8. Cokriging. GenStat will predict the values of any one variate from values of it and others coregionalized with it at points or in blocks by ordinary cokriging using the models fitted by the coregionalization directive. 9. Nested analysis. Geostatisticians are recognizing the merits of nested sampling and analysis for revealing the gross spatial structure of regionalized variables (see Chapters 2 and 9). The data can be analysed by ANOVA, but the solution is not unique; they are better analysed by residual maximum likelihood (REML) as devised by Patterson and Thompson (1971). GenStat has a comprehensive suite of facilities for analysing data from such designs. Webster and Oliver (2007) provide examples of GenStat code for these tasks. The standard operations can be done either by written commands in the GenStat lan- guage or by menu. In the latter case GenStat generates the code and saves it in an output log. This means that you have a full record, step by step, of what you have done and that you can modify the code to create more comprehensive programs.
A.2 VESPER
Budiman Minasny, Alex B. McBratney and Brett M. Whelan
Australian Centre for Precision Agriculture, The University of Sydney, NSW 2006, Australia
A.2.1 Background
VESPER (Variogram Estimation and Spatial Prediction plus Error) is a user-friendly
PC-windows software program that can calculate and model global local variograms and do global and local kriging in either punctual or block form. VESPER was developed to deal with the large volume of intensive data col- lected by on-the-go proximal soil and crop sensors (approximately 5000–65 000 data points per km2 ). The purpose is to represent the data (yield data or soil electri- cal conductivity, etc.) as a digital map at a regular grid interval. In most geostatistical software, spatial interpolation usually involves two separate steps: calculating and modelling of the variogram for the whole area followed by prediction at unsampled points by kriging on a regular grid over the area. There are several shortcomings to this approach: first the time taken to calculate an empirical variogram of the whole area can be excessive (e.g. a variogram for 100 000 data points can take hours to calculate), and secondly information is lost by assuming a single variogram model for the whole area, and which results in a smooth map. VESPER can accommodate the large number of data and take into account the local spatial structure. 316 Appendix: Software
Kriging with local variograms, also known as kriging and automated variogram modelling within a moving window, involves searching for the closest neigh- bourhood for each prediction site, estimating the empirical variogram from the neighbourhood, fitting a variogram model to the data automatically by a non-linear least squares approach, kriging with local neighbourhood and variogram parameters and calculating the uncertainty of kriging prediction. The program adapts itself spa- tially in the presence of distinct differences in local structure over the whole field. Local variogram estimation and kriging can preserve the true local spatial variation in the predictions. In most cases, local variograms could circumvent the problems of anisotropy and the need for trend analysis.
A.2.2 The Software
VESPER comes as executable files consisting of an interface (written in Visual ba-
sic) and a main computation program (written in Fortran). The execution of the program is through the interface. Figure A.1a shows the main interface panel, where input and output files are controlled. Input data containing Cartesian coordinates and the variable of interest are required as an ASCII text file. The output files record the specific session details, variogram model parameters, the prediction locations, values and associated prediction error. The variogram panel (Fig. A.1b) provides the choice of global (whole-area) or local variogram estimation. The variogram is estimated by Matheron’s method- of-moments. A comprehensive range of models (Fig. A.1c) can be fitted to the empirical variogram using four possible weighting procedures (Fig. A.1c). Nonlin- ear least-squares estimation is used in the model fitting process, minimizing (Jian et al. 1996): n X 2 RD wi O .hi / O .hi / ; (A.1) i D1
where wi is the weighting option, O .hi / is the estimated semivariance at distance hi
and O .hi / is the semivariance predicted by the model. The ‘goodness of fit’ of the models can be compared with the Akaike Information Criterion, AIC, and sum of squared errors. If a global variogram is required, the ‘Fit Variogram’ button provides access to an interactive calculation and modelling panel (Fig. A.1d); model parameters can be extracted for subsequent kriging. The global modelling panel also allows subjec- tive model fitting through interactive parameter control bars. This is useful with small data sets and applications where emphasis is needed at particular sample separations. Local variograms are calculated automatically for each neighbourhood during the local kriging process, but the maximum distance and number of lags required for estimating them may be set through the Variogram panel. Experience at the Australian Centre for Precision Agriculture has shown that an exponential model is Appendix: Software 317
