How accurate is the NWM?
Before discussing how accurate the NWM is, it is important to define what we mean by ‘accurate’ for a model such as this. This is because potential suitability for any particular woodland type is an inherently untestable quality, at least in the short term, because there is so little semi-natural woodland left in Scotland.
Potential sources of error and uncertainty are discussed below, but it should be noted that most of these are difficult to quantify. Three ways of analysing the approach and the output are to:
- Identify potential sources of error in the model development
- Examine how well the datasets predict site conditions, and how the inclusion of other data might improve the predictions
- Test the predictions against the distribution and type of existing native woodland.
Potential sources of error and uncertainty in the NWM
Errors in source data
Land cover and soils maps are, by their nature, simplifications of reality. Classes merge into each other, and boundaries can be uncertain. Some features are more accurately identified from these datasets than others, e.g. arable or alluvial soils are easier to identify than bracken land or soils with subtle drainage distinctions.
Uncertainties within the literature regarding optimum site conditions for different woodlands
Exact requirements for establishing woodland/scrub types cannot yet be defined, because the precise limits for growth and survival of many native tree and scrub species are not fully understood.
Limitations of woodland classifications
The NVC is the standard UK vegetation classification system. It is, however, based on relatively small samples, and there are some inadequacies in its application to some Scottish woodlands. Other classifications, especially that of McVean and Ratcliffe (1962) provide valuable additional information to alleviate some of these issues.
Impacts of ‘external’ pressures
Grazing and other land uses can make it difficult to compare actual vegetation on the ground to predictions for vegetation potential based only on the biophysical attributes of a site.
The NWM’s prediction of site conditions
In addition to soils, three components of climate have a strong influence on woodland growth. We considered each in turn and discussed whether their inclusion would improve the model prediction or not, as described below (see Hester et al. 2003 for more detail).
Accumulated temperature
The predictions of the NWM were found to compare very well with the limits of accumulated temperature for growth of different tree species (Pyatt & Suarez 1997), without this information being explicitly included in the model. Assessment of the upper tree limit gives almost 100% accuracy. This is to be expected, because the development of different soils requires specific moisture and temperature regimes. As a result, the NWM predicts species with demanding requirements for growth, such as oak, on soils which themselves develop under the same temperature regime. Therefore, it was concluded that inclusion of accumulated temperature data would not significantly improve the NWM output.
Moisture deficit
A comparison between the NWM and estimated soil moisture deficit limits for tree growth found that moisture deficit is not a limiting factor for growth of woodland and scrub in the Scottish uplands. Therefore, inclusion of soil moisture deficit data would not be expected to improve the model within Scotland.
Exposure
NWM predictions were compared with the Detailed Aspect Method Scoring (DAMS), a widely used measure of exposure (Dunham et al. 2000). A score of 24 is considered to be the maximum limit of forest growth (Hale et al. 1998), and the NWM also predicted no suitability for closed woodland above this threshold. Therefore, it was also concluded that the explicit inclusion of DAMS data would not significantly improve model predictions.
In summary, these tests suggest that the combined soils and land cover datasets provide a robust surrogate for the main climatic variables at the resolution of the model output; the model output would be unlikely to be significantly improved by their addition. This gives good support for the chosen approach for this model.
Validation of NWM predictions against existing native woodland
It may be expected that, if the model accurately predicts potential woodland type, a randomly selected piece of semi-natural woodland should conform to the type predicted for that area. However, it is unfortunately not that simple, as there are several reasons why the NWM’s expression of the most appropriate potential-natural woodland for a particular site may not fit with the composition of the existing woodland:
- The current woodland may have been profoundly altered by human actions. Historical management will often have altered the woodland type from that which would be expected from knowledge of the underlying soil type. For example, if oak has been selectively felled from a woodland (as often occurred in the past), then its current absence does not indicate that the soil is unsuitable for its growth.
- Alternatively, if years of grazing have led to the loss of a ‘key species’, such as ash, then the current woodland type will also not be a true representation of what the site could potentially support.
Therefore, care must be taken to understand management history when comparing current woodland with model predictions.
Despite these difficulties, direct comparisons of the NWM output with actual NVC field survey of semi-natural woods have demonstrated the model to be robust, with at least 70% accuracy for the main woodland types at the target scales of resolution (Hester et al. 2003).
D. Stone and H. Gray (unpublished data) reviewed NWM outputs for the Highlands and compared them with actual NVC surveys for three key woodland types. They assessed differences between the two data sources and the implications of such differences. Their findings, and those from other studies, are summarised below.
