A Multi-Modal Deep-Learning Air Quality Prediction Method Based on Multi-Station Time-Series Data and Remote-Sensing Images: Case Study of Beijing and Tianjin

Hanzhong Xia, Xiaoxia Chen, Zhen Wang, Xinyi Chen, Fangyan DongEntropy 2024 · DOI

TL;DR

This paper introduces Res-GCN, a novel multi-modal deep learning model that combines time-series air quality data with remote-sensing imagery to predict PM2.5 concentrations. By leveraging both data sources, Res-GCN significantly improves prediction accuracy, offering a more comprehensive understanding of air pollution dynamics.

What Problem Does This Paper Address?

Air pollution, particularly PM2.5, poses significant threats to human health, ecological balance, and economic stability. Accurate forecasting of air quality is crucial for implementing preventive measures and safeguarding public well-being. While traditional methods rely heavily on time-series data from monitoring stations, the potential of remote-sensing images, which offer rich spatial information about pollutant dispersion, has been underexplored. This paper aims to bridge this gap by developing a model that effectively integrates these two distinct data modalities for enhanced air quality prediction.

Methods

The core of the proposed Res-GCN model lies in its multi-modal architecture, which comprises two main components:

  • ResNet for Remote-Sensing Image Feature Extraction: A modified Residual Network (ResNet) is employed to extract visual features from remote-sensing images. This component is designed to capture fine-grained details of pollutant dispersion, providing a spatial context that is often missing in time-series-only models. The ResNet architecture is tailored to the specific geographical scope of the dataset, avoiding excessive depth.
  • DSTGCN for Time-Series Data Feature Extraction: A Dynamic Spatio-Temporal Graph Convolutional Network (DSTGCN) is utilized to process the time-series air quality data from multiple stations. DSTGCN builds upon existing spatio-temporal graph convolutional networks (STGCN) but introduces a key innovation: it constructs dynamic graphs using Dynamic Time Warping (DTW) distance. This approach allows the model to better adapt to the temporal dynamics and varying spatial correlations between stations, which are often influenced by meteorological conditions. The DSTGCN integrates graph convolutional layers (GCN) for spatial feature extraction, temporal convolutional blocks (TCN) for capturing long-range temporal dependencies, and a temporal attention mechanism to focus on pivotal time steps.
  • Feature Fusion: The features extracted from ResNet and DSTGCN are fused using a parameter matrix approach. This fusion layer combines the visual and spatio-temporal information, which is then fed into a fully connected layer and a 1D convolution layer to generate the final PM2.5 concentration predictions.

Datasets

The study utilizes two primary datasets:

  • Time-Series Data:
    • Beijing: Hourly air quality data (including PM2.5) and meteorological data from 34 monitoring stations, spanning from January 1, 2018, to January 1, 2021.
  • Tianjin: Hourly air quality data (including PM2.5) and meteorological data from 27 monitoring stations, covering the period from May 1, 2014, to May 1, 2015.
  • Missing values in the time-series data were handled through imputation using preceding time steps, arithmetic sequences, and linear interpolation. Data was standardized using Z-score normalization.
  • Remote-Sensing Images:
    • The study uses daily 1-km ground-level PM2.5 remote-sensing images from the ChinaHighPM2.5 (CHAP) dataset.
  • For Beