基于综合多尺度特征决策树模型的土地利用变化分析

针对当前土地利用变化模型研究中较少考虑多尺度融合的问题,尝试构建综合空间多尺度特征的决策树模型(decision tree model,DTM),并与典型数据挖掘方法模拟效果进行对比。选取1995年、2000年和2005年的南京市土地利用现状为数据源,并以1995年和2000年两期数据为训练样本,以空间剖分的土地利用单位(9.5km2大小的格网单元)为模拟单元,构建综合模拟单元内部特征、邻域特征和全局特征的空间多尺度DTM,利用此模型对2000-2005年南京市土地利用变化进行模拟与分析,同时将该模型与朴素贝叶斯(naive Bayes,NB)、BP神经网络(back propagation neural network,BPNN)和支持向量机(support vector machine,SVM)模型进行模拟精度的对比。结果表明:DTM模拟精度为88.97%,高于NB、BPNN和SVM模型的模拟精度(分别为84.44%、87.13%和83.46%)。本文提出的综合多尺度特征的DTM具有与典型数据挖掘方法相类似或更好的精度,而且还具有运算效率高、可解释性强以及简单易扩展等突出特点,有利于LUCC模型的异地复用和决策支持。

Analysis on land use change based on decision-tree model with comprehensive multi-scale characteristics

Abstract: In present model on land use and land cover change (LUCC) research, the fusion of multiple scales has been less considered. To solve this problem, a decision tree model (DTM) synthesizing multi-scale features was proposed in this paper; additionally, the simulation result was compared to those from the typical data mining methods. Firstly, the space was subdivided to land use unit with appropriate grid cell size of 9.5 km2, which was used as simulation unit, and land use intensity (LUI), landscape shape index (LSI), and dominant land use type (DLT) were selected as inner indices, proportion of construction land (POC) and neighborhood average intensity of land use (NAI) were selected as neighborhood index, and the city-suburb index (CSI) was selected as global index. All the indicators of three scales were used as the property terms of the DTM. Then, taking the multi-scale evaluation indices of land use in Nanjing City, Jiangsu Province in 1995 as the test properties of the space instance, in accordance with the method of natural split point, land use intensity was classified into 4 levels, i.e. highest, higher, lower, lowest; and taking the actual levels of land use intensity in 2000 as the prediction of land-use-intensity type (PLT), the DTM was generated based on the sample sets of 816 space instances by using 10-fold iterative method. And the obtained model after the training was expressed to a tree structure. Thirdly, by using spatial and statistical analysis of GIS, the corresponding index values of feature attributes at different scales and conceptual levels were extracted and calculated from land use data. Further, the semantic information at a high abstraction level in different scales was constructed, and the law restraints to the model with the features of different scales were expressed. Based on the land use data in 2000, the simulated classification of land use intensity in 2005 was conducted by using the previous generated DTM. And then, by performing 10-fold cross-training for the 1995 and 2000 sample sets of land use change, the simulated land use intensity in 2005 was obtained and compared with the actual land use intensity in 2005. Finally, the simulated result was compared to those from the Naive Bayes (NB) method, the back propagation neural network (BPNN) method and the support vector machine (SVM) method in model simulation accuracy. The comparative experiments showed that the simulation accuracy of the DTM proposed in this paper was 88.97%, which was higher than the NB, BPNN and SVM methods whose simulation accuracies were respectively 84.44%, 87.13% and 83.46%. Since the simulation results from the DTM could be expressed as a tree structure with semantic clarity, when compared to the NB and SVM models, the DTM had better explanation; on the other hand, the DTM was easy to be constructed and expanded. In conclusion, the proposed DTM with synthesized multi-scale features in this study has similar or better accuracy compared with typical data mining methods, has the advantage of high computing efficiency, is easy to be extended and explained and can be reused at other places, and hence can contribute to support decision in LUCC management.