Assessment of the accuracy of profile available water and potential rooting depth estimates held within New Zealand's fundamental soil layers geo-database.
The Fundamental Soil Layers (FSL) contain a range of New Zealand's soil attribute information within a Geographic Information System (GIS) accessible spatial database. The development of the FSL relied heavily on professional assessment of regional soils, often using limited data, old soil maps of differing quality and different soil classification systems (Hewitt 2008; Lilburne et al. 2012). Whilst the FSL provides estimates for a range of useful soil attributes, there are questions around the accuracy and lineage of the data used to generate these estimates (Newsome et al. 2000), and a new statistically based soil map-S-M AP (Lilburne et al. 2012) is being developed to replace the FSL. In the interim the FSL remain a primary source of soil data for interested users, and the data have been used in a variety of applications.
An important recent example is the development of the NIWA CLUES tool. This tool brings together a suite of existing models, developed by various research agencies, for assessing the effect of land use on water quality. Many of the key models included in CLUES use inputs from the FSL to drive the results (Woods et al. 2006). Included in CLUES is the OVERSEER nutrient budget model program (OVERSEER 2013), which has been used to generate maps of nutrient runoff with a high level of spatial detail (100m resolution) (Dymond et al. 2013). To generate these nutrient runoff maps the OVERSEER model was parameterised with data obtained from the FSL to identify shallow soils.
In other examples FSL data have been used to develop a surface of soil C : N ratios for New Zealand (Watt and Palmer 2012), and to identify environmental factors influencing wood properties in plantation forests across New Zealand (Palmer et al. 2013).
The spatial scale of the FSL data used in these examples varies from point data extracted for specific locations (Watt and Palmer 2012; Palmer et al. 2013), to catchment level data in the case of the CLUES models. The FSL is largely based on small scale soil maps, with properties assigned to parcels of soil through a process of interpretation and judgement by regional soil experts (Hewitt 2008; Lilburne et al. 2012). With little formal testing of the data at any scale the suitability of the FSL data for purposes other than regional assessments and generalisations for large areas remains unclear to users of the data.
This study provides an accuracy assessment of the FSL attributes PAW and PRD, and informs users of potential data limitations in certain use cases. In addition, we examine the utility of FSL data lineage and variability estimates as indicators of data reliability. Using recent data measured from soil profiles across the New Zealand landscape we compare accurate observations of two key soil properties, Profile Available Water (PAW) and Potential Rooting Depth (PRD) against the corresponding estimates found in the FSL. These FSL estimates were accompanied by estimated lineage information indicating the source and quality of the estimates. We evaluate the effect of the FSL estimate lineage on the accuracy of predictions by examining the impact of each lineage estimate class on the relative accuracy of the FSL predictions.
Materials and methods
Fundamental soil layers data
FSL data were acquired from the Land Resource Information System (LRIS) (Landcare Research NZ Ltd 2000a, 2000b). These data are attached to the New Zealand Land Resource Inventory (NZLRI) polygon database. Each polygon defines a parcel of land judged to be homogenous in terms of five physical factors, including soil, which form the basis of the inventory (Lynn et al. 2009). The resulting parcels are typically small, with an average area of 2.5 [km.sup.2], and the national coverage NZLRI contains ~100 000 polygons. Each NZLRI polygon is assigned a single set of values for all FSL attributes, including PAW and PRD class estimates (Newsome et al. 2008).
PAW class provided an estimated range for profile total available water. Estimates were derived by measuring the volumetric change in water content between water potentials of -10kPa and -1500kPa for the profile, to a depth of 0.9 m or to the estimated rooting depth (whichever is the lesser) (Newsome et al. 2008). There are six classes, covering the range from 0 to 350 mm PAW.
PRD class provided estimates of the approximate depth to a layer that may be restrictive to root growth. Assessment of PRD is based on the work of Webb and Wilson (1995) whereby a maximum Estimated Rooting Depth (ERD) is reduced to account for limiting soil properties such as: poor aeration, low soil water and penetration resistance. There are six classes, covering the range from 0.15 to 1.5 m PRD.
