This article tackles a practical paradox in urban mobility: micro-mobility is promoted as a clean, efficient alternative, yet the safety footprint of bicycles, e-bikes, and e-scooters is still not fully mapped at a scale that planners can actually use (Kaya & Kabakuş, 2026). Using Türkiye as a national laboratory, the authors define micro-mobility as traditional bicycles, electric bicycles, and electric scooters, and position the study around a clear need: moving beyond fragmented crash descriptions toward an integrated, risk-focused understanding of where crashes are likely to concentrate and which factors meaningfully differentiate collision types (Kaya & Kabakuş, 2026).
Empirically, the work stands on a large macro-scale crash dataset: 139,025 police-reported micro-mobility crashes across all 81 provinces between 2015 and 2023 (Kaya & Kabakuş, 2026). The authors structure the risk space using 17 independent variables and 102 sub-variables, integrating them into a GIS workflow so that “risk” can be expressed as a spatial surface rather than only as regression coefficients (Kaya & Kabakuş, 2026).
Methodologically, the signature move is the hybridization of generative AI with spatial decision-making and classical statistical validation. Risk-factor “impact levels” are first weighted via six Large Language Models (DeepSeek, GEMINI, Perplexity, ChatGPT, Copilot, Poe), and these weights are then carried into map algebra to build a national crash suitability map (Kaya & Kabakuş, 2026). The authors treat the LLM stage as a reproducible measurement process rather than a one-off prompt experiment: they standardize the prompt, repeat every query three times per model, and enforce a 5% divergence threshold to flag instability and re-run prompts until convergence (Kaya & Kabakuş, 2026). This is a subtle but important design choice because it makes the AI step behave more like a controlled instrument than a “vibes-based” black box.
They then test whether the LLM weighting is internally consistent across models. Using ANOVA and a Friedman test, they report no statistically significant differences in how the models weight or rank criteria, supporting the claim that the AI-driven weights are stable enough to be aggregated into a single weighting scheme (Kaya & Kabakuş, 2026). The payoff is interpretability: the paper can argue that the resulting crash suitability surface is not an artifact of one model’s quirks, but a convergent output across multiple LLMs.
On the spatial side, the GIS workflow is explicit and replicable. The authors apply Inverse Distance Weighting (IDW) interpolation in ArcMap (power=2, variable search radius with 12 neighbours, automatic cell size) using continuous provincial crash frequencies assigned to provincial centroids (Kaya & Kabakuş, 2026). They normalize sub-variable surfaces to the [0,1] range to enforce comparability across heterogeneous criteria, then combine layers via Raster Calculator and weight overlap logic to build 17 maps and finally a single crash suitability map (Kaya & Kabakuş, 2026).
For inferential validation, the study uses multinomial logistic regression because the dependent variable is collision type with three categories: vehicle-to-vehicle collisions, single-vehicle crashes, and collisions with fixed or moving objects (Kaya & Kabakuş, 2026). They also report low multicollinearity across predictors (VIF values close to 1 across listed variables), reducing the usual concern that effects are being inflated by redundant inputs (Kaya & Kabakuş, 2026). Model fit is presented as strong using pseudo-R² values, with Cox and Snell R² at 56.6%, Nagelkerke R² at 72.8%, and McFadden R² at 55.6% (Kaya & Kabakuş, 2026).
Substantively, the national crash suitability map is translated into an intuitive risk geography. Warm colours (red, orange) identify high crash potential areas, which the authors describe as metropolitan and dense urban areas, university districts, areas with a high young population, and touristic areas; cold colours (blue, green) indicate low crash potential, more common in rural regions of Eastern Anatolia, Central Anatolia, and the Black Sea, where micro-mobility use and traffic density are lower and road geometry differs (Kaya & Kabakuş, 2026). The paper is careful to frame these as macro-scale tendencies rather than local causal diagnoses, which is exactly the kind of restraint reviewers like and scooters dislike.
The implementation logic is direct: the suitability maps are presented as decision-support outputs for local governments to prioritize infrastructure and safety investments, while shared micro-mobility operators are pointed toward operational controls such as geofencing, speed restrictions, and seasonal service adjustments; the authors also connect risk patterns to enforcement and preparedness needs, including helmet use and emergency response planning (Kaya & Kabakuş, 2026). In short, the study treats “where is risk likely to be high” as a planning product, not just an academic result.
Mini glossary of key concepts (all definitions grounded in the study)
Micro-mobility vehicles: In this study, micro-mobility refers specifically to traditional bicycles, electric bicycles, and electric scooters, treated as the main categories involved in Türkiye’s micro-mobility crashes during the study period (Kaya & Kabakuş, 2026).
Crash suitability map: A composite, GIS-based risk surface created by combining multiple crash-related factor maps with their corresponding weights to indicate areas with higher or lower potential risk of micro-mobility crashes (Kaya & Kabakuş, 2026).
AI-driven criterion weighting: A procedure where multiple LLMs are used to assign “effect degrees” or weights to crash-related sub-variables, and these weights are validated for consistency before being used in map combination (Kaya & Kabakuş, 2026).
Prompt design and consistency control: The study’s mechanism for making LLM weighting reproducible by standardizing prompts, repeating each query three times per model, and enforcing a 5% divergence threshold to re-run unstable outputs until convergence (Kaya & Kabakuş, 2026).
Inverse Distance Weighting (IDW): A spatial interpolation method used to convert provincial crash-frequency inputs into continuous risk surfaces; the study applies IDW in ArcMap with default parameters (power=2, variable search radius with 12 neighbours, automatic cell size) (Kaya & Kabakuş, 2026).
Normalization to [0,1]: A preprocessing step that rescales map values into a common range to ensure homogeneity and integrity when combining heterogeneous criteria into a single suitability surface (Kaya & Kabakuş, 2026).
Raster Calculator map combination: The GIS operation used to combine normalized criterion layers into higher-level maps (first at the independent-variable level, then into the final crash suitability map) (Kaya & Kabakuş, 2026).
Collision type (dependent variable): The outcome in the multinomial logistic regression model, consolidated into three categories for analysis: vehicle-to-vehicle collisions, single-vehicle crashes, and collisions with fixed or moving objects (Kaya & Kabakuş, 2026).
High-risk zones (map interpretation): Areas shown in warm colours (red/orange) on the national map indicating higher crash potential, described as metropolitan cities and dense urban areas, university districts, high-young-population areas, and touristic areas (Kaya & Kabakuş, 2026).
Reference
Kaya, Ö., & Kabakuş, N. (2026). Mapping micro-mobility risk: AI-powered geospatial analysis and predictive modelling. Safety Science, 196, 107107. https://doi.org/10.1016/j.ssci.2025.107107