1Australian Centre for Precision Agriculture
2Australian Cotton CRC
3Department of Agricultural Chemistry and Soil Science
The University of Sydney
McMillan Building, A05, The University of Sydney, NSW, 2006, Australia
Phone: +61 2 9351 3706, Fax: +61 2 9351 3706
E-mail: firstname.lastname@example.org, Alex.McBratney@acss.usyd.edu.au
Many agroecosystems face the dilemma of managing potential environmental risk at a farm scale whilst monitoring the same risk at a catchment scale. A model is proposed that enables the use of both broad-scale and fine-scale data sources in modelling local environs. A case study from the lower Hunter Valley is presented as an example of the application of the model in the real world. The study aims to predict digital terroir (local environs) through a) the deconstruction of existing broad-scale soil information and recombination with derived landform attributes and b) the aggregation and extrapolation of fine-scale data. The two models are used to refine each other and are further verified with independent ground-truthing. Results are still preliminary however strong spatial patterns are evident in the preliminary models.
Information technologies have found various applications in Australian agriculture, ranging from broad-scale catchment studies e.g. the Murray-Darling basin to fine-scale site-specific on-farm studies (precision agriculture). While the technologies used in these various studies have greatly enhanced our knowledge, the true value of the data is not realised until it has been ground-truthed. The ability to effectively ground-truth remotely and proximally sensed data is one of the biggest impediments to the successful adoption of information technologies in Australia. Unfortunately ground-truthing, especially on a large-scale, can be costly in terms of time and money, thus impractical in many situations. Information users must then rely on existing data to ground-truth their data.
In Australia existing environmental data sets contain their own impediments to use. The cost of intensively surveying a country the size of Australia, especially with a small population base, is not economically viable. Thus existing data is often base on sparse data sets and a lot of "expert knowledge". The density of the data sets varies greatly (usually proportionally to population density) as does the quality of data. Raw data is often unavailable leaving users at the mercy of errors in the original analysis. Finally environmental data is often on a scale incompatible with the newly sensed data resulting in a need to either aggregate or disaggregate existing data to make it compatible with the new data.
Aggregation (or disaggregation) of data presents several considerations for data analysis (see McBratney, 1998 for a more detailed discussion). One of the main considerations is what scale of data - raw or aggregated/disaggregated - is the model to be run on. Modelling prior to aggregation may produce a different result to modelling after aggeregation (Van Beurden and Douven, 1999). Modelling followed by aggregation appears the more sensible option as it allows analysis at the appropriate scale for the data. However environmental modelling, which draws on several baseline data sources, is usually done after aggregation/disaggregation (Van Beurden and Douven, 1999). This ensures all data are at the same scale and produces faster results. McBratney (1998) suggests a third approach whereby the modelling process also scales the output. This allows for the input of raw data and minimising the propagation of errors in the initial aggregation/disaggregation of data.
The availability of remote and proximal sensed data allows humankind to monitor and model their environment at greater detail than they ever have before. There is no reason to suggest that this ability to monitor and model will not continue to increase. However what is lacking is the knowledge of how to incorporate this multisensor data into a model to form an overall picture of environmental health. A wide variety of different scale environmental data are now available for use in the agriculture industry. These data are particularly useful in irrigation areas where there is a threat of land degradation from irresponsible application of water. To use this data effectively environmental variables need to be monitored at the catchment scale but the management performed on-farm at a finer spatial scale. Figure 1 attempts to reconcile these data to produce a local environ model that enables the environmental health of an agricultural production system to be monitored and maintained.
