Towards a Continental Sampling Frame for the NFI Jerry Vanclay Southern Cross University, PO Box 157, Lismore NSW 2480
[email protected] Fax (02) 6621 2669
Summary In an inventory designed to deliver information for the Montreal Process, flexibility and stakeholder confidence may be more important than questionable estimates of statistical efficiency. One candidate for a continental sampling frame could involve stratified random sampling with point samples, supplemented with existing data updated using an appropriate model. However, consistent with spirit of the Montreal Process, stakeholders should participate in formulating the sampling frame. Introduction Specification of a sampling frame involves three elements: 1. What units to sample (e.g, trees, plots, points); 2. How to select those units (e.g., randomly, subjectively); and 3. Procedures for selecting units (e.g., all at once, or successively depending on the outcome of a previous stage or phase of sampling). Many alternatives can be chosen for each of these three elements, so the number of combinations is considerable. An appropriate choice depends on the purpose of the inventory. Design criteria Objectives The stated objectives are to (see Terms of Reference): 1. Provide a responsive flexible system that can provide authoritative national-level data to meet the reporting requirements of Australia’s Montreal Process Category A indicators (plus indicators in Categories B and C where cost-effective) for all forests. 2. Produce data at a resolution equivalent to Australia’s 80 Interim Biogeographic Regions of Australia (IBRA) regions, represented at 14 Eco-regional zones and at the national level. 3. Provide a level of precision commensurate with comparable member countries participating in the Montreal Process. 4. Establish a standardised and repeatable system to conduct successive inventories and make statements about change, relevant for 5-yearly reporting. The following observations are also pertinent (notes from meeting of 9/1/01): ♦ There should be something to report despite the limited budget. ♦ There is no need for independent estimates, so that existing data and procedures could be used to achieve best result for a given outlay.
♦ The emphasis is on National estimates and on regions crossing State boundaries (e.g., Murray-Darling basin), and should not encroach on State issues Design imperatives ♦ Stakeholder support. Must have broad support amongst stakeholders including those in industry and conservation, and within Australia and abroad. ♦ Multiple-use. Must offer insights into both timber and broader conservation issues. Should as a minimum, be able to report stand structure and composition. This presumably requires, as a minimum, sampling to report a list of trees with their species and sizes – however, it would help foster stakeholder ownership if stakeholders were consulted and invited to nominate intermediate outcomes before detailed inventory design commences. Stakeholders Gaining the support of a broad group of stakeholders is critical to the Montreal Process. If a few inventory specialists develop a plan in isolation and present it to stakeholders only after it has been set in stone, it is unlikely to be well received (unlike Moses’ commandments! ). The best way to gain the support of stakeholders is to involve them in all stages of the design, from setting the objectives and priorities, through the design negotiations, right through to the peer review of results. Unresolved questions Lund (1990) posed several insightful questions that can help to focus an inventory design, and which do not appear to have been articulated by the NFI steering committee: ♦ What decisions are going to be made on the basis of this inventory? ♦ What information is really needed to make these decisions? ♦ To what area should these estimates apply? ♦ When are the results needed, and in what form? ♦ What are the constraints on funding and people to perform the inventory? ♦ What impact will information errors have on the decisions being made? ♦ What impact will incorrect decisions have on the resources? Sampling for Montreal indicators Table 1 offers a cryptic summary of selected Montreal criteria relevant to the design of the continental sampling frame. A fuller exploration of these issues is given in the Appendix. Several aspects relating to category A indicators are noteworthy ♦ Suitable data may exist for many locations and could be used to reduce the sampling effort. ♦ GIS is critical for estimating efficiently the areas required for several indicators. ♦ Spatially explicit representation of forest types and tenure classes is needed. ♦ Some indicators can be provided efficiently only by forest owners/managers.
