Adapting the SCBUFRA Framework for Spatio-Temporal Urban Fire Risk Classification in Dubai

Abstract

Predictive urban fire risk models are typically developed and tested on clean and complete datasets. Operational fire records are often incomplete and arrive unevenly over time. How such models perform under these conditions has rarely been evaluated and has not been examined for the operational records of Dubai Civil Defence. This dissertation adapts the scenario- and case-based urban fire risk assessment framework (SCBUFRA) of Cui et al. (2024) from continuous regression to binary weekly classification across 219 Dubai analysis areas, and uses it to measure the effect of three conditions, historical training-window length, training-time feature completeness, and inference-time source availability. These fire records were combined with six auxiliary contextual sources on a time-ordered panel of 68,547 area-week observations spanning 2020 to 2025. An ensemble of four tree learners, weighted by precision-recall performance, was evaluated against a logistic-regression baseline. On a held-out test year the ensemble reached an area under the precision-recall curve (PR-AUC) of 0.0766, 0.0160 above the baseline, though it did not outperform a historical-frequency ranking, which indicates that the predictive signal is spatial and static. Points of interest carried most of the predictive signal, near 47% of feature importance; a recent one-year training window performed at least as well as longer histories; and the model lost little when individual sources were withheld at prediction time. The contribution is a controlled evaluation of operational data conditions that prior work had left unmeasured, together with a working tool that prioritises areas for inspection rather than predicting individual incidents.