基于SVR-LUR模型的城市道路PM10空间浓度分布模拟

针对传统土地利用回归模型(Land use regression, LUR)未考虑影响因子与大气污染物之间非线性复杂关系 和易出现多重共线性的问题。以PM10为例,采用支持向量回归机(Support vector machine regression, SVR) 改进土地利用回归模型的建模方法构建SVR-LUR模型,对上海市南浦大桥周边区域PM10空 间分布进行模拟。研究结果表明: 1)研究区域PM10浓度与100 m缓冲区内的空地面积, 150 m缓冲区内的建筑工地面积、 空地面积、河流面积, 200 m缓冲区内的绿地面积和河流面积,以及湿度、交通流量和背景浓度相关性较高。 2) SVR-LUR模型可较好地对研究区PM10浓度进行空间分布预测。SVR-LUR模型与LUR模型相比, SVR-LUR模型预测精 度较高,其测试集比LUR模型测试集的平均绝对误差(Mean absolute error, MAE)及均方根误差(Root mean squares error, RMSE)分别减小了22.92\%、33.51\%,拟合指数(Index of agreement, IA)值增加了13.20\%。 相较于普通克里金插值模型所得到的单一梯度空间分布预测结果, SVR-LUR 模型能够更有效揭示小范围内的空间差异。 3)研究区PM10浓度空间分布呈现出西高东低的总格局,在建筑物和路网密集的地方浓度较高,而在靠近江面和空地 的区域浓度相对较低。模拟结果与实际情况相符。

Simulation of Spatial Concentration Distribution of Urban Road PM10 Based on SVR-LUR Model

The traditional land use regression(LUR) model does not consider the nonlinear complex relationship between impact factors and atmospheric pollutants. Taking PM10 as an example, Support Vector Machine Regression (SVR) has been used to improve the land use regression model modeling method to construct SVR-LUR model, and then the spatial distribution of PM10 around Nanpu Bridge in Shanghai, China is simulatedbased on the model. The results show that: 1) There is a high correlation between the PM10 concentration and the empty area in the 100 m buffer zone, the construction area, the empty area and the river area in the 150 m buffer zone, the green area and the river area in the 200 m buffer zone, as well as the humidity, traffic flow and background concentration. 2) the SVR-LUR model can better predict the spatial distribution of PM10 concentration in the study area. Compared with LUR model, SVR-LUR model has higher prediction accuracy. Compared with LUR model, the mean absolute error (MAE) and root mean squares error (RMSE) oftest set of SVR-LUR model reduces by are22.92\%, 33.51\% less than those of LUR model, MAE and RMSE respectively, and whilethe index of agreement (IA) value increases by 13.20\%. Compared with the prediction results of single gradient spatial distribution obtained by ordinary Kriging interpolation model, SVR-LUR model can more effectively reveal the spatial differences in a small range. 3) The spatial distribution of PM10 concentration in the study area shows a general pattern of high concentration in the west and low in the east. The concentration is higher in the areas with dense buildings and road network, but relatively lower in the areas near the river and open space. The simulation results are consistent with the actual situation.