Table 1. Summary of Montreal Category A Indicators and possible estimation approaches. Item 1.1.a/c 1.1.b/d
Description Area by forest type and tenure Area by growth stage
Method GIS GIS
1.2.a 1.2.b 2.1.a
List of forest dwelling species Status of species at risk Area excluded/available for timber production
2.1.d 2.1.d 2.1.f
Annual removal of wood products Sustainable annual production Area and proportion of new plantation adequately stocked after 1 year Area and proportion of harvested native forest regenerated adequately Area affected by processes/agents that may change health/vitality – productivity measurably impaired Area with significant soil erosion – area where soil erosion addressed Number of visits Employment in the forest sector (various) Selected entries from Categories B-C Fragmentation Area with changed biological, physical or chemical components Stream length in which stream flow deviates from historic trends Area with diminished soil organic matter or changed soil chemical properties Area with soil compaction or changed physical properties Total forest ecosystem biomass and carbon
GIS narrative GIS, modelling report modelling report
Possible sources of data forest type maps, cadastral data existing data, remote sensing, field survey existing lists, species distributions NPWS existing data, adjustments for unmappable fragments. harvest data provided by others owners, BIOCLIM owners
report
owners
GIS
infra-red imagery
report
owners
report report narrative
owner other agencies. other agencies
GIS sampling
remote sensing remote sensing, field survey
sampling, modelling sampling
stream gauges
sampling
soil samples
sampling, modelling report
stand table, canopy height, understorey, soil samples owners
2.1.g 3.1.a 4.1.a 6.2.c 6.5.a 7.1-5 1.1.e 3.1.c 4.1.c 4.1.d 4.1.e 5.1.a 6.4.a
Area managed to protect cultural and other values
soil samples
To satisfy reporting obligations under the Montreal Process, different procedures may be required to assess, for all forest lands: 1. Areas within various categories (type, tenure, management intent, etc) – use existing data, remote sensing and GIS. 2. Tree numbers and sizes, other plant species and relative density, selected site and soil properties – use existing data supplemented with additional field surveys. 3. Forest-dwelling species, both on land and in water – where existing data are inadequate, independent surveys are required since efficient inventory of trees may be incompatible with effective fauna survey (noise, time of day, season of year, etc). 4. Forest products and services, and their production, quality and use – unlike trees, most products and services have a flow, so can be assessed efficiently at points of concentration (car parks, campsites, access points, streams, weigh bridges, processing plants, etc). 5. Management objectives and systems – the effectiveness of a stated management system may be gauged in the field, but the statement of intent must be provided by the manager.
Existing Information Literature A review of literature offers some helpful insights in clarifying inventory objectives and design criteria, and offers several cautionary tales, but offers no obvious model for a sampling frame for the NFI. The alternatives canvassed in the literature remain unresolved, and all appear to have both supporters and critics. Many alternatives have been used at one time for national or regional inventories, and an distillation of current favourites and trends offers no clear guidance for our situation. Support for, and criticism of, the alternatives seems to be context dependent and all too often rather subjective. For example, Finland has changed from transects to fixed-area L-shaped plots and more recently to point samples; the USA has changed from clusters of point samples to fixedarea circular plots, while France continues with concentric circular plots. Canada and Sweden both use systematic samples on a square grid, while the USA uses a triangular grid. Canada and the USA both use double-sampling while Finland uses a more complex multistage system (Tomppo 1996). Perhaps only two clear trends emerge: ♦ During the 1970-80s, when inventories often had a single objective (e.g., timber volumes), there was a tend toward more complex statistically-efficient estimators based on fewer field data – perhaps as a response to increasing field costs and more efficient computing. ♦ During the past decade, the need for multi-resource inventories has led to simpler, more flexible designs. Several approaches that were promoted during the 1970s (e.g., SPR, 4-stage sampling) have become exceptions rather than favourites. A few selected quotes from Schreuder, Gregoire and Wood’s (1993) definitive text sum up the situation: ♦ [any requirement for] “the estimation of multiple resource characteristics … may reduce the gains in efficiency from SPR designs.” (p. 182) ♦ “The results of AIRIS [Alaska Integrated Resource Inventory System, a 4-phase sample for multiple resources] were disappointing. The effort was too ambitious for the available resources and objectives changed over time. The correlation between dependent variables of interest and covariates were not satisfactory …” (p. 191) ♦ “emphasis over time shifted from primarily estimating timber volume … to estimating a multitude of parameters, several of which are not correlated with timber volume. The need for a quick and accurate response to requests for estimates for smaller areas … also requires a simpler design. Simple designs facilitate poststratification for subpopulation estimation.” (p. 200) ♦ P.383: “Model-based sampling procedures can be quite efficient but … are less defensible in potentially controversial situations than probabilistic sampling. ” (p. 383) Nature of resource and existing data Given the constraints on the NFI, it is useful to consider the availability and utility of existing data within several categories of forest:
