融合多源时空数据的地下水硫酸盐预测模型

Prediction Model of Groundwater Sulphate Based on Combined Multi-source Spatiotemporal Data

准确预测地下水SO_(4)^(2-)空间变化趋势对改善地下水质量、提高区域地下水管理水平具有重要意义.以2011、2014、2017和2020年叶尔羌河流域平原区土地覆盖数据、土壤参数数据、数字高程数据等多源时空数据和地下水p H值为特征变量,分析其与地下水SO_(4)^(2-)浓度的相关性,利用贝叶斯优化算法优化随机森林回归,建立BOA-RFR模型,并基于BOA-RFR模型对特征变量进行重要性分析,对模型预测精度进行评价,最后生成地下水SO_(4)^(2-)预测图.结果表明,p H值、地面高程(GE)和贡献区荒地(BAR)面积占比作为影响地下水水化学组分的重要参数,与地下水SO_(4)^(2-)浓度均呈现极显著负相关,对地下水SO_(4)^(2-)浓度预测的重要度均大于25%;地统计插值方法作为空间分布预测建模的辅助手段,加入辅助样本后的BOA-RFR模型,地下水SO_(4)^(2-)浓度预测的R^(2)均大于0.96,且多辅助样本构建模型的RMSE和MAE最大值较少样本模型的最小值分别降低了4.7%和23.8%;在地下水SO_(4)^(2-)浓度预测中,高SO_(4)^(2-)地下水向叶尔羌河流域平原区东北部富集,且面积呈扩张趋势.

The accurate prediction of spatial variation trends in groundwater SO_(4)^(2-)is of great significance for improving groundwater quality and regional groundwater management level.The multi-source spatio-temporal data such as land cover data,soil parameter data,digital elevation data,and groundwater pH value in the plain area of the Yarkant River Basin in 2011,2014,2017,and 2020 were used as characteristic variables to analyze their correlation with groundwater SO_(4)^(2-)concentration.To enhance the prediction accuracy,the Bayesian optimization algorithm(BOA)was used to optimize the random forest regression(RFR).Based on the BOA-RFR model,the importance of the characteristic variables was analyzed,the prediction accuracy of the model was evaluated,and the groundwater SO_(4)^(2-)prediction map was generated.The results showed that pH value,ground elevation(GE),and percentage of bare land(BAR)in the contribution area were important parameters influencing groundwater hydrochemical composition,which were significantly negatively correlated with groundwater SO_(4)^(2-)concentration,and the importance of impact factors for predicting groundwater SO_(4)^(2-)concentration exceeded 25%.The geostatistical interpolation method was used as an auxiliary tool for the predictive modeling of spatial distribution.After adding auxiliary samples,the R^(2)of groundwater SO_(4)^(2-)concentration prediction of the BOARFR model was greater than 0.96,and the maximum values of RMSE and MAE were reduced by 4.7%and 23.8%,respectively,compared with the minimum values of the model with fewer samples.The SO_(4)^(2-)concentration prediction map showed that high SO_(4)^(2-)groundwater was enriched in the northeast of the plain area of the Yarkand River Basin,an area that was expanding.