- Email: [email protected]
Chapter 29 Rule-Based Land Unit Mapping of the Tiwi Islands, Northern Territory, Australia
Developments in Soil Science, volume 31 P. Lagacherie, A.B. McBratney and M. Voltz (Editors) r 2007 Elsevier B.V. All rights reserved
401
Chapter 29
RULE-BASED LAND UNIT MAPPING OF THE TIWI ISLANDS, NORTHERN TERRITORY, AUSTRALIA Ian D. Hollingsworth, Elisabeth N. Bui, Inakwu O.A. Odeh and Phillip McLeod
Abstract We have applied a decision tree analysis (DTA) to map soil, plant community and land unit classes across the Tiwi Islands (7320 km2), located in northern Australia. Our survey substituted environmental analysis and DTA for traditional air photo interpretation to provide a continuous land unit coverage over the islands. DTA was used to derive mapping rules for land units, their component vegetation classes and soil families from secondary survey site observations and distributed environmental data. The mapping was tested on a legacy data set. The environmental variables used are: elevation, slope, latitude, longitude, landform pattern class, wetness class, static wetness index, erosion or deposition index, Landsat TM band 5:7 and vegetation cover class extracted from digital topographic 1:50,000-scale mapping. We needed to reduce the number of the resulting land unit classes derived from historical surveys into nine classes so as to produce a more meaningful mapping. This was achieved by generalising the component vegetation and soil units of the land unit classes thus producing the broader land unit classification. This procedure dovetails well with the current survey approach in which surveyors often develop more detailed classification systems than can be accurately mapped. We recommend a predictive mapping approach based on explicit mapping rules that integrate available land resources information and facilitate production of upgradeable maps.
29.1 Introduction 29.1.1 Background Thematic map products of land resources survey in Australia have been criticized for not making use of the range of distributed environmental information that is now available and for not producing the land information of a quality required by modern users (Cook et al., 1996; McBratney et al., 2003; McKenzie and Austin, 1993). Land resource assessment in the Northern Territory (NT) is based on air photo interpretation of map units and descriptive surveys (McDonald et al., 1996) to represent soil, landform and habitat diversity in a landscape where the native vegetation is largely intact. The land unit map
402
Ian D. Hollingsworth et al.
product of this process is widely used for development and conservation planning. However, there are gaps and inconsistencies in the resulting coverage because much of the survey work has been project based. Opportunities now exist to upgrade and integrate the land resource maps using extensive and ubiquitous digital elevation models across northern Australia. Digital soil mapping (DSM) methods could potentially be used to improve the quality and extent of land resource mapping in an economic environment where public funding of field survey work is declining. Several DSM techniques have been reviewed extensively by McBratney et al. (2003). A most relevant technique to our work here is, decision tree analysis (DTA), which could be used to generate rule-based mapping analogous to expert systems used by traditional land resource surveyors. Guisan (2000) describes a broad range of applications using this technique for quantitative modelling of ecosystem. DTA has been used to map vegetation properties (Franklin et al., 2000) and to model habitat distribution (Guisan and Zimmermann, 2000) but it is prone to error when extrapolated to unfamiliar landscapes (Gahegan, 2000). DTA was used in Australia to extract soil mapping rules from extensive geology and DEM-derived attributes (Bui et al., 1999) and to produce continuous soil property maps from disparate soil survey data (Bui and Henderson, 2003; Henderson et al., 2005). However, although the results of the traditional land resource approach are generally thematic maps, DTA has not been used at local-to-regional landscape scales to map land resources including soil. The underlying main aim of the work reported here was to make a continuous and consistent assessment of land resources of the Tiwi Islands, comprising Bathurst and Melville Islands (7320 km2) for the purpose of strategic development planning. The islands are located in tropics at 121 S, 1301 E (Fig. 29.1). Disparate land unit maps had been created over parts of the Tiwi Islands in the 1970s to assess land capability for forestry and agriculture (Olsen, 1980; Van Cuylenburg and Dunlop, 1973; Wells and Cuylenburg, 1978; Wells et al., 1978). Since this time, digital 1:50,000 topographic mapping (capable of deriving a 50-m DEM) over the Tiwi Islands has become available and a range of DSM techniques based on the analysis of DEM data has been developed. In this chapter, we utilised a DTA mapping approach in an analogous manner to conventional land resource assessment as practiced in the NT making use of legacy soil survey data, digital topographic mapping and Landsat TM data. The aim was to demonstrate the technique in a way that could be readily used in the routine survey program to rationalise map units and extrapolate existing mapping. Because climate and lithology are relatively uniform across Tiwi Islands (Nott, 1994) we focused on sampling physiographic factors controlling soilscape diversity. We utilized digital mapping methods in a first-pass
Rule-based land unit mapping of the Tiwi Islands, Australia
403
Figure 29.1. Location of the Tiwi Islands in Australia. assessment of the land resources over the full extent of Tiwi Islands, a remote region where traditional land resource mapping has been difficult to apply. 29.1.2 Environment The Tiwi Islands have an equatorial savannah climate that is dominated by northwest monsoon and the southeast trade weather patterns. The islands experience seasonal drought during the southeast trade season between May and September. The surface hydrology exerts considerable control over habitat diversity in this climate because water supply is a limiting factor for plant growth (Eamus, 2003). The geology is relatively simple and consists of quaternary alluvium, flat bedded, tertiary Van Diemen sandstone and the underlying Cretaceous Mookinu mudstone and Wangarlu mudstone members of the Bathurst Island formation. Tertiary Van Diemen sandstone is a fine-to-medium grained quartzose sandstone of fluvial and partly littoral origin. The sandstone covers most of Melville and Bathurst Islands and varies in thickness to a maximum of 80 m and dips gently to the northwest. The origin of the clastic sediments in the sandstone were highlands, located to the south of the Tiwi Islands during the Tertiary Age (Nott, 1994). Dissected plateau remnants 150-m high form the current highlands in the centre of the islands, which are fringed by coastal mangroves and
404
Ian D. Hollingsworth et al.
