappa statistic of 0.480, surpassing the results of LR, CART, and SVM according to the reported metrics.

Key Findings

The RF‑based susceptibility map identified the low‑lying Kano Plains near Lake Victoria as the area with the greatest flood vulnerability, a pattern that aligns with historical flood records and the impacts observed during the May 2024 event.

Practical Implications

These results demonstrate the utility of integrating SAR data with ensemble machine‑learning methods for flood susceptibility mapping in data‑scarce regions, offering valuable inputs for disaster risk reduction strategies, land‑use planning, and the development of early‑warning systems.

Future Directions

The authors suggest that incorporating additional temporal SAR observations and expanding the set of environmental predictors could further refine model accuracy and support more granular risk assessments.This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.