Abstract

Adverseweather conditions, such as heavy rainfall and associated flooding, pose significant challenges to road safety and emergency response in Dubai, UAE. The existing road infrastructure, primarily designed for arid conditions, often struggles to cope with these infrequent yet severe events, leading to increased traffic accidents and hindered access to critical services. While current research in the UAE has explored traffic accidents, there is a notable gap in predictive modeling that integrates granular meteorological data with specific spatial characteristics to quantify accident risk with high spatiotemporal precision. To address this, this research investigates the comparative performance of ensemble machine learning models in predicting hourly accident frequencies and isolates the specific environmental and infrastructural features that drive localized accident risk. The study utilizes historical traffic accident data published by Dubai Police, detailed meteorological data from Open-Meteo, and road network characteristics from OpenStreetMap, integrated via Level 6 Geohashing. To address the extreme class imbalance of spatial incident data, a rigorous 1:10 negative sampling ratio was employed. The evaluation revealed that gradient-boosted architectures (XGBoost and LightGBM) are highly effective at prioritizing recall, successfully identifying nearly 90% of true accident conditions. Crucially, the study demonstrated that a leaner, Infrastructure-Only architecture was operationally superior, achieving a 2.5x Top-1% Operational Lift. SHAP interpretability analysis highlighted that high-velocity thresholds, concentrated tourism infrastructure, and temporal commute rhythms are the primary catalysts for accident volatility, while continuous meteorological variables introduced predictive noise rather than causal signal. These outcomes support a highly scalable framework for proactive traffic management and optimized emergency service deployment in Dubai, recommending that future systems integrate real-time traffic volume telemetry to further improve predictive precision.

Publication Date

4-29-2026

Document Type

Thesis

Student Type

Graduate

Degree Name

Professional Studies (MS)

Department, Program, or Center

Graduate Programs & Research

Advisor

Ayman Ibrahim

Comments

This thesis has been embargoed. The full-text will be available on or around 5/3/2027.

Campus

RIT Dubai

Download

Share

COinS