Section 12 Discussion

In the original proposal we envisaged that WPA would be based on ground-based data that were already available, whilst WPB would provide new data based on remote sensing. We proposed to compare land-use change estimates based on these individually, and combined. We anticipated WP-B would produce at least two data sets which provide reliable, high-resolution data on land-use change, with the detail to provide the Beta matrices at least every few years back to 1990. However, based on the analysis of the false positive rate in previous sections, land-use change from the remotely sensed data appears to be wrong 80-90 % of the time. Furthermore, the new woodland maps gave no information on the Beta matrices, and did not provide a plausible time series for new planting/deforestation, compared with the existing FC statistics. Basing an analysis solely on the to WPB data taken at face value therefore seems not very useful, and we restrict the comparison to WPA versus the combination of WPA and WPB, where the latter are corrected for estimated false positive rates using ground-based data (denoted WPAB).

In the results of WP-A, the data assimilation estimates remain close to the Countryside Survey data because these are the only data that fully specify the beta matrix. The decadal-scale trends in these data remain clear in the assimilated estimates. The Agricultural Census data only specify the net change, and there is considerable year-to-year noise in the time series. The assimilation algorithm effectively smooths out a lot of this noise, but follows the general trend. The Forestry Commission data for afforestation and deforestation are followed quite closely because these are specified with high precision.

Contrasting these with results from WP-AB, the main effect is further smoothing of the time series. This is because it is now a weighted average of several more data sets, and these additional datasets do not show a strong coherent pattern. Particularly, these datasets do not show the sharp decadal trends seen in the CS data, so this is smoothed out. All of these data sets have uncertainties considerably larger that the Agricultural Census data, so their weighting is relatively low. The overall effect of this is that the WP-AB estimates are smoother, but the effect on the general pattern and the absolute magnitude of change is relatively small.

In terms of assessing which of these combinations should be used in the inventory, the answer is not entirely clear. Using only WPA data is obviously simpler, as it requires fewer data sources. Given the analysis of the false positive rates in the remotely sensed data, it is clear these are dominated by spurious differences, and most of the apparent change is simply noise. It is therefore questionable whether there is any real value in including these. The argument for including these data is that it is plausible that they do include some true signal as well as noise – their accuracy in detecting change is low but not zero. Given that we can account for the false positive rate and down-weight those datasets with highest random uncertainty, it makes some sense to use all the information that we have available. The practical effect is to smooth out features in the WPA data which we know to be artefacts (the sharp decadal trends in CS data), so given that the analysis has been done already, there is no immediate problem with using the additional data.

One assumption that is implicit is that the biases in the WPB data are fixed, known constants, so we can reliably correct the data with the calculated false positive rates. The false positive rates are almost certainly not fixed, but quite how they vary, and whether this is systematic is hard to know. Further work would be needed to estimate this. A more elegant approach would be to include the false positive rates as parameters to be estimated in the data assimilation algorithm, at the cost of increasing the complexity and computation time.

Whilst we have improved the representation of the CS data by using linear interpolation rather than decadal step changes, this still leaves an artefact in the data (a different linear trend each decade). This could be improved with a smoothing routine (e.g. LOESS or GAM) which would give a better approximation to what we expect is the true pattern of change over time. The same point can be applied to the Agricultural Census data, where much of the year-to-year noise seems implausible, and smoothing this out to some degree would seem more realistic.

The data assimilation routine still gives relatively high weight to the CS data, based on the unique way it monitors change. However, exactly how much uncertainty is introduced when extrapolating from the sample 1-km squares to national scale is unclear, and the methodology for doing this is now rather obscure. Revisiting this, with an appropriate method for quantifying the uncertainty in national-scale estimates, would be worthwhile if these data remain central to estimates of change.

One effect of introducing the WPB datasets is that the spatial attribution of land-use change becomes a more confused picture: we have a number of data sources which give conflicting (and largely erroneous) information on where land use has changed. The effect of this on the time series of maps produced by the data assimilation algorithm is to fragment the distribution of land use types. For example, if we used only LCM as the basis of past change spatially, the algorithm would reproduce the spatial pattern contained in LCM data. If we mix several inconsistent data sets, we get a change in the spatial pattern which is largely incoherent. More sophisticated techniques for combining the spatial data are required to address this.