For the purposes of this study two sets of class boundaries (minimum and maximum values) for PAW and PRD were defined for each soil polygon. The 'narrow' class boundaries were defined using only the maximum and minimum values for the class value assigned to the polygon. Using PRD as an example, PRD class values, and their minimum and maximum PRD (PRD MIN and PRD MAX respectively) are shown in Table 1.
For each polygon, the PRD class value (with 'narrow' boundaries specified as in Table 1) is accompanied by an assessment of the variability of the class boundaries (Table 2). Applying assessed variability to the 'narrow' classes increases the class interval width by an amount specified in Table 2, to generate 'wide' class boundaries (Newsome et al. 2008). For example, a soil with PRD class 3 but a variability of 1 has 'narrow' class boundaries of 0.6-0.89 m (Table 1), but since it 'straddles' the class above and the class below, the lower 'wide' class boundary is the lower boundary for PRD class 4 (0.45 m) and the upper 'wide' class boundary is the upper boundary for PRD class 2 (1.19 m).
For PAW, similar logic was used to generate 'narrow' and 'wide' class boundaries (see Supplementary Materials as available on journal's website).
Minimum, maximum and midpoint values presented in the FSL data layers typically refer to the 'wide' class boundaries. However, observed soil data were tested for their occurrence within expected FSL class boundaries, for both 'wide' and 'narrow' class intervals. This was done to assess the degree to which apparent FSL accuracy is affected by the generous widths of the 'wide' interval.
PAW and PRD estimates in the FSL are also accompanied by lineage information describing the source and quality of the data used to produce each estimate (Table 3). The most reliable estimates are coded 'm' (estimated from analyses or measurements of the named soil) and the least reliable estimates are estimated from 1 : 253 440 soil maps, and are coded 'uf.'.
A dataset consisting of 35 soil profiles was made available by Dr Michael Watt of Scion (New Zealand Forest Research Institute Limited). The Scion dataset was designed to cover eight of the 15 major New Zealand soil orders, and was intended to be representative of soils used for forestry across New Zealand. Following the method of Gradwell (1972), saturated core samples were drained using a hanging column of water, followed by further extraction in a pressurised vessel. The dataset included a summary value for Available Water Capacity (AWC), defined as the difference in soil water content between -10kPa and -1500kPa water potential. This definition is consistent with the values used to define PAW in the FSL (Webb and Wilson 1995).
Potential Rooting Depth was assessed in the field using a combination of soil morphology and observable rooting depths, or the depth to impenetrable rock or pans or dense layers. A field penetrometer was sometimes used as a guide for assessing dense subsoil layers (M. Watt, pers. comm., September 7, 2012). These techniques incorporate all criteria used to assess PRD for the FSL (Newsome et al. 2008).
GPS coordinates for each pit were imported into to ArcGIS (ArcMap10, ESRI Redlands, CA 92373-8100) and converted to match the FSL geodatabase projection (NZTM2000). A spatial join was performed to extract FSL estimates for PRD and PAW at each pit location.
A second dataset of 173 profiles was obtained from the Sustainable Land Management and Climate Change (SLMACC) project 'Forestry systems for difficult sites' (M. Bloomberg, M. D. Laffan, unpubl. data) aimed at evaluating 'difficult' land targeted for afforestation in New Zealand. Land meeting this criterion is difficult to farm sustainably due to climate or soil limitations, and the areas sampled included steep hill country and erosion prone land, identified using the New Zealand Land Use Capability (LUC) system (Lynn et al. 2009). Profiles were exposed to at least 1 m depth, or until an impenetrable layer was reached, with the deepest pit being 1.5 m. This dataset did not contain measured values for PAW.
A GPS receiver was placed at the head of each pit for the duration of the sampling, and the average of the calculated positions was recorded on completion of sampling.
The methods for overlaying and extracting data used in the Scion dataset were repeated to extract FSL estimates for the SLMACC PRD observations. However, some PRD observations were recorded only as >1,0m, which meant that they might fall into one of two FSL PRD classes (PRD class 1, 1.2-1.5 m or PRD class 2, 0.9-1.19 m). To address this, the two deepest PRD classes (1 & 2) were merged before analysis. Only 11 points were affected by this merger, and the effect on the results was assessed to be minimal.
