ASI-GCNss

Limited borehole sampling at contaminated sites results in sparse and unevenly distributed data on soil pollutants. Traditional interpolation methods, such as ordinary kriging (OK) and inverse distance weighting (IDW), may obscure local variations in soil contamination when applied to such sparse data, thus reducing the interpolation accuracy. To overcome this challenge, we propose an adaptive spatial interpolation graph convolutional network (ASI-GCN) model. This model synergistically integrated the principles of spatial auto-correlation with the capabilities of graph convolutional networks (GCN). By merging these elements, the ASI-GCN model effectively constrained the transfer of pollutant concentrations. It adeptly captures nuanced variations in spatial structure, thereby enhancing the precision of soil pollution characterization. We applied our model to a coking plant in Beijing based on 215 soil samples collected from 15 boreholes. To evaluate the robustness of the model, three pollutants with distinct volatilization characteristics were employed, including arsenic (As, non-volatile), benzo(a)pyrene (BaP, semi-volatile), and benzene (Ben, volatile). The leave-one-out cross-validation procedure was used to estimate the performance of ASI-GCN and its derivatives, ASI-GCN_RC_G and ASI-GCN_RC_K, when they are used to make predictions at unsampled sites.The results showed that the ASI-GCN model demonstrated improved performance with R² values of 0.52, 0.67, and 0.57, and RMSE values of 5.041 mg/kg, 1.699 mg/kg , and 167.710 mg/kg for As, BaP, and Ben, respectively, outperforming traditional models like OK and IDW. Further analysis revealed the significant influence of pollutant volatility on vertical migration patterns. Non-volatile As was primarily distributed in the fill and silty sand layers, semi-volatile BaP concentrated in the silty sand layer, while volatile Ben was predominantly found in the clay and fine sand layers. The ASI-GCN model provides key insights for risk assessment and remediation strategies by capturing complex contaminant distributions with limited borehole data, demonstrating its potential for application in various contaminated sites.污染场地的有限钻孔取样会导致土壤污染物数据呈现稀疏且分布不均的特点。传统插值方法，例如普通克里金法（ordinary kriging, OK）与反距离加权法（inverse distance weighting, IDW），在处理此类稀疏数据时，可能会掩盖土壤污染的局部变异特征，进而降低插值精度。 为解决这一难题，本研究提出一种自适应空间插值图卷积神经网络（adaptive spatial interpolation graph convolutional network, ASI-GCN）模型。该模型将空间自相关原理与图卷积神经网络（graph convolutional networks, GCN）的建模能力进行协同融合，通过融合上述两类要素，ASI-GCN模型可有效约束污染物浓度的空间传递过程，精准捕捉空间结构的细微变异特征，进而提升土壤污染表征的精度。 本研究以北京某焦化厂为研究区域，基于15个钻孔采集的215组土壤样品，对所提模型进行了验证应用。 为评估模型的鲁棒性，本研究选取了3种具有不同挥发特性的污染物作为研究对象，分别为砷（arsenic, As，非挥发性）、苯并(a)芘（benzo(a)pyrene, BaP，半挥发性）以及苯（benzene, Ben，挥发性）。本研究采用留一交叉验证法，对ASI-GCN及其衍生模型ASI-GCN_RC_G、ASI-GCN_RC_K在未采样点位的预测性能进行评估。 结果表明，针对砷、苯并(a)芘与苯，ASI-GCN模型的决定系数（coefficient of determination, R²）分别为0.52、0.67与0.57，均方根误差（root mean square error, RMSE）分别为5.041 mg/kg、1.699 mg/kg与167.710 mg/kg，其性能优于普通克里金法与反距离加权法等传统插值模型。 进一步分析表明，污染物的挥发特性对其垂向迁移模式具有显著影响：非挥发性砷主要赋存于填土与粉砂层中，半挥发性苯并(a)芘集中分布于粉砂层，而挥发性苯则主要分布于黏土与细砂层中。 ASI-GCN模型可通过有限的钻孔数据精准刻画复杂的污染物空间分布特征，为污染场地的风险评估与修复策略制定提供关键依据，展现出在各类污染场地中应用的潜力。