基于蜂群算法的城市土地利用变化建模

通过引入人工蜂群算法用于构建土地利用变化的驱动力模型,分析土地利用变化的驱动力机制。算法原理通过模仿蜜蜂采蜜行为,自动搜索和提取土地利用变化样本中不同土地变化类型所对应的驱动力分类规则。分类规则的构建采用“IF…THEN”形式,并选取3种不同的适应度函数分别进行模拟验证。研究案例基于UCI实验数据集和美国纽卡斯尔市真实土地利用变化数据集。由实验结果可知,采用蜂群算法模型的总体精度和Kappa系数评价优于其它算法,表明蜂群算法应用于土地利用变化建模具有可行性。

Urban Land Use Change Modelling Based on Artificial Bee Colony Algorithm

This article applied artificial bee colony (ABC) algorithm to modelling the urban land use change. The principle of ABC algorithm is to discover the optimum nectar source by imitating bees’ behavior of gathering honey. Hence, ABC is a new kind of optimization algorithm same as genetic algorithm, neural network, which have been wildly implemented in the modelling of land use cover/change (LUCC). This study focus on the implementation at three aspects: 1) Land use transition rules: In order to analyzes the effects between land use dynamic and its driving forces, we defined the format of land use transition rules, which are composed of land use type, the fitness function, the cover percentage and the corresponding upper and lower values of each factor. Therefore these transition rules, which are obtained via automatic searching from samples, are able to demonstrate the relationship between land use dynamic and driving forces in the form of “IF…THEN”. 2) Fitness functions: In order to eliminate the distractions caused by imbalanced samples, we chose three different fitness functions to evaluate the classification results. These functions, including the G-mean function, the F-measure function and the Percentage of Correct Predicted (PCP), are well suit to various requirements such as sensitivity preferred rather than the overall accuracy. 3) Performance testing: two experimentations, which are based on UCI datasets and reality land use change of New Castle County between 1984 and 2002, were provided. The classification results were also compared with logistic regression, neural network, SVM and Ada-boost algorithms. According to the UCI datasets, the ABC which adopt the PCP as fitness function has acquired the best overall accuracy. They were 98.00% for Iris and 97.51% for Breast-Cancer-W. On contrary, from the results of New Castle County datasets, the ABC using G-mean instead of PCP as fitness function has obtain better results, which were 78.53% of overall accuracy and 0.569 3 of Kappa coefficient. These differences indicates that the G-mean function are more suitable for imbalanced samples. In addition, we also provided a discussion about the optimal parameter selection problem of ABC. In this article, when the number of bee was 100, the maximum of integration time was 500 and the limitation of unimproved-times was 50, the ABCs had achieved the best performance. In conclusion, the comparison of two cases shows that the overall accuracy and Kappa coefficient of ABC algorithms are comparable to the other algorithms. Hence, the ABC algorithm is appropriate for the modeling of urban land use change.