cheniers. Soils on sandstone are typically Oxisols (Ustoxs) with deep sandy to sandy loam surface horizons. Soils originating from mudstone are typically Ultisols (Ustults) with strongly acid subsoils overlain with a layer of pisolithic gravel. 29.2 Materials and methods 29.2.1 Definition of land units Land units, as described in the NT land resources survey, represent unique subclass combinations of geomorphic unit, lithology, vegetation and soil classification that occur as a repetitive pattern within broader land systems. The land units are delineated from stereographic air photo interpretation and field survey observations onto 1:100,000-scale topographic base maps. The complexity of the unit concept and the subjectivity of subclass definitions in different surveys often lead to disparity in the quality and types of unit being used in different surveys. We correlated historical soil and vegetation survey reports of the Tiwi Islands to produce a legend of 27 land units describing 10 soil family classes and 26 plant community types (Brocklehurst, 1998; Olsen, 1980; Wells and Cuylenburg, 1978; Wells et al., 1978) an excerpt of which is produced in Table 29.1. This approach was consistent with the current land unit classification system used in the NT, which concatenates landform class, soil class and vegetation class to construct a unique land unit class. 29.2.2 Landscape analysis The ANUDEM, program, was used to generate 50-m grid by fitting a splined elevation surface to 1:50,000 digital contours interspersed with spot heights, drainage network and coastline data. TAPES-G (Gallant and Wilson, 1996; Wilson and Gallant, 2000) was used to derive slope, steady-state wetness index and a sediment–erosion index to characterise the spatial distribution of soil water content and erosion–deposition processes. The steady-state wetness index is a derivative of a specific catchment area that could be used as a surrogate for subsurface flow. This is most appropriate to a humid environment (Troch et al., 1993). The steady-state wetness index (o) is defined as: As ¼ ln tan b where As is the specific catchment area (catchment area draining across a unit width of contour; m2/m) and b is the slope angle (in degrees). This index is similar to the specific catchment area, or upslope area per width of contour that has been used widely in soil property mapping from digital terrain data (Wilson and Gallant, 2000). The wetness index was divided into six quantiles to
Rule-based land unit mapping of the Tiwi Islands, Australia
405
Table 29.1. Land unit, vegetation and soil classes used. Mapped land unit
Land unit legend code
Landform element
(a) 85
Gec084
Foot slopes
(a) Eucalypt forest
Gaq085
Hill crests
(a) Eucalypt forest
Gfc085
Slopes
(a) Eucalypt forest
1a: E. miniata, E. tetrodonta and E. nesophila openforest
Udf085
Slopes
(a) Eucalypt forest
1a: E. miniata, E. tetrodonta and E. nesophila openforest
Gaq086
Summit surfaces
(a): Eucalypt forest
1b: E. miniata and E. tetrodonta open forest/ woodland
Laq086
Summit surfaces
(a): Eucalypt forest
1b: E. miniata and E. tetrodonta open forest/ woodland
Lfc087
Plains
(a) Eucalypt forest
1b: E. miniata and E. tetrodonta open forest/ woodland
Gec088
Fan
(a) Eucalypt forest
Uec088
Fan
(a) Eucalypt forest
1b: E. miniata and E. tetrodonta open forest/ woodland 1b: E. miniata and E. tetrodonta open forest/ woodland
(b) 86
(c) 88
Mapped vegetation unit
Described vegetation community 1a: E. miniata, E. tetrodonta and E. nesophila openforest with Chrysopogon fallax grassland understorey 1a: E. miniata, E. tetrodonta and E. nesophila openforest
Soil family
K11 Koolpinyah: deep, gravelly, imperfectly drained, yellow sandy loam over sandy clay loam K9 Hotham: deep, gravelly, well drained, red, sandy loam over sandy clay K9 Hotham: deep, gravelly, well drained, red, sandy loam over sandy clay K9 Hotham: deep, gravelly, well drained, red, sandy loam over sandy clay K8 Berrimah: very deep, well drained, red, sandy loam over acidic, sandy clay K7 Berrimah: deep, slightly or nongravelly, well drained, red, sandy loam over sandy clay loam K8 Berrimah: very deep, well drained, red, sandy loam over acidic, sandy clay T6 Cockatoo: deep, nongravelly, well drained, red sandy T6 Cockatoo: deep, nongravelly, well drained, red sandy soils
Ian D. Hollingsworth et al.
406
Figure 29.2. Static soil wetness index derived from 50-m DEM and classified into six quantiles. represent the distribution of runon and runoff elements in the landscape (Fig. 29.2). While no direct association with traditional landform elements was made, the visible distribution of wetness index quantiles in the landscape indicated a close association with the surface drainage system and hillslope topography. The erosion–deposition index (DTc), a dimensionless sediment transport capacity, was also computed using TAPES-G as a nonlinear function of specific discharge and slope, expressed as h n n i DTcj ¼ Am Am sj sin bj sj sin bj where b is the slope (in degrees) and As is the specific catchment area or drainage area per unit width orthogonal to a flow line (m2/m). DTc represents the change in sediment transport capacity across a grid cell, and can be used as a measure of the erosion or deposition potential in each grid cell (Wilson and Gallant, 2000). Uniform excess rainfall conditions were assumed. The published 1:250,000-scale geological map did not accurately discriminate outcrops of mudstone from extensive flat-bedded sandstone. As an alternative, we used a clay index calculated as a ratio of Landsat TM bands 5 and 7 to assess the variation in lithology. Vegetation cover was assessed from vegetation layer in the 1:50,000 digital topographic mapping. The mapped classes are: mediumdensity woodland, scattered woodland, saline coastal flats and marine swamps, pine plantation, dense vegetation, mangrove, inland water, intertidal flat foreshore, bare areas, lakes, dunes and cheniers.
Rule-based land unit mapping of the Tiwi Islands, Australia
407
29.2.3 Survey design The stratified survey design covering 120 sites included 4 replicate observations in each of 6 elements of 5 landscape patterns. Landform element (50-m grid) and landform pattern (100 ha) variation was premapped using classification of the static soil wetness index at fine and coarse levels of resolution. First, six quantiles of the frequency distribution of static wetness index was used to represent landform element variation. Secondly, a network of regular 100-ha hexagons was overlaid on the wetness class grid and the areas of each wetness index class were cross-tabulated for each hexagon. Using an agglomerative classification, ALOC in the PATN software program (Belbin, 1987) five landform classes representing coarse (100-ha resolution), landform pattern variation were created. The landscape pattern variation and the survey site coverage are depicted in Figure 29.3. Prior to field work, survey site locations were selected in elements of each landscape pattern. Access considerations, landform element size and overall coverage of the islands also influenced wherever survey sites were placed. 29.2.4 Survey analysis DTA models were fitted using the See5 program (http://www.rulequest.com/) to predict the land unit, and their component soil family and vegetation classes from latitude, longitude, elevation, slope, wetness index, wetness class, landform class, erosion–deposition index, vegetation cover and clay index (Landsat TM band 5:7). To restrict the size of the tree that was generated, at least four observations were required for each the DTA leaf (final node). DTA models were developed on a training dataset (120 sites in the recent survey) and tested on
Figure 29.3. Landform pattern variation and field survey site coverage.