Figure 1 - A model to predict local environs for environmental management given a variety of different scaled data
The model uses a combination of disaggregation ("top-down") and aggregation ("bottom-up) techniques to predict the local environ. As mentioned before the collection of fine-scale data is expensive and time consuming but produces high quality data. Broad-scale data collection is inexpensive and relatively fast to obtain however it is usually of poor quality. The model in Figure 1 aims to utilise the broad-scale data as a template onto which the fine-scale data can be extrapolated. The model is really a feedback loop. Separate predictions of the local environ are produced by both disaggregation of broad-scale data and aggregation of fine-scale data. Both predictions are checked with ground-truthing, with the hypothesis being that the bottom-up prediction is more accurate. The similarity of the two predictions can be tested using statistical techniques such as the Simple Similarity Index (SSI) or Map Cross Correlation (MCC) (Van Beurden and Douven, 1999). Once validated and refined the bottom-up prediction, rather than the straight ground-truthed values, is extrapolated onto broad-scale data to refine the regional prediction of local environs. Site-directed sampling and/or the accumulation of further data over time can then further refine both predictions.
Once a suitable regional prediction has been produced the model can be used to produce maps of potential environmental risk e.g. salinity, soil acidification, subsurface water contamination. By identifying environmental sensitive areas in the risk assessments, a monitoring scheme can be established to maximise the value of the data obtained. These sites are termed environment control points (ECP's). If the development of land degradation is considered in terms of a production dynamics these ECP's are analogous to points in the production line where problems first arise. As environmental degradation has different sources the spatial location of ECP's for different types of environmental degradation will vary.
The Case Study
The model outlined above is currently being applied to describe the local environ (terroir) in a new irrigation scheme for viticulture in the lower Hunter Valley.
The Hunter Wine Country (HWC) is located in the lower Hunter valley approximately 200km north of Sydney. The region has a long tradition of viticulture dating back to the 1820's and the establishment of the oldest existing winery in Australia at Wyndham Estate, Dalwood. The valley may experience long hot dry spells during summer that may effect the crop at critical phenological stages e.g. anthesis, veraison. While several of the larger vineyards have irrigation set up from on-farm water stores or irrigation licenses from the Hunter river, the valley has had no large scale irrigation scheme until recently.
Over the past decade the Australian viticulture industry has seen an enormous increase in vineyard plantings. Reflecting this, new plantings as well as replantings have increased in the HWC. The increase in viticultural activity lead to a proposal for and eventually the construction of an irrigation scheme to "drought-proof" the lower hunter valley. A statutory body, the Private Irrigation District (PID), was established by local growers to fund and construct the Pokolbin Pipeline Project. The pipeline services 384 properties over 500km2.
The Australian continent has a long history of land degradation under irrigation. Salinity is the most prominent cause of degradation however sodicity, erosion and soil acidification are also a real threat. An environmental audit of the PID catchment by Creelman (2000) has revealed salinity in the supply water as well as the presence of salt in subsurface water. The siltstone and sandstone underlying much of the PID are contributing salt to the soil profile when weathered. Creelman (2000) also identified soil acidification from acidic rainfall from local heavy industry and long term fertiliser use in the valley as a potential problem.
In recognition of these dangers the DWLC in approving the irrigation scheme has stipulated that the PID (the licensee) is to "conduct soil, surface water and ground water quality monitoring...and to monitor the impact on soil and water resources as a result of development".
These guidelines pose several problems for the PID:
- Firstly there is a need to establish what the current environmental situation is too enable future generations to monitor whether there has been a degradation (or improvement) in the environment since the inception of the PID i.e. the construction of a current Geographic Information System (GIS).
- Secondly there is a need to design a monitoring system that enables the PID to monitor the local environment, that gives an early warning of any potential problems and is cost effective in the long run.
Before the second point can be addressed efficiently the GIS needs to be established and validated. The Department of Soil Science and Agricultural Chemistry at the University of Sydney has undertaken the construction of a GIS for the PID.
Constructing the GIS
The first step in creating the GIS was the collection of baseline data for the area bounded by the PID. This data included a DEM derived from existing topographic maps, a section of the DLWC's soil landscapes of the Singleton 1:250000 Sheet (Kovac and Lawrie, 1991), Landsat 7 +ETM imagery (August 1999 and Dec 2000), Spotlite panchromatic imagery, a commissioned EM 34 survey on an approximate 1km grid and cadastre data on roads, land subdivision and details of the PPP. All baseline data was registered using either WGS84 or MGA94. The datums are interchangeable.
