Bridging Temporal Gaps in Thermal Remote Sensing: A Generative Adversarial Network Approach to Land-Cover-Stratified Diurnal Heat Retention in Peri-Urban Landscapes

Existing urban heat island (UHI) indicators are typically coarse in resolution, limited to daytime acquisitions, or computed without reference to land cover, and cannot resolve field-scale diurnal behavior. This study introduces the Relative Diurnal Thermal Index (RDTI), a dimensionless, land-cover-stratified anomaly computed from daily 10 m day–night land surface temperature (LST) pairs. A conditional generative adversarial network fused daily 1 km MODIS day/night LST with Landsat-8/9 and Sentinel-2 imagery to generate continuous 10 m LST across four temperate French metropolitan peripheries (Paris, Lille, Nantes, and Bordeaux). RDTI expresses the observed diurnal temperature range as a standardized anomaly against a 25-year (2000–2025) MODIS climatology stratified by land-cover class, so each pixel is referenced to the historical behavior of its own surface type. The generated LST was validated against same-date Landsat with centered RMSE of 1.03–2.38 °C, against 3.18–3.36 °C for interpolation of the coarse input, and outperformed TsHARP and random-forest downscaling in three of four cities; the nighttime field was validated independently against ECOSTRESS at 70 m, reproducing the built–cropland thermal contrast to within 0.052 °C. Once normalized against class-specific climatology, built-up surfaces showed suppressed nocturnal cooling relative to cropland in all four cities (Cohen’s d 0.54–1.67). RDTI correlates at −0.29 to −0.80 with relative-warmth SUHII surfaces while remaining distinct, by design, from absolute-temperature hot-spot methods, providing a field-scale, climatologically grounded diurnal complement to existing UHI indicators.