Geoscience Reference
In-Depth Information
only perform better than the null model of no change over longer time periods,
where they predicted approximately 70 percent of the landscape correctly on a
pixel-by-pixel basis. Although they suggest how the landscape might change in the
future should the
status quo
be maintained, these models epitomise the predictive
'black-box' approach frequently critiqued by geographers and others (e.g., Sayer,
1992; Mac Nally, 2000). Nevertheless, statistical tools are being developed that
help explore relationships between a suite of predictor variables and the observed
data. For example, hierarchical partitioning (used by Millington et al., 2007) esti-
mates the contribution of each predictor to the total variance both in isolation and
in conjunction with all other variables. Using such methods shifts the emphasis
from producing the 'best' predictive model to isolating the variance explained by
each predictor (Mac Nally, 2000). Such approaches are far better suited to hypoth-
esis formulation than is the (often blind) search for the single 'best' predictive
model.
Simulation models
Simulation models are used for prediction (e.g., forecasting of response to change
using 'what if . . . ?' scenarios), synthesis and integration of data, and heuristic
insight. They range in representational detail from very simple cellular-automata
models to detailed agent-based representations of the decision-making process in
spatio-temporally dynamic landscapes. There is a tension in simulation modelling
of LUCC between models emphasising the ecological heterogeneity of the landscape
at the expense of representing the actors engaged in decision making and
vice
-
versa
.
This divide between landscape and actor has perhaps arisen due to the different foci
of the various disciplines modelling LUCC (Veldkamp et al., 2001). In the social
sciences the emphasis is on understanding the micro-level motivations of decision
makers, whereas in ecology it is more on aggregate macro-level patterns of land use
and habitat, with the hope that the socio-economic drivers of change are subsumed
within the transition rules or probabilities. However, as Bockstael (1996) points
out this means that the nature of these drivers is not transparent; public versus
private and exogenous versus endogenous effects, for example, cannot easily be
disentangled.
The two most widely adopted types of simulation model are grid-based and
agent-based. Grid-based models (sometimes also called cellular or raster models)
have been much used for spatial modelling of LUCC, especially, but not exclusively,
by ecologists. In such models, the landscape is typically conceived of as a 2D
m
×
n
lattice, whose cells are internally homogeneous, with their state described by either
a categorical (e.g., habitat type) or continuous (e.g., land value) variable. The cell
size used will vary depending on the problem being addressed and may range from
sub-meter (e.g., individual plants) to km
(e.g., broadscale landscape pattern).
Representations of change in grid-based models take a variety of forms including
transition matrix approaches, simple quantitative neighbourhood rules, or more
complicated hybrid semi-qualitative approaches (Perry and Enright, 2006).
Jenerette and Wu (2001) used a grid-based model to explore patterns of urban
LUCC near Phoenix, Arizona. They employed a spatially explicit Markov approach
in which transitions were a function of neighbourhood conditions. They developed
models at two spatial grains: 250
+
75 m (fi ne). Jenerette
and Wu used a parameterisation based on observed transitions and another one
×
250 m (coarse) and 75
×