408
Ian D. Hollingsworth et al.
legacy survey data (108 sites) that had been collected to support forestry development projects in the southeast of Bathurst Island and the western end of Melville Island in the 1970s. The quality of the resulting maps was assessed based on the frequency of correct and incorrect allocations of these sites to classes. 29.3 Results The 9 aggregated classes, extracted from 27 land unit classes obtained from historical survey reports, exhibit a reasonable mapping result. We used a simplified vegetation classification (7 classes instead of 26) to achieve generalization in this case. A selection of the resulting land unit classes, vegetation classes and soil classes are presented in Table 29.1 along with the more detailed classifications that were extracted from the historical survey legends. The tabulated results for training and test datasets from models for land unit classes, vegetation classes and soil classes are shown in Tables 29.2–29.4. There is a considerable increase in the error of prediction between the training data and the test data. This is due to the fact that the test data selected from previous surveys were less accurately located (7100 m compared with 710 m) and were concentrated in a part of the islands (Plate 29a, see Colour Plate Section) not well represented in the training dataset, thus adding to errors in prediction. For example, in the test survey area, land unit 109 comprises undulating landscapes with underlying cretaceous clay sediments that were misclassified as the more extensive land unit 88, a similar landscape but underlain with unconsolidated sandstone and sand colluvium. These misclassifications were manually corrected in the map so as to agree with site observations. The land unit map is shown in Plate 29b (see Colour Plate Section). The red and orange colours indicate extensive areas of Ustoxs. The soil prediction model could not distinguish between gravelly and nongravelly phases of the soils (described as Hotham and Berrimah soil families in the previous land unit mapping). The grey colours represent Ustults and the blue colours Aquic soils. In the instance of the test survey area, the mapping model did not distinguish Ustoxs formed in residual sandstone from Ustults formed in cretaceous clay sediments in a reliable way. These inaccuracies in the mapping were ascribed to shortcomings in the ability to discern lithological boundaries using the clay index imagery. Woodland vegetation with a thick understory and grass cover would have obscured soil spectral signature except where land had been recently burnt. Deep Ustoxs on flat-to-undulating terrain occur in land units 85, 86, 88 and 90 showed surface horizon texture variation from sandy (Kiluppa) to sandy loam (Berrimah) and gravel content variation from nongravelly (Berrimah) to
Training data Total
Errors
(a)
(b)
(c)
33 11 25 5 0 18 3 23 2 120
3 2 1 1 0 1 0 8 0 16
23 2 1 1
1 7
2
(d)
(e)
(f)
(g)
5 1 1
1 22
(h)
(i)
2 1
4
2
1
1
17
1
2
3 2
15 2
Land unit classified as (a) 85 (b) 86 (c) 88 (d) 90 (e) 95 (f) 105 (g) 112 (h) 109 (i) 114 Error rate ¼ 13%
Test data Total 15 8 35 0 12 23 15 0 0 108
Errors
(a)
(b)
9 7 25 0 12 21 10 0 0 84
6 2 6
1
1 6 1
1
(c)
(d)
3
(e)
(f)
(g)
(h)
1 1 4
1 2 5
1
10
3 2 8
2
1 8
1 1
2 2 1
2 3 5
5 2 8
(i)
Land unit classified as
Rule-based land unit mapping of the Tiwi Islands, Australia
Table 29.2. Land unit classification tree evaluation.
(a) 85 (b) 86 (c) 88 (d) 90 (e) 95 (f) 105 (g) 112 (h) 109 (i) 114 Error rate ¼ 77%
409
410
Table 29.3. Soil classification tree evaluation. Training data Total
Errors
(a)
(b)
(c)
(d)
34 44 9 9 2 6 3 2 10 1 120
11 9 1 3 2 3 0 2 5 1 37
22 4
9 35 1 1
1 1 8
1 4
(e)
(f)
(g)
(h)
(i)
(j)
Soil family classified as
0
(a) Hotham (b) Berrimah (c) Mirrikau (d) Ramil (e) Irgil (f) Killuppa (g) Wangitti (h) Rinnamatta (i) Marrakai (j) Koolpinyah Error rate ¼ 31%
(j)
Soil family classified as
0
(a) Hotham (b) Berrimah (c) Mirrikau (d) Ramil (e) Irgil (f) Killuppa (g) Wangitti (h) Rinnamatta (i) Marrakai (j) Koolpinyah Error rate ¼ 73%
1
6
2
2 2
3
1
2
1
32
50
13
1 3 1 1
13
0
5
7
0
1 5 1 9
(g)
(h)
(i)
Test data Total
(a)
(b)
(c)
7 10 12 13 2 13 0 2 9 11 79
3 7 4 5 1 6
3 14 1 3 1 4
4 1 1 1
1 6 6 13
2 4 32
1
8
(d)
(e)
(f)
2 1 1
4
1
0
1
2 0 1 0 0 1 0 1 1 6
1 2 7 3 1
4 0
4
Ian D. Hollingsworth et al.
11 28 13 14 2 14 1 1 13 11 108
Errors
Training data Total
Errors
(a)
(b)
(c)
92 7 6 10 0 4 1 120
2 6 2 0 0 0 1 11
90 2 2
1
1 4
(d)
(e)
(f)
(g)
(h)
(i)
(j)
2 3
Vegetation classified as (a) Eucalypt forest (b) Melaleuca forest (c) Mangrove forest (d) Vine Forest (e) Sparse woodland (f) Grassland/sedge (g) Coastal woodland Error rate ¼ 9%
10 4 1
Test data Total
Errors
(a)
81 14 0 9 3 0 1 108
14 14 0 7 3 0 1 39
67 11 7 2 1
(b)
(c)
(d)
3 1
11 2 2 1
(e)
(f)
(g)
Vegetation classified as (a) Eucalypt forest (b) Melaleuca forest (c) Mangrove forest (d) Vine Forest (e) Sparse woodland (f) Grassland/sedge (g) Beaches/cheniers Error rate ¼ 36%
Rule-based land unit mapping of the Tiwi Islands, Australia
Table 29.4. Vegetation classification tree evaluation.
