Sensitivity of urban structure-temperature relationships to grid parameterization

Grid-based aggregations are widely used in geographic analyses to summarize spatial data, yet the sensitivity of these methods to grid size and orientation remains poorly understood, a challenge known as the Modifiable Areal Unit Problem. Here, we examined how changes in grid cell size and orientation affect the representation of urban landscape structure and surface temperature in New York City using summer 2008 data. We applied the Structure of Urban Landscapes (STURLA) classification to Landsat 7 thermal data (60 m resolution), testing grid resolutions from 10 m to 10,000 m and orientations from 0° to 180°. We used linear regression models to quantify structure-temperature relationships across spatial scales. Our results show that grid size significantly influences both urban structure representation and its relationship with surface temperature, while grid orientation has minimal impact because urban complexity at intermediate scales buffers rotation effects. STURLA classes decreased exponentially with increasing grid size (from >140 to <20 classes), while temperature variance explained peaked at 400 m resolution (R 2 = 0.884). For this NYC case study, an optimal range of 200-800 m was identified, where classifications maintained both high urban complexity and strong structure-temperature relationship. The 200-800 m optimal range is specific to NYC and requires cross-city validation. This range corresponds to urban functional units from building complexes to neighborhood blocks. These findings demonstrate that grid size, not orientation, is the critical parameter in urban analysis, and that intermediate resolutions offer the best balance between capturing urban heterogeneity and structure-function relationships. • Grid size profoundly affects urban analysis; orientation has no impact • Optimal resolution range identified: 200-800 m for urban climate studies • Urban structure diversity decreases exponentially with coarser grids • Temperature prediction peaks at 400 m resolution (R 2 = 0.884) • 200-500 m captures both urban complexity and thermal relationships