The application of landscape productivity and environmental modelling to improve plantation site selection and yield prediction

The area planted to Eucalyptus globulus (Labill.) in southern Australia has increased rapidly. However, robust site selection and productivity prediction models have only become available recently for this species. The silvicultural and management questions being explored with these models cover a range of temporal and spatial scales. The models themselves are complex, data intensive and highly non-linear in their function. This raises a number of important issues that affect model application. These issues are explored with a series of applied examples. Firstly, what is the effect of patchy input data of limited extent? Models designed using sampling plot data on which soils data are recorded are now applied across broad landscape extents to make site selection decisions and regional productivity estimates. The work in this thesis shows in separate examples how terrain analysis can be used to supplement input surfaces, including surfaces related to waterlogging and soil nutrition. An approach is developed for dealing with poorly defined soils variables such as soil depth and used to categorise regional forest productivity based not only on the expected yield but the co-efficient of variation in that yield. In a further example this approach is used to define appropriate site surveying procedures for plantation land assessment. In addition to examining the effect of model predictive accuracy on the quality of spatial inputs, input data grain size is examined to identify the scale sensitivity of the model. This is important because the same forest growth models are used to predict the yield of individual forest plots and to make continental scale predictions. The results indicate that forest productivity models like PROMOD are not scale dependent and even with the use of broad-scale input data provide unbiased estimates of mean productivity. However, a comparison using fine-scale data (10 x 10 metre) shows that sub-grid variance at large scales (1000 x 1000 metre) can be very large, particularly for soil variables that can change abruptly over short distances. Just as spatial input variables vary in quality and scale so may the temporal (particularly weather) variables supplied to models. Using rainfall representation in models as an example, an analysis of the intrinsic input data variability allows a more sophisticated consideration of site selection that includes risk evaluation as a criteria. Furthermore, it was shown that incorporating the distributional characteristics of rainfall markedly affects the prediction of the risk associated with a particular plantation development.