411
412
Ian D. Hollingsworth et al.
gravelly (Hotham). The Hotham, Berrimah and Kiluppa soils mapped relatively consistently with most of the misclassifications occurring as a result of confusion between each of these units. 29.4 Discussion and conclusions The traditional soil survey method produces a knowledge base that is expressed in soil map legends and can be formalized in terms of sets of mapping rules based on distributed environmental attributes (Bui, 2004). DTA produces sets of rules that are analogous to traditional survey methods. However, with this DSM technique, it is possible to upgrade the mapping, or knowledge base, as more field survey data become available. This approach would appear to have many advantages in the digital age: the ability to integrate with the knowledge accumulated over the years being not the least. However, statistical mapping methods such as those applied in the current Tiwi Islands study may not readily reproduce the level of map legend complexity used in routine survey work undertaken in the NT. This is because traditional surveys have a tendency to over specify land unit classes in relation to the field survey datasets, even in low relief landscapes with fairly monotonous lithology such as the Tiwi Islands. This has also been identified as a map quality issue in Australian land resources mapping (McKenzie and Austin, 1993). Consequently, the expectations of end users of land resources mapping may need to be modified or attuned to the need for further field survey work to meet specific information requirements. Also, alternative approaches to the one used in our study to defining test and training datasets could be considered. The location accuracy of legacy soil survey data is a recognized issue when this data have been used to develop test predictive models (Bui and Moran, 2001). Our test dataset comprised site information extracted from legacy soil surveys conducted 20–30 years ago. A cross-validation testing method would keep more information in the training set. However, because of the different methods used to select and locate sites, we decided to separate legacy and current survey data in our analysis. A general finding was that there is less error in the test and training datasets for vegetation class predictions than for soil and land unit class predictions (Tables 29.2 and 29.3). Plant communities tend to compensate for edaphic variation by changing their topographic positions (Guisan and Zimmermann, 2000). The simplification that we needed to make to apply data-based mapping techniques on the Tiwi Islands demonstrates the level of inference that can be drawn reliably from the amount of available survey information. The mapping that we have produced is general relative to the land unit classification system in
Rule-based land unit mapping of the Tiwi Islands, Australia
413
use in the NT and a review of the field survey database in the NT in relation to map unit descriptions and some test area surveys may be warranted to rationalize the land unit framework and to check and develop formalized mapping rules. The advantages of developing explicit rule-based mapping systems are that they would formalize current knowledge of soil and ecosystem variation and would integrate the current, disparate land resources information base using continuous environmental coverages, to create continuous land resources coverages and measurements of their reliability. We also found a much higher error of prediction for the test area than for the more extensive training area, which is probably explained by the poor delineation of different lithologies at the local scales. High-resolution lithological information and additional sampling of different lithologies will be needed to map land units according to the system in use elsewhere in the NT. The key outcome of our work is to demonstrate the use of digital topographic data (1:50,000) for land resources assessment and environmental analysis in the NT and to point out the limits to mapping inferences that can be made from field survey work. No assessment of the reliability of the traditional land unit survey information has been made. However, it is likely that land unit map legends overstate the inferences that can be made. A review of the NT database and map legend structure to produce an explicit set of knowledge-based mapping rules that can be tested and upgraded by the ongoing survey program may be worthwhile. DSM methods such as DTA could be integrated into this approach.
Acknowledgements We need to acknowledge the support and guidance of the Tiwi Land Council, Kate Haddon, Sylvatech Pty Ltd. and the Northern Territory Department of Business Infrastructure and Resource Development for this work.
References Belbin, L., 1987. The Use of Non-Hierarchical Allocation Methods for Clustering Large Sets of Data. Australian Comp. J. 19, 32–41. Brocklehurst, P. 1998. The History and natural Resources of the Tiwi Islands, Northern Territory. Chapter 4 – Vegetation, Parks & Wildlife Commission, N.T. Bui, E.N., 2004. Soil survey as a knowledge system. Geoderma. 120 (1–2), 17–26. Bui, E.N., Henderson, B.L., 2003. Vegetation indicators of salinity in northern Queensland. Austral Ecol. 28, 539–552. Bui, E.N., Loughhead, A., Corner, R., 1999. Extracting soil-landscape rules from previous soil surveys. Australian J. Soil Res. 37, 495–508. Bui, E.N., Moran, C.J., 2001. Disaggregation of polygons of surficial geology and soil maps using spatial modelling and legacy data. Geoderma 103, 79–94.
414
Ian D. Hollingsworth et al.
Cook, S.E., Corner, R.J., Grealish, G., Gessler, P.E., Chartres, C.J., 1996. A rule-based system to map soil properties. Soil Sci. Soc. Amer. J. 60, 1893–1900. Eamus, D., 2003. How does ecosystem water balance affect net primary productivity of woody ecosystems? Functional Plant Biol 30, 187–205. Franklin, J., McCullough, P., Gray, C., 2000. Terrain variables for predictive mapping of vegetation communities in Southern California. In: J. Williams and J. Gallant (Eds.), Terrian Analysis: Principals and Applictions. John Wiley and Sons, New York, p. 381. Gahegan, M., 2000. On the application of inductive machine learning tools to geographical analysis. Geogr. Anal. 32, 113–139. Gallant, J.C., Wilson, J.P., 1996. TAPES-G: a grid-based terrain analysis program for the environmental sciences. Comp. Geosci. 22, 713–722. Guisan, A., Zimmermann, N.E., 2000. Predictive habitat distribution models in ecology. Ecol. Model. 135, 147–186. Henderson, B.L., Bui, E.N., Moran, C.J., Simon, D.A.P., 2005. Australia-wide predictions of soil properties using decision trees. Geoderma 124, 383–398. McBratney, A.B., Mendonca Santos, M.L., Minasny, B., 2003. On digital soil mapping. Geoderma 117, 3–52. McDonald, R.C., Isbell, R.F., Speight, J.G., Walker, J., Hopkins, M.S., 1996. Australian Soil and Land Survey Field Handbook. Inkata Press, Melbourne, 250 pp. McKenzie, N.J., Austin, M.P., 1993. A quantitative Australian approach to medium and small scale surveys based on soil stratigraphy and environmental correlation. Geoderma 57, 329–355. Nott, J., 1994. Long-Term Landscape Evolution in the Darwin Region and Its Implications for the Origin of Landsurfaces in the North of the Northern-Territory. Australian J. Earth Sci. 41, 407– 415. Olsen, C.J., 1980. A Report on the Land Resources of South East Bathurst Island. LC 80/2, Land Conservation Unit. Conservation Commission of the Northern Territory, Darwin, N.T. Troch, P.A., Detroch, F.P., Brutsaert, W., 1993. Effective Water-Table Depth to Describe Initial Conditions Prior to Storm Rainfall in Humid Regions. Water Resour. Res. 29, 427–434. Van Cuylenburg, H.R.M. and Dunlop, C.R., 1973. Land Units of the Seventeen Mile Plain, Melville Island. 14, Animal Industry and Agricultural Branch Department of the Northern Territory, Darwin NT. Wells, M.R. and Cuylenburg, H.R.M.v., 1978. Land Units of Areas Adjacent to the Tuyu and Yapilika Forestry Plantations, Melville Island, N.T. LC78/9, Land Conservation Unit, Territory Parks and Wildlife Commission. Wells, M.R., Cuylenburg, H.R.M.v. and Dunlop, C.R., 1978. Land Systems of the Western Half of Melville Island, NT. LC78/10, Land Conservation Unit Report, Territory Parks and Wildlife Commission, Darwin. Wilson, J.P., Gallant, J.C., 2000. Terrain Analysis Principles and Applications. John Wiley & Sons, New York, 469 p.
