Results

ClimateNA estimates correlate highly with weather station observations

When comparing weather station observations with ClimateNA estimates for the locations of the given weather stations, we found strong correlations for both Mean Annual Temperature (Figure 1A) and Mean Annual Precipitation (Figure 1B), with R-squared values of 0.991 and 0.906 respectively. 

Using sampled level 4e ecozone averages has a minimal impact on correlation between estimates and observations

When aggregating sampled 5km grid points by level 4e ecozone and then comparing to ecozone means of weather stations, we observe less than 0.015 differences in the R-squared values of the correlation between observations and ClimateNA estimations when compared to the point-to-point correlations. We observe strong correlations between weather station observed ecozone means and ClimateNA predicted ecozone means for both MAT (Figure 1C) and MAP (Figure 1D), with R-Squared values of 0.993 and 0.895 respectively. 

Discussion

The correlation observed between weather station observations and ClimateNA estimations of the weather station locations (figure 1A and 1B) verifies that ClimateNA can generate climate data that closely matches recorded data. This is an important first step, as the rest of the project hinges on the assumption that the climate data generated by ClimateNA is accurate. 

The correlation between the ecozone means of weather station observations and sampled 5km grid points (figure 1C and 1D) provides evidence that ecozone averages may be used as a representation of local climate. We observed only minor differences  between these correlations and the correlations between just ClimateNA estimates and weather station observations, suggesting that aggregating by ecozone costs very little predictive power in a prospective climate matching model. This supports the use of ecozones to drastically reduce the database size required for the assisted migration tool. The correlation seen in Mean Annual Temperature is considerably stronger than that observed for Mean Annual Precipitation, which appears to have some observations that stray considerably from the trendline. 

To further investigate the potential issues in the Mean Annual Precipitation data, a map of Mean Absolute Error between weather station observed ecozone means and ClimateNA estimated ecozone means was created (figure 2). This highlights areas where the ClimateNA ecozone means are the most different from the observed means. In particular, regions of coastal BC and southern Mexico appear to not be well represented by the model. This may be due to the design of the ecozones (which tend to be long, narrow bands due to the elevation gradients), latitudinal extent of ecozones, distribution of weather stations, or other factors: further investigation is needed in order to better integrate these areas into the assisted migration tool. 

The correlation matrices of annual climate variables (figure 3) show which climate variables are highly associated with each other, and that these relationships are not predicted to change much in future climates. This will help in selecting climate variables to refine our climate matching model and reduce the computational power needed. For example, if MAT correlates 100% with MCMT, there may be no need to include MCMT in our model. 

The PCA performed on annual climate variables may be used to further refine our climate matching model by reducing the suite of annual climate variables into only a few multivariate components. In this case, we can use 5 composite variables to explain nearly 95% of the variation in our dataset. Reducing our variables to be matched in our ecozone database from 27 to 5 will drastically improve efficiency, for roughly a 5% reduction in predictive power. The loadings of each variable (ie. their relative contributions to PC1 and PC2) can also be seen in the PCA biplot (figure 4). This shows the relative importance of each variable to the principal components. For example, we can see that several temperature-based metrics (MAT, MWMT, MCMT) are major contributors of PC1, while precipitation metrics (MAP, MSP) appear to be major contributors of PC2. This can be further visually confirmed in the maps of the PC value gradients across North America (Figure 5). We also investigated whether the environmental gradients of our PCs tracked well with ecozone delineations. In coastal BC, we can see that the larger valleys to the east are tracked well by PC2, but the finer scale ecozone delineations are too small to accurately track with our 5km resolution (Figure 6). Future work may aim to increase the resolution to better track with fine ecozone delineations. 

The results of this study provide promising evidence for the support of using ecozones as proxies for local annual climate variables, especially MAT. The study also highlights some potential problem areas for the use of this approach with respect to MAP, which may need to be addressed before accurate assisted migration guidance can be given for these areas. Additionally, this study finds that the relationships between annual climate variables is not predicted to significantly change in the near future, which means we will not need to account for changes in the relationships between climate variables when making predictions. Finally, the study generated a suite of multivariate climate metrics to reduce dimensionality of the annual climate data, allowing for a reduction in variables from 25 to 5 while still explaining 93.8% of the variation within the data. These observations will assist further development of a North America-wide assisted migration framework which integrates current and projected climates.