Scheduled maintenance: 8-10 am ACT/NSW each Wednesday. Expect shutdowns and restarts.

Contact us: data at auscover.org.au  |  Disclaimer: Please read
Outreach » Expert Workshop on Sampling Strategy and Selection of Ground Cover Control Sites

Expert Workshop on Sampling Strategy and Selection of Ground Cover Control Sites

Last modified by Matt Paget on 2013/06/17 18:47

Report:

A report detailing the sampling strategy and site selection criteria is now available:

Malthus et al. (2013). Groundcover monitoring for Australia: Sampling strategy and selection of ground cover control sites. CSIRO Australia

1.           The Expert Workshop

An Expert Workshop on Sampling Strategy and Selection of Ground Cover Control Sites was held at CSIRO Land and Water, Black Mountain, Canberra on the 17th - 18th of August 2010. The workshop brought together 13 experts from Federal and State Agencies (Appendix 1) to discuss the development of a statistically robust sampling strategy for ground cover validation. The meeting agenda is given in Appendix 2.

At the outset of the meeting, it was intended that sampling strategy would:

¡      address the spatial and temporal variability observed in ground cover over Australia (in particular its major agricultural systems and rangelands)

¡      consider the current focus on the rangelands

¡      identify the number and likely locations of validation sites needed over the next three years

¡      prioritise areas to target for field validation

¡      consider existing suitable validation sites in the rangelands and the proposed TERN validation sites

¡      have flexibility to incorporate additional sites if additional resources become available (i.e. monitoring ground cover levels under broadacre cropping)

¡      note the current available budget (~400 site visits including training sites planned with additional resources being sought) and the timeframe of the current project for field site data collection (until May 2013). However, available budget was not a driver in the development of the sampling framework.

¡      Note that the field sampling method for site description and measurement will be based on the Queensland DERM ‘SLATS’ modified discrete point sampling method along 100 m transects, using the star-shaped transect approach (Scarth et al. 2006) for pastoral environments and the cross-transect method (Schmidt et al, 2010) for agricultural crops sown in lines.

 

 

Additionally, in wide ranging discussion, the meeting also considered:

¡      The limitations of the current version of the MODIS FC product

¡      The limitations of the models that will use the FC data (e.g. CEMSYS, SedNet)

¡      The desired reliability of the FC product

¡      The often conflicting criteria for validation site selection and the need for statistical rigour versus available resources

¡      The advantages and disadvantages of more sites versus multiple visits per site

¡      The practicalities of implementing the sampling strategy

 

1.1.   Summary of workshop discussion

The workshop began with perspective setting, in the context of the importance for ground cover and soil condition information to assess impacts of the Caring for our Country programme and of the importance of wind and soil erosion modelling in that process.  The MODIS fractional cover product is seen as the first step in monitoring ground cover at higher resolution in allowing for the initial assessment of variation in ground cover through time and space. The need for a ground validation effort to establish confidence in the data was recognized given the limited validation of the product to date. The aims of the sampling strategy were outlined.

The conceptual basis for the fractional cover algorithm (Guerschman et al. 2009, Section 2) was both outlined and its value as a tool for assessing changes over time was discussed. The qualitative flag analysis (Section 3) was used to highlight known limitations in the product. The spectral and spatial limitations of its practical implementation using MODIS sensor data were recognized. It was acknowledged that the index may overlook useful spectral information available in Landsat and MODIS sensor data and utilized in other ground cover unmixing approaches. The appropriateness of data from the Northern Territory being used to establish endmembers for other parts of the Australian continent was acknowledged as needing evaluation; i.e. there is a need to establish the current algorithm’s representivity. Similarly, the suitability of the triangle being fixed with respect to time and space also needed evaluation. 

Alternative approaches to the development of a sampling strategy were discussed which included:

¡      An index-based site selection approach that used the fractional cover triangle as its starting point. This could highlight areas of low cover and assess their distribution spatially across the continent, and from which proximity and access could then be assessed. However, it was recognized this approach might overlook other spectral variation not represented in the fractional cover triangle itself. 

¡      A stratified approach enabled through the use of a geographical information system  the Multi-Criteria Analysis Software shell (MCAS), a spatial decision support system for mapping and visualisation of different scenarios (Lesslie et al. 2008). The value of this approach as a means for communicating the process to others was recognized.