Plate 29. Tiwi Islands, Northern Territory, Australia (a) Locations of training and test survey sites, (b) Landform units – reddish colours indicating Oxisols, grey Ultisols and blues Aquic soils.
401
Chapter 29
RULE-BASED LAND UNIT MAPPING OF THE TIWI ISLANDS, NORTHERN TERRITORY, AUSTRALIA Ian D. Hollingsworth, Elisabeth N. Bui, Inakwu O.A. Odeh and Phillip McLeod
Abstract We have applied a decision tree analysis (DTA) to map soil, plant community and land unit classes across the Tiwi Islands (7320 km2), located in northern Australia. Our survey substituted environmental analysis and DTA for traditional air photo interpretation to provide a continuous land unit coverage over the islands. DTA was used to derive mapping rules for land units, their component vegetation classes and soil families from secondary survey site observations and distributed environmental data. The mapping was tested on a legacy data set. The environmental variables used are: elevation, slope, latitude, longitude, landform pattern class, wetness class, static wetness index, erosion or deposition index, Landsat TM band 5:7 and vegetation cover class extracted from digital topographic 1:50,000-scale mapping. We needed to reduce the number of the resulting land unit classes derived from historical surveys into nine classes so as to produce a more meaningful mapping. This was achieved by generalising the component vegetation and soil units of the land unit classes thus producing the broader land unit classification. This procedure dovetails well with the current survey approach in which surveyors often develop more detailed classification systems than can be accurately mapped. We recommend a predictive mapping approach based on explicit mapping rules that integrate available land resources information and facilitate production of upgradeable maps.
29.1 Introduction 29.1.1 Background Thematic map products of land resources survey in Australia have been criticized for not making use of the range of distributed environmental information that is now available and for not producing the land information of a quality required by modern users (Cook et al., 1996; McBratney et al., 2003; McKenzie and Austin, 1993). Land resource assessment in the Northern Territory (NT) is based on air photo interpretation of map units and descriptive surveys (McDonald et al., 1996) to represent soil, landform and habitat diversity in a landscape where the native vegetation is largely intact. The land unit map
402
Ian D. Hollingsworth et al.
product of this process is widely used for development and conservation planning. However, there are gaps and inconsistencies in the resulting coverage because much of the survey work has been project based. Opportunities now exist to upgrade and integrate the land resource maps using extensive and ubiquitous digital elevation models across northern Australia. Digital soil mapping (DSM) methods could potentially be used to improve the quality and extent of land resource mapping in an economic environment where public funding of field survey work is declining. Several DSM techniques have been reviewed extensively by McBratney et al. (2003). A most relevant technique to our work here is, decision tree analysis (DTA), which could be used to generate rule-based mapping analogous to expert systems used by traditional land resource surveyors. Guisan (2000) describes a broad range of applications using this technique for quantitative modelling of ecosystem. DTA has been used to map vegetation properties (Franklin et al., 2000) and to model habitat distribution (Guisan and Zimmermann, 2000) but it is prone to error when extrapolated to unfamiliar landscapes (Gahegan, 2000). DTA was used in Australia to extract soil mapping rules from extensive geology and DEM-derived attributes (Bui et al., 1999) and to produce continuous soil property maps from disparate soil survey data (Bui and Henderson, 2003; Henderson et al., 2005). However, although the results of the traditional land resource approach are generally thematic maps, DTA has not been used at local-to-regional landscape scales to map land resources including soil. The underlying main aim of the work reported here was to make a continuous and consistent assessment of land resources of the Tiwi Islands, comprising Bathurst and Melville Islands (7320 km2) for the purpose of strategic development planning. The islands are located in tropics at 121 S, 1301 E (Fig. 29.1). Disparate land unit maps had been created over parts of the Tiwi Islands in the 1970s to assess land capability for forestry and agriculture (Olsen, 1980; Van Cuylenburg and Dunlop, 1973; Wells and Cuylenburg, 1978; Wells et al., 1978). Since this time, digital 1:50,000 topographic mapping (capable of deriving a 50-m DEM) over the Tiwi Islands has become available and a range of DSM techniques based on the analysis of DEM data has been developed. In this chapter, we utilised a DTA mapping approach in an analogous manner to conventional land resource assessment as practiced in the NT making use of legacy soil survey data, digital topographic mapping and Landsat TM data. The aim was to demonstrate the technique in a way that could be readily used in the routine survey program to rationalise map units and extrapolate existing mapping. Because climate and lithology are relatively uniform across Tiwi Islands (Nott, 1994) we focused on sampling physiographic factors controlling soilscape diversity. We utilized digital mapping methods in a first-pass
Rule-based land unit mapping of the Tiwi Islands, Australia
403
Figure 29.1. Location of the Tiwi Islands in Australia. assessment of the land resources over the full extent of Tiwi Islands, a remote region where traditional land resource mapping has been difficult to apply. 29.1.2 Environment The Tiwi Islands have an equatorial savannah climate that is dominated by northwest monsoon and the southeast trade weather patterns. The islands experience seasonal drought during the southeast trade season between May and September. The surface hydrology exerts considerable control over habitat diversity in this climate because water supply is a limiting factor for plant growth (Eamus, 2003). The geology is relatively simple and consists of quaternary alluvium, flat bedded, tertiary Van Diemen sandstone and the underlying Cretaceous Mookinu mudstone and Wangarlu mudstone members of the Bathurst Island formation. Tertiary Van Diemen sandstone is a fine-to-medium grained quartzose sandstone of fluvial and partly littoral origin. The sandstone covers most of Melville and Bathurst Islands and varies in thickness to a maximum of 80 m and dips gently to the northwest. The origin of the clastic sediments in the sandstone were highlands, located to the south of the Tiwi Islands during the Tertiary Age (Nott, 1994). Dissected plateau remnants 150-m high form the current highlands in the centre of the islands, which are fringed by coastal mangroves and
404
Ian D. Hollingsworth et al.