Several tests have been run to assess the accuracy of FSL estimates for PAW and PRD from the Scion and SLMACC data. First, the accuracy of 'wide' and 'narrow' intervals for each variable was assessed by calculating the proportion of observed values that fell within the 'narrow' and 'wide' class intervals and producing exact binomial 95% confidence intervals (Cl) for each. Because the two proportions are not independent (falling within the 'narrow' interval always means falling within the wider interval), no statistical tests were run to see whether the proportion of observations which fall within the 'wide', but not within the 'narrow' interval was significantly different from 0.
In order to see whether there is a linear relationship between the FSL midpoint estimates for PAW and PRD and the respective observations in Scion and SLMACC datasets, a regression was fitted to these data. The use of FSL midpoint estimates for PAW in geospatially-based models such as the Soil Water Balance Model (SWatBal) (Palmer et al. 2009) made these values a logical choice for testing the accuracy of the FSL estimates. To test for the effect of lineage estimate on the above relationship, a non-parametric ANOVA was run with lineage as the fixed effect. The non-parametric bootstrap ANOVA was chosen to take into account non-normality of residuals. The unbalanced design was taken into account by calculating Type 111 statistics.
For the SLMACC dataset, the spatial hierarchy of the samples (samples were spatially clustered by region-forest-forest inventory plot) was taken into account by fitting the mixed effects regression model, rather than the fixed effects regression model, with various levels of the spatial hierarchy as random effects. For the Scion dataset, there was no apparent spatial hierarchy, with individual plots widely distributed throughout New Zealand.
Diagnostics were run for all the regression models to check for the fit, and residual and random effects normality.
Where sufficient data (i.e. more than two observations) were available the Pearson's product moment correlation coefficient between midpoint and observed data for PAW and PRD for each lineage estimate category was calculated. Because non-normality of observations was a potential concern when evaluating the statistical significance of the estimated Pearson's correlation coefficient, a non-parametric permutation test was also run. The results, which are available on request, did not differ from those obtained from the standard test.
Finally, a logistic regression was used to assess the effect of lineage on the likelihood of the observed values falling within the 'wide' FSL class interval. This was repeated for the 'narrow' interval. For the SLMACC dataset, a mixed-effects logistic regression model was fitted to account for the spatially hierarchical structure of the data.
All statistical calculations were carried out using the R statistical software package (R Core Team 2012). The mixed effects models were estimated using the lme4 package (Bates et al. 2011).
The proportion of observed values within FSL intervals
Table 4 shows the proportion of observed PAW and observed Scion and SLMACC PRDs which fall within the 'wide', but not 'narrow' intervals.
In general, the proportion of observed values falling within 'wide' or 'narrow' FSL intervals did not exceed 52%. The exception was for the SLMACC PRD, where 70% of observed values fell within the 'wide' interval. Note that the average 'wide' FSL interval width for PRD was wider for the locations in the SLMACC dataset (0.8 m) when compared with the locations in the Scion dataset (0.5 m) (data not shown). This greater 'wide' interval width has a large influence on the apparent accuracy of the FSL class estimates when compared with the observations in the SLMACC dataset. Using the 'narrow' interval, the probability of success (observed value falling within the class interval) fell to 36%.
Although not tested for statistical significance, the proportions of observations falling within 'wide' but not 'narrow' intervals were sizeable, therefore the choice of interval width had a large bearing on the apparent accuracy of the FSL estimates.
Testing for linear relationship between FSL midpoint estimates and observations
The results of the regression models fitted to the observed data vs estimated (FSL class midpoint) data are shown in Fig. 1, with the thick line corresponding to the fitted model and the dashed line corresponding to a perfect fit between observed and estimated data.
