Using multi-timescale methods and satellite-derived land surface temperature for the interpolation of daily maximum air temperature in Oregon

There is a growing demand for fine-grained (e.g. 1 km) daily historical meteorological data. Traditional approaches interpolate between observations collected at existing meteorological stations; however, the demand for ever increasing fine resolutions has begun to outstrip the available ground-station data, especially for daily timescales. One promising technique known as the climate-aided interpolation (CAI) has the advantage of incorporating climatological data to improve fine-grained spatial detail while still producing a daily product. By estimating spatial structure at a coarser timescale than the target timescale, more ground-station data can be used, leading to improved accuracy in these multi-timescale procedures. Another approach to improve fine-grained spatial structure is the incorporation of additional environmental covariates that are measured at fine grains. Elevation has long been used in this role, but other covariates, especially remotely sensed measures of land cover and land surface temperatures (LST), are also promising. In this paper, we provide a comprehensive evaluation of three interpolation methods and rigorous evaluation of multi-timescale procedures that do and do not include additional environmental covariates including LST. We perform this evaluation on daily maximum air temperature predictions using a region with widely varying climates, the state of Oregon, USA. We find that multi-timescale procedures provide a roughly 10% improvement in prediction accuracy over single-timescale procedures. For multi-timescale approaches, we find that universal kriging has the lowest root mean square error (RMSE) but has a lower predictive accuracy when the number of station decreases in the study area. Thus, generalized additive models (GAM) and geographically weighted regression (GWR) may be superior interpolation methods in areas with sparse station data. Most covariates other than elevation do not improve accuracy, but LST may be useful during summer when elevation displays low correlation with T_max. Analyses also show that elevation and LST provide increased spatial heterogeneity.