|
|
|
Land Use/Land Cover
|
Reconstruction of long term land cover changes by a maximum likelihood interpolation method using genetic algorithm
Data
- Input Data
1) History database of the global environment (HYDE)
In this study, History Database of the global Environment (HYDE) is used for input data, such as potential
and actual land cover data and fragmentary observational data. This datar, HYDE, has natural
background vegetation based on the BIOME model (Prentice et al, 1992). The biome model of Prentice
et al. (1992) is the first used to select which plant types may potentially be present at a particular site.
This rule-base captures the effects of minimum tem perature tolerances and chilling requirements on
determining the distributions of different plant types.
2) Land cover data at the start year and the end year:
In this study, year 1700 of HYDE is used as a land cover data at the start year and year 1990 of HYDE
are used as a land cover data at the end year. In land cover data at the start and end year, agriculture
land, pasture, human s ettlement, and intensive agriculture are considered as the grass class. To
generate the simulation data of land cover changes from the start year to the end year, land cover data of
HYDE at the start year and the end year are modified in to five classes that is explained in 2.2.
3) Point based observational data:
Point based observational data were collected from the HYDE. Poi nts of human activities such as
agriculture, pasture, and human settlements are picked up. These collected point based observational
data are the year of 1750, 1800, 1850, 1900, 1950, and 1970 of HYDE.
4) Cultivation intensity
Cultivation intensity data are overlaid to represent the impact of agricultural activities. Cultivation intensity
data is used to modify the actual vegetation data. In this study, if the pixel has value more than 50% in
cultivation intensity data, the class in the representing area in actual vegetation will be modified to
intensive agriculture.
5) Total area of agriculture area:
Total area of agriculture area is strongly related with the total population. In the interpolation, a restricted
condition should be taken into account in which the total agricultural area should be proportional to the
total population. However, the computational work can be very heavy if the restricted condition is clearly.taken into account. To avoid this, the interval of the time-slice in the interpolation neighboring time slices
which results in an almost constant growth rate of agriculture area expansion between the neighboring
time slice. Therefore, knowledge on the land cover changes can be very much simplified.
6) Transitional probability:
Knowledge on land cover change is given in terms of transitional probability from one class to another.
Transitional probability changes according to regional condition. In areas which are climatologically
suitable for high agriculture, the transitional probability from forest or grassland to agriculture areas is
relatively high. In areas where the possibility of wind erosion is high, the transitional probability from
grassland to barren or desert area is relatively high. In areas with a very high suitability of agriculture,
ordinary agricultural areas are likely to change to intensive agriculture area.
- Class category and time interval
In this study, class category is divided in to five land cover classes, forest, grass, APH, IA, barren-l and,
and water. APH means Agriculture land / Pasture / Human Settlement, and IA means Intensive
Agriculture. This class category is different from the categories used in the input data, “ History Database
of the global Environment (HYDE)”. These data should be reclassified to match the new five class
category. In this study, result of long term land cover changes is from 1900 to 1990. The time interval of
the result is 10 years.
Processing
- Three dimensional representation of an individual (coding)
In this study, a three dimensional array is defined to represent an individual in a space and time domain.
The horizontal plane represents two dimensional spaces, and the vertical dimension represents temporal
dimension.
- Initialization of population
An initial population for a gene tic algorithm is selected unsystematically. One random trial is made to
produce each individual. On the other hand, value of each member of initial population is the same
because all members of the initial population are automatically selected by the same procedure.
- Definition and computation of an individual’s fitness
1) Spatio-temporal behavioral models of class variable data
In the GA based interpolation, any types of behavioral models can be applied if they can determine the
probability of every possible behavior and transition of nominal or class variables. For nominal variable
data, possible changes in a class at one pixel are basically defined by the probability of the changes from
one class to another. In this study, transitional probability is determined by the combination of classes in
the neighborhood. Spatial and temporal relations affected the transitional probability in three ways. The
first is spatial continuity which is based on the assumption that the same class data tend to continue in
the spatial demotion. The second is temporal continuity which is an extension of the spatial continuity to
the temporal domain. The third is expansion contraction relations which is based on the assumption that
some data class have a higher possibility of expanding their area at the next time slice while others tend
to contract.
2) Definition and computation of fitness of an individual
Fitness of an individual is defined by the combination on behavioral fitness and observational fitness.
Behavioral fitness is the combined probability of a change in events of nominal variables under the
condition that these changes follow a given probabilistic behavioral model or rule. Observational fitness
is the combined probability that the observational nominal values occur under probabilistic functions of
observational error or uncertainties. Observational probability can be determined by accuracy, resolution
and frequency of observation. Overall fitness can be computed by multiplying behavioral fitness and
observational fitness. Thus, behavioral or structural models and observational data can be integrated by
optimizing the overall fitness.
- Definition of operators
1) Reproduction
Reproduction is a process in which individual strings are copied according to their objective function
values or the fitness values. Copying strings according to their fitness values means that strings with a
higher value have a higher probability of contributing one or more offspring in the next generation.
2) Crossover
The crossover operator first randomly mates newly reproduced individuals in the mating pool. It then
randomly locates a window of random size for a pair of individuals. Finally, the contents of the individual
within the window are swapped to create new individuals.
3) Mutation
Mutation is a genetic operator that alters one or more gene values in a chromosome from its initial state.
This can result in entirely new gene values being added to the gene pool. Mutation is an important part of
the genetic rearch to prevent the population from stagnating at any local optima.
- Improvement of the search
1) Hill-climbing method
If the complex space of problem resolutions becomes larger and larger, the population size and the
generation size have to be increased bigger and bigger at same time. The efficiency of GA is one of the
weak point to real world application of the GA. Hill- climbing is a good method of a search strategy that
exploits the best among know possibilities for finding a improved solution. In this study, the potential for
combining the Hill-climbing strategy with GA was investigated.
2) Population diversity
Premature convergence is caused by early emergence of an individual that is better than the others in
the population, although far from optimal. To avoid premature convergence , one has to avoid the loss of
population diversity. Although reducing the reproduction number cannot always eliminate premature
convergence, it can be used as a simple way to reduce rapid convergence. In this study, the duplicated
number of individuals was limited less than two. If the individual’s expected duplicated number is larger
that two, it was set equal to two.
|
|
|
|
|
|
|