The required accuracy and reliability of the fractional cover data was also discussed. In the absence of sufficient and adequate validation data, the accuracy of the current fractional cover product was not known, but that those who had worked with the product felt that it was accurate in relative terms, i.e. in showing changes in ground cover fractions over time.  As the ground cover data are ultimately to be used for input into wind and water erosion modelling, the model requirements therefore drive the need for reliability. For CEMSYS the precision of the data required was established as +/- 15% of the bare ground component (J. Leys, pers. comm.). If an estimated FC value for a pixel falls within this range of precision it is regarded as accurate.

For SedNet there is a need to calculate the cover fractions (particularly differentiation between bare ground and non-photosynthetic vegetation) for prediction of sheet and rill erosion. Previously, this has been based on 2001 NLWRA cover data estimated using a satellite-based time series but of uncertain accuracy leading to over-estimation of erosion in northern and central Australia (Hairsine et al. 2009). The precision of the data required was similarly established as +/- 15% (P. Hairsine, pers. comm.).

The working group agreed that there was no developed approach for allocating samples for broad-scale spectral unmixing which could be adopted for the fractional cover validation effort. Without existing information about the variation in spectra of the different components it was also not scientifically feasible to calculate explicit sample sizes to achieve a particular product precision. Thus, in the absence of this information the expert group felt no feel for the adequacy of the ~400 sites that (at the time) current funding allowed as a suitable sample size. Instead, the workshop used the practical experience of the participants to consider what number of validation sites may be necessary. It was proposed to focus validation effort on areas where it is known the algorithm works well (‘validation’, to establish levels of confidence in the data) and on specific land classes shown to be problematic in qualitative validation (Section 3, ‘calibration’, to allow for algorithm improvement). Primarily, this includes rangeland areas dominated by grassland and shrublands, with gibber, red, black, and bright soils. It was agreed that validation over wide spatial scales was the initial priority and, as validation information increased, to latterly focus on spatio-temporal sampling. An initial validation analysis based on existing SLATs data, available for Queensland and parts of NSW, was recommended.

The expert group acknowledged the need to focus effort on priority areas of rangelands (as key sources of wind erosion), with additional effort on croplands (which would require higher temporal resolution ~4 times a year to capture the nature of their variability in cover fractions) (Stewart et al. 2011). Moreover, the need for temporal sampling, at some stage, was recognized, to capture growth response cover changes (e.g. in agricultural areas and following recent rains) to determine temporal reliability of the data. However, in the absence of sufficient spatial validation data, the initial prioritization of sampling sites would be on spatial coverage, with site revisits planned for later sampling seasons.

Thus, the workshop proposed to use stratification to ensure sampling effort was representatively spread across the identified areas of interest. The key initial agreed criteria for site selection included:

¡      A 9:1 split of effort between rangelands and agricultural areas.

¡      To focus effort on areas with less than 20% tree cover.

¡      To initially stratify sampling effort by soil type (gibber, red, black, bright, other).

The practicalities of implementing a sampling strategy were also discussed. The need for sampling sites to be suitable for the resolution of the MODIS product (500 m) was recognized.  Discussion also focused on the robustness of the SLATS sampling method and the contributions to error from its extent (100 x 100 m compared to MODIS pixel size), radial pattern and from operator error. The value in using Landsat data to assist the site selection process (through assessment of homogeneity) was recognized as was allowing field teams the flexibility in determining the ultimate location of a particular sampling site on the basis of greater local knowledge.

The workshop recommended that the sampling strategy be adaptive, with reviews undertaken on an annual basis to modify effort as validation information increased. An ongoing analysis would allow the assessment of soil, vegetation and spectral classes that may have been overlooked as well as allow for expansion of effort should more resources become available for the sampling effort.

 

1.2.   Key expert workshop outputs and recommendations

¡      The absence of existing validation data for nearly all states (with the exception of Queensland and a limited set of sites in New South Wales), makes it a non-trivial task to establish in a purely statistical sense the number of validation sites required to establish confidence in the data.

¡      On the basis of the approach to understanding reductions in error outlined in Section 4, a preliminary analysis of existing Queensland and New South Wales validation data, and any new validation measurements should be undertaken to lead to a more informed idea of the actual number of sites required statistically. As the validation data set increases, any reductions in variance can be assessed.

¡      Reliability of the fractional product is driven by the models that will ultimately use the data (e.g. CEMSYS wind erosion and SedNet sediment erosion models). An aspirational aim of +/- 15% of any of the bare ground component was agreed with the focus on the relativity of the signal over time.

