Cellular Automaton Model Study For Simulating Spatio-Temporal Evolution Of Zhalong Wetland

The simulation of spatio-temporal evolution of wetlands is useful for monitoring and analyzing wetland pattern changes. It can vividly assist in managing the wetland resources. Recently most research of wetland pattern changes has focused on landscape indices analysis. The spatio-temporal characteristics of wetlands are seldom represented and the change process in the specific location cannot be monitored, thus, it is difficult to carry on a pertinent management plan. A spatio-temporal evolution simulation can make up the deficiency of the landscape indices analysis. Considering the characteristics of wetland pattern changes and spatio-temporal data, a cellular automaton (CA) model is established to integrate intelligent methods, including neural network, fuzzy logic and rough set to simulate the spatio-temporal evolution of wetlands. The main methods and results are shown as follows.(1) Because CA has spatio-temporal characteristics and it can simulate the evolution of complex systems, a spatio-temporal evolution of wetlands (STEW) model based on CA is proposed. This model integrates some intelligent modeling methods to realize micro-state transition of CA and acquire macro-adjusting rules automatically. In the model the centroid point movement of cells with the same state is chosen as a macro-adjusting rule. It can lead state change of cells and offsets the limitation of the quantitative macro-adjusting rule, which jusf constrains the number of cells of every state. To evaluate the results of CA reasonably, a comprehensive comparison method is proposed. In the method not only the pixel to pixel comparison but also the influences of the shape and array of patches to the pattern are considered.(2) The wetland pattern changes are complex and the data monitoring wetlands are fuzzy and multi-scale, so a Takagi-Sugeno fuzzy neural network based on samples distribution characteristics (TSFNN-SDC) is put forward to complete the micro-state transition. Membership functions of linguistic values of variables are obtained automatically based on the distribution characteristics of samples. So the fuzzy space is partitioned reasonably. The clustering centers are fuzzified based on the membership functions to get initial rules and the same rules merge into one. In effect, the final rules with their degrees of importance are obtained. According to the antecedent conditions of the rules the fuzzification layer and inference layer are connected so that the structure of network is simplified. Simulation results indicate that the proposed fuzzy neural network has better simulation accuracy and better generalization ability than traditional T-S fuzzy neural network.(3) Because the socio-economic information of wetlands is absent and it is difficult to design formula to get quantitative macro-adjusting rules, a Rough-Neural-Network (Rough-NN) model is constructed. Rule set are reduced and simulated to acquire the linguistic macro-adjusting rules. By these rules the micro-state transition potentials are adjusted. To improve the generalization of rules obtained by rough set, the paper proposes a parameter-representing membership function method which turns antecedent conditions and consequent conclusions of rules into input and output data respectively, at the same time, a three-layer feedforward neural network is applied to simulate the rules. By the parameter-representing method the network has higher training speed and better reasoning results for testing samples than that of using the usual vector-representing method or interval-representing method.(4) Zhalong wetland is taken as a case study of this STEW model. Landscape indices of landscape pattern of Zhalong wetland in the period of 1986-2002 are calculated and analyzed to explore the driving factors. And then based on the patterns in 1986-1999 STEW model is applied to forecast the spatio-temporal evolution of Zhalong wetland from 2000 to 2002. Meanwhile the two models applied to simulate the expansion of urban are modified to simulate the spatio-temporal evolution of Zhalong wetland. Simulation results show that the proposed STEW model has higher prediction accuracy than the modified models. The mean prediction accuracy with the STEW model is about 70%, which is comparable with urban expansion prediction accuracy obtained by cellular automata recently.In summary, the STEW model based on CA is proposed. In this model, the TSFNN-SDC is designed to calculate the micro-state transforming potentials of CA; and the rough set theory and neural network are integrated to establish the Rough-NN model which is applied to obtain the macro-adjusting rules represented by linguistic values to adjust the potentials. Good prediction effects are achieved when the STEW model is applied to simulate the pattern changes in Zhalong wetland.