The relationships between observed data (Scion PAW, Scion PRD and SLMACC PRD) and FSL 'wide' midpoint estimates were weak. Pooled correlations were 0.17, 0.13 and -0.08 between FSL midpoint estimates and PAW, PRD Scion and PRD SLMACC data respectively (Table 5).
Correlations for models classified by data lineage ranged from -0.40 to 0.99. None were statistically significant except for the negative correlation between the FSL 'm' data lineage and the SLMACC PRD data (Table 5). The correlation of 0.99 between FSL 'm' data lineage and the PAW data was not significant, because it was estimated from only 5 data.
Non-parametric ANOVA with data lineage as the fixed effect showed that lineage had no significant effect on the accuracy of the FSL midpoint estimates (P = 0.739, 0.927 and 0.532 for the regressions for PAW, PRD Scion and PRD SLMACC respectively).
Testing for the effect of data lineage on probability of data falling within class intervals
A logistic regression was used to assess the effect of FSL data lineage on the likelihood of the observed values falling within FSL class intervals. For the PAW data in the Scion dataset, the estimated probabilities showed the expected trend, with decreasing probabilities from the most reliable ('m') to the least reliable ('uf') lineages (Fig. 2). However, the trend was not strong and lineage proved to have no statistically significant effect on the probability of the observation being contained in the 'wide' or in the 'narrow' interval (P = 0.20 and P=0.07 respectively).
Similarly, for the Scion dataset, the lineage had no effect on the probability of a PRD observation falling within the 'narrow' or the 'wide' interval (P = 0.36 and P = 0.38 respectively). In contrast, for the SLMACC dataset, after adjusting for the hierarchy of the data the lineage was found to have a statistically significant effect on the probability of a PRD observation falling within the 'narrow' or the 'wide' interval (P = 0.02 for both). However, the trends in the probabilities did not conform to expectations. The trend for the 'narrow' interval was the inverse of the expected trend, with probability of a PRD observation falling within the 'narrow' interval actually increasing from lineages m to u, before declining sharply for the uf lineage. For the 'wide' interval, probabilities for lineages m to u were similar, with only the probability for the least reliable uf lineage being markedly lower than the rest.
It is clear that the FSL failed to provide useful estimates of PAW for the Scion dataset. The use of class estimates with minima and maxima is an attempt to overcome the inherent uncertainty of PAW estimates. However, the results show that even with the application of the generous 'wide' interval, which makes allowance for the likely variability of the estimates, only slightly more than half the observed values occurred within the 'wide' interval. In addition to this concern, the over-generous nature of the provided intervals masks the true quality of the data. Use of the 'narrow' class interval provides a better indication of the actual quality of PAW estimates.
While the data lineage did display the expected declining trend in accuracy from the most reliable 'm' lineage, to the least reliable 'u' and 'uf lineages, lineage did not have a statistically significant effect on the probability of an observed PAW value falling within either 'wide' or 'narrow' class intervals.
Finally, regression relationships between observed PAW and FSL midpoint estimates were generally weak or even the inverse of the expected positive correlation.
In describing the techniques for deriving the PAW values for the FSL, Webb and Wilson (1995) make reference to a variety of methods including pedo-transfer techniques and mean values for soil types. These techniques vary in accuracy, but similar approaches have been shown to produce reasonable estimates of PAW (Wosten et al. 2001). In the New Zealand context, Webb (2003) and Cichota et al. (2013) demonstrated the use of basic soil information to predict soil water release properties for grouped soils--suggesting that PAW in particular can be successfully predicted at a landscape level using basic soil information. It is thus unlikely that the low level of accuracy within the FSL can be attributed to the failings of these techniques alone. A more likely explanation is the limited availability of high quality measured soil data for New Zealand. A review of the literature revealed a great deal of uncertainty and high error levels in the creation of the soil maps and the assignment of a named soil type to each polygon within the FSL. Webb et al. (2000) found the soil data used for the LR1S to be unreliable and variable. They identify problems with poor geographic distribution, frequency and representation in the sampling methodology as well as changing soil series names and maps without much new sampling (Webb et al. 2000). This difficulty is not unique to the FSL and has caused a proliferation of approaches to soil mapping (McBratney et al. 2003).