¡      Sampling sites should be representative at the MODIS resolution scale (500m).

¡      Key criteria for validation site selection:

o        90% of the validation effort should be focused on the rangelands where information is most urgently required, with 10% effort on croplands (Figure 5).

o        Predominantly tree covered areas in these regions should be ignored such that validation effort is targeted on the priority non-woody vegetation types of grasslands and shrublands.

o        Sampling sites should subsequently be allocated on a key set of soil colours widespread in the rangelands which are known to cause problems in application of the FC algorithm (gibber) or where qualitative validation has indicated poor performance (red, black and bright soils).

¡      There is a need to spread validation sites spatially but to prioritise on what States can contribute and can reasonably access, including the use of existing suitable sites (Figure 10).

¡      An iterative sampling strategy should be implemented, initially with wide spatial coverage and limited temporal coverage. Sampling effort should be annually reviewed and adapted to meet identified information needs (e.g. sites poorly represented in field data already obtained) and the inclusion of new sampling sites as further resources become available. Site revisits to assess temporal variability should be planned for later sampling seasons.

 

1.           Priority datasets For sampling site selection

On the basis of the recommendations of the Expert Workshop, the key datasets selected for sampling site stratification are:

1.    Priority soils – the four key soil types to be focused on (Figure 12). This data is based on 1:2.5 million scale maps of the Northcote soil classification and the Interim Biogeographic Regionalisation for Australia (IBRA) land surface of Australia (used to identify gibber).

2.    Rangelands boundary based on the Australian Collaborative Rangeland Information System (ACRIS) boundary definition (Figure 9).

3.    Agricultural / non agricultural areas classification, generated from catchment scale land use to mask out non agricultural areas (Figure 13)[1].

4.    Forest cover data set derived from the Forests of Australia 2008 dataset (Figure 14), used to eliminate areas of dominant tree cover from the process.

5.    Road network and SLATS sampling team locations - to define accessibility.

1.           Sampling Strategy

 

1.1.   Background

In simple statistical sampling there is a clear relationship between the accuracy of an estimate, the inherent variation in the population and the sample size.  For example if we wish to calculate the mean of a population with variance file:///Z:/tmp/msohtml1/01/clip_image002.gif, the accuracy of our estimate file:///Z:/tmp/msohtml1/01/clip_image004.gif, the sample mean is simply

file:///Z:/tmp/msohtml1/01/clip_image006.gif

Thus if we know two quantities we can derive the third. Typically we have some estimate of the population variance and have a required precision which we then use to determine the required sample size.

The workshop highlighted a number of difficulties in using this direct approach. First, quantities of interest are not simple means and sums of population quantities but are the result of a spectral un-mixing process. The workshop identified that there is no developed methodology for calculating sample allocation for broad-scale spectral un-mixing.

The second issue was that there was an absence of existing validation data in easily accessible form for nearly all states (with the exception of Queensland and a limited set of sites in New South Wales). This makes it difficult to establish in a statistical sense the variation in spectral signature of the components of interest spatially and temporally and the impact of this on the un-mixing algorithm.

Third, there was no clear expression of the precision that would be required from the product for decision making. Without these components it was not scientifically feasible to calculate explicit sample sizes to achieve particular precisions. Instead the workshop focussed on determining particular land classes that had been problematic in earlier pilot work as areas that would provide maximum information as well as considering sampling protocols at local areas to deal with spatial and temporal scaling issues.

In this section we consider what a credible sampling scheme could look like. We approach it consistently with the workshop discussions. We note that the workshop endorsed that the process of developing any product of this size will be iterative. As data is collected and resources committed to its analysis greater clarity will be achieved about the information needs for a national product.

 

1.2.   Indicative sampling scheme

We begin the development of an indicative sampling scheme by considering what can be learned from other existing national remote sensing products. The primary example in Australia is the land cover product used in the National Carbon Accounting System (NCAS, Furby 2002). We note that this product uses sophisticated classifiers rather than spectral un-mixing so comparisons have to be made cautiously. Given this, it is still the only operational continental remote sensing monitoring product currently in operation.

At its inception the NCAS project collected multiple samples from 800 aerial photos across all 85 IBRA bioregions in Australia (Lowell et al. 2002). This set of initial samples was found to not produce sufficiently accurate estimates over all of Australia so an additional sample 1000 10km by 10Kkm IKONOS satellite images were obtained and multiple calibration samples derived from these. This produced a dataset of around 3000 samples to be used for calibration (P Caccetta, pers com). 

