AI-Assisted Measurement Choices Shape Surface Urban Heat Island Estimates: Rural Reference, Spatial Support, and Product Differences in Ahmedabad, India

Abstract

Surface urban heat island (SUHI) intensity is commonly reported as a single daytime value from one thermal product and one rural reference. Using Landsat daytime land surface temperature at 100 m and MODIS daytime and night-time LST at 1 km for the 2020 and 2025 hot seasons in Ahmedabad, India, with an urban mask defined independently of any thermal input, we varied the rural reference class, spatial support, temporal aggregation and thermal product one element at a time. Three results follow. The rural reference dominates the daytime estimate: holding epoch, product and season fixed, the 2025 daytime intensity was +2.29 °C against natural vegetation, +1.02 °C against cropland and −0.74 °C against bare and sparsely vegetated ground, a range of 3.03 °C that includes a change of sign. Aggregation attenuates different statistics at very different rates, and the contrast can be made at a single support: moving the 100 m field to 1 km leaves the mean intensity unchanged to within 2 % but retains none of the urban area exceeding the fine-grid 99th percentile. That surface occupies 4.5 km² in 158 patches averaging about three pixels, so its loss is close to what block-averaging must produce and quantifies fragmentation rather than revealing a new effect. Extending the rescaling to 2.9 km reduces the mean to 64 % of its fine-grid value, a decline we attribute tentatively to mixing of urban and rural land within coarse cells and do not resolve here. Placing the two products on a common 1 km grid left a mean absolute monthly difference of 1.24 °C, numerically equal to the daytime intensity being measured. Two qualifications apply. The residual is not attributable to the instrument, since overpass time, platform, retrieval, screening and compositing differ between products and were not separated; and the day–night comparisons reported here use fields of different spatial support and are not quantitative. The findings indicate that SUHI values are not comparable across studies unless the reference class, spatial support and product are reported together.