Derivation of new environmental surfaces
Both the DEM and the Landsat data were provided on 25m grids. The DEM was reprojected onto the Landsat grid in ARC/INFO® to produce a common grid. Primary landform attributes, such as slope, aspect, profile curvature and plan curvature were derived directly from the DEM data using the methods of Zevenbergen and Thorne (1987).
Secondary landform attributes were also derived. Water flow attributes e.g. flow accumulation and upslope and downslope flow length were determined within the GRID module of ARC/INFO®. The secondary landform characteristics Compound Topographic Index (CTI) (Gessler et al, 1995), relative position on slope (rpos) and a topographic index (ATI) (Taylor et al, 2001) were also derived. Solar radiation and its' derivatives were calculated from the DEM and climatic data using the SRAD program (Moore, 1997)
Landsat data was used to derive landcover classes. At the time of analysis only the off-season image from August 1999 was available. The Landsat image was divided into 6 Landcover classifications using a supervised classification in ERDAS Imagine® based on bands 1-5 and 7 of the Landsat 7 +ETM image. Band 6 (thermal image) and band 8 (panchromatic) were omitted from the analysis as they were on different scales and contained little additional data on reflectance properties. A Normalised Differences Vegetation Index (NDVI) was also derived in Imagine® from Bands 2, 3 and 4 of the Landsat image.
Figure 2 - Soil landscape units of the PID (from Kovac and Lawrie, 1991)
Predicting a new soil map
Deconstruction of DLWC landscape units
The area bounded by the PID contains 9 soil landscape units from the Singleton sheet (Figure 2). Each pixel of the DEM/Landsat data set was assigned to a landscape unit using the INTERSECT function in ARC/INFO®. Of these 9 units the Branxton, Rothbury and Pokolbin units make up 93% of the PID area. Each soil landscape unit contains a typical catenary sequence and a typical profile description for each soil profile in the catenary sequence. The catenary sequence for the Branxton unit is displayed in Figure 3. The soil profiles specified in the soil landscape catenary sequences were expressed as Great Soil Groups (GSG). Using the profile descriptions in the Singleton sheet and descriptions from the Factual Key for the Recognition of Australians Soils (Northcote, 1979) these GSG's were keyed to the suborder level of the Australian Soil Classification system (Isbell, 1996). Each landscape unit was individually run through Fuzme (Minasny and McBratney, 2000) and segregated in classes based on slope degrees, rpos and ATI. The number of classes specified in Fuzme was equal to the number of soil profiles identified in the soil landscape catenary sequence.
Construction of a new soil map for the PID
Slope, ATI and rpos were used to allocate each Fuzme class to an ASC suborder. For example, in the Branxton sequence (Fig. 3) the class with a low rpos, low mean slope and the lowest ATI will fall into the Chromosol order (indicative of soils at the crest of hills). Whilst the class with a high rpos, low mean slope and highest ATI (indicative of soils at the base of a slope) will be classed as Tenosols. The resultant soil map is shown in Figure 4. Figure 4 is only expressed to the order level to allow definition between soil types.
Figure 3 - Catenary sequence for the Branxton soil landscape unit (from Kovac and Lawrie, 1991)
Figure 4 - Predicted soil map to ASC order for the PID
Prediction of Soil Attributes
Each soil profile in the catenary sequence contains a soil profile description and a reference to a profile description in the Factual key for the Recognition of Australian soils (Northcote, 1979) enabling the mapping of soil properties as well as soil type. Soil properties with numerical descriptions were taken as read. Theses properties included topsoil pH, soil salinity, topsoil and subsoil erodibility, profile thickness (depth to bedrock) and depth to watertable. Soil properties with linguistic descriptions were converted to a numerical scale for ease in analysis. The method of Minasny (2000) was used to express texture numerically using centroid values of the Australian texture triangle (Marshall, 1947).