Whether 3000 is an appropriate number of samples is debateable. Given this level of sampling there are still issues with the accuracy of the NCAS data for a number of applications.  Against this, there are potential efficiencies that can be achieved when using spectral un-mixing approaches such as that used in the fractional cover algorithm, as opposed to the hard classification approach adopted in NCAS.  Firstly, unsupervised methods can be used to extract additional signal information from the entire image.  Secondly, certain spectral signals may lack variation over larger regions; this tells us something about the spectral consistency of objects across the space – if an object has more consistent variation across space then it may require a lower intensity of sampling effort.  Thirdly, error in spectral signal may have less of an effect when detecting changes rather than current state (in other words, it may be easier to detect the relative changes in an object  and with less error than in performing object classification first and using that as the basis for the change detection).  On the basis of this, and in discussions with Mark Berman, an international expert on spectral un-mixing with experience in its application to remote sensing it was considered that 1000-1500 sites would be a credible sample, provided that they contained a reasonable number of relatively pure pixels.  Given this, in the following we will assume a sample of 1500 sites.  We do not believe it is likely that a broadly credible national product could be constructed with fewer samples, and the requirement may be higher.

Given that the workshop identified reasonably quickly that there was no way of calculating “required” sample sizes, and that initial funding provided for a relatively low number of sites, discussion concentrated on prioritising locations for sampling and the protocols for the different jurisdictions.  To envisage a national sampling regime requires additional considerations.  There are two considerations.

¡      On a purely statistical basis it is recommended that the sample be distributed systematically across the in-scope region.   This has a number of advantages.  Firstly, it gives maximum spatial representation, and in the absence of information about the spatial patterns of variation this is prudent. A systematic sample, however, has the weakness of not providing spatial information locally to a sample point, but this should not be an issue in this case. A systematic sample should also be approximately self weighting, i.e. each pixel is equally “representative”.  A systematic spatial sample will allocate samples approximately proportional to the area of any spatial strata.

¡      The alternative consideration is that some pixels are harder to un-mix than others.  The workshop identified a number of soil types that had proven problematic from qualitative evaluations of the fractional cover product and recommended focussing attention on these.  Moreover, as new information is collected it can contribute to continuing reanalysis to determine new priorities.

The results of allocating the 1500 sample on a spatially systematic basis is shown in Table 3.  In this table 10 percent of the sample is allocated to non-rangeland agricultural regions as recommended by the workshop.  Forested areas in both regions have been excluded from the analysis. This table should be seen as an aspiration, but information gained as samples are collected will potentially change where the best return on investment occurs.

 

Table 3.  Sample allocations proportional to area based on a 90:10 split between rangelands / non-rangelands as recommended by the workshop. Forested areas excluded.

 

 

New South Wales

Victoria

Queensland

South Australia

Western Australia

Northern Territory

Total

Agriculture

36

24

25

23

41

1

150

Rangelands

108

0

331

142

596

173

1,350

 

The workshop identified a number of soil types that should be targeted to improve the quality of the national product. An allocation proportional to the areas of these different soil types in the rangelands is given in Table 4. An alternative allocation, which fixes the state sample sizes but allocates evenly across the soil classes is given in Table 5.  This allocation gives equal attention to each soil class.

 

It is not possible to choose between any of these allocations on the basis of the information available and the discussions from the workshop.  The primary issue identified that priority should be given to collecting data in certain soil classes.  Thus samples should be targeted in these areas in the initial sampling effort.  Once sufficient data is collected in each soil class as well as spatially a rigorous analysis can be undertaken to:

¡      Assess the validity of the assumption that the 1500 sample size will provide enough information to provide an informative validation of the fractional cover product.

¡      Prioritise further sampling, and confirm existing sites.

¡      Assess the impact of current ground cover sampling effort on the overall uncertainty of the product and to consider i) new priority areas for targeted sampling ii) improvements in the fractional cover algorithm itself.

¡      Consider the utility of the current product and raw data as new information from new sensors potentially becomes available. 

These tasks are not trivial and would require significant resourcing. 

 

In conclusion, at this initial stage of the process an indicative sample allocation is given in Table 3.  Following reanalysis after initial sampling efforts, and with available funds, samples should be targeted into the soil classes identified by the workshop as a priority. When this data is collected a rigorous analysis needs to be completed to further refine the sampling program.

