Scoring Istanbul’s buildings for residential conversion through a 15-minute city lens

Istanbul’s office towers and shopping malls may hold part of the answer to the city’s deepening housing shortage, and a new computational framework now offers a way to work out which of them stand in neighbourhoods that could actually support residential life. The framework, called PARS, short for Potential Adaptive Reuse Score, was developed by Esma Selen Aksoy and Meryem Birgül Çolakoğlu and published in the journal Computational Urban Science. Rather than examining the buildings themselves, it evaluates the surrounding urban context, asking whether everyday services, amenities and transit are within a 15-minute walk of each candidate structure. The approach responds to a recognised gap in adaptive reuse research: most existing evaluation tools concentrate on building-level characteristics such as physical condition, adaptability and material circularity, while the capacity of the surrounding environment to support everyday residential life remains weakly integrated into decision logic.

The empirical analysis covers 1,454 buildings across Istanbul, drawn from the 2024 building inventory of the Istanbul Metropolitan Municipality. Of these, 948 are registered as office or plaza buildings and 506 as shopping malls. For each building, the researchers generated a 15-minute walking isochrone using the OpenRouteService pedestrian network, creating a building-centred catchment that defines the urban environment a hypothetical resident could reach on foot. All subsequent measurements were computed within these individual isochrones, meaning that two buildings on the same street could receive different assessments if their walkable surroundings differ. The authors note that many of these buildings remain physically intact but increasingly show functional underperformance as economic restructuring, hybrid work and changing consumption patterns alter how office and retail space is used.

Four indicators capture complementary dimensions of residential suitability. The first, Population–Amenity Balance, quantifies whether amenity provision is proportionate to local population intensity, computed as the amenity count per 1,000 inhabitants within the isochrone. The second, Amenity Diversity, uses the Shannon entropy index to measure how evenly amenities are spread across nine functional categories, including art and culture, education, emergency services, grocery, health, leisure, restaurants and cafés, general services and shops. These categories were derived from the Geofabrik OpenStreetMap classification, with amenity data obtained from OpenStreetMap itself. The third indicator, Connectivity, measures metro and rail accessibility through a stop-based adaptation of Katz–Bonacich centrality, in which a station’s importance depends on the number of stops on its line and on transfer stations reachable along the network, with influence decaying exponentially by hop distance. The attenuation coefficient was calibrated at 0.75, the midpoint of a range the authors identified as balancing nearby and distant transfer opportunities.

The fourth indicator, Amenity–Rent Alignment, addresses an economic dimension that proximity measures alone miss. Housing rent levels, obtained at neighbourhood scale from the real estate analytics platform Endeksa, were compared with the price levels of nearby amenities, derived from Google Places API cost tiers ranging from free to very expensive. The indicator is formulated as a symmetric mismatch measure: the absolute deviation from equilibrium between normalised rent and normalised amenity prices counts against a building’s score regardless of direction, since substantial divergence between residential economic conditions and surrounding consumption environments may reduce effective compatibility for future residents. Price-level information was available only for a subset of amenities, and 21 buildings could not be assigned a valid value; as a conservative assumption these received the maximum mismatch penalty.

The four indicators were aggregated into a single composite score under two alternative configurations. The first, Equal Weight Aggregation, assigns each indicator a 0.25 weight and serves as a transparent, assumption-minimising baseline. The second, Typology Differentiation Aggregation, derives weights from the data itself. K-Means clustering first segmented the building stock into six urban-context typologies, and a Random Forest classifier was then trained to predict cluster membership from the four indicators, with Gini-based feature importance values, averaged across 30 independent model runs of 700 trees each, converted into indicator weights. The typologies proved highly distinguishable within the four-indicator feature space, achieving a macro-F1 score of 0.991. Amenity Diversity contributed most strongly to typology separation, followed by Connectivity and Amenity–Rent Alignment, while Population–Amenity Balance played a comparatively smaller role.

The clustering itself revealed a strongly imbalanced structure. Cluster 6, the most common contextual condition, dominated the dataset with 756 buildings, representing functionally diverse but quantitatively thin environments with limited transit access. Cluster 1 contained 269 buildings with high diversity despite low amenity density, and Cluster 5 held 261 peripheral but somewhat mixed environments. Cluster 2, with 98 buildings, exhibited the weakest conditions, combining near-zero amenity balance, extremely low connectivity and bimodal diversity. Cluster 3, with 45 buildings, captured diverse but quantitatively thin areas with the highest Amenity–Rent Alignment values, while Cluster 4, just 25 buildings, displayed the strongest performance: the highest amenity balance in the dataset, high homogeneous diversity and connectivity surpassing all other clusters, marking highly central, amenity-rich and exceptionally well-connected locations.