1. Plantations (State, private industrial and small private plantations): There are other initiatives to gather information on this resource, and these should make data available. 2. Native forest – State – Timber production: Forest Services should be able to provide data. 3. Native Forest – State – Recently converted to conservation or private use: Forest Services may still have relevant data that could be updated by simulation models. 4. Native Forest – State – Conservation: It is unlikely that much suitable data are available. 5. Native Forest – Private – Timber production: Owners are often enthusiastic, are likely to have data, and could probably be induced to make it available. 6. Native Forest – Private – Not managed for timber: Unlikely that much data will be available. Increasing interest in these forests may renew calls to quantify this resource, but owners may be suspicious and reluctant to cooperate. Thus, if existing data are to be used, much of any supplementary sampling effort may be directed at private native forests and the conservation estate. Mapping versus Sampling Some commentators make a distinction between mapping and sampling of forest attributes. However, the process of inference and mapping from satellite data is not dissimilar from ground-based sampling. Inferences based on satellite data rely on samples (ground-truth) extrapolated to other pixels deemed similar (e.g., by supervised classification). Conventional forest inventory is also extrapolated from discrete samples to strata deemed homogeneous with respect to parameters of interest. Thus, it is possible at least in principle, to prepare maps presenting any forest attribute assessed, provided that we have samples in each stratum. However, the quality of such maps will depend on the intensity of sampling and our ability to denote homogeneous strata. This caveat also applies to maps prepared from satellite data. Sampling flora and fauna Montreal indicators address both flora and fauna, but these cannot efficiently be assessed simultaneously, and separate inventory procedures are appropriate. For instance, efficient mensuration may require good light, and should involve verbal communication between team members (i.e., the booker should “calling-back” data while recording). In contrast, assessing wildlife may be best done at dawn, in silence. In addition, perennial woody vegetation can be assessed at any time (preferably during the dry season), but other plants and animals may be seasonal, and timing may be critical to obtain an accurate record. Longevity of inventory data Forest inventory data are generally durable and offer useful insights for several years – indeed for some forest attributes, for several decades. Many forest attributes change slowly, or in predictable ways, and in the absence of major disturbance (e.g., wildfire, harvest), need not be assessed frequently. Topography, forest type, site productivity may remain essentially constant for long periods, or may change too slowly to be detected readily. Trees grow in size, but in a predictable way, so that models can be used to update data. Thus forest inventory data may remain useful for a long time,
provided gross changes are monitored. Such changes may include wildfire, harvesting, dieback, etc. Costs and other constraints The timeframe (report by 2003) and funds available for the inventory pose real constraints. A national inventory with similar objectives in the USA (which has about the same land area as Australia) has an on-going budget of US$ 100 million a year, indexed for inflation (Anon2000). A corresponding effort in Denmark (with about 1/200th the area of Australia) has a budget equivalent to AU$1 million/year. An Optimal Sampling Frame? The concept of an optimal approach may be misleading, as several alternatives may deliver the required information in a way satisfactory to stakeholders, and differences in cost and statistical efficiency may be small. There is no doubt that an optimal procedures can be established for specific inventories where there is a single clear objective and reasonable estimates of population variability, travel costs and fieldwork costs are available. However, any attempt to establish an optimum may be hypothetical at best, if objectives may be refined, additional information requested, or the variability and costs remain uncertain … Optimizing the design The quality of any information depends substantially on two factors: the nature of the underlying data (viz. Number and size of plots and the way they were measured), and the way in which inferences were made (viz. Simple averages and totals versus more sophisticated techniques relying on ancillary information). Inventory designs are usually optimized by using fewer plots and making better inferences (by stratifying, using multi-stage or model-based approaches). This is easily done when an inventory has a single objective, but becomes complex when there are multiple objectives. There is no single best approach for undertaking an inventory with diverse requirements such as the Montreal criteria. It is possible, in principle, to optimize an inventory design, if we attach a weight to each of the criteria, and have prior data about the variation in the environment and the cost of inventory work (including travel, etc). However, estimates of such weights, variation and costs are likely to be rather subjective, and any bias will be reflected in the so-called optimum. In addition, an “optimal” procedure for a particular set of criteria may create obstacles to the estimation of additional parameters. Thus it may be preferable to design the sampling frame, not around a pseudo-optimum, but rather on flexibility and on stakeholder confidence. Some considerations are presented in Table 2. Sampling considerations Table 2 presents some considerations influencing the choice of inventory design. Two broad alternatives are considered, contrasting a formal and statistically defensible design (whether systematic, multi-stage, or otherwise) with a more pragmatic stratified approach allowing the use of existing data. There are two important questions about the use of existing data. 1. One relates to the possibility of bias: can we be sure that the units were not established because the forest “looked good” or were otherwise atypical of their stratum? In many cases, this may be resolved by referring to the procedures that
applied when the data was originally collected. For instance, was there an edict that plots should not be established within stream buffers and other unloggable areas? 2. The other issue relates to estimates of precision: if we know that we have the best possible estimate, do we need to know exactly how good that estimate is? Many inventory practitioners feel that the estimates of precision are largely of academic interest, and that users of the data never refer to them (e.g., Lund 1990). This question about precision is fundamental to sample design, so an example is appropriate. Suppose a formal sample told us that we probably had around about 90 whatsits, and that we could be 99% sure there were between 70 and 110. A less formal approach utilizing more data might tell us that we could be pretty certain there was 99 whatsits, but be unable to give us reliable confidence limits (“perhaps 96-102 but don’t count on it”). Which is the most useful result? Many managers happily forgo the confidence limits in order to secure a better estimate… Table 2. Some considerations on sampling design. Formal sampling design Adaptive, using existing data Technical implications: Limited ability to use existing data – thus budget restrictions may mean fewer data and a weaker basis for inferences, so estimates may be limited by data rather than technique.