 

 

 

Table 4.  Sample allocations in rangelands proportional to soil classes identified by workshop. Soil regions are taken from the 1:25 million soil map as shown in Figure 12. Non-woody area refers to area of each soil after forested areas have been excluded.

 

Soil colour

Non-woody area

(ha)

New South Wales

Victoria

Queensland

South Australia

Western Australia

Northern Territory

Gibber

9,909,337

-

-

17

19

-

-

Bright

109,964,548

2

-

28

23

264

78

Red

32,093,327

12

-

17

26

54

7

Black

65,394,640

48

0

116

7

18

45

Other

158,946,085

47

0

153

68

260

44

Total

376,307,937

108

0

331

142

596

173

 

 

 

Table 5.  Sample allocations in rangelands equally in each soil class identified by workshop. Non-woody area refers to area of each soil after forested areas have been excluded.

 

Soil colour

Non-woody

New South Wales

Victoria

Queensland

South Australia

Western Australia

Northern Territory

Gibber

9,909,337

0

0

66

28

0

0

Bright

109,964,548

27

0

66

28

149

43

Red

32,093,327

27

0

66

28

149

43

Black

65,394,640

27

0

66

28

149

43

Other

158,946,085

27

0

66

28

149

43

Total

376,307,937

108

0

331

142

596

173

 

 

2.           Ground team specific sampling site selection

 

A sampling protocol for the ground teams in various states has been developed to further aid the selection of suitable field validation sites (full version Appendix 3). This outlines selection based on:

1.    Key soil colours of interest in individual jurisdictions and proportion of sites needed for each soil colour.

2.    Range of fractional cover, given the seasonal conditions (Low brown/green cover; High green cover; High brown cover).

3.    Land cover on the basis of <20% Foliage Projected Cover, and grazing lands in the rangelands.

4.    Site accessibility (distance from base, distance from roads, existing sites, tenure, and availability of Landsat data).

5.    Sites that are homogeneous at the scale of 500 m MODIS pixels, are open, away from boundaries or obvious ecotones and with no obvious evidence of fire scars.

ArcMap project datasets provided to states and territories will include the key datasets outlined in Section 6, along with road data, existing sites and tenure. Landsat data will be used to assist site selection particularly for points 4 and 5 above. 

 

3.           Annual review and temporal sampling

At the beginning of each sampling year (2011, 2012 and 2013), adaptive reviews of the data collected and progress in validation will be required to determine the adequacy of the validation measurements undertaken in areas of priority cover ranges and soil colours (investment will be required to support this ongoing analysis).

The aim here is to ascertain if a sufficient spectral range of data has been acquired (in the so-called spectral space) and to identify if sampling effort needs to be redirected to either meet requirements for spectral or temporal coverage or areas where the algorithm is known to need improvement. As the multi-criteria approach to stratification is not an exact optimization, issues of the spatial representativeness of the validation can be assessed at this stage.

The first two years of sampling effort focus on spatial coverage of validation effort across the priority cover ranges and soil colours. Whilst some analysis of temporal variation can be assessed using information from multiple-revisited sites in the Queensland DERM’s SLATS data analysed in year 1, the need for repeated sampling of sites already visited will be an additional component of the annual reviews in subsequent years (i.e. years 2 and 3).

In cropping areas 3 to 4 repeat visits will be required to map changes over the crop growing cycle; revisiting other sites already measured allows the potential to capture growth response (e.g. to recent rains) and to get a sense of the temporal reliability of the FC product by removing the spatial dimension. In the rangelands, temporal sampling allows for the validation of the precision of changes in the green and dry vegetation fractions through the seasonal cycle (i.e. with what precision can the FC product detect change in the same place?).

Review analysis may also highlight concerns on timing of sampling where insufficient validation effort in the rangelands has been concentrated at particular times of year and have failed to adequately sample other periods (e.g. when dominated by green versus dry vegetation components).

 

 

4.           Expanding the validation dataset

 

In addition to further resources to expand the number of sampling sites proposed in this report, other nationwide measurement efforts that could contribute to the validation of the fractional cover dataset include initiatives developing out of TERN, the Terrestrial Ecosystem Research Network, funded from NCRIS and EIF sources. Of these, the most promising is AusPlots, a subcomponent of the TERN Long-term Australian Multi-scale Plot System (LAMPS) facility. AusPlots is a series of replicated continental-scale surveillance plots implemented in rangelands and forest biomes, to be monitored at low intensity via standardised methodologies. AusPlots-Rangelands is establishing a network of 100 x 100 m plots across the rangelands, building on previous monitoring efforts by state and territory governments, including ACRIS (the Australian Collaborative Rangelands Information System).

