Comparison of methods for the production of high resolution global daily climate layers for species modeling

Background/Question/Methods

Fine grained global climate layers are crucial for many applications in global environmental change studies including conservation, species modeling and agricultural applications.  Yet, most layers are limited in their temporal and spatial resolution as well as in temporal extent.  Here we report on the methodological underpinnings of the production of daily global climate layers at ca. one kilometer resolution over the 2001-2010 time-period. Initial work on this project focused on comparing interpolation methods for predicting daily maximum temperature using the state of Oregon as a case study. We compared kriging, generalized additive models (GAMs) and geographically weighted regressions (GWRs) with seven covariates including distance to coast, percent forest cover, latitude, longitude, aspect, elevation and MODIS derived land surface temperature (LST) utilizing single and multi-timescale methods. Our accuracy assessment methods include multiple holdouts, spatial transects, and examination of the spatial autocorrelation with Moran’s I. We then scaled up this approach to the global scale and provided an accuracy assessment which includes daily holdouts for global layers. 

Results/Conclusions

We found that multi-timescale (climate-aided) methods improved accuracy more than single time-scale approaches, with a decrease of up to 0.2C in RMSE. Further, while kriging and GAMs performed better than GWR, kriging has a tendency to overfit and is more strongly affected by a decrease in the number of weather station records. We therefore utilized multi-time scale GAMs to predict air maximum temperature globally for the year 2010.  Results also show that LST and elevation provide much of the spatial fine grained structure in the layers. In addition, LST captures seasonal variation and may improve predictions during the summer. Initial results for 2010 show high daily prediction accuracy with a median value of about 2 C in RMSE and with higher accuracy of 1.5 C or less in some regions. Next steps include expanding the temporal extent to the full 2001-2010 time period and generating bioclimatic variables for use in species modeling.