Use of existing data means that funds can be directed to under-sampled areas, so that this approach is likely to provide the best possible database.
Enables reliable estimates of precision (“We’ll Estimates of precision will be approximate know precisely how good – or bad – our (“Our estimate may be the best, but we won’t estimate really is”). know for sure just how good it is”). Should allow the strongest inferences, producing the best possible estimates from a given database, at least for certain classes of indicators.
Data may restrict the way inferences can be drawn, limiting the use of statistically efficient estimators – trade-off specialist performance to get a good “all-rounder”.
For a given number of data, specially collected For a given budget, this may offer the best allfor a single specific purpose, this is round performance, especially if existing data undoubtedly the best approach. holdings are substantial.
Political and procedural implications: Estimates may deteriorate if disaggregated below designed resolution. Estimates may contradict estimates provided by other agencies – correctly or otherwise.
Quality of estimates always dependent on data used and strata assumed. Shared data implies compatible estimates or reveals a need to check estimation procedures
Independent data and procedures, independently managed.
Sharing data raises issues about access, confidentiality, divergent copies, etc.
Suggested Sampling Frame The considerations expressed above lead to a recommendation of one possible sampling frame that could be presented to stakeholders for their endorsement. Sampling Frame In theory, sampling should embrace three tenets:
1. an unbiased estimate (if we take enough samples we’ll eventually discover the “truth”), 2. a precise estimate (we want consistent and repeatable results, even with few samples), and 3. an ability to assess the precision of the estimate (even though we’ll never know what the “truth” really is, we’d like to judge how far from it we’re likely to be). However, precision is the “Holy Grail”, and sacrifices are commonly made to attain it (e.g., ridge regression introduces bias to improve precision; systematic sampling compromises estimates of precision – and risks bias – in an attempt to improve precision). The art of defining a sampling frame is to find the most efficient compromise between cost and precision, without too many compromises on bias and assessing precision. Sampling Units Because the Montreal Process addresses biodiversity (e.g., Criterion 3.1) and biomass (5.1), inventory should record a list of trees species and sizes occurring within a stand, even if these data are used only to compute summary statistics. Thus the sampling unit should not be the tree (or organism), but rather a point or plot within a forest stand. Whilst plots and point samples are equally capable of providing tree lists, nominating a reference point as the sampling unit may offer greater flexibility for non-tree parameters. This should be compatible with the used of concentric plots (including point samples), nearest neighbour and other approaches that may be considered for sampling non-tree aspects of the environment. Point samples are statistically efficient because they sample trees with probability proportional to size, thus spreading the sampling effort more evenly between large and small trees. Point samples are most efficient when the probability of sampling is such that 4-10 trees tend to be sampled at each point. Such small samples may be adequate in a monospecific plantation, but clearly clusters of point samples are necessary to get meaningful tree lists in rainforests and other species-rich forests. Further gains in efficiency may be gained through censorship (e.g., the common practice of sampling only trees larger than a specified size; and the Finnish practice of imposing a maximum radius of 12.45 metres; Tomppo 1996) and through supplementary sampling (e.g., adding a small concentric fixed-radius plot to sample smaller plants). However, these efficiencies may be less compelling than the need to deliver in a timely and costeffective manner. Accordingly, I advocate a broad definition of sampling unit to encompass any plot- or point- based approach that is repeatable (but not necessarily permanent), georeferenced, and provides a list of tree species, sizes and relative abundance (so that corresponding stems/ha can be inferred). This is consistent with Canadian procedures proposed for Montreal Process inventory. Selection of Units If estimates of precision are needed, sampling units should be selected objectively, perhaps randomly, perhaps according to the size or importance of the unit, or according to the variability in the (sub-)population. If estimates of precision are not required, units may also be located subjectively: deliberate placement of a unit to represent a stratum can improve precision (but obstruct the ability is estimate that precision), but deliberate
and consistent placement of sampling units in the “best” or “worst” parts of a stratum will introduce bias that may not easily be detected. For a single objective (e.g., timber volumes), there are efficiencies to be gained by sampling with probability proportional to the expected volumes and/or with proportional to the variability. However, this may confer no advantage for multiple objectives, unless the various parameters of interest are correlated. Given the range of Montreal criteria, and the likelihood that procedures for assessing these may change, simple random sampling within defined strata offers greatest flexibility and seems expedient. This is consistent with Canadian procedures, except that the latter restricts random sampling to a 4x4 km grid. Sampling Procedures If it was 1788 and we were commencing the first forest inventory in terra nullis, a systematic inventory would be appropriate. Today however, we know