Soil moisture estimates for each horizon were calculated using the pedotransfer functions of Minasny et al (1999). Available water content (AWC) was calculated assuming soil moisture potentials of -10 to -1500 kPa whilst readily available water (RAW) used potentials between -8 and -60 kPa). Profile moisture estimates were then estimated by multiplying AWC or RAW values by horizon thickness and summing horizons.
The concept of Terroir and its prediction
Terroir is a unique term specific to viticulture and oenology. It is used to classify vineyards according to the social and historical background of the vineyard as well as its physical environment (Wilson, 1998). For this exercise we have classified terroir solely on the physical environment. The objective is to identify similar discrete environs in the lower Hunter Valley. Thus the local environ in Figure 1 may be considered to be a digital terroir. If premium grapes can be identified as being produced within a certain “digital terroir”, other locations exhibiting the same digital terroir may be worth investigating for vineyard potential.
The prediction of digital terroir is based on the derived soil map. Each individual soil profile located within each DLWC landscape is isolated. If the particular soil profile has only a small distribution (<10ha), the profile is considered to be a discrete terroir within itself and no further analysis is performed. Data sets greater than 160 points are further subdivided using k-means clustering in JMP. The criteria for this classification were CTI, the predicted EM survey values at 10m spacing in the vertical position (EM_10V) and net radiation. CTI was selected to further characterize the soils in terms of catenary sequences. The EM_10V values were used as an alternative indication of soil type. The prediction map was interpolatd using regression kriging with primary and secondary landform attributes. Net radiation was included as a climatic influence on plant development and grape maturation. Clustering produces centroid (mean) values for each of the classes and assigns data points to one of the classes based on how well the data fit the centroid values
The number of classes (clusters) assigned to each profile by the operator was arbitrary. Larger data sets tended to have more classes however the classification was also based on how the classes were spatially orientated and how different the centroid values for the different classes were. The minimum number of classes was 1 and the maximum 10.
In all the nine landscape units yielded 41 discrete soil profiles which in turn produced 149 potential digital terroirs. A nomenclature was derived for naming the digital terroirs to minimize confusion. The digital terroirs were labeled as
DLWC Landscape id – ASC order and suborder – cluster class
Therefore the second cluster class in the yellow dermosol (AC DE) of the Pokolbin landscape unit is labeled
The large number of units (149) in the digital terroir map means that a full representation of the map is difficult (even in colour). Thus a section of the map that covers a vineyard currently being ground-truthed has been displayed in Figure 5. This has also been further simplified by reclassifying in Arcview® to remove smaller terroirs and conversion of the raster image into a shapefile.
Figure 5 - Predicted digital terroir map for a section of a vineyard within the PID
Prediction of Risk Assessment and Environmental Control Points.
The deconstruction of the soil landscape units and reconstruction of the data with landform attributes and profile description data has produced a more detailed picture of the soil and digital terroirs in the PID. If we assume that the soil and digital terroir map have been validated and are accurate, this data can now be used to identify areas where there is a risk of environmental degradation. To date three risks have been identified and calculated. These are soil acidification, soil salinity and soil erosion. (NB these predictions are currently based on the soil map thus contain many carry through errors from the deconstruction of the landscape units. While they have been derived, these risk predictions are of little value until the soil and/or digital terrior map has been refined and validated). Once risk assessments have been made it is possible to identify ECP's for each risk. The lack of confidence in current risk assessments has precluded the prediction of ECP's to date.
The aim of this study was to derive digital terroirs using properties of the physical environment for environmental management. To this extent both the soil and digital terroir maps show continuity and reasonable spatial structure rather than random noise. Terroirs within the vineyard shown are of sufficient size (>2ha) to potentially allow differential management. The spatial structure exhibited gives us some hope that the digital terroir prediction may hold some merit.
As indicated in the report this case study is only a preliminary investigation. If the GIS is to be an effective tool in monitoring and managing the local environment all steps in the modelling process need considerable improvement. Firstly the predicted surfaces, especially the soil and soil property maps require ground-truthing within the PID to establish how accurate the predictions are. The broad-scale predicted surfaces also need to be improved by extrapolation from the fine-scale prediction model. Many of the current prediction contain carry through errors from the DLWC soil landscape units. The influence of the soil landscapes units can be seen in the soil map with distinct edgeing still apparent between units. Ground-truthing has already begun with two honours projects currently being undertaken within the PID as well a detailed soil survey will be undertaken as part of the University's undergraduate soil science course.
