7.2 Land use/cover change modelling

Objectives

• To describe the techniques available for modelling land use/cover change
• To implement an example of one such technique (Markov chains)

Overview – modelling land / use cover change

Modelling of changes in land use / cover over time is important for several reasons. Firstly, such models can be used to make predictions of future land use / cover. These predictions can then be used in many application areas, such as hydrological modelling where impermeable surfaces are particularly important, in assessing landscape changes and their impact on conservation and biodiversity, and in assessing likely future carbon budgets. Secondly, retrospective modelling of historic land use / cover change can provide valuable insights into underlying processes, as well as the effectiveness of land use planning policies. Various techniques are available for modelling land use/cover change over time. Several examples of such approaches are outlined below, but this is not an exhaustive list of all the various approaches available and many more exist.

Markov chains

A Markov chain is a relatively simple means of modelling how land use/cover changes over time. In a classic Markov chain, the probability that a given area will change its land use/cover between one time period and the next depends on only one thing: its land use / cover in the initial time period. In order to model the relationship between land use/cover in an initial time period, and land use/cover in a subsequent time, a transition probability matrix is often used (see example below). A transition is a change in land/use cover. Such a matrix shows the probability of a land use/cover change from one state to another taking place within a specified time period.

To model land use/cover change over a given period, random numbers can be generated for each grid cell or polygon within a landscape. In the example in the table above, a random number of 0.97 generated for a grid square that is currently cropland would suggest that the grid square would become woodland in the next period. Looking at the ‘cropland’ row (depicting current land use / cover state), random numbers 0.95 would indicate a transition to woodland, as in our example value of 0.97.

Cellular Automata

Whilst Markov chains are attractive because of their simplicity, they do make a number of assumptions that are unlikely to be met in ‘real world’ situations:

• the probability of a transition occurring in one grid cell or polygon is independent of corresponding probabilities in neighbouring grid cells. In reality, a change in one grid cell may well be accompanied by changes in surrounding grid cells and the pattern of land use / cover change will show spatial structure.
• the probability of a given transition occurring is constant across a landscape. In reality, some areas may be less prone to change – for example, where planning restrictions such as ‘green belts’ are in place.
• the probability of transition is invariant over time. Numerous studies suggest that this is not the case, and that transition probabilities may vary considerably from one period to the next. More generally, the Markov chain approach assumes that future land / use cover change will mirror past observed change.
• observed differences in land use/cover over time are affected by classification accuracy as well as genuine change on the ground. Land use / cover classifications derived through remote sensing are very rarely completely accurate. As a result, in comparing land use / cover at two points in time, the observed changes in land use / cover will be amplified by classification inaccuracy. A transition probability matrix that does not correct for classification accuracy is therefore likely to over-estimate the extent of change when predicting into the future.

For this reason, other approaches have been adopted that relax some of these assumptions. One such approach is the cellular automata methodology, in which the future state of a given grid square or parcel depends not only on its current state, but also on the current state of surrounding grid squares. The approach can be further modified so that – for example, transition probabilities depend on the suitability of a given grid square for a particular land use. Lau and Kam (2005) use this approach to model land use / cover change in Melbourne, Australia.

Land use scenarios

Even after taking the above issues into account by adopting approaches such as cellular automata, there remains a fundamental problem with approaches to land use / cover change modelling that are based on previous land use / cover histories: just because a pattern of change has been observed in the past, there is no guarantee that this trend will continue into the future. For this reason, several other approaches have been developed that do not predict future land use / cover change from past experience.

One such approach is the land use scenarios approach, classically outlined in a paper by Veldkamp and Fresco (1997). Under this approach, a panel of experts is convened to consider possible ‘alternative futures’ affecting the underlying drivers of land use / cover change. This expert panel will develop a set of possible concrete future scenarios for the underlying drivers that could affect land use / cover in a given area. For example, one scenario might be a massive surge in global demand for biofuels, with associated widespread conversion of land into crops such as maize and oil seed rape. A second scenario might be rapid urban economic growth, with associated rises in demand for land for housing and industry. A third scenario might be one of high government regulation of land use / cover, with development being tightly restricted in specific subregions. These scenarios are then translated into a land use / cover pattern by a GIS specialist, and the resultant land use / cover prediction referred back to the panel for review. The translation of scenarios affecting drivers of land use into predicted land use / cover requires considerable skill and specific software packages have been developed for this purpose. One such package is CLUE, available for free download (see readings below).

Land use / cover change, globalisation and commodity chains

Such approaches to modelling land use change often involve considerable simplification, for example in assuming that population increases inevitably lead to accelerated land use change. In a key paper, a set of leading land use experts highlighted some of the myths surrounding land use change modelling, such as the assumption that urban areas had little impact on rural land use change, or that global commodity chains were unimportant in driving patterns of land use / cover (Lambin et al, 2001). In an increasingly globalised world, it is now recognised that – for example – demand for timber from developed nations can be a major determinant of land use change in developing nations. Similarly, changes in environmental regulation in one nation, such as the introduction of policies to limit deforestation, can lead to the displacement of forest loss into neighbouring countries. More recently, therefore, land use change modelling has combined econometric modelling of global commodity chains and trading patterns with measurement of land use / cover change (Meyfroidt, Rudel, and Lambin, 2010).