much about the forest estate, and know that it varies greatly in nature and extent in different parts of Australia. Because of this variation, great efficiencies can be made by taking into account prior information about our forests, and tailoring the nature and intensity of the sampling to the characteristics within each stratum. Thus we are well placed to heed Schumacher and Chapman’s (1954) observation that “precision may be gained by dividing the population into as many blocks as expedient, even though the number of random sampling units taken from each may be the minimum of two”. In most cases, we may further gain precision by dividing the population into homogenous strata using prior information, rather than simply using geometric blocks. Thus I advocate the two-stage approach of stratified random sampling. Additional stages or phases (i.e., multi-stage and multi-phase sampling) appear to offer few advantages and significant disadvantages (notably, loss of flexibility) because of the multiple objectives involved and the likelihood of low correlation between the several variables of interest. This suggestion departs from trends in the USA and Canada, where systematic samples have been adopted. How many strata? More strata will provide better estimates, but will require more samples and involve greater cost. Existing data may provide estimates for some strata, but even a modest number of strata will require the establishment of many new plots. Thus there should be the maximum number of strata that can sensibly be created with existing data and the funds available for fieldwork. Existing data The approach advocated does not preclude the use of existing data, provided that ♦ Plot establishment and measurement was consistent with guidelines prevailing at the time, and those guidelines remain compatible with currently prescribed procedures; ♦ Plots are georeferenced sufficiently well that the stratum to which they belong can be established unambiguously; ♦ It is reasonable to assume that data remain representative of that stratum and were not deliberately located in “good forest” or other atypical parts of the stratum; ♦ No major changes (fire, storm, logging, dieback) have been observed within the stratum since last measurement; and ♦ A suitable model exists and can be used to update the data to account for growth, etc.
Additional issues arise if many data exist in a stratum, particularly if they derive from more than one point in time, or if some appear atypical of the stratum. However, some guidelines can be offered. The enduring principle is to obtain the most realistic representation of the conditions within a stratum. At least two data are needed within each stratum to allow estimates of the mean and the variance, but there is no upper limit on the number of data within a stratum. If there are many data within a stratum, it is reasonable to discard the oldest data (if data derive from more than one date), but not to discard data on the basis of the variable of interest (e.g., highest or lowest volume), as that would bias estimates of both the mean and its precision. Each datum, even if “atypical”, is representative of part of the stratum, if obtained in a way consistent with established procedures. It is OK to discard the full set of data obtained at one time, but individual data may only be discarded if there is strong evidence of measurement error or incorrect procedure, no matter how unrepresentative they may appear. Further stratification is a more satisfactory way to deal with apparently unrepresentative data. The danger of introducing bias through the use of data derived from plots not randomly located is small, especially within a stratified approach. If strata are reasonably homogeneous with respect to the parameters of interest, then there is little scope to distort estimates by deliberately locating plots in the “best” (or “worst”) parts of a stratum. If a stratum is heterogeneous, and large variation is confirmed by data at hand, then an efficient solution is further stratification, possible when there are sufficient georeferenced data. Thus a problem may arise only when a stratum is heterogeneous and all existing data occur at one extreme of the conditions represented, a situation that should be evident if means and variances are compared with those from comparable strata. Data management The sampling frame suggested above raises several data management issues, especially if existing data are to be used. Three kinds of data are involved: 1. Spatial data, tessellating forest land into strata (i.e., so that every scrap of forest land is in one and only one stratum), and allowing georeferenced samples to be assigned to strata. This should probably take the form of a computerized GIS, but could have been paper-based in former times. 2. Attribute data, documenting the attributes of each stratum (tenure, forest type, area, land use history, dates of major changes such as fires and harvesting, etc). This could take the form of a database integrated with, or independent of the GIS above, provided that identifiers in the spatial and attribute databases have a one-to-one correspondence. 3. Survey data, including details of tree species and sizes, and other parameters recorded on sample units during field surveys. This can be integrated with, or independent of the attribute database above, provided that the attribute database includes pointers to survey data belonging to each stratum. Much of the necessary data will already exist in the hands of forest owners and managers, and some care will be needed to use this information efficiently while avoiding unnecessary duplication of effort. In addition, questions need to be resolved for each of these data sets about ownership and confidentiality; read and write permission; maintaining master, duplicates and any summary files; how to prevent divergent copies, and so on.
