基于神经网络的元胞自动机及模拟复杂土地利用系统

Cellular automata for simulating complex land use systems using neural networks

本文提出了基于神经网络的元胞自动机 (CellularAutomata), 并将其用来模拟复杂的土地利用系统及其演变。国际上已经有许多利用元胞自动机进行城市模拟的研究, 但这些模型往往局限于模拟从非城市用地到城市用地的转变。模拟多种土地利用的动态系统比一般模拟城市演化要复杂得多, 需要使用许多空间变量和参数, 而确定模型的参数值和模型结构有很大困难。本文通过神经网络、元胞自动机和GIS相结合来进行土地利用的动态模拟, 并利用多时相的遥感分类图像来训练神经网络, 能十分方便地确定模型参数和模型结构, 消除常规模拟方法所带来的弊端。

This paper presents a new method to simulate the dynamics of multiple land uses based on the integration of neural networks,cellular automata and GIS. Recently, cellular automata (CA) have been increasingly used to simulate urban growth and land use dynamics. However, simulation of multiple land use changes using CA models is difficult because numerous spatial variables and parameters have to be utilized. Conventional CA models have problems in defining simulation parameter values, transition rules and model structures. In this paper, a three-layer neural network with multiple output neurons is designed to calculate conversion probabilities for competing multiple land uses. The neural-network-based CA model is directly developed in a GIS environment by using ARC/INFO GRID AML. The GIS provides both data and spatial analysis functions for constructing the neural network. Real data are conveniently retrieved from the GIS database for calibrating and testing the model. The GIS functions are also used for the neural network calculations. The neural network has multiple output neurons to generate conversion probabilities at each iteration. Land use conversion is decided by comparing the conversion probabilities. The model is carried out by iterative looping the neural network for simulating multiple land use changes. Complex global patterns can be generated from local interactions through the neural network. The simulation results are not deterministic because a stochastic variable is used and site attributes are dynamically updated at the end of each loop. The proposed method can overcome some of the shortcomings of the currently used CA models in simulating complex urban systems and multiple land use changes by significantly reducing the tedious work in defining parameter values, transition rules and model structures. The model has been successfully applied to the simulation of land use dynamics in the Pearl River Delta.