Following extensive consultation, a repeatable, quantifiable, and standard sampling method has been devised. 1,000 permanent baseline assessment sites are envisaged in the rangelands with locations stratified on the basis of bioregion, vegetation community type, condition and historical disturbance. The specific location of each plot is ultimately determined in the field.

The intention is to sample the plots once to three times per decade (but only once under NCRIS/EIF funding). Key parameters measured will include plant taxonomy, vegetation structure, Leaf Area Index, cover, soil parameters, plant genetics, spatial patterning and a measure of homogeneity. These measurements are achieved by a combination of point and transect measures across the plots (B. Sparrow, pers. comm.).

Stratification of the plot locations is based on different criteria than those identified as high priority in this report. Nevertheless, although adopting a different sampling approach to the SLATS method adopted here, the AusPlots dataset may have considerable value for fractional cover product validation in providing a complementary dataset with additional spatial coverage across the rangelands. Moreover, the appropriateness of each site to MODIS scale validation can be assessed on the basis of the site homogeneity measure. The low sampling intensity will not, however, meet the shorter-term needs for temporal resolution data.

Additional TERN initiatives of relevance include:

¡      Long-Term Ecological Research Network (LTERN), designed to continue measurements at a core set of existing extensive plot networks or plot transects in key national ecosystems and biomes. Most sites are networks of long archives of plots of variable size, but the AusPlots measurement method will be adopted.

¡      TERN AusCover, the primary aim of which is to provide Australian-wide biophysical map products and remote sensing data time-series, but for which field calibration and validation effort is included. Validation will be targeted at up to seven intensive sites (5 x 5 km) covering a range of key biomes, selected for homogeneity, for intensive airborne hyperspectral and lidar studies, and combined with intensive field measurement programmes. As such, sampling sites will be few and less concentrated in the rangelands, but will provide data to assess accuracy of fractional cover estimation at a range of scales. 

 

5.           Conclusions

An ideal sampling strategy has been proposed aimed at pragmatic validation of the MODIS FC product over and above the very limited validation presented in the original paper (Guerschman et al. 2009). Validation is critical to establish a) confidence in the reliability of the data for use in erosion models and b) to highlight areas of low confidence (high error) as useful guidance for algorithm improvement and c) reliability in monitoring fractional cover components in themselves. In addition to inputs into the CEMSYS and SedNet models, the FC product also has great potential for providing an indication of ground cover changes for land managers. To have confidence in the data these managers need to know the reliability of the indicated cover.

In the absence of existing validation data, the sampling scheme proposed addresses the need for rigour to achieve confidence in the fractional cover data whilst acknowledging the practical challenges in achieving a validation of the dataset at continental scale. A stratified approach to sampling at 1500 sites is proposed, focused on key areas for which accurate cover data is seen as most urgent to inform wind and water erosion studies (rangelands and croplands). It is further focused on grass and shrubland cover types and high priority soils (black, red, gibber and bright).

During Year 1, existing Queensland (~500 sites) and NSW (~30 sites) field estimates of ground cover, obtained using the SLATS method, will be used to perform an initial analysis of the MODIS FC product stratified by soil colour/type, and type and amount of vegetation cover/structure. The output of this analysis, performed by CSIRO Land and Water, will be a preliminary identification of areas where the FC algorithm is performing well and less well and allows initial estimates of variability. Consistency in the results across the two states can also be assessed as to a certain extent can temporal variation. Results will further inform the development of the sampling strategy in subsequent years (2 and 3) during annual review, for both further validation and algorithm improvement.

Current resourcing does not meet the needs for field sampling visits to adequately validate the fractional cover product. Further resources are required and additional TERN-funded surveillance (AusPlots) data could be incorporated. Resourcing will also be required to allow for annual review and analysis such that the sampling strategy can be adaptively implemented to focus effort on prioritised information needs.

In the context of this study rigour is addressed by consistency, encapsulating:

¡      A standard methodology for the identification of priority areas for site visits

¡      Guidance to field teams on the selection of specific sites to sample

¡      Standardisation of the methodology to be used for acquiring field data by the State-based validation teams (including consistent training in the method and common guidelines).

 

Tags:
Created by Peter Scarth on 2012/10/19 16:52

This wiki is licensed under a Creative Commons 2.0 license
XWiki Enterprise 4.5.3 - Documentation