An additional complication is that a single soil polygon represents a parcel of land which is likely to be heterogeneous in some respects. Extracting an estimate from a single point for testing purposes, or use in a model, provides only a narrow reflection of the data accuracy within the larger polygon. Nonetheless, many users of the FSL will require only point based estimates. In addition to the examples provided in the introduction, Gibb et al. (1999) note that land resource inventory maps in New Zealand are 'increasingly being used in raster-based spatial process models of climate change, carbon sequestration, hydrology, catchment erosion, and forest ecosystems' and this comment holds true for the FSL (for further examples see: Barringer and Lilburne 1999; Palmer et al. 2009; Hock et al. 2014).
Finally, models of soil properties and processes are often scale-dependent, and aggregation or disaggregation of results may require selection of a different model. Reasons for this include: i) The aggregation level, since sampled points are usually averaged at smaller map scales, it) input data--few soil surveys offer the same detail at increasing scales, and iii) scale dependent changes in the relative importance of sub-processes -point sampling will emphasise local processes whilst missing broader trends (Heuvelink and Pebesma 1999). In contrast to this approach, development of the FSL relied on expert judgement to interpret a variety of regional soil maps developed at different scales, using inconsistent methods and for a variety of purposes.
Results for PRD appeared encouraging with 70% of the measured SLMACC dataset values falling within the 'wide' interval. These results were tempered by those from the Scion dataset, where only 31% of measured values fell within the 'wide' interval. In addition, the width of the average 'wide' interval (80 cm) was so great that class intervals failed to establish meaningful boundaries on estimates. For these reasons the 'narrow' interval is considered a better indicator of data accuracy. For both the SLMACC and Scion datasets the small fraction of observed values that lie within the 'narrow' interval confirms that as for PAW, FSL class estimates are poor indicators of the location of the observed PRD.
Compared with the Scion dataset, the application of 'narrow' intervals to the SLMACC dataset had a much greater effect on the proportion of values considered accurate. It is possible that the merging of two FSL class intervals where PRD >1 m may have caused this result. Closer inspection revealed that only 11 profiles were affected by this change and could not account for the large number of SLMACC dataset values falling into the FSL 'wide' class interval. An alternative explanation may lie in the sampling methodology used. Inspection showed that many more of the SLMACC data points fell into the high variability FSL classes than for the Scion dataset. Applying the adjustments for higher variability estimates to the class estimate intervals resulted in much wider intervals in the FSL data used for comparison with the SLMACC dataset, and hence the greater proportion of measured values that were counted as 'In'.
Given the poor results for the accuracy of FSL interval data, the regression of observed PRD values vs FSL estimates for midpoint PRD had little chance of producing accurate results. The correlation coefficients from the SLMACC dataset were the only statistically valid results, and they strongly contradicted the FSL documentation, with the theoretically most reliable 'm' lineage having a negative correlation between measured values and FSL midpoint estimates (Table 5). The logistic regression confirmed the fact that estimate lineage was in most cases either unrelated or even contra-indicative of the accuracy of PRD values. For example, using the 'narrow' interval, the FSL estimate with two of the less reliable lineages (r and u) would appear to provide the best estimates. This suggests lineage estimate is an unreliable predictor of data quality.
Ichii et al. (2009) note that soil surveys may often underestimate the true rooting depth but the similarity in techniques used to derive PRD values for both the FSL estimates and the two observed datasets makes it unlikely that measurement error is the source of the discrepancy.
Examples where success has been achieved in modelling PRD in a landscape have relied on topographic inputs such as slope and catchment curvature (Ziadat 2010). Other sophisticated studies have used spatiotemporal evapotranspiration models derived from satellite based remote sensing data (Ichii et al. 2009). In contrast, the PRD estimates for the FSL have relied heavily on soil surveys and professional interpretation of land form. This suggests that the FSL methodology may explain the poor performance of PRD estimates. This assumption could be tested if newer techniques substituting topographic information from digital sources were used to derive alternate estimates of PRD for selected regions where estimates could be tested with field observations.
