针对传统 DRASTIC模型在参数权重确定过程中主观性强问题,以粤北某地区浅层地下水为研究对象,利用采集的地下水相关数据和新增土地利用类型参数,在优化 BP 神经网络算法和构建 DRASTICL 模型基础上,借助地下水NO-3 浓度进行模型验证,建立耦合 BP 神经网络算法的 BP-DRASTICL 模型,进而根据地下水脆弱性空间分布特点提出了地下水污染风险管控对策.结果表明:训练函数为trainlm、学习率为 0.1、隐含层神经元节点数为 6 时,BP 神经网络算法效果最好,相应地获得的最优 BP-DRASTICL模型参数权重依次为 0.1420(地下水埋深,D)、0.1151(净补给量,R)、0.0791(含水层介质,A)、0.1833(土壤介质,S)、0.0908(地形,T)、0.1574(包气带介质影响,I)、0.0891(渗透系数,C)和 0.1433(土地利用类型,L).D、S、T 和 L 对评价结果的影响最大.与 DRASTIC 模型、DRASTICL 模型相比,BP-DRASTICL模型的 Pearson和Spearman相关系数最高,分别达到 0.615 和 0.656,表明硝酸盐浓度与脆弱性指数之间具有很高的线性关系.研究区地下水脆弱性总体属于极低脆弱性和低脆弱性,高脆弱性和极高脆弱性地区主要分布在研究区中部.针对研究区脆弱性的空间分布特征,差异化提出了地下水污染风险管控对策.利用 BP 神经网络算法确定 DRASTICL模型参数权重,因减少了人为主观性的影响,比传统专家打分法更准确.
GROUNDWATER VULNERABILITY EVALUATION AND RISK CONTROL IN A CERTAIN AREA IN NORTHERN GUANGDONG PROVINCE BASED ON BP NEURAL NETWORK
Aiming at the high subjectivity of the traditional DRASTIC model in the process of the parameter weight determination,taking shallow groundwater in a certain area in northern Guangdong Province as a study target,the BP neural network was optimized and the DRASTICL model was constructed by using the collected shallow groundwater-related data and adding land use type parameter.On this basis,groundwater NO-3 concentration was used to verify the models,and further the BP neural network and the constructed DRASTICL model were coupled to establish a BP-DRASTICL model.Finally,risk control strategies were suggested according to the spatial distribution characteristics of groundwater vulnerability.The results showed that when the training function was trainlm,the learning rate was 0.1,and the number of hidden layer neurons was 6,BP neural network performed best,and accordingly the optimal BP-DRASTICL parameter weights were 0.1420(groundwater depth,D),0.1151(recharge,R),0.0791(aquifer media,A),0.1833(soil medium,S),0.0908(topography,T),0.1574(influence of vadose zone media,I),0.0891(hydraulic conductivity,C)and 0.1433(land use type,L).D,S,T and L had the greatest influence on the evaluation results.Compared with the DRASTIC models and the DRASTICL model,the BP-DRASTICL model had the highest Pearson(0.615)and Spearman(0.656)correlation coefficients,indicating a high linear correlation between the actual nitrate concentration and the vulnerability index.The groundwater vulnerability was generally in the extremely low,and low level,across the study area,and the areas with the high and extremely high vulnerability level were mainly distributed in the middle of the study area.According to the spatial distribution characteristics of the vulnerability,differentiated strategies were proposed for groundwater pollution risk control.Using the BP neural network to determine the parameter weights of the DRASTICL model is more accurate than using the traditional expert scoring method,because it reduces the influence of human subjectivity.
NETL
NSTL
万方数据
