Scholarlink

<p>Aerosols have impact on human health, agriculture, and global climate. In this study, we present MultiscaleTCNGraphSAGE, a graph-based neural network designed for the spatiotemporal prediction of aerosol optical depth (AOD550) on a global scale. This model combines monthly aerosol data from the CAMS system with meteorological variables from ERA5. Using an architecture that integrates temporal convolutions at multiple scales (kernels of size 3, 5, and 7 with dilations 1, 2, and 3) and GraphSAGE spatial convolutions, our approach effectively captures the spatial and temporal dependencies present in global aerosol data. Compared to the best benchmark model, Transformer+GCN, MultiscaleTCNGraphSAGE achieves a 5.4% reduction in RMSE, a 6.4% decrease in MAE, a 2.7% increase in the coefficient of determination, and a 1.3% improvement in Pearson correlation. These results consistently demonstrate the ability of the proposed model to capture both short- and long-term aerosol dynamics, improving the prediction of extreme aerosol-related events.</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

Global forecasting of aerosol optical depth through a deep learning spatiotemporal modeling

  • 期刊论文

  • 发表日期：

    2025年12月1日

查看全文
  • M. Semper,
  • M. Curado,
  • J. F. Vicent

摘要

Aerosols have impact on human health, agriculture, and global climate. In this study, we present MultiscaleTCNGraphSAGE, a graph-based neural network designed for the spatiotemporal prediction of aerosol optical depth (AOD550) on a global scale. This model combines monthly aerosol data from the CAMS system with meteorological variables from ERA5. Using an architecture that integrates temporal convolutions at multiple scales (kernels of size 3, 5, and 7 with dilations 1, 2, and 3) and GraphSAGE spatial convolutions, our approach effectively captures the spatial and temporal dependencies present in global aerosol data. Compared to the best benchmark model, Transformer+GCN, MultiscaleTCNGraphSAGE achieves a 5.4% reduction in RMSE, a 6.4% decrease in MAE, a 2.7% increase in the coefficient of determination, and a 1.3% improvement in Pearson correlation. These results consistently demonstrate the ability of the proposed model to capture both short- and long-term aerosol dynamics, improving the prediction of extreme aerosol-related events.