基于随机森林的元胞自动机城市扩展模拟——以佛山市为例

本文提出一种基于随机森林的元胞自动机城市扩展(RF-CA)模型。通过在多个决策树的生成过程中分别对训练样本集和分裂节点的候选空间变量引入随机因素,提取城市扩展元胞自动机的转换规则。该模型便于并行构建,能在运算量没有显著增加的前提下提高预测的精度,对城市扩展中存在的随机因素有较强的容忍度。RF-CA模型可进行袋外误差估计,以快速获取模型参数;也可度量空间变量重要性,解释各空间变量在城市扩展中的作用。将该模型应用于佛山市1988-2012年的城市扩展模拟中,结果表明,与常用的逻辑回归模型相比,RF-CA模型进行模拟和预测分别能够提高1.7%和2.6%的精度,非常适用于复杂非线性特征的城市系统演变模型与扩展研究;通过对影响佛山市城市扩展的空间变量进行重要性度量,发现对佛山城市扩张模拟研究而言,距国道的距离与距城市中心的距离具有最重要的作用。

Urban expansion simulation by random-forest-based cellular automata: a case study of Foshan City

Cellular automata (CA) has been frequently used to investigate the logical nature of self-reproducible systems and simulate the evolution of complex geographical phenomena such as urban expansion. The core of cellular automata is to define transition rules. Traditionally, approaches for defining the transition rules of cellular automata had difficulty to balance the interpretability, accuracy, and convenience. This article presents a new cellular automata model for simulating urban expansion based on random forest algorithm. The proposed model extracts CA transition rules of urban expansion by introducing random factors in training samples and candidate spatial variables that split nodes during the multiple decision trees building process. One significant advantage of our approach is that it can be easily adopted for parallel implementation and has high prediction accuracy and tolerance to random factors in urban expansion. Another strength of the proposed approach is that it can estimate out-of-bag errors to obtain model parameters quickly and measure the importance of spatial variables and explain the contribution of each variable in urban expansion. The model was applied to simulate urban expansion in Foshan City, Guangdong Province. We used the urban land change of 1988 and 2000 as the dependent variable and the spatial variables as the independent variables to construct the CA model based on random forests, then simulate and predict urban expansion of 2000 and 2012. The results show that random forest model can improve the simulation and prediction accuracies by 1.7% and 2.6%, respectively, when compared to the logistic regression model commonly used in CA simulation. This suggests that random forest model is superior for modeling complex nonlinear urban evolution. Urban expansion of Foshan City in 2024 was also predicted according to its urban development trend. Through measuring the importance of some spatial variables that affect urban expansion, we found that distance to national roads and to the city center are the two most important spatial variables for urban expansion simulation in Foshan City.