Scholarlink

<p>This article aims to predict the concentration of air pollutants at any unmonitored location based on sparse monitoring points in the monitoring area, thereby achieving the goal of fine-grained air pollution mapping. To learn the spatial distribution characteristics of air pollutants from sparse monitoring data, this article proposes a novel Graph Neural Network (GNN) model called Graph Convolutional Neural Networks on K Neighbors (KN-GCN). Additionally, a data augmentation method is employed to enhance the sparse monitoring data and prevent overfitting of the KN-GCN model during the training process. Moreover, since the ground truth concentration value is unavailable at unmonitored locations, the accuracy of the prediction cannot be measured. Therefore, a training strategy is designed to reflect the unmeasurable accuracy on the metrics of the KN-GCN model. To evaluate the proposed method, a Computational Fluid Dynamics (CFD) simulation experiment and a public dataset experiment are conducted. The results reveal that the proposed method outperforms the baseline methods by an average of 65% and 17.8% in the CFD experiment and public dataset experiment, respectively.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A novel graph convolutional neural network on K neighbors model for fine-grained air pollution distribution mapping based on sparse monitoring

  • 期刊论文

  • 发表日期：

    2026年1月21日

查看全文
  • Qing Liu,
  • Lei Cheng,
  • Daiwei Tang,
  • Essel Emmanuel

摘要

This article aims to predict the concentration of air pollutants at any unmonitored location based on sparse monitoring points in the monitoring area, thereby achieving the goal of fine-grained air pollution mapping. To learn the spatial distribution characteristics of air pollutants from sparse monitoring data, this article proposes a novel Graph Neural Network (GNN) model called Graph Convolutional Neural Networks on K Neighbors (KN-GCN). Additionally, a data augmentation method is employed to enhance the sparse monitoring data and prevent overfitting of the KN-GCN model during the training process. Moreover, since the ground truth concentration value is unavailable at unmonitored locations, the accuracy of the prediction cannot be measured. Therefore, a training strategy is designed to reflect the unmeasurable accuracy on the metrics of the KN-GCN model. To evaluate the proposed method, a Computational Fluid Dynamics (CFD) simulation experiment and a public dataset experiment are conducted. The results reveal that the proposed method outperforms the baseline methods by an average of 65% and 17.8% in the CFD experiment and public dataset experiment, respectively.