4.2.1 National forest inventories
Most countries hold at least some NFI data that can be used to support emissions estimation for REDD+. Well-designed NFIs are based on probabilistic samples with well-understood statistical properties which helps estimation of confidence intervals. NFIs are a valuable source of information for emissions and removals estimation, particularly with respect to above-ground biomass, and by extension below ground biomass. Though traditionally established for forest resource assessment, most NFIs, (often in close collaboration with forest research institutions) also gather information on ecosystem-related variables, and through field-based interviews may help provide information on drivers of forest change. Participation in an NFI provides excellent experience with the challenges and practicalities of forest monitoring, and NFI field experience is extremely useful in understanding the relationship between ground-based and remotely sensed data.
As well as biomass, NFIs increasingly include the dead wood pool, and some have started to acquire information on soil organic carbon and litter, although measuring temporal change in these pools is challenging. Where the sampling design is suitable (or can be augmented) NFIs can be used to estimate REDD+ activities directly. Nevertheless:
- existing NFI sampling designs are unlikely to be optimized to estimate REDD+ activities such as deforestation or forest degradation, which thus increases uncertainties in estimating emissions and removals, and could require augmentation of the sampling as discussed below.
- although NFI sample plots are usually geo-located, and can provide useful indication of where sampling should be intensified, they generally do not deliver information sufficient(1) to track REDD+ drivers or to direct policy responses to deforestation or degradation.
- although an NFI for an entire country might be desirable, it is often logistically complex and expensive in large countries, especially those with large areas of non-commercial forest. It may also take 10 years or more to establish a complete NFI time series. Alternatives to estimating change during this period need to be considered when designing a NFI based system to monitor and estimate the GHG outcomes of REDD+ activities.
Tropical forests differ significantly from temperate forests in terms of diversity of tree species, the presence of very large trees, and the rate of recovery after forest disturbance. This makes it more challenging to estimate forest biomass, and change in biomass, across spatial scales ranging from local to landscape, region and national. Forest inventory experience is much less in the tropics than in temperate forests (Burslem & Ledo, 2015; Saatchi, 2015). These authors in their review of inventory experience in tropical forests provide useful guidance relating to the design and conduct of inventories for biomass, coarse wood debris and soil carbon. They highlight the limitations of earlier inventories and identify pitfalls to be avoided. They stress that approaches need to be adequate to deal with the very high spatial variability in mature forests and those recently disturbed by logging or fire. Young regenerating natural forests and plantations contain less biomass and are more homogeneous, thus simplifying the inventory task. The reliability of biomass estimates is lower at finer spatial scales.
NFI data directly estimate carbon stock change and can also be used to support the gain-loss method. Firstly, observations of biomass and carbon change on NFI plots between points in time can be used to estimate emission and removal factors, or help develop Tier 3 models of forest growth, debris and soil carbon. Secondly, under appropriate sampling designs, NFI plot-level land use and land-use change data can provide estimates of areas of particular land-use change categories. Thirdly, where models are used to enhance estimation of REDD+ activities, NFIs plus existing data can be used in model establishment and model verification.
Typically NFIs consist of an array of plots (or clusters of sub-plots) established in a systematic fashion across entire countries. Plot size is generally in the range 0.01 to 1 ha. For tropical inventories plots may need to be as large as 1 ha to reduce potentially large variance due to large variability rather than increasing the number of plots and associated costs. Larger plot sizes also help make the link to remotely sensed data. Observations and measurements on these plots vary but always include amount of forest cover and sufficient tree-level data such as species, diameter, and height which can be used with allometric models to predict volumes and biomass of individual trees (Lawrence et al., 2010). The tree-level predictions are aggregated to estimate plot-level tree volume or biomass and carbon stock. In addition, NFIs often acquire data on tree and shrub species diversity and general topography. Less commonly, observations or measurements will also include aspects of litter and other dead material, site history(2), soil, and canopy characteristics. These NFI data are typically used to estimate forest population parameters – including production or development related - at a precision considered relevant to national level planning, taking cost and NFI practice into account.
When measurements on the same plots are obtained at multiple points in time, annual change (and associated carbon change) can be estimated for each plot(3). The timing of plot re-measurements within an NFI varies from only a couple of years in fast growing environments to 5-10 years in slower growing environments. Frequency may be less for environments that are more expensive to access and measure or for forests with low commercial value. A proportion of all plots may be measured each year so that the entire system is measured over a 5-10 year period. In an interpenetrating panel system, plots measured in any particular year (a panel) are systematically intermixed with plots measured in other years (panels) so that estimates for the entire area may be obtained each year. Heikkinen et al., 2012, describe methods for making more precise estimates using panel data and other data.
NFIs commonly use probability sampling in the form of simple random, systematic, or stratified random sampling designs. Probability sampling requires that each potential plot location has a probability greater than zero of being selected for the sample and that a randomization scheme is used to actually select the sample. The resulting data are used with unbiased estimators to calculate estimates of totals, changes and variances. Estimates for sub-sets of the original forest area are possible if sufficient plots can be grouped into domains or strata and all points within the domain have a probability greater than zero for inclusion in the original sample. The number of plots required depends on variability of the population, the precision required, and the need to estimate rare events, such as deforestation. Increases or decreases in the area considered forest could violate design-based sampling principles and thus compromise the unbiased nature to the estimators. This problem may be avoided by expanding the NFI design to other land use types, and unless this is done the NFI will not be sensitive to afforestation or reforestation, and will always detect loss of forest area.
If the NFI plots were distributed using a systematic grid, the same grid spacing can be used to extend the sample into areas that were not included in the original NFI (e.g., to include forests on privately managed land or within land classified as crop- or grazing land or, settlements where they meet the adopted definition of forest). Similarly, intensification (increasing the number of plots per unit area) can be implemented to improve estimates in areas of particular interest (e.g. where change (deforestation or degradation) is happening or likely to happen.) In addition, estimates based on data from independent probability sampling designs can be combined to produce more precise estimates. If the boundaries of areas of interest change over time, it can become very complicated with repeated measurements to manage selection probabilities of plots in dynamic strata. For this reason if pre-stratification of the study area for variance reduction is performed, the stratum boundaries should be defined by features that do not change such as ecoregions, topography, or climate zones, or well-defined socioeconomic factors such as access to infrastructure (Box 20).
Where NFI data are (or can be) grouped according to strata being used for REDD+ estimation they are likely to be valuable sources of data to estimate emissions/removals factors for REDD+ activities, or to develop Tier 3 models of forest growth, debris and soil carbon. If the land area associated with the NFI does not correspond spatially with the area of land to which the MRV is meant to apply, or if the NFI is not well-designed, the use of NFI data for the MRV could be called into question. In these cases, it might be more appropriate to use the NFI data for calibration and verification of remote sensing maps, or other estimation procedures like model-based estimation, in which case the data may be used purposively to parameterize models that are then used to make inferences, rather than basing inference directly on a probability sample.
Properly implemented, NFI-based methods satisfy Tier 3 requirements for above-ground biomass as set out in the GPG2003. Long-established NFIs are well-documented with respect to the validity and completeness of the data, assumptions, and models. Although new tropical NFIs do not have such long histories, and may face additional difficulties with locating and re-measuring plots in hard-to-reach areas, lessons learned from forest inventories in non-tropical countries can be used to improve sampling designs, field protocols, and statistical estimators.
Box 20: Stratification and statistics
In statistics, stratification subdivides a population into sub-populations, called strata, for two primary purposes, namely to:
- identify important sub-populations such as primary versus modified natural forest or deforested versus undisturbed forest area for which separate estimates are required,
- reduce the uncertainty of estimates for population parameters and/or selected sub-population parameters.
The two purposes are not necessarily mutually exclusive.
Stratification as a process aggregates individual population units such as forest stands or image pixels into strata. If the primary purpose of stratification is reduction of uncertainty, then population units assigned to the same stratum should be more similar to each other than to units assigned to other strata.
Two approaches to stratification are common, one characterized as stratified random sampling (also called pre-stratification) and the other characterized as post-stratification. The primary distinction between the two approaches is whether the sampling depends on, or is independent of, the stratification.
With stratified random sampling, the stratification is established before the sampling, primarily so that desired within-strata sampling intensities, and hence within-strata sample sizes, can be ensured. Therefore, in this case, the sampling depends on the stratification. As an example, greater sampling intensities may be desired for forest land subject to human activities than for remote and inaccessible forests generally not subject to human activities. As a second example, stratified random sampling can ensure sufficient sample sizes to achieve desired levels of precision for strata defined by rare activity classes such as deforestation (Olofsson et al., 2013). Under these circumstances, the stratification should be established before the sampling so that sufficient with-strata sample sizes can be ensured. Within-strata sample sizes for stratified random sampling are fixed because they are determined prior to the sampling.
With post-stratification, the sampling is conducted independently of and often before the stratification is imposed. Therefore, because stratified random sampling is not possible, the sampling intensities cannot be varied to accommodate desired within-strata sample sizes. An example is an NFI that uses a combination of permanent plots and a sampling design that does not change over time. One result is that sufficiently large within-strata sample sizes cannot be ensured. Nevertheless, stratifications imposed independently of (often subsequent to) the sampling can still increase the precision of population estimates. For example, if sampling intensities and/or strata sizes are sufficiently large, then the within-strata sample sizes may still be large enough to produce sufficiently precise within-strata estimates (McRoberts et al., 2013). Further, if the strata are relatively homogeneous, then the within-strata variances will be smaller than the overall population variance with the result that the uncertainty of the population mean will be reduced. Within-strata sample sizes with post-stratification depend on the sampling design and the stratification imposed and are generally not known beforehand.
The stratified random sampling and the post-stratified estimators of the population mean are identical and are unbiased in the sense that on average, over all possible samples, the estimate of the population mean will equal the true value. However, the estimate obtained with any particular sample may deviate substantially from the true value. One consequence of the fact that the sample sizes are fixed with stratified random sampling, but unknown initially with post-stratification is that the variance estimators differ slightly (Cochran, 1977, 134 - Eq. 5A.40).
Although NFI data can be used to satisfy criteria for Approaches 1 and 2 to land representation (chapter 2.3.2 of GPG2003 ), sampling intensities rarely exceed 1 plot/km2 (Tomppo et al., 2010, Table 2.3), which is very low spatial resolution. Effective tracking of REDD+ activities on the ground requires a higher spatial resolution than 1km2.
For example, evidence of past disturbance.
Use of permanent plots increases precision of change estimation – see section 126.96.36.199 of GPG2003 .