cheniers. Soils on sandstone are typically Oxisols (Ustoxs) with deep sandy to sandy loam surface horizons. Soils originating from mudstone are typically Ultisols (Ustults) with strongly acid subsoils overlain with a layer of pisolithic gravel. 29.2 Materials and methods 29.2.1 Definition of land units Land units, as described in the NT land resources survey, represent unique subclass combinations of geomorphic unit, lithology, vegetation and soil classification that occur as a repetitive pattern within broader land systems. The land units are delineated from stereographic air photo interpretation and field survey observations onto 1:100,000-scale topographic base maps. The complexity of the unit concept and the subjectivity of subclass definitions in different surveys often lead to disparity in the quality and types of unit being used in different surveys. We correlated historical soil and vegetation survey reports of the Tiwi Islands to produce a legend of 27 land units describing 10 soil family classes and 26 plant community types (Brocklehurst, 1998; Olsen, 1980; Wells and Cuylenburg, 1978; Wells et al., 1978) an excerpt of which is produced in Table 29.1. This approach was consistent with the current land unit classification system used in the NT, which concatenates landform class, soil class and vegetation class to construct a unique land unit class. 29.2.2 Landscape analysis The ANUDEM, program, was used to generate 50-m grid by fitting a splined elevation surface to 1:50,000 digital contours interspersed with spot heights, drainage network and coastline data. TAPES-G (Gallant and Wilson, 1996; Wilson and Gallant, 2000) was used to derive slope, steady-state wetness index and a sediment–erosion index to characterise the spatial distribution of soil water content and erosion–deposition processes. The steady-state wetness index is a derivative of a specific catchment area that could be used as a surrogate for subsurface flow. This is most appropriate to a humid environment (Troch et al., 1993). The steady-state wetness index (o) is defined as: As ¼ ln tan b where As is the specific catchment area (catchment area draining across a unit width of contour; m2/m) and b is the slope angle (in degrees). This index is similar to the specific catchment area, or upslope area per width of contour that has been used widely in soil property mapping from digital terrain data (Wilson and Gallant, 2000). The wetness index was divided into six quantiles to
Rule-based land unit mapping of the Tiwi Islands, Australia
405
Table 29.1. Land unit, vegetation and soil classes used. Mapped land unit
Land unit legend code
Landform element
(a) 85
Gec084
Foot slopes
(a) Eucalypt forest
Gaq085
Hill crests
(a) Eucalypt forest
Gfc085
Slopes
(a) Eucalypt forest
1a: E. miniata, E. tetrodonta and E. nesophila openforest
Udf085
Slopes
(a) Eucalypt forest
1a: E. miniata, E. tetrodonta and E. nesophila openforest
Gaq086
Summit surfaces
(a): Eucalypt forest
1b: E. miniata and E. tetrodonta open forest/ woodland
Laq086
Summit surfaces
(a): Eucalypt forest
1b: E. miniata and E. tetrodonta open forest/ woodland
Lfc087
Plains
(a) Eucalypt forest
1b: E. miniata and E. tetrodonta open forest/ woodland
Gec088
Fan
(a) Eucalypt forest
Uec088
Fan
(a) Eucalypt forest
1b: E. miniata and E. tetrodonta open forest/ woodland 1b: E. miniata and E. tetrodonta open forest/ woodland
(b) 86
(c) 88
Mapped vegetation unit
Described vegetation community 1a: E. miniata, E. tetrodonta and E. nesophila openforest with Chrysopogon fallax grassland understorey 1a: E. miniata, E. tetrodonta and E. nesophila openforest
Soil family
K11 Koolpinyah: deep, gravelly, imperfectly drained, yellow sandy loam over sandy clay loam K9 Hotham: deep, gravelly, well drained, red, sandy loam over sandy clay K9 Hotham: deep, gravelly, well drained, red, sandy loam over sandy clay K9 Hotham: deep, gravelly, well drained, red, sandy loam over sandy clay K8 Berrimah: very deep, well drained, red, sandy loam over acidic, sandy clay K7 Berrimah: deep, slightly or nongravelly, well drained, red, sandy loam over sandy clay loam K8 Berrimah: very deep, well drained, red, sandy loam over acidic, sandy clay T6 Cockatoo: deep, nongravelly, well drained, red sandy T6 Cockatoo: deep, nongravelly, well drained, red sandy soils
Ian D. Hollingsworth et al.
406
Figure 29.2. Static soil wetness index derived from 50-m DEM and classified into six quantiles. represent the distribution of runon and runoff elements in the landscape (Fig. 29.2). While no direct association with traditional landform elements was made, the visible distribution of wetness index quantiles in the landscape indicated a close association with the surface drainage system and hillslope topography. The erosion–deposition index (DTc), a dimensionless sediment transport capacity, was also computed using TAPES-G as a nonlinear function of specific discharge and slope, expressed as h n n i DTcj ¼ Am Am sj sin bj sj sin bj where b is the slope (in degrees) and As is the specific catchment area or drainage area per unit width orthogonal to a flow line (m2/m). DTc represents the change in sediment transport capacity across a grid cell, and can be used as a measure of the erosion or deposition potential in each grid cell (Wilson and Gallant, 2000). Uniform excess rainfall conditions were assumed. The published 1:250,000-scale geological map did not accurately discriminate outcrops of mudstone from extensive flat-bedded sandstone. As an alternative, we used a clay index calculated as a ratio of Landsat TM bands 5 and 7 to assess the variation in lithology. Vegetation cover was assessed from vegetation layer in the 1:50,000 digital topographic mapping. The mapped classes are: mediumdensity woodland, scattered woodland, saline coastal flats and marine swamps, pine plantation, dense vegetation, mangrove, inland water, intertidal flat foreshore, bare areas, lakes, dunes and cheniers.
Rule-based land unit mapping of the Tiwi Islands, Australia
407
29.2.3 Survey design The stratified survey design covering 120 sites included 4 replicate observations in each of 6 elements of 5 landscape patterns. Landform element (50-m grid) and landform pattern (100 ha) variation was premapped using classification of the static soil wetness index at fine and coarse levels of resolution. First, six quantiles of the frequency distribution of static wetness index was used to represent landform element variation. Secondly, a network of regular 100-ha hexagons was overlaid on the wetness class grid and the areas of each wetness index class were cross-tabulated for each hexagon. Using an agglomerative classification, ALOC in the PATN software program (Belbin, 1987) five landform classes representing coarse (100-ha resolution), landform pattern variation were created. The landscape pattern variation and the survey site coverage are depicted in Figure 29.3. Prior to field work, survey site locations were selected in elements of each landscape pattern. Access considerations, landform element size and overall coverage of the islands also influenced wherever survey sites were placed. 29.2.4 Survey analysis DTA models were fitted using the See5 program (http://www.rulequest.com/) to predict the land unit, and their component soil family and vegetation classes from latitude, longitude, elevation, slope, wetness index, wetness class, landform class, erosion–deposition index, vegetation cover and clay index (Landsat TM band 5:7). To restrict the size of the tree that was generated, at least four observations were required for each the DTA leaf (final node). DTA models were developed on a training dataset (120 sites in the recent survey) and tested on
Figure 29.3. Landform pattern variation and field survey site coverage.
