Future Directions

This project represents a work in progress. Currently, a first milestone has been reached in the project: the creation of the first set of paleoclimate ecosystem maps. There are many future directions that can be taken with this project. There are also a number of refinements and extensions that I would like to pursue immediately.

1) Re-examine the PCA for a better reduction of the data to its ‘key’ variables. (Top)

It would be advantageous to refine the numerous climate variables to fewer key variables that represent the entire dataset. This can be accomplished through a principal component analysis (PCA). However, the results of the PCA in this case are somewhat inconclusive. Ideally, for variable reduction, the PCA eigenvectors would show most variables with negligible loadings—leaving most of the variance explained in a few key variables (Manly 2005). For this to occur, the original variables must be very highly correlated. While some correlation is evident (and expected), there is a great deal of independence among the variables. In this case, it is necessary to use multiple variables to incorporate a reasonable amount of the variance explained into the further statistical analyses. While some variables in the first principal components were well correlated, I believe the variables can be reduced by a further inspection of and experimentation with the principal components. By applying this to the subsequent discriminant analysis, I am hopeful that a more efficient and generalizable modelling procedure can be developed.

2) Re-structure the existing ecosystem classes so that redundent Alberta and B.C. classes are eliminated. (Top)

This is an interesting problem, and probably more difficult that it first appears. There are certainly areas within the two provinces that are homogoneous across the political bounadry. Also, there are ecosystem zone classificaitons that differ in each province but represent essentailly identical species disctributions, stand structures, etc. However, identifying these redundent ecosystem classifications will be difficult without an intimate knowledge of the preocedure used to develop the ecosystem classes themselves.

At this time, I believe a "bottom-up" procedure might work better: modelling in variants as before, but assigning similar variants to the same ecosystem sub-zone. This may require the creation of new alternative subzones that include both Alberta and B.C. variants. However, it may be possible to retain most of the original sub-zones, eliminate a few of the redundant classes, and include the now 'sub-zone-less' variants into the remaining sub-zones.

3) Better understand & display the probability outputs of the discriminant analysis. (Top)

The probabilities for each classification generated by the discriminant analysis (DISCRIM) procedure in SAS present a remarkable opportunity to comprehend and visualize the ‘classification confidence’ for each of the ecosystem areas in question. This is an excellent tool for understanding and displaying the strengths and weaknesses in the classification procedure. As discussed in the 'Statistical Analysis' portion of this site, the areas of low confidence appear to be located along the ecosystem boundaries in the classifications. This seems intuitive, as these transition (or ‘ecotone’) areas would be some of the more difficult to predict with any degree of certainty—falling in between two potential classifications. Determining a statistically reasonable method or procedure for generating appropriate confidences for these probabilities would be critical to a meaningful display of the DISCRIM outputs.

4) Refine the classification of ‘outlier’ climatic data from the paleoclimate data sets. (Top)

As discussed in 'Statistical Analysis', the outlier paleoclimate data represents an interesting problem—the lack of an appropriate ecosystem classification for certain climatic conditions. In order to prevent unreasonable classifications in this area (such as the CWH area in the northern prairies at 21,000 YBP - click here for image), I would like to develop a procedure to set boundaries on outlier data classification.

It would seem that the key to this procedure lies in the classification probabilities produced in the SAS DISCRIM procedure. Because these classifications have such an unreasonably high confidence associated with them (for reasons described in 'Statistical Analysis'), it should be possible to set a threshold probability, beyond which the data would be considered outside the range of reasonable classification confidence (see figure to the left - click for larger view). Ironically, this threshold would have to be at the extreme high confidence levels. I believe it is possible to properly define this threshold. Though, caution will have to be taken not to discard other valid classifications where the paleoclimate data very closely matches the modern baseline data. I would hypothesize that the greatest potential for this to occur would be along the coastline ecosystems where climatic changes are minimized (especially seasonally), despite dramatic continental climatic fluctuations. While a trial-and-error approach will inevitably be used at some point, the visualizations of the DISCRIM probabilities should provide an excellent tool for determining a reasonable starting point for this procedure.

5) Incorporate major geologic changes throughout the Holocene such as isostatic rebound, etc. (Top)

Because abiotic conditions are equally important to ecological niche modelling, it is important to incorporate topographic and other geologic/geographic features into the ecosystem models. While this project already incorporates topographic variables in the form of a DEM input to determine climatic conditions, there are other abiotic factors that have changed throughout the Holocene. One example is isostatic rebound—the rising of the continental landmass in response to the gradual retreat (and subsequent mass loss) of the continental ice sheets throughout the Holocene period. In the time periods on which I am focusing, isostatic rebound would have certainly affected the coastline topography (Peltier 1998). This has dramatic implications for the social sciences also, especially concerning the migration of early North Americans and their adaptations to a shifting coastline and an ice-free landscape.

6) Compare the modelled ecosystem classifications to available fossil data. (Top)

This is a long-term goal for this project that is still some distance away. However, the comparison of the paleo-ecosystem maps with existing paleo-species distribution data is the key to assessing the validity and accuracy of this modelling process. Data sources such as fossil tree and pollen data will be key to comparing the produced ecosystem maps with the actual ranges of individual species throughout the Holocene period. 

<<<< Back to Results & Discussion | Top of this page | Continue on to References & Links >>>>

---

References:

Manly, B. F. J. (2005). Multivariate Statistical Mehtods: A Primer. 3^rd Ed. Chapman & Hall / CRC Press, Boca Raton.

Peltier, W. R. (1998). "Postglacial variations in the level of the sea: Implications for climate dynamics and solid-earth geophysics." Reviews of Geophysics 36(4): 603-689.

© 2007 - David Roberts, University of Alberta