About U of O
Prospective Students
News & Events
Alumni & Friends
FranšaisLibrariesMapsKeyword Search and DirectoriesCoursesU of O Home








Towards Improved Paleoclimate Estimates: the modern pollen-vegetation relation

By: M. Sawada, K. Gajewski University of Ottawa, Department of Geography and P. Richard Université de Montréal, Département de géographie

Presented at the 1999 Annual Meeting of the Canadian Geophysical Union, Climate System History and Dynamics Section (Contribution to CSHD-4), Banff, Alberta, May 14th.


In this talk we will describe our exploration of pollen source area and will:

  1. Outline the importance of understanding pollen source area and its contributions to paleoclimatology,
  2. We will then outline the GIS database we have compiled to study this question, and
  3. We will then present our initial attempts to address one of the major assumptions in the empirical estimation of pollen source area, namely, how to improve the estimation of pollen source area using appropriate distance weighting functions that best approximate the pollen samples view of the surrounding vegetation.

First, pollen can be used to reconstruct past climates at a number of spatial scales. Accurate mesoscale paleoclimate reconstructions and their interpretations require the precise definition of both the spatial scales and types of vegetation best represented by different pollen taxa. Determining a suitable spatial scale is dependent on defining pollen source area. Pollen source area is defined as the vegetated area surrounding a pollen sample site that contributes the majority of pollen. For example, the vegetated radius surrounding a pollen sampling site that that cumulatively contributes 70% of the pollen detected at a collecting basin (See Jackson 1994 for a full review). However, the degree to which pollen samples from lake sites represent local to regional scale vegetation abundance is debatable, in part due to the discrepancies between empirical and theoretical studies of pollen source areas for lake sized pollen sampling basins. Furthermore, lack of detailed records of vegetation abundance at local scales surrounding pollen sampling sites make it difficult to test theoretical models of pollen source areas. In order to reconstruct past climates at a regional scale from a network of fossil pollen sites we must first determine the scale at which pollen records the vegetation of its origin.

Determining those taxa that are best related to vegetation is important for qualitative and quantitative climate reconstruction using fossil pollen. For example, including only those pollen taxa that are well correlated with vegetation in modern-analog type climate reconstructions may increase the signal to noise ratio and lead to more accurate climate reconstructions.

Moreover, by determining the pollen source area for sites with different pollen sampling properties, such as lakes, bogs or moss polsters, we can begin to determine estimates of the optimal spatial scale for gridding climate reconstructions based on pollen data and thereby allow for better comparisons with GCM output.

Finally, if we can determine the pollen samples view of the surrounding vegetation we can begin to look at the controls on vegetation dynamics at the regional scale, and in particular, begin to better filter climatic signals from abiotic controls on vegetation change during Holocene migrations. For example, if we can determine the scale at which pollen records vegetation then we can determine the nature of the network of proximate pollen sampling sites required to test different hypotheses of plant migration, such as the diffusive front versus long distance outlier model of migration during the Holocene (cf. Clark et al. 1998).

In summary, as a starting point, understanding the pollen vegetation relation will allow us to eventually:

1. Determine and understand the relative role of landscape heterogeneity and climate in driving vegetation dynamics during the Holocene and thereby allow more accurate paleoclimate reconstructions.
2. Contribute to ecological theory and contribute to processed based modeling of future changes in vegetation in response to global warming.
3. Understand the nature and rates and scale of response during past migrations which will contribute to adaptive forest management strategies to mitigate the effects of future environmental change.

LPCOur GIS database contains the location and attribute information compiled from a number of sources. First, we have modern pollen samples from T. Webb (Webb et al. 1994), the NAPD and P.Richard at both lakes and bogs in Southern Quebec. Second, we have fossil pollen samples from the Quebec Pollen Database created by Pierre Richard at the University of Montreal. Third, for comparison with modern pollen, the Quebec Ministry of Forests has allowed us to use data collected at over 24000 sites on edaphic parameters, and more importantly on absolute crown coverage of all major tree taxa in the region. Absolute crown cover is the best proxy of pollen source strength (Jackson 1994) and therefore the best variable to be compared with pollen. To our knowledge no other study has used crown cover at this scale of study because it is difficult to collect.

LPCA PCA of the 17 major tree taxa in southern Quebec distinguishes on the first component between northern mixed hardwood forest and boreal forest. This adequately summarizes the large-scale physiognomic vegetation zonation in which our study of pollen source area is situated. We find mixed or southern type forests in the south and boreal forests in the north. Overlaid in the top plate are the locations of the modern pollen samples from bog and lake sites. In this talk we will only be discussing lake sites.



LPCThe mean vegetation proportions surrounding each modern pollen sampling site at radii from 10 to 100 km tend to indicate a rather homogenous percentage composition up to 100km around each site. The most abundant types are Spruce (Picea), Fir (Abies), Birch (Betula), Maple (Acer), Popular (Populus), and Pine (Pinus), consistent with the boreal forest dominance in the study region. Some types such as Carya, Tilia and others are not abundant in the regional vegetation and make up less than 1% on average of the vegetation at all sites and so cannot be used for determination of accurate pollen vegetation relations in the current study. We will only present results for the major and most abundant tree taxa.


To study pollen source area we require pollen spectra (the percentage of pollen for n taxa for a given site) for a number of sites and 'vegetation spectra' (the percentage of vegetation for n taxa for a given site - the same n taxa as in the pollen spectra) within different radii surrounding each site. We then apply, as a first approximation, geometric-mean regression (Webb et al. 1991) to the pollen spectra and the different vegetation spectra at each radius for all sites. Generally, the vegetation sampling radius which explains the most variance in the pollen spectra and the best overall visual fit (e.g., random residuals and small intercept) is empirically defined as the pollen source area for a given taxa. Pollen source areas are theoretically expected to differ among taxa because each taxa has different pollen production and dispersal characteristics.

Theoretically, the pollen samples sees the surrounding vegetation in a distance-weighted manner. A current and critical challenge is to determine the nature of the distance weighting (Jackson and Kearsley 1998, Jackson and Lyford 1999). Some distance weighting functions should produce better fits between pollen and vegetation. The weighting function that produces the optimal fit should best approximate the pollen samples view of the surrounding vegetation. We assume here that if pollen sees vegetation in a distance-weighted fashion then distance weighing should consistently produce better fits then unweighted vegetation proportions surrounding pollen sample sites. We therefore tested a number of different distance weighting scenarios against the unweighted scenario to determine which best approximates the pollen samples view of vegetation in southern Quebec.

Slide 7 - Method 1 -

LPCOur first distance weighting method considers unweighted versus inverse distance and inverse distance squared at different vegetation sampling radii for lake sites in southern Quebec. If the vegetation surrounding pollen sampling sites is heterogeneous and anisotropic then inverse distance and inverse distance squared will produce different results. If vegetation was randomly distributed across the landscape in constant proportions then the two weighting methods will produce different results where pollen source area is concerned. The difference between the two weighting methods is illustrated. The green line represents the proportional weight given to vegetation at distance d in meters from the lake center. The inverse distance squared weight model decays faster than the inverse distance model and provides less weight to vegetation a given distance from the pollen sample site.

Slide 8 -

LPCThe results of applying these two methods of weighting were almost identical. So we present only inverse distance compared to unweighted vegetation for a 40-km vegetation sampling radius. Only proportions for Spruce, Fir, Maple, Birch and Pine are shown because these are the most abundant types in our study are. Pollen proportion is on the y-axis and vegetation proportion on the x-axis. The scales for all taxa are equivalent for comparisons. The R2 for unweighted vegetation vs. pollen proportions minus the R2 for weighted vegetation vs. pollen proportions are presented.
The difference in R2 illustrates the difference in explained variance between the two vegetation weighting scenarios. A positive result indicates a greater explained variance for the unweighted vegetation vs. pollen proportions at the 40-km sampling radius. Overall, one can see that regression model fits and explained variance were very similar for the distance weighted and unweighted vegetation and pollen proportions.

Slide 9

LPCThe second distance weighting model used for vegetation surrounding each pollen sample site is based on Sutton's formulation for diffusion from a ground level source. This is appropriate for pollen in forested regions because most pollen is released and entrained in the forest canopy (Prentice 1988). The formulation we used was modified by Prentice (1988) to include increasing source areas with increasing basin size. The equation as can be seen in the figure predicts a leptokurtic pattern of pollen dispersal, with deposition highest near the point source, such as an individual tree, and a gradual trailing off with distance. This model has not to our knowledge been applied previously at this spatial scale.

Here we allowed each tree taxon a different weight according to the fall velocities of its pollen. These curves illustrate the proportion of particulate remaining airborne at distance d m from the source. The x-axis is log of distance. The green line shows one of the heaviest pollen types, Fir, which by 10 km has very little airborne pollen according to the model, less than .01 percent. Alternatively, Pine, a light pollen type, still has greater than 45 percent airborne pollen at 10km from the source. Translating this into vegetation weights, the pollen sample should see 45% of the abundance of pine 10 km away but less than 0.01 percent of the abundance of Abies. This model we are using assumes equal pollen production for all taxa and we assume a mean wind-speed of 3 m/s and a lake radius of 300 m.