The EM-34 survey needs to be updated. From the regression kriging model some landscape units produced apparent soil electrical conductivity predictions outside the range of actual values due to a lack of initial survey points within these areas. This is particularly noticeable in the digital terroir analysis where the outlying data points have been removed from the data sets before analysis and mapped in black. The model is particularly inaccurate in rugged terrain and along waterways where few measurements were taken. By identifying these areas a directed sampling can be devised to take measurements and improved predictions within these areas.
The identification of the digital terroir needs to be automated to remove the emphasis on expert knowledge. The model will also benefit from additional variables that are influential on plant productivity, e.g. rootzone depth, frost prediction and soil moisture storage as well as the improvement in the predicted environmental surfaces discussed above.
The GIS is restricted at the moment to essentially modelling soil and some limited soil-water interactions. More emphasis needs to be given to other data sources especially the Landsat image now that a within season image is available. The proper understanding of the impact of the irrigation system will also require a subsurface water model to monitor and model local watertable fluctuations.
Finally there is a need to devise a continuous monitoring system for the soil and water resources in the PID. The GIS can be used to identify potential “hotspots” where salinity, sodicity, acidity and erosion problems may first appear. Again further ground-truthing will be needed for the final sampling scheme. The development of a standardized monitoring kit will also be required to ensure the data gain from monitoring can be correctly analysed and incorporated into the GIS.
The proposed model is a way of reconciling data from a variety of different scales to enable environmental management to be coordinated on a regional scale but implemented on a farm size scale.
The environmental model outlined in the first part of the paper is being applied to a catchment study in the lower Hunter Valley. Results from the case study are encouraging however they are still preliminary and require further ground-truthing and model correction before any definite conclusions can be made.
Creelman, R.A., (2000). Geological and Geochemical framework for assessing salinity in the Hunter Valley. Water and Salinity Symposium, Sydney.
Gessler, P.E., Moore, I.D., McKenzie, N.J., and Ryan, P.J., (1995). Soil-landscape modelling and spatial prediction of soil attributes. Int. J. Geog. Info. Sys. 9, 421-432.
Hartigan, J.A. 1975. Clustering Algorithms, Academic Press, New York
Isbell, R.F., 1996. The Australian Soil Classification. Australian Soil and Land Survey Handbook. CSIRO Publishing, Australia.
Kovac, M. and Lawrie, J.W., 1991. Soil Landscapes of the Singleton 1:250 000 Sheet. Soil Conservation Service of NSW, Sydney.
Marshall, T.J. (1947). Mechanical composition of soil in relation to field descriptions of texture. Council for Scientific and Industrial Research Bulletin 224
McBratney, A.B. (1998). Some considerations on methods for spatially aggregating and disaggregating soil information. Nutrient Cycling in Agroecosytems 50, 51-62.
Minasny, B. and McBratney, A.B. 2000. FuzME Ver. 2.1. Australian Centre for Precision Agriculture. The University of Sydney. (www.usyd.edu.au/su/agric/acpa)
Minansy, B. (2000). Efficient methods for predicting soil hydraulic properties. PhD Thesis. University of Sydney.
Moore, I.D. (1997). Radiation Program; SRAD 3.1. Centre for Resource and Environmental Studies. The Australian National University, Canberra, ACT, Australia.
Northcote, K.H. (1979) A factual key for the recognition of Australian Soils 4th Edition. Rellim Technical Publications Pty. Ltd., Adelaide, South Australia
Van Beurden, A. U. C. J. and Douven, W. J. A. M. (1999) Aggregation issues of spatial information in environmental research. Int. J. Geographical Information Science. 13:5, 513-527
Wilson, J.E., (1998). Terroir: The role of geology, climate, and culture in the making of French wines. Mitchell Beazley, London