EDA – From cakes towards maps
EDA stands for extended downscaling approach. We developed the approach to explore how plants have been arranged in landscapes in the past. EDA is a forward modelling approach that combines pollen data from a number of sites with modern landscape patterns.
Why is it useful?
Pollen data are point data. A single pollen sample from a large lake does not give any information about the spatial arrangement of taxa in the surroundings of a site – it just provides mean vegetation composition in the region and runs into the homogeneity problem. Sets of pollen records from large lakes can elucidate patterns on a large, continental scale, as recent applications using NEOTOMA and the EPD have shown (although some caveats remain). Yet, spatial resolution remains too low to answer questions raised in e.g. ecology, conservation or archaeology:
Where did species invade a landscape?
Which species co-occurred?
Where were humans active?
Samples from very small sites can provide suitable information on a much smaller spatial scale, when corrected with MARCO POLO or LOVE. These methods have their limitations as well, however, as many sites would be needed to reconstruct (past) vegetation patterns in a large area and suitable sites cannot always be found where wanted or needed.
The EDA refines vegetation reconstruction by addressing medium-scale landscape-vegetation patterns using the more widely available pollen records from large lakes and peatlands. In two earlier examples we have shown the potential of the approach in simple settings with only few species during the Lateglacial and Early Holocene. We have now implemented the EDA in R and tested whether the approach is also applicable in more complex settings.
How does it work?
The EDA combines modelling of pollen deposition with optimisation. Pollen deposition is modelled for multiple sites (lakes), for which also empiric pollen data is available. Modelling is based on a given (relatively) stable landscape pattern, e.g. the distribution of soil types. At the start, each unit (soil type) is assigned a random vegetation composition. Pollen deposition at each of the sites (lakes) is then modelled from the distance weighted abundance of each taxon, its pollen productivity and a dispersal model. Optimisation then searches for that vegetation composition on each unit (soil type), with which modelled pollen deposition best matches empiric pollen deposition at all sites. In our tests, the EDA already worked reasonably well with only 5 sites – if they cover the full environmental gradient in a landscape. For reliable results, a minimum of 20 sites is needed.
What’s the future?
Our EDA examples so far deal with just one landscape pattern – soil types – because soils are often the main driver for species patterns in flat areas like NE Germany. The situation is certainly more complex in landscapes where relief and exposition play a larger role. The EDA is flexible with respect to such additional stable landscape patterns. A combination of factors (e.g. soil types and slopes) can be implemented by feeding the model with an overlay of landscape patterns.
How can I use it?
The EDAinR function will soon be available in the DISQOVER package, contact us for urgent request.