The results of this model were comparable between unweighted and Sutton weighted vegetation for most pollen types. Exceptions, were as indicated here, Abies, where in the unweighted situation Abies is underrepresented and in the Sutton model extreme overrepresentation is evident. Abies would have to be extremely abundant in the vegetation closer than 200 meters from the lake shore in order to show proportional representation under the Sutton model. In general, the Sutton weighting model did not provide better overall fits than the unweighted vegetation. Part of this problem with heavy wind dispersed taxa may be due to poor estimates of fall velocity.

Slide 10

LPCOur final approach to the source area estimation was to use an anisotropic definition of source area - considering that most pollen should come from upwind of the accumulating lake. Here we took the most frequent May wind directions at the World Meterological Organization Stations in southern Quebec and surrounding region. We then gridded this directional data and extracted the most frequent wind directions at each of the modern pollen sampling sites. We then defined an ellipse with a major axis of 40 km and minor axis of 30 km around each pollen sample site and oriented the origin of the ellipse 15-km downwind in the dominant direction as illustrated. This was the vegetation sampling "radius" which we then compared against unweigted vegetation within a 40-km circular radius.

LPCAgain, this method did not produce significant improvements over the unweighted situation except with Betula where the elliptical source area produced an 11% improvement in explained variance at this scale. However, pine shows a 25% decrease in explained variance under the elliptical model when compared to unweighted vegetation within 40 km.



Tentative Conclusion
At this time we can see no obvious improvement in using distance, dispersal (Sutton) or anisotropic source areas over unweighted vegetation for determining the optimal pollen source area. Consequently, we used unweighted vegetation with pollen proportions to provide some first approximates of the reliable grainsize and spatial representation of vegetation by pollen in Southern Quebec.

Slide 11 - Idividulal

LPCBecause of the failure of distance weighting techniques to produce obvious imporvements in model fits, a different approach was then used do determine average pollen source areas that considers only individual sites. First, each pollen site has pollen spectra of 17 types. Second, for ten vegetation sampling radii (10 to 100 km by 10 km increments) around each site we can produce ten vegetation spectra with the same 17 types. We did this for each pollen site and choose as the pollen source area, the vegetation sampling radius which provided maximum correlation (measured by Pearson's R - a magnitude invariant measure) with its respective pollen site spectra. The results of the different radii are illustrated. The yellow radii are the optimal source radii for bog sites and blue circles are for lake sites. The optimal vegetation sampling radii are not globally spatially autocorrelated. This method assumes that the vegetation spectra should covary with the pollen spectra regardless of the magnitude of the individual taxon percentages.

The frequency distribution for the optimal vegetation sampling radii suggests that most sites, approximately 63 % of them, have optimal radii less than or equal to 20 km. However, this technique is new and requires some refining of the optimal source radii criterion and more exploration of the validity of the assumptions. Possibly, only selecting source radii that have statistically significant improvement over previous radii would help.

Slide 12 - Picea

LPCBecause unweighted vegetation sampling radii showed the better fits compared to weighted vegetation, we used these to compare optimal source areas with those predicted by our individual site source radius technique to see if the average source areas would concurr.

For example, this graph shows Picea pollen against Picea vegetation proportions for all lake sites at sampling radii from 10 to 100km. Here we find the best fit between Picea pollen and vegetation first at 20 km. This would be the source area for Picea. There are further improvements but not until 90 km and the slopes and the intercept at that radius is not a significant improvement.

Slide 13 - Abies

LPCFor Abies we find the optimal source radius at 40 km which is in disagreement with its theoretical expectation of less than 300 m.




Slide 14 - Acer

LPCAcer suggests optimal correlation at 20 km and is severely underrepresented in the pollen record.




Slide 15 - Betula

LPCBetula indicates an optimal correlation at 20 km and is over-represented. However, Betula pollen is quite light and we would expect a larger source area.




LPCBetula is problematic because in our region we have a mix of yellow and white birch and when all sites are considered the relation is tenuous at best.




Slide 17 - Pinus

LPCPinus is also problematic using in our case because we find optimal correlation at 10km whereas theoretical and other empirical studies find optimal radii for pine at 100km or greater.




Unweighted vegetation tends to produce the best linear relations between pollen and vegetation for the major tree taxa in Southern Quebec. However, our estimates may suffer non-linear constraints due to the fact that we are dealing with percentage data. Our next step will be to do these same comparisons using extended R-value models (e.g., Jackson et al. 1995) based on maximum likelihood in order to avoid the percentage constraint. This will determine if our first approximations of pollen source area are robust and thereby allow us to move on to the next step and use the information we have gained for climate reconstruction and the interpretation of vegetation spatial structure during Holocene vegetation migrations.


Bradshaw, R.H.W. and Webb, T. 1985. Relationships between contemporary pollen and vegetation data from Wisconsin and Michigan, USA.     Ecology 66(3):721-737
Calcote, R. 1995. Pollen source area and pollen productivity evidence from forest hollows. Journal of Ecology, 83:591-602
Fall, P.L. 1992. Spatial patterns of atmospheric pollen dispersal in the Colorado Rocky Mountains, USA. Review of Palaeobotany and Palynology 74:293-313
Jackson, S.T. 1990. Pollen source area and representation in small lakes of northeastern United States. Review of Palaeobotany and Palynology, 63:53-76
Jackson, S.T. and Kearsley, J.B. 1998. Quantitative representation of local forest composition in forest-floor pollen assemblages. Journal of Ecology 86:474-490
Jackson, S.T. and Wong, A. 1994. Using forest patchiness to determine pollen source areas of closed-canopy pollen assemblages. Journal of Ecology 82:89-99.
Jackson, S.T., Webb, T., Prentice, I.C., and Hansen, J.E. 1995. Exploration and calibration of pollen/vegetation relationships: a PC program for the extended R-value models. Review of Palaeobotany and Palynology 84:365-374
Jacobson, G.L., and Bradshaw, R.H.W. 1981. The selection of sites for paleovegetational studies. Quaternary Research 16:80-96.
Janssen, C.R. 1973. Local and regional pollen deposition in Birks, H.J.B. and West, R.G. eds. Quaternary Plant Ecology, Blackwell, Oxford, pp.31-42.
Kermack, K.A. and Haldane, J.B.S. 1950. Organic correlation and allometry. Biometrika, 37:30-41.
Lynch, E.A. 1996. The ability of pollen from small lakes and ponds to sense fine-scale vegetation patterns in the central rocky Mountains, USA. Review of Palaeobotany and Palynology 94:197-210
Parsons, R.W. and Prentice, I.C. 1981. Statistical approaches to R-values and the pollen-vegetation relationship. Review of Palaeobotany and Palynology 32: 127-152.
Prentice, I.C. 1985. Pollen representation, source area, and basin size: toward a unified theory of pollen analysis. Quaternary Research 23:76-86
Prentice, I.C. 1986. Forest-composition calibration of pollen data, in B.E. Berglund ed., Handbook of Holocene Palaeoecology and Palaeohydrology. John Wiley and Sons Ltd. pp. 799-816
Prentice, I.C. 1988. Records of vegetation in time and space: the principles of pollen analysis. In B. Huntley and T. Webb, Vegetation History. Kluwer Academic Publishers, pp. 17-42
Prentice, I.C., and Parsons, R.W. 1983. Maximum likelihood linear calibration of pollen spectra in terms of forest composition. Biometrics 39: 1051-1057
Prentice, I.C., and Webb, T. 1986. Pollen percentages, tree abundance and the Fagerlind effect. Journal of Quaternary Science 1(1):35-43
Prentice, I.C., Berglund, B.E., and Olsson, T. 1987. Quantitative forest-composition sensing characteristics of pollen samples from Swedish lakes. Boreas, 16:42-54
Riggs, D.S., Guarnieri, J.A., and Addelman, S. 1978. Fitting straight lines when both variables are subject to error. Life Sciences, 22:1305-1360.
Sokal, R.R., Oden, N.L., and Thomson, B.A. 1998. Local spatial autocorrelation in a biological model. Geographical Analysis, 30(4):331-354
Sugita, S. 1993. A model of pollen source area for an entire lake surface. Quaternary Research 39:239-244
Sugita, S. 1994. Pollen representation of vegetation in Quaternary sediments: theory and method in patchy vegetation. Journal of Ecology 82:881-897
Webb, T., Howe, S.E., Bradshaw, R.H.W. and Heide, K.M. 1981. Estimating plant abundance from pollen percentages: the use of regression analysis. Review of Palaeobotany and Palynology 34:269-300



About U of O | Prospective Students | Students | Services | Academics | Research | News and Events | Alumni and Friends

System requirements | Feedback | Privacy Policy | Accessibility

© University of Ottawa
If you are looking for additional information, please contact us.
Technical questions or comments about this site? Last updated: 2009.11.23