Logistical considerations
3.6.5 Impact of sampling intensity on scheme performance A set of sample coordinates for each of the sampling options at each of the sampling
intensities were generated. However, it is not intended that these are taken as actual sampling sites for a preferred sampling scheme selected by UKSIC for the following reasons:
• Because of the logistical and computational demands of this task, three sampling intensities were selected to test and compare the different design options. Having agreed on a general strategy, these sampling intensities should be adjusted to the exact criteria required from any preferred sampling scheme (for example to 3500 points, or 4250 rather than 4000 exactly).
• The list of Land Cover classes used in each country was determined in part by what data was available. This should not necessarily prevent the inclusion of other classes (land uses), nor indeed the exclusion of some classes to improve estimation in others, if there are clear policy reasons for these choices. These would require revising the sampling intensities and locations accordingly in the preferred sampling option.
• Similarly, UKSIC may consider that some relatively infrequent land uses are important for specific reasons and therefore essential in a Tier 1 network. Further adjustment of sample allocation would be needed to improve the precision with which they could be monitored. By the same token, there could be scope to reduce effort in some of the larger classes. This could be judged using additional data, but will depend on UKSIC priorities. These would require revising the sampling intensities and locations accordingly in the preferred sampling option. It is worth noting that our results suggest that the precision of estimates (design- or model-based) across any reporting unit decreases rapidly when there are fewer than 100 observations within that unit.
• The division of effort between England, Scotland, Wales and Northern Ireland was proportional to land area in the selected classes.
between countries, ensuring that local requirements for information on particular classes is met, or if different levels of funding are available in the different administrations.
• Some supplementary sampling could be accommodated for certain indicators, over and above that required for the canary indicators, should this be deemed necessary.
Workshop Box 15: Further discussion of design options
Our findings are in line with other published research, in that estimation of means of large areas from model-based sampling from a grid is generally less efficient than design-based estimation from stratified random sampling. Complete sampling of a regular grid is not suitable for Tier 1 soil monitoring because of the requirement to provide estimates of status and change in soil indicators for a range of land uses within countries (the reporting units). The size and spatial pattern of these reporting units differ greatly and each may only be sampled adequately if the regular grid is distorted as in our study (see, for example, Figure 3.1), or if the grid is incompletely sampled. Our results suggest that the precision of estimates (design- or model- based) across any reporting unit decreases rapidly for less than 100 observations within that reporting unit (Fig 3.11). Therefore even further distortion of the regular grid would be required to achieve adequate sampling of all reporting units.
If a completely new monitoring scheme were to be set up, we would recommend a stratified random sampling scheme in which stratification is used to ensure the best spatial coverage possible and is consistent with adequate sampling of all classes.
The exact sample sizes required would depend on acceptable tolerances for specific reporting requirements and their action levels. Although varying the sampling intensity to achieve adequate sample sizes within all reporting classes achieves the principal design, there are also disadvantages of such schemes. First, it should be noted that if sampling is not in proportion to area, design-based estimation of reporting classes that have not been included in the design (secondary reporting classes, for example soil type) will be more complicated. Second, the number of sampling points lying within a secondary reporting class will depend not just on the area of this class but also on the distribution of this area across the primary reporting classes. If the monitoring scheme is required to allow adequate estimation for secondary reporting classes, more samples will be required than the minimum needed solely to give adequate estimation in the primary reporting classes. By incorporating peaty/non peaty soils types into the design phase, we demonstrated that the inclusion of secondary information can also improve the ultimate sampling design but this will always be dependent on the spatial coverage and quality of pre- existing information.
Similarly, the initial design needs to make some allowance for potential future changes in land classes at sampling locations and for potential future unavailability of the initial sampling locations, although these latter points apply to all possible schemes. We recommend a minimum contingency of 20 per cent, however this should be assessed when the specifications for each country are finalised as individual countries may require greater contingencies if there are significant requirements for secondary reporting or interpretation.