Examples of Similar Systems Although statistical purists may be reluctant to support the approach advocated, it is workable, can rapidly deliver a cost-efficient outcome, and has been demonstrated elsewhere. Two examples of this approach are offered. Queensland’s Native Forest Inventory System The Queensland Department of Forestry adopted its Native Forest Inventory System following a major review of resource estimation and yield prediction in the mid-1980s (Vanclay et al 1987, Vanclay 1990), and the general approach has remained in use since, with its successor agencies. All native forests were tessellated in a hierarchy involving administrative and statistical strata (essentially District, Logging Area, Management Unit, homogeneous subunit). The spatial representation of each subunit was maintained on paper in the District office and a photocopy was provided to Resources Branch in Head Office. In the Queensland system, subunits were contiguous, but this need not be mandatory. The attribute database was maintained on PC in the District Office, using standard software developed by Resources Branch and implemented in Clipper (a DBase lookalike, compiled, and able to restrict modifications by Districts). Districts provided a new copy of the attribute database to Resources Branch after each update. Resources Branch created a database of all georeferenced field inventory data collected during the previous 20 years, and worked with Districts to identify the subunits to which these data belonged, and implemented growth models able to update plot data on demand. These data included 1-acre rectangular, 0.5-ha circular plots, and point samples established with optical wedges. Districts progressively obtained new field inventory for unsampled subunits, for units in which major changes had occurred, and for subunits for which they felt current estimates were unreliable. Inventory was obtained in conjunction with other field activities to minimize travel costs. The data was sent to Resources Branch for data entry and inclusion in the database. Resource estimates were prepared assuming stratified random sampling, and were used in operational planning at the District level, in setting the annual allowable cut, as well as in discussions leading to the Wet Tropics World Heritage Area and the South-East Queensland Regional Forest Agreement. The system worked because of the partnership and trust that was developed between the Districts and Resources Branch. Resources Branch adopted the role of a service agency, providing information useful to Districts that could be used in operational planning, as well as satisfying strategic and other information needs in Head Office. The system also drew on the local knowledge available at District level, and gave Districts a sense of ownership of data and involved them in delivery of information. Canada’s Forest Inventory Until recently, Canada’s National Forest Inventory was compiled by collating existing management-level information provided by each of the Provinces. This follows from the fact that the provinces have the jurisdiction for managing and monitoring forest lands and forest resources. The federal government does not conduct a national forest
inventory, but relies on available data for national and international reporting. Canada’s National Forest Inventory summarizes and reports these data for 47,000 cells assigned within a heirarchy of ecological strata (15 ecozones, 194 ecoregions, 1030 ecodistricts; Hirvonen and Lowe 1996; cf. Australia’s 14 ecozones and 80 IBRA regions) in a national database known as CanFI (Anon 1999a). The Inventory has been conducted in this way every five years since (and including) 1986, however, the principal objective was reporting current status rather than revealing changes. The approach was not very effective at revealing change as it was not a true time-series. Users were advised that mathematical differences between successive inventories seldom reflected real change during the 5-year period. Increasing demands for additional forest resource attributes, for policy, national and international reporting, and for reports on indicators of sustainable forest management led to the development of a new approach A new version of the NFI (Anon 1999b) is based on a systematic multi-stage 1% sample involving 50 plots within each ecotypes. The use of 50 plots assumes that volume (and presumably other variables of interest) will have a coefficient of variation of 70%, and that a sampling error of 10% will yield a satisfactory precision. A range of plot types is envisaged (fixed or variable in area) with the objective of sampling 30 trees/plot (with the proviso that no major component of a ground plot should inventory less than 100 m2). The majority of plots will be permanent plots re-measured on a 10-year cycle, but only when significant change has been revealed by management records or remote sensing. The new approach forgoes precision in the estimation of the current resource in order to obtain better estimates of change. In doing so, they also obtain better estimates of precision, even though the precision itself may well be worse. They have not specifically addressed the issue of permanent plot bias. The first national summary based on this design is anticipated in 2005. Limitations of the approach The approach suggested involves several limitations: Statistical rigor Statistical purists will object to the lack of statistical rigor, and point out that the lax definition of a sampling unit introduces uncontrolled variation, that the use of units not randomly located allows bias and obstructs estimates of precision, and that the decision to avoid multi-stage sampling lacks efficiency. Pragmatic practitioners will counter these claims with the view that gains in flexibility and practicality outweigh the slight loss of efficiency, that any bias will be negligible if stratification is good, that estimates of precision do not need to be exact, and that any uncontrolled variation due to choice of sample units is likely to be less than errors from “trying to fit a square peg in a round hole”. I know of no way to objectively weigh up tradeoffs in statistical rigor versus practicality in application, and merely suggest that beauty is in the eye of the beholder. Bias It is likely that any bias will be small, but it is impossible to rule out the possibility that the use of some existing data could introduce bias into the estimates. Any bias may be reduced by stratifying to the maximum extent possible, not by post-stratification of the data in question, but by stratifying with independent prior data (maps, remote sensing, tenure, topography, etc.).