The FSL can be seen as an example of soil mapping relying largely on Bui's (2004) framework of an expert based approach. In contrast, McBratney et al. (2003) suggest that quantitative approaches should be favoured in the development of new soil maps. In outlining a framework for the creation of digital soil maps, McBratney el al. (2003) emphasise the importance of accurate soil data as both an input to the framework, and for validation of the final soil map. In light of this, the same lack of soil data underlying the weakness of the FSL may limit any future soil mapping efforts without the acquisition of new field data.
It may be argued that the FSL data are appropriate only at regional scales. However, it is clear that regardless of the intended uses of the FSL, the data are being used at a variety of spatial scales, and our results raise questions about the suitability of the FSL data extracted at large map scales.
While we cannot generalise the results from this study to the whole of New Zealand, note that: i) the Scion dataset was designed to cover 8 of the 15 major NZ soil orders, and was intended to be representative of a wide range of soils under forests across New Zealand; if) the SLMACC dataset, in contrast, was focussed on 'difficult' soils with limitations to pastoral farming; and Hi) both the Scion and SLMACC datasets gave similar results when compared with the FSL.
It is possible that by focusing on 'difficult' and/or forested sites our sample locations coincided with heterogeneous regions not well defined by the NZLRI methodology. While we cannot entirely rule this out, the average area of the NZLRI polygons in this study was quite small, and was consistent with the average size for all NZLRI polygons (2.5 [km.sup.2]). This observation, combined with the poor performance of the lineage estimate, suggests our results should hold for non-forested land also.
The results of this study showed that FSL estimates for PAW and PRD are unreliable when compared against observed data. Where specified class intervals contain a high proportion of the actual values, the intervals are excessively wide. The lineage of data used to determine FSL estimates has negligible effect on the estimation accuracy.
The poor performance of the FSL observed in this study highlights that the users of this data should be cautious when operating at larger map scales. Recommendations for future research include: identification of the source of the errors identified in this research, and the suitability of FSL data for regional applications.
This project was funded in part by the Ministry for Primary Industries SLMACC UOCX0901 'Forestry Systems for Difficult Sites' contract. Dr Michael Watt provided access to Scion data, as well as helpful guidance.
Barringer J, Lilburne L (1999) Scale issues in developing regional-scale soil water balance surfaces. In 'Proceedings of the 11th Annual Colloquium of the Spatial Information Research Centre.'(University of Otago: Dunedin, New Zealand)
Bates D, Maechler M, Bolker B (2011) lme4: Linear mixed-effects models using S4 classes. R package Version 0.999375-42. Available at: http://CRAN.R-project.org/package=lme4
Bui EN (2004) Soil survey as a knowledge system. Geoderma 120, 17-26. doi: 10.1016/j.geoderma.2003.07.006
Cichota R, Vogeler I, Snow VO, Webb TH (2013) Ensemble pedotransfer functions to derive hydraulic properties for New Zealand soils. Soil Research 51, 94 111. doi: 10.1071/SR12338
Dymond J, Ausseil A-G, Parfitt R, Herzig A, McDowell R (2013) Nitrate and phosphorus leaching in New Zealand: a national perspective. New Zealand Journal of Agricultural Research 56, 49-59. doi: 10.1080/00288233.2012.747185
Gibb RG, Johnston MR, Harmsworth GR (1999) Estimating spatial thematic uncertainty in the NZLRI. In 'Paper presented at the SIRC 99. The 11th Annual Colloquium of the Spatial Information Research Centre.' (University of Otago: Dunedin, New Zealand)
Gradwell MW (1972) Methods for physical analysis of soils. Department of Scientific and Industrial Research, Lower Hutt, New Zealand.