Pinewood
NWM predictions clearly show how the composition of pinewoods changes from the east to the west of Scotland. Stands of pure W18 dominate in the east with some W18/W17. Moving further west, mosaics with W4, scattered trees/scrub and peatland become more common, with pine an increasingly minor component, until in the far west of Scotland W17/W18 is almost the only pine-containing category predicted.
Pine woodland in the Cairngorms
Correlations between NWM predictions and NVC surveys at Rothiemurchus were very good. Of the 453.4ha of pinewoods identified in the North Rothiemurchus NVC survey, 432.3ha (95.4%) occur on sites identified by NWM as suitable for W18 alone or for W17/18.
Breckenridge (2001) also suggested that the NWM is very accurate in its predictions for pinewood. She visited 22 sites predicted as suitable for W18 or W18 mosaics, in Wester Ross, Sutherland and the central Highlands. Fourteen sites were on ancient woodland sites within the pine zone and the other eight were on randomly selected sites outwith the pine zone.
Of the 14 ancient woodland sites:
- six were currently native pinewoods
- two were largely semi-natural broadleaved woodlands with areas of pine or pinewood ground flora
- five were conifer plantations with remnant areas of pine, pinewood ground flora or other signs that the site may formerly have supported pinewood
- one was a pine plantation with no evidence of native pines, although it was adjacent to an even-aged, pure oak wood with pinewood ground flora.
Of the eight sites outwith the ‘pine zone’, a comparison with the optimal precursor vegetation in Rodwell and Patterson (1994) suggested that all had potential for W18 development to a greater or lesser degree.
Oakwoods
The most extensive stands of pure oakwoods (W11 and W17, the core NVC types) are predicted in the eastern straths of the Highlands. Further west, oakwoods are increasingly predicted as a minor component of a mosaic and, in this form, become almost ubiquitous on lower slopes and coastal fringes.
Oak woodland
D. Stone and H. Gray (unpublished data) found a fairly good correlation between NWM outputs and NVC data from Loch Sunart SAC. NVC survey shows 71% of the site to be covered by W11 and W17. Of this area, 60% is predicted by the NWM to be suitable for W11, W17 or mosaics containing these types.
Ashwoods
The potential for ashwoods is much less than for oak or pine, because the basic soils they require are not common in Scotland. It should also be noted that the resolution of the NWM is too coarse to pick up all the many small fragments of ashwoods in ravines and gorges. The only areas where the model predicts concentrations of ashwoods are Skye, Morvern, Mull, Islay, Argyll and the lowland fringes in Buchan and Caithness.
In comparisons of the model output with NVC surveys of north Mull, D. Stone and H. Gray (unpublished data) found a much poorer correlation than expected, with most sites dominated by oak. The site requirements for ashwoods (base-rich soils on sites with low-medium exposure) are fairly distinctive, and the correlation with ashwoods in other areas (e.g. Rassal) are good. The NVC surveys are of good quality so it is unlikely that the existing habitats have been misinterpreted. Further investigation suggested a likely explanation for these disparities: recent site visits indicate that young regeneration in many Mull woods is dominated by ash, rather than oak. There is widespread evidence of conversion of woodland in the western Highlands to oak during the 17th/18th Centuries (Thompson et al. 2001). Given the chance, such woods would be expected to revert over time to types more suited to local soils and exposure – in this case ashwoods, as indeed predicted by the model.
Accuracy of the NWM - Summary
When using a predictive model it is important to be aware of all caveats affecting the robustness of the predictions. However, from the tests carried out so far, it seems that the advantages of the NWM as a strategic tool far outweigh the uncertainties involved.
When tested against those components of climate which are believed to be particularly important for tree growth, the NWM has been shown to predict site conditions well. The soil and land cover datasets appear to act as robust surrogates for the climatic data considered and so it is unlikely that significant advantage would be gained by incorporating such data into the model.
Where the model has been tested against NVC data for existing woodland, excellent agreement has been found. The only major exception found so far is on Mull, where the actual woodland type (oakwood) was quite different from that predicted by the NWM as being most suitable for the soil type (ashwood). This seems to be because the woodland’s composition has been drastically altered from its natural state by human management. It will be possible to test this theory further as the woodlands continue to regenerate naturally, and other areas are restored to native woodland. It is also important to realise that the NWM is a dynamic model, with the potential to be updated when better data become available, or our understanding of woodland and scrub requirements is improved by new research.