Precision The use of existing data means that estimates of precision will be approximate, unless all sample units were randomly located, sufficiently recently that no updating is required. Estimates of precision will probably tend toward underestimates (i.e., the sampling error may be bigger than we think) if many data were updated with models, if many units were deliberately located to be “typical” of the stratum, or if post-stratification was done on the basis of the data themselves. Estimates of precision may tend towards overestimates if units were deliberately placed to sample extremes within a stratum. If a precise knowledge of precision is critical, then bounds can be placed on precision estimates. Generally, the estimate obtained with the usual formula (for stratified random sampling) will tend to be a lower bound (for a more conservative lower bound, censor the greatest outlier in each stratum with >2 units). An estimate of precision calculated assuming simple random sampling without stratification (and based on data not updated) provides a reasonable upper bound (except in the rare case where within-strata variation exceeds the between-strata variation, a situation more likely to arise in pedagogy than in practice). Change detection The approach suggested should provide the best possible estimates of the current situation. The approach is repeatable, and could be repeated in the same way in five years or so. If there are big changes in our forests during that time, it is likely that the approach would reveal those changes adequately. However, if there are no changes, or changes are small, it is possible that the approach advocated may not reflect the (absence of) change correctly. In this situation, at least in theory, estimates of precision should alert users that changes are not significant and may not be real. However, estimates of precision with this method will be approximate, and may not alert users to spurious apparent change. If change detection is critical, there are two alternatives. One is to use remeasured permanent plots. This works well, provided that forest users (and organisms) do not become aware of these “special plots” and modify their behaviour in the vicinity of the plots. Sampling with partial replacement (SPR) addresses some of the problems of “modified behaviour”, but is apparently less effective when multiple resource characteristics are to be estimated (Schreuder et al. 1993). The other alternative is to use the approach advocated, with care in the selection of sampling units when the inventory is repeated. By resampling only strata subject to gross change, and elsewhere using the same data as previously, updated with a model, it should be possible to get consistent estimates of change. Like SPR, this procedure is relatively straightforward when applied over one reassessment period, but becomes complex (even unworkable) if required for a long sequence spanning several reassessments. Moving Forward Stakeholder support is important for the success of the Montreal Process, and such support may be forthcoming only if stakeholders have a real opportunity to participate in the formulation of the sampling frame. Thus I suggest that the next step should
involve a workshop to which a broad group of influential stakeholders are invited, and at which they have the opportunity to consider this proposal, and hear a range of alternatives promoted and criticized. Careful facilitation could help a general consensus emerge. Selected References Anon., 1999a. Canada’s National Forest Inventory. Anon., 1999b. A Plot-based National Forest Inventory Design For Canada. Anon., c. 2000. Strategic Plan for Forest Inventory and Monitoring. http://www.srsfia.usfs.msstate.edu/wo/strategy.htm (7 Feb 2001). Hirvonen, H. and J.J. Lowe, 1996. Integration of Canada’s Forestry Inventory with the National Ecological Framework for State of Sustainability Reporting. In: R. Päivinen, J. Vanclay & S. Miina (eds) New Thrusts in Forest Inventory. EFI Proceedings No 7, pp. 11-26. Lund, H.G., 1990. The platonic verses and inventory objectives. In: V.J. LaBau and T. Cunia (eds) State-of-the-art Methodology of Forestry Inventory. USDA For Serv Gen Tech Rep PNW-GTR-263. Pp. 1-7. Schumacher, F.X. and R.A. Chapman, 1954. Sampling Methods in Forestry and Range Management. Duke University School of Forestry, Bulletin 7 (revised). Schreuder, H.T., T.G. Gregoire and G.B. Wood, 1993. Sampling Methods for Multiresource Forest Inventory. Wiley, NY. Tomppo, E., 1996. Multi-source National Forest Inventory of Finland. In: R. Päivinen, J. Vanclay & S. Miina (eds) New Thrusts in Forest Inventory. EFI Proceedings No 7, pp. 27-41. Vanclay, J.K., 1990. Design and implementation of a state-of-the-art inventory and forecasting system for indigenous forests. In: H.G. Lund and G. Preto (eds) Global Natural Resource Monitoring and Assessment: Preparing for the 21st century, Proceedings of the international conference and workshop, Sept 24-30, 1989, Venice, Italy. American Society for Photogrammetry and Remote Sensing, Bethesda, USA, p. 1072-1078. Vanclay, J.K., N.B. Henry, B.L. McCormack and R.A. Preston, 1987. Report of the Native Forest Resources Task Force. Queensland Department of Forestry.
