Genetic Land - Modeling land use change using evolutionary algorithms

Future land use configurations provide valuable knowledge for policy makers and economic agents, especially under expected environmental changes such as decreasing rainfall or increasing temperatures, or scenarios of policy guidance such as carbon sequestration enforcement. In this paper, modelling land use change is designed as an optimization problem in which landscapes (land uses) are generated through the use of genetic algorithms (GA), according to an objective function (e.g. minimization of soil erosion, or maximization of carbon sequestration), and a set of local restrictions (e.g. soil depth, water availability, or landscape structure). GAs are search and optimization procedures based on the mechanics of natural selection and genetics. The GA starts with a population of random individuals, each corresponding to a particular candidate solution to the problem. The best solutions are propagated; they are mated with each other and originate “offspring solutions” which randomly combine the characteristics of each “parent”. The repeated application of these operations leads to a dynamic system that emulates the evolutionary mechanisms that occur in nature. The fittest individuals survive and propagate their traits to future generations, while unfit individuals have a tendency to die and become extinct (Goldberg, 1989). Applications of GA to land use planning have been experimented (Brookes, 2001, Ducheyne et al, 2001). However, long-term planning with a time-span component has not yet been addressed. GeneticLand, the GA for land use generation, works on a region represented by a bi-dimensional array of cells. For each cell, there is a number of possible land uses (U1, U2, ..., Un). The task of the GA is to search for an optimal assignment of these land uses to the cells, evolving the landscape patterns that are most suitable for satisfying the objective function, for a certain time period (e.g. 50 years in the future). GeneticLand develops under a multi-objective function: (i) Minimization of soil erosion – each solution is validated by applying the USLE, with the best solution being the one that minimizes the landscape soil erosion value; (ii) Maximization of carbon sequestration – each solution is validated by applying atmospheric CO2 carbon uptake estimates, with the best solution being the one that maximizes the landscape carbon uptake; and (iii) Maximization of the landscape economic value – each solution is validated by applying an economic value (derived from expert judgment), with the best solution being the one that maximizes the landscape economic value. As an optimization problem, not all possible land use assignments are feasible. GeneticLand considers two sets of restrictions that must be met: (i) physical constraints (soil type suitability, slope, rainfall-evapotranspiration ratio, and a soil wetness index) and (ii) landscape ecology restrictions at several levels (minimum patch area, land use adjacency index and landscape contagion index). The former assures physical feasibility and the latter the spatial coherence of the landscape. The physical and landscape restrictions were derived from the analysis of past events based on a time series of Landsat images (1985-2003), in order to identify the drivers of land use change and structure. Since the problem has multiple objectives, the GA integrates multi-objective extensions allowing it to evolve a set of non-dominated solutions. An evolutive type algorithm – Evolutive strategy (1+1) – is used, due to the need to accommodate the very large solution space. Current applications have about 1000 decision variables, while the problem analysed by GeneticLand has almost 111000, generated by a landscape with 333*333 discrete pixels. GeneticLand is developed and validated for a Mediterranean type landscape located in southern Portugal. Future climate triggers, such as the increase of intense rainfall episodes, is accommodated to simulate climate change . This paper presents: (1) the formulation of land use modelling as an optimization problem; (2) the formulation of the GA for the explicit spatial domain, (3) the land use constraints derived for a Mediterranean landscape, (4) the results illustrating conflicting objectives, and (5) limitations encountered.