An Assessment of Methods and Remote-Sensing Derived Covariates for Regional Predictions of 1 km Daily Maximum Air Temperature

The monitoring and prediction of biodiversity and environmental changes is constrained by the availability of accurate and spatially contiguous climatic variables at fine temporal and spatial grains. In this study, we evaluate best practices for generating gridded, one-kilometer resolution, daily maximum air temperature surfaces in a regional context, the state of Oregon, USA. Covariates used in the interpolation include remote sensing derived elevation, aspect, canopy height, percent forest cover and MODIS Land Surface Temperature (LST). Because of missing values, we aggregated daily LST values as long term (2000–2010) monthly climatologies to leverage its spatial detail in the interpolation. We predicted temperature with three methods—Universal Kriging, Geographically Weighted Regression (GWR) and Generalized Additive Models (GAM)—and assessed predictions using meteorological stations over 365 days in 2010. We find that GAM is least sensitive to overtraining (overfitting) and results in lowest errors in term of distance to closest training stations. Mean elevation, LST, and distance to ocean are flagged most frequently as significant covariates among all daily predictions. Results indicate that GAM with latitude, longitude and elevation is the top model but that LST has potential in providing additional fine-grained spatial structure related to land cover effects. The study also highlights the need for more rigorous methods and data to evaluate the spatial structure and fine grained accuracy of predicted surfaces.