Appendix: Overview of selected Montreal Indicators Category A 1.1.a/c Area by forest type and tenure – either estimate with GIS using existing forest type maps overlaid with cadastral data and accept risks of misalignment and slivers, or survey forest owners and accept risk of missing data and non-additivity. 1.1.b/d Area by forest type, growth stage, and tenure – existing data should be available for many sites, and could be inferred for other sites from remote sensing. 1.2.a List of forest dwelling species – probably not very informative at the continental scale, but could be presented as GIS overlays of species distribution maps that exist or are commissioned independently. 1.2.b Status of species at risk – since this may be determined by legislation, it poses reporting rather than sampling obligations. 2.1.a Areas of forest land excluded from and available for timber production – use existing data, supplemented by GIS-based estimates that may include adjustments to account for inaccessible forest not readily mapped. 2.1.d Annual removal of wood products compared to sustainable production – report harvest data gathered by others; cross-check and extrapolate existing estimates of sustained yield using environmental data (e.g., BIOCLIM). 2.1.f Area and proportion of new plantation adequately stocked after one year – What is a plantation? What is adequate? Is self-reporting acceptable or is an independent estimate needed? 2.1.g Area and proportion of harvested native forest that has regenerated adequately – What is native? What is adequate? Is self-reporting acceptable or is an independent estimate needed? 3.1.a Area of forest affected by processes/agents that may change ecosystem health/vitality – All of our forests are affected by anthropogenic activities and emissions directly or indirectly. What kind of processes and agents, what magnitude of change, and what measure of health/vitality do we consider? I propose that we redefine this criteria, at least for the purposes of category A, as “Area of forest in which productivity is measurably impaired” and we assess it using infra-red satellite imagery only in areas considered or reported to be at risk. 4.1.a Area of forest land with significant soil erosion – The interim measure is the converse (Area where soil erosion has been addressed) and may cause confusion. Both are hard to measure. Canada will attempt to estimate this indicator using air photos to systematically sample 1% of their forest. Why not explore other alternatives such as sediment loads in major streams draining forested catchments? 6.2.c Number of visits – report data collected by other agencies. 6.5.a Employment in the forest sector – report data collected by other agencies. 7.1-5 Narratives, not addressed through sampling. Selected Category B and C indicators that may require sampling 1.1.e Fragmentation – estimate using GIS.
1.3.a Genetic variation in forest dwelling species – cannot measure this without exacerbating indicator 3.1.a 3.1.c Area of forest land with diminished or improved biological, physical or chemical components – exotic species (plant, animal or fungal), site preparation, fertilizer … let’s just say all the forest! 4.1.c Stream length in which stream flow deviates from historic trends – could probably address this by modelling. This indicator seems to be a “blunt instrument”, especially if past trends highly variable. Seems likely that the intent is to target dry-season base flow, and the nature of storm-related peak flows; if so, a more direct indicator may be warranted. 4.1.d Area of forest land with diminished soil organic matter or changed soil chemical properties – This may involve sampling that could be done in conjunction with NFI. 4.1.e Area of forest land with soil compaction or changed physical properties – a may involve sampling that could be done in conjunction with NFI, especially if it can be assessed simply with a probe. 5.1.a Total forest ecosystem biomass and carbon – could estimate from a stand table (stem diameters and numbers), especially if details of canopy height and understorey are also included. 6.4.a Forest area managed to protect cultural and other values – Probably needs advice from managers.