Heuvelink GBM, Pebesma EJ (1999) Spatial aggregation and soil process modelling. Geoderma 89, 47-65. doi : 10.1016/SOO16-7061 (98)00077-9
Hewitt AE (2008) 'Assessment of available soil and land resource information for the Manawatu-Wanganui Region.' (Landcare Research, New Zealand and Manawatu-Wanganui Horizons Regional Council: Palmerston North, New Zealand)
Hock B, Payn T, Heaphy M (2014) Issues in the re-use of non-forestry specific spatial data sources for national environmental reporting on planted forests in New Zealand. Journal of Spatial Science 59, 237-252. doi: 10.1080/14498596.2014.913508
Ichii K, Wang W, Hashimoto H, Yang F, Votava P, Michaelis AR, Nemani RR (2009) Refinement of rooting depths using satellite-based evapotranspiration seasonality for ecosystem modeling in California. Agricultural and Forest Meteorology 149, 1907-1918. doi: 10.1016/j.agrformet.2009.06.019
Landcare Research NZ Ltd (2000a) New Zealand Fundamental Soil Layer, North Island. Available at: http://lris.scinfo.org.nz/layer/136/(accessed 2 September 2012).
Landcare Research NZ Ltd (20006) New Zealand Fundamental Soil Layer, South Island. Available at: http://lris.scinfo.org.nz/layer/137/(accessed 2 September 2012).
Lilburne LR, Hewitt AE, Webb TW (2012) Soil and informatics science combine to develop S-map: A new generation soil information system for New Zealand. Geoderma 170, 232-238. doi:10.1016/j.geoderma. 2011.11.012
Lynn IH, Manderson A, Page M, Harmsworth G, Eyles G, Douglas G, Mackay A, Newsome PFJ (2009) 'Land use capability survey handbook: A New Zealand handbook for the classification of land.' pp. 10-12. (Manaaki Whenua Press: Lincoln, New Zealand)
McBratney AB, Mendonca Santos ML, Minasny B (2003) On digital soil mapping. Geoderma 117, 3-52. doi : 10.1016/SOO 16-7061 (03)00223-4 Newsome PFJ, Wilde RH, Willoughby EJ (2000) 'Land and resource information system spatial data layers. Vol. 1 Label format.' (Landcare Research: Palmerston North, New Zealand)
Newsome PFJ, Wilde RH, Willoughby EJ (2008) 'Land and resource information system spatial data layers: Data dictionary.' (Landcare Research: Palmerston North, New Zealand)
OVERSEER (2013) AgResearch. Available at: http://overseer.org.nz/
Palmer DJ, Watt MS, Hock BK, Lowe DJ, Payn TW (2009) A Dynamic Framework for Spatial Modelling Pinus radiata Soil Water Balance (SWatBal) across New Zealand. Forest Research Bulletin No. 234. Scion Research, Rotorua.
Palmer DJ, Kimberley MO, Cown DJ, McKinley RB (2013) Assessing prediction accuracy in a regression kriging surface of Pinus radiata outerwood density across New Zealand. Forest Ecology and Management 308, 9-16. doi:10.1016/j.foreco.2013.07.024
R Core Team (2012) 'R: A Language and environment for statistical computing.' (R Foundation for Statistical Computing: Vienna, Austria) Available at: www.R-project.org
Watt MS, Palmer DJ (2012) Use of regression kriging to develop a Carbon: Nitrogen ratio surface for New Zealand. Geoderma 183-184, 49-57. doi: 10.1016/j.geoderma.2012.03.013
Webb TH (2003) Identification of functional horizons to predict physical properties for soils from alluvium in Canterbury, New Zealand. Australian Journal of Soil Research 41, 1005-1019. doi:10.1071/SR01077
Webb TH, Wilson AD (1995) 'A manual of land characteristics for evaluation of rural land.' (Manaaki Whenua Press: Lincoln, New Zealand)
Webb TH, Claydon JJ, Harris SR (2000) Quantifying variability of soil physical properties within soil series to address modern land-use issues on the Canterbury Plains, New Zealand. Australian Journal of Soil Research 38, 1115-1129. doi: 10.1071/SR99091
Woods R, Elliot S, Shankar U, Bidwell V, Harris S, Wheeler D, Clothier B, Green S, Hewitt A, Gibb R, Parfitt R (2006) The CLUES Project: Predicting the effects of land-use on water quality--Stage II. Report prepared for the Ministry of Agriculture and Forestry. NIWA Client Report HAM2006-096. Available at: http://maxa.maf.govt.nz/mafhet/rural-nz/sustainable-resource-use/clues/stage-2/Stage- II.pdf (accessed 11 September 2015).
