Generating high-resolution spatial interpolations of temperature processes is a vital task for studying urban climate anomalies and their various consequences. Such processes often constitute a complex and demanding data environment: Anthropogenic and natural conditions of the urban landscape result in anisotropic spatial dependencies and trend patterns that often vary in diurnal and seasonal cycles. Two-step geostatistical methods such as residual kriging take spatial heterogeneity into account but ignore the temporal dimension, which can result in a significant loss of potentially useful information. In this study, we propose nonparametric spatial detrending to obtain a process that fulfills the assumptions of ordinary kriging. In our application to urban air temperature series from monitoring sites distributed over the urban and suburban area of Augsburg, Germany, we provide an in-depth analysis of time-varying spatial heterogeneity. By using sub-sampling to account for diurnal, seasonal, and spatial trends, we produce interpolation maps with a resolution of 100 m × 100 m. The validation in a narrower sense is based on cross-validation and shows favorable behavior of the proposed method even when sub-samples are neglected. The broader sense validation is based on hold-out monitoring sites and provides further empirical support for our proposal.