The resulting score distributions differ markedly between the two aggregation schemes. Under Typology Differentiation Aggregation, the mean PARS score is 49.06 with a median of 51.02, and most buildings cluster between roughly 45 and 60 points. Under Equal Weight Aggregation the distribution shifts downward, with a mean of 40.13 and a median of 40.83. The comparison shows more than a simple displacement: buildings were redistributed throughout the score hierarchy, with some gaining far more than others. Cluster 4 achieved a median score exceeding 80 points under the typology-differentiated scheme, while Cluster 2 recorded the lowest median at approximately 20 points, with limited overlap between them. Cluster 1 and Cluster 6 formed a second tier with medians around 60 and 52 points respectively. The typology-differentiated weighting also improved discriminatory resolution: the standard deviation of scores rose from 11.48 to 12.88, an increase of 12.2 percent, mean pairwise score separation rose from 11.54 to 12.74, and the mean distance to the nearest quartile threshold increased from 4.48 to 5.09.

Crucially, the data-driven weighting did not overturn the underlying hierarchy. When buildings were stratified into four priority quartiles, 88.4 percent retained the same quartile membership under both aggregation schemes, and most transitions occurred between adjacent quartiles rather than between extremes. Spatial autocorrelation analysis using Global Moran’s I confirmed that the highest-priority and lowest-priority quartiles form geographically coherent clusters, with values ranging between 0.445 and 0.517 for the top quartile and 0.460 and 0.505 for the bottom, while intermediate quartiles showed substantially weaker clustering at 0.255 to 0.285. On the map, high-readiness buildings concentrate along the Bosphorus corridor and the coastal bands of the Sea of Marmara, with readiness weakening towards peripheral districts, though isolated high-scoring buildings appear in inland and secondary centres where local combinations of accessibility and amenity provision remain favourable.

Indicator-level decomposition shows what separates high-priority from low-priority buildings. Amenity Diversity exhibits the strongest and most consistent relationship with prioritisation, with median values rising from 0.550 in the lowest quartile to 0.779 in the highest. Connectivity shows a similarly pronounced gradient, from 0.026 to 0.239, while Population–Amenity Balance follows a weaker trend from 0.013 to 0.069 with considerable overlap between quartiles. Amenity–Rent Alignment contributes least, with mismatch medians ranging only from 0.098 to 0.077. The authors interpret this as evidence that readiness emerges through the interaction of multiple complementary urban functions rather than the abundance of any single amenity type, a pattern consistent with the underlying logic of the 15-minute city concept, which defines residential sustainability through access to daily needs within walking distance.

The framework’s plausibility was assessed through spatial consistency rather than outcome validation, which the authors note was not feasible given the limited number of completed residential adaptive reuse cases in Istanbul. The twenty highest-scoring buildings concentrate in Fatih and Kadıköy, with smaller numbers in Pendik and Çekmeköy, districts where mixed-use activity and metro corridors such as the M4, M5 and M8 lines support strong indicator values. The twenty lowest-scoring buildings cluster in peripheral districts including Arnavutköy, Başakşehir and Avcılar, areas previously discussed in planning reports as places where housing production has outpaced supporting urban functions. Notably, buildings in Istanbul’s central business districts mostly scored in the 0 to 50 range under equal weighting and rarely reached the highest priority band under typology-differentiated weighting, a finding consistent with Istanbul Planning Agency concerns about spatial constraints on large-scale conversion in the CBDs. The authors stress that PARS is an early-stage screening tool, not a feasibility assessment: architectural suitability, structural adaptability, ownership, financial and regulatory factors were deliberately excluded, and results depend on the completeness of open datasets such as OpenStreetMap. Because it relies on widely available amenity, transport, population and housing-market data, the methodological structure is designed to be transferable to other metropolitan contexts.

Subject of Research: A computational framework evaluating the residential adaptive reuse potential of office buildings and shopping malls in Istanbul through urban-context indicators

Article Title: PARS: an urban context framework for residential adaptive reuse

Article References: Aksoy, E. S., & Çolakoğlu, M. B. (2026). PARS: an urban context framework for residential adaptive reuse. Computational Urban Science, 6(1), Article 57. https://doi.org/10.1007/s43762-026-00292-3

Image Credits: AI Generated

DOI: 10.1007/s43762-026-00292-3

Keywords: adaptive reuse, 15-minute city, Istanbul, urban analytics, housing shortage, office-to-residential conversion, shopping malls, GIS, K-Means clustering, Random Forest, walkability, composite indicator