408
Ian D. Hollingsworth et al.
legacy survey data (108 sites) that had been collected to support forestry development projects in the southeast of Bathurst Island and the western end of Melville Island in the 1970s. The quality of the resulting maps was assessed based on the frequency of correct and incorrect allocations of these sites to classes. 29.3 Results The 9 aggregated classes, extracted from 27 land unit classes obtained from historical survey reports, exhibit a reasonable mapping result. We used a simplified vegetation classification (7 classes instead of 26) to achieve generalization in this case. A selection of the resulting land unit classes, vegetation classes and soil classes are presented in Table 29.1 along with the more detailed classifications that were extracted from the historical survey legends. The tabulated results for training and test datasets from models for land unit classes, vegetation classes and soil classes are shown in Tables 29.2–29.4. There is a considerable increase in the error of prediction between the training data and the test data. This is due to the fact that the test data selected from previous surveys were less accurately located (7100 m compared with 710 m) and were concentrated in a part of the islands (Plate 29a, see Colour Plate Section) not well represented in the training dataset, thus adding to errors in prediction. For example, in the test survey area, land unit 109 comprises undulating landscapes with underlying cretaceous clay sediments that were misclassified as the more extensive land unit 88, a similar landscape but underlain with unconsolidated sandstone and sand colluvium. These misclassifications were manually corrected in the map so as to agree with site observations. The land unit map is shown in Plate 29b (see Colour Plate Section). The red and orange colours indicate extensive areas of Ustoxs. The soil prediction model could not distinguish between gravelly and nongravelly phases of the soils (described as Hotham and Berrimah soil families in the previous land unit mapping). The grey colours represent Ustults and the blue colours Aquic soils. In the instance of the test survey area, the mapping model did not distinguish Ustoxs formed in residual sandstone from Ustults formed in cretaceous clay sediments in a reliable way. These inaccuracies in the mapping were ascribed to shortcomings in the ability to discern lithological boundaries using the clay index imagery. Woodland vegetation with a thick understory and grass cover would have obscured soil spectral signature except where land had been recently burnt. Deep Ustoxs on flat-to-undulating terrain occur in land units 85, 86, 88 and 90 showed surface horizon texture variation from sandy (Kiluppa) to sandy loam (Berrimah) and gravel content variation from nongravelly (Berrimah) to
Training data Total
Errors
(a)
(b)
(c)
33 11 25 5 0 18 3 23 2 120
3 2 1 1 0 1 0 8 0 16
23 2 1 1
1 7
2
(d)
(e)
(f)
(g)
5 1 1
1 22
(h)
(i)
2 1
4
2
1
1
17
1
2
3 2
15 2
Land unit classified as (a) 85 (b) 86 (c) 88 (d) 90 (e) 95 (f) 105 (g) 112 (h) 109 (i) 114 Error rate ¼ 13%
Test data Total 15 8 35 0 12 23 15 0 0 108
Errors
(a)
(b)
9 7 25 0 12 21 10 0 0 84
6 2 6
1
1 6 1
1
(c)
(d)
3
(e)
(f)
(g)
(h)
1 1 4
1 2 5
1
10
3 2 8
2
1 8
1 1
2 2 1
2 3 5
5 2 8
(i)
Land unit classified as
Rule-based land unit mapping of the Tiwi Islands, Australia
Table 29.2. Land unit classification tree evaluation.
(a) 85 (b) 86 (c) 88 (d) 90 (e) 95 (f) 105 (g) 112 (h) 109 (i) 114 Error rate ¼ 77%
409
410
Table 29.3. Soil classification tree evaluation. Training data Total
Errors
(a)
(b)
(c)
(d)
34 44 9 9 2 6 3 2 10 1 120
11 9 1 3 2 3 0 2 5 1 37
22 4
9 35 1 1
1 1 8
1 4
(e)
(f)
(g)
(h)
(i)
(j)
Soil family classified as
0
(a) Hotham (b) Berrimah (c) Mirrikau (d) Ramil (e) Irgil (f) Killuppa (g) Wangitti (h) Rinnamatta (i) Marrakai (j) Koolpinyah Error rate ¼ 31%
(j)
Soil family classified as
0
(a) Hotham (b) Berrimah (c) Mirrikau (d) Ramil (e) Irgil (f) Killuppa (g) Wangitti (h) Rinnamatta (i) Marrakai (j) Koolpinyah Error rate ¼ 73%
1
6
2
2 2
3
1
2
1
32
50
13
1 3 1 1
13
0
5
7
0
1 5 1 9
(g)
(h)
(i)
Test data Total
(a)
(b)
(c)
7 10 12 13 2 13 0 2 9 11 79
3 7 4 5 1 6
3 14 1 3 1 4
4 1 1 1
1 6 6 13
2 4 32
1
8
(d)
(e)
(f)
2 1 1
4
1
0
1
2 0 1 0 0 1 0 1 1 6
1 2 7 3 1
4 0
4
Ian D. Hollingsworth et al.
11 28 13 14 2 14 1 1 13 11 108
Errors
Training data Total
Errors
(a)
(b)
(c)
92 7 6 10 0 4 1 120
2 6 2 0 0 0 1 11
90 2 2
1
1 4
(d)
(e)
(f)
(g)
(h)
(i)
(j)
2 3
Vegetation classified as (a) Eucalypt forest (b) Melaleuca forest (c) Mangrove forest (d) Vine Forest (e) Sparse woodland (f) Grassland/sedge (g) Coastal woodland Error rate ¼ 9%
10 4 1
Test data Total
Errors
(a)
81 14 0 9 3 0 1 108
14 14 0 7 3 0 1 39
67 11 7 2 1
(b)
(c)
(d)
3 1
11 2 2 1
(e)
(f)
(g)
Vegetation classified as (a) Eucalypt forest (b) Melaleuca forest (c) Mangrove forest (d) Vine Forest (e) Sparse woodland (f) Grassland/sedge (g) Beaches/cheniers Error rate ¼ 36%
Rule-based land unit mapping of the Tiwi Islands, Australia
Table 29.4. Vegetation classification tree evaluation.
411
412
Ian D. Hollingsworth et al.
