基于Apache Spark机器学习的生态安全格局构建方法

区域生态安全格局构建往往是通过对各类环境因子不同层次、不同深度的分析以生成规划结果,少有学者直接对生态安全格局与环境因子之间的关系进行分析,研究通用模型。利用大数据计算框架Apache Spark机器学习库的Logistic Regression(LR)模型对佛山市高明、三水和顺德区已有生态安全格局规划数据与岩性、土壤、用地类型、NDVI、海拔、坡度、道路距离、河流距离、年均降雨量、人口密度等多个变量的相互关系进行了训练学习,得到回归模型,用以预测广东省生态安全格局,结果显示:1)基于Spark-LR的保障生态安全格局模型(GSPM)精度达到90.58%,其预测的广东省保障安全格局高概率区比例为50.56%,在实际应用中,有一定的参考价值;2)总体上GSPM预测的生态安全格局分布与已有规划类似,但是模型容易受到样点分布均匀性的影响;3)GSPM预测结果更加切合生态资源保护的需求,而常规方法构建的结果则需要进一步优化;4)Spark-LR机器学习对生态安全格局中城市扩张的预测具有一定的优势。

An ecological security pattern construction method based on Apache Spark machine learning

The construction of a regional ecological security pattern (ESP) is often based on the analysis of different levels and different depths of various environmental factors. Few scholars directly analyze the relationship between ecological security patterns and environmental factors and instead, tend to study only the general model. In this paper, using the logistic regression (LR) model of the big data computing framework Apache Spark machine learning library, we trained and tested the relationship between the ESP planning data and environmental factors and obtained the guaranteed security pattern model (GSPM). The ESP planning data were obtained from Gaoming, Sanshui, and Shunde districts in Foshan City. The environmental factors include 17 variables: lithology, soil texture, soil type, land cover, vegetation normalization index, elevation, slope, yin and yang aspect, curvature, distance to fault, distance to road, distance to river, distance to construction land, annual average rainfall, annual average temperature, annual average wind speed, and population density. GSPM was used to predict the ESP of Guangdong province. The results showed that (1) the accuracy of the GSPM reached 90.58%, the proportion of the high-probability area of the guaranteed ESP in Guangdong Province was 50.56%, and there is a certain reference value of this model; (2) The overall ESP of the prediction results is similar to the existing plans, but the model is susceptible to the uniformity of sample distribution; (3) GSPM prediction results are more in line with the needs of ecological resource protection, while the results of conventional method construction need further optimization; and (4) SPARK-LR machine learning has certain advantages on the prediction of urban expansion.