Wosten JHM, Pachepsky YA, Rawls WJ (2001) Pedotransfer functions: bridging the gap between available basic soil data and missing soil hydraulic characteristics. Journal of Hydrology 251, 123-150. doi: 10.1016/S0022-1694(01 (00464-4
Ziadat FM (2010) Prediction of soil depth from digital terrain data by integrating statistical and visual approaches. Pedosphere 20, 361-367. doi : 10.1016/S1002-0160(10)60025-2
Grant Pearse (A), Elena Moltchanova (B), and Mark Bloomberg (A,C,D)
(A) New Zealand School of Forestry, University of Canterbury, Christchurch, New Zealand,
(B) Department of Mathematics and Statistics, University of Canterbury, Christchurch, New Zealand.
(C) Present address: Department of Land Management and Systems, Lincoln University, Christchurch, New Zealand.
(D) Corresponding author. Email: firstname.lastname@example.org
Table 1. Definitions of 'narrow' PRD class boundaries Reproduced from Newsome et al. (2008) PRD CLASS PRD_MIN (m) PRD_MAX (m) Description 1 1.2 1.5 Very deep 2 0.9 1.19 Deep 3 0.6 0.89 Moderately deep 4 0.45 0.59 Slightly deep 5 0.25 0.44 Shallow 6 0.15 0.24 Very shallow Table 2. Definitions of 'wide' class boundaries Reproduced from Newsome et al. (2008) Variability FSL 'wide' class boundaries and location of mean class value 0 Occurs mostly within the nominated class. The middle of the nominated class is a good approximation for a numerical value. 1 Straddles the class above and below. The mean is the middle of the nominated class. 1 - Straddles this class and the class below. The mean is taken at the class boundary. 1 + Straddles this class and the class above. The mean is taken at the class boundary. 2 Straddles 2 classes above and below. The mean is the middle of the nominated class. Table 3. Definitions for lineage information for the FSL Lineage codes are fisted in decreasing order of reliability. Reproduced from Newsome et al. (2000) Value Description m Estimated from analyses or measurements of the named soil. r Estimated from relationships with other soils but this estimate is considered to be reliable. p Deduced from soil profile morphology. This code has been used only in the South Island. u Estimated from relationships with other soils but with an unknown level of accuracy. uf Estimated from General Soil Survey data (scale 1:253 440). In general, the quality of the estimate is less reliable than class 'u' above. Table 4. The proportion (%) of observed PAW and PRD values within FSL class intervals Observed n Wide FSL class Narrow FSL class datasets interval interval Proportion 95% CI Proportion 95% CI PAW 35 51.4 34.0-68.0 31.7 21-55 PRD (Scion) 35 31.4 16.9-9.3 17.1 6.6-33 PRD (SLMACC) 173 70.0 63.0-77.0 36.0 29-44 Table 5. Correlation between the FSL mid-point values and observed values (p-value in brackets) Lineage PAW (Scion) PRD (Scion) PRD (SLMACC) m 0.99(0.0598) 0.56(0.0564) -0.34(0.0261) p 0.38(0.6163) 0.06(0.8687) -0.17(0.2802) r -0.40(0.7397) 0.12(0.8183) 0.26(0.2612) u 0.01(0.9703) -0.04(0.9328) -0.15(0.9190) uf -- -- -0.02(0.9391) pooled 0.17(0.3314) 0.13(0.4440) -0.08(0.2983)
|Printer friendly Cite/link Email Feedback|
|Author:||Pearse, Grant; Moltchanova, Elena; Bloomberg, Mark|
|Date:||Oct 1, 2015|
|Previous Article:||Measuring soil organic carbon: which technique and where to from here?|
|Next Article:||Effect of abandonment of olive orchards on soil organic carbon sequestration in Mediterranean Lebanon.|