gravelly (Hotham). The Hotham, Berrimah and Kiluppa soils mapped relatively consistently with most of the misclassifications occurring as a result of confusion between each of these units. 29.4 Discussion and conclusions The traditional soil survey method produces a knowledge base that is expressed in soil map legends and can be formalized in terms of sets of mapping rules based on distributed environmental attributes (Bui, 2004). DTA produces sets of rules that are analogous to traditional survey methods. However, with this DSM technique, it is possible to upgrade the mapping, or knowledge base, as more field survey data become available. This approach would appear to have many advantages in the digital age: the ability to integrate with the knowledge accumulated over the years being not the least. However, statistical mapping methods such as those applied in the current Tiwi Islands study may not readily reproduce the level of map legend complexity used in routine survey work undertaken in the NT. This is because traditional surveys have a tendency to over specify land unit classes in relation to the field survey datasets, even in low relief landscapes with fairly monotonous lithology such as the Tiwi Islands. This has also been identified as a map quality issue in Australian land resources mapping (McKenzie and Austin, 1993). Consequently, the expectations of end users of land resources mapping may need to be modified or attuned to the need for further field survey work to meet specific information requirements. Also, alternative approaches to the one used in our study to defining test and training datasets could be considered. The location accuracy of legacy soil survey data is a recognized issue when this data have been used to develop test predictive models (Bui and Moran, 2001). Our test dataset comprised site information extracted from legacy soil surveys conducted 20–30 years ago. A cross-validation testing method would keep more information in the training set. However, because of the different methods used to select and locate sites, we decided to separate legacy and current survey data in our analysis. A general finding was that there is less error in the test and training datasets for vegetation class predictions than for soil and land unit class predictions (Tables 29.2 and 29.3). Plant communities tend to compensate for edaphic variation by changing their topographic positions (Guisan and Zimmermann, 2000). The simplification that we needed to make to apply data-based mapping techniques on the Tiwi Islands demonstrates the level of inference that can be drawn reliably from the amount of available survey information. The mapping that we have produced is general relative to the land unit classification system in
Rule-based land unit mapping of the Tiwi Islands, Australia
413
use in the NT and a review of the field survey database in the NT in relation to map unit descriptions and some test area surveys may be warranted to rationalize the land unit framework and to check and develop formalized mapping rules. The advantages of developing explicit rule-based mapping systems are that they would formalize current knowledge of soil and ecosystem variation and would integrate the current, disparate land resources information base using continuous environmental coverages, to create continuous land resources coverages and measurements of their reliability. We also found a much higher error of prediction for the test area than for the more extensive training area, which is probably explained by the poor delineation of different lithologies at the local scales. High-resolution lithological information and additional sampling of different lithologies will be needed to map land units according to the system in use elsewhere in the NT. The key outcome of our work is to demonstrate the use of digital topographic data (1:50,000) for land resources assessment and environmental analysis in the NT and to point out the limits to mapping inferences that can be made from field survey work. No assessment of the reliability of the traditional land unit survey information has been made. However, it is likely that land unit map legends overstate the inferences that can be made. A review of the NT database and map legend structure to produce an explicit set of knowledge-based mapping rules that can be tested and upgraded by the ongoing survey program may be worthwhile. DSM methods such as DTA could be integrated into this approach.
Acknowledgements We need to acknowledge the support and guidance of the Tiwi Land Council, Kate Haddon, Sylvatech Pty Ltd. and the Northern Territory Department of Business Infrastructure and Resource Development for this work.
References Belbin, L., 1987. The Use of Non-Hierarchical Allocation Methods for Clustering Large Sets of Data. Australian Comp. J. 19, 32–41. Brocklehurst, P. 1998. The History and natural Resources of the Tiwi Islands, Northern Territory. Chapter 4 – Vegetation, Parks & Wildlife Commission, N.T. Bui, E.N., 2004. Soil survey as a knowledge system. Geoderma. 120 (1–2), 17–26. Bui, E.N., Henderson, B.L., 2003. Vegetation indicators of salinity in northern Queensland. Austral Ecol. 28, 539–552. Bui, E.N., Loughhead, A., Corner, R., 1999. Extracting soil-landscape rules from previous soil surveys. Australian J. Soil Res. 37, 495–508. Bui, E.N., Moran, C.J., 2001. Disaggregation of polygons of surficial geology and soil maps using spatial modelling and legacy data. Geoderma 103, 79–94.
414
Ian D. Hollingsworth et al.
Cook, S.E., Corner, R.J., Grealish, G., Gessler, P.E., Chartres, C.J., 1996. A rule-based system to map soil properties. Soil Sci. Soc. Amer. J. 60, 1893–1900. Eamus, D., 2003. How does ecosystem water balance affect net primary productivity of woody ecosystems? Functional Plant Biol 30, 187–205. Franklin, J., McCullough, P., Gray, C., 2000. Terrain variables for predictive mapping of vegetation communities in Southern California. In: J. Williams and J. Gallant (Eds.), Terrian Analysis: Principals and Applictions. John Wiley and Sons, New York, p. 381. Gahegan, M., 2000. On the application of inductive machine learning tools to geographical analysis. Geogr. Anal. 32, 113–139. Gallant, J.C., Wilson, J.P., 1996. TAPES-G: a grid-based terrain analysis program for the environmental sciences. Comp. Geosci. 22, 713–722. Guisan, A., Zimmermann, N.E., 2000. Predictive habitat distribution models in ecology. Ecol. Model. 135, 147–186. Henderson, B.L., Bui, E.N., Moran, C.J., Simon, D.A.P., 2005. Australia-wide predictions of soil properties using decision trees. Geoderma 124, 383–398. McBratney, A.B., Mendonca Santos, M.L., Minasny, B., 2003. On digital soil mapping. Geoderma 117, 3–52. McDonald, R.C., Isbell, R.F., Speight, J.G., Walker, J., Hopkins, M.S., 1996. Australian Soil and Land Survey Field Handbook. Inkata Press, Melbourne, 250 pp. McKenzie, N.J., Austin, M.P., 1993. A quantitative Australian approach to medium and small scale surveys based on soil stratigraphy and environmental correlation. Geoderma 57, 329–355. Nott, J., 1994. Long-Term Landscape Evolution in the Darwin Region and Its Implications for the Origin of Landsurfaces in the North of the Northern-Territory. Australian J. Earth Sci. 41, 407– 415. Olsen, C.J., 1980. A Report on the Land Resources of South East Bathurst Island. LC 80/2, Land Conservation Unit. Conservation Commission of the Northern Territory, Darwin, N.T. Troch, P.A., Detroch, F.P., Brutsaert, W., 1993. Effective Water-Table Depth to Describe Initial Conditions Prior to Storm Rainfall in Humid Regions. Water Resour. Res. 29, 427–434. Van Cuylenburg, H.R.M. and Dunlop, C.R., 1973. Land Units of the Seventeen Mile Plain, Melville Island. 14, Animal Industry and Agricultural Branch Department of the Northern Territory, Darwin NT. Wells, M.R. and Cuylenburg, H.R.M.v., 1978. Land Units of Areas Adjacent to the Tuyu and Yapilika Forestry Plantations, Melville Island, N.T. LC78/9, Land Conservation Unit, Territory Parks and Wildlife Commission. Wells, M.R., Cuylenburg, H.R.M.v. and Dunlop, C.R., 1978. Land Systems of the Western Half of Melville Island, NT. LC78/10, Land Conservation Unit Report, Territory Parks and Wildlife Commission, Darwin. Wilson, J.P., Gallant, J.C., 2000. Terrain Analysis Principles and Applications. John Wiley & Sons, New York, 469 p.
Plate 29. Tiwi Islands, Northern Territory, Australia (a) Locations of training and test survey sites, (b) Landform units – reddish colours indicating Oxisols, grey Ultisols and blues Aquic soils.