Reducing grid artifacts in air temperature downscaling with a Volume-Preserving interpolation preprocessing step

Statistical downscaling, including machine learning based downscaling, is widely used to produce high spatiotemporal resolution air temperatures for monitoring extreme heat events. However, coarse baseline data often exhibit sharp discontinuities at grid boundaries, which propagate into downscaled outputs as unnatural artifacts and spatial discretization errors. These discontinuities arise from the aggregation of complex terrain, land cover, and atmospheric variability within large grid cells. To address this while aligning with the central assumption of statistical downscaling that a coarse baseline represents the average of all underlying fine cells, we innovatively apply volume-preserving pycnophylactic interpolation as a preprocessing step prior to downscaling. Specifically, we first apply pycnophylactic interpolation to ERA5-Land 2-m temperature baselines and then downscaled them to generate hourly 1-km air temperatures for London in the summer of 2022. The downscaling incorporated data from satellite products, reanalysis datasets, weather radar, and crowdsourced stations. Five widely used machine learning based downscaling algorithms were adopted to evaluate the general applicability of our proposed approach. To assess the performance of our method, we further applied three representative mean-bias-corrected point interpolation techniques as alternative preprocessing schemes for comparison. Furthermore, to improve preprocessing efficiency, we developed a Gaussian filtering variant of pycnophylactic interpolation and compared it with the conventional method. The results show that, across all downscaling models, our approach could improve the spatial quality of the downscaled outputs while respecting the core downscaling assumption, and it can also enhance accuracy to a certain extent. Moreover, the proposed Gaussian filtering variant reduces computational time by 36.8 % compared with conventional pycnophylactic interpolation, providing an efficient and scalable trade-off for large-scale workflows.