Fig. A.1 Operational panels of VESPER: (a) file input/output control panel, (b) variogram panel showing available models, (c) weighting options for model fitting and (d) global variogram opera- tion window
usually the best for local variogram estimation of yield data. More complex models, e.g. Matérn models, often become unstable with automatic local fitting (the covari- ance matrix can become non positive-definite in some local areas). We recommend limiting model selection to either exponential or spherical models. The kriging panel (Fig. A.2) provides punctual or block options. It is possible to define the block size (if relevant), set neighbourhood limits based on radial distance or number of data points and manipulate the kriging region. For yield data, a block size close to the swath width is recommended (a 10 m2 block is sufficient for most combine harvester fronts at present). 318 Appendix: Software
Fig. A.2 Kriging panel of VESPER – showing controls for punctual or block kriging, neighbour- hood definition, boundary, grid and advanced options
An interpolation grid is used to define the location of the points to which the data will be predicted. The interpolation grid allows data that are collected at different in- tervals to be collocated and analysed further. The interpolation grid can be specified in one of the following options: When the field has a rectangular shape, a grid with regular distance can be defined. When the field has an irregular shape, the boundary can be manually defined and a grid with regular distance can be generated that is confined to the boundary area. A file containing a pre-defined grid can be specified. Users can define the spatial coordinates of the prediction. For most precision agriculture applications, the field boundary will provide the lim- its of the kriging region. VESPER provides the option of importing an existing boundary file or describing the field boundary using an interactive drawing tool. The prediction grid (at user-defined distances) may then be produced with the soft- ware or an existing grid file can be imported. These features are important for the continuity of prediction sites through time within a field. In addition, for specific applications, VESPER can also perform: Lognormal Kriging – transforms lognormal data before performing the interpo- lation process. Appendix: Software 319
Non-negative weight – used to ensure ‘extreme’ values do not produce irrational
results e.g. negative yield values. VESPER uses the method of Deutsch (1996) for correcting negative weights. Sigma2 (data uncertainty) – is a user-defined estimate of the variance or uncer- tainty in the data. In operation, VESPER provides a window displaying the operational progress (Fig. A.3). For all forms of kriging a prediction progress map is produced together
Fig. A.3 Local variogram, data neighbourhood and prediction point progress map for an area with: (a) low variability and (b) greater variability 320 Appendix: Software
with a count of visited versus total prediction sites. For local kriging, individual var- iograms and the fitted models are displayed for the search neighbourhood around each prediction point. Note in Fig. A.3a, b that this local method allows changes in local variability to be reflected in the variogram parameters for each prediction. The graphical progress facilities can be disengaged to increase the speed of the predic- tion process. The output for all kriging operations is an ASCII text file containing the pre- diction point location coordinates, the predicted value and the kriging variance. An input file detailing the exact settings for each prediction session is also saved to- gether with a report file logging global variogram parameters or the parameters of each local variogram depending on the operation. Other details of the data and krig- ing session are also recorded in this file for future reference. Maps of estimates and prediction variances (Fig. A.4) can be obtained at the end of kriging. VESPER is available as freeware from the ACPA at www.usyd.edu.au/su/agric/ acpa The CSIRO Precision Viticulture group produces a PostVesper tool which au- tomates the process of converting file output from the Vesper kriging program into raster format in ArcGIS.
A.2.3 Applications
VESPER has been used both for research and practical applications. Google Scholar identified about 135 papers that mentioned the use of VESPER for kriging. In Aus- tralia, it is being used routinely for making yield maps with data from commercially available yield monitors for cereals and grapes, for example. VESPER is used rou- tinely to format multi-year and multi-sensor data onto a single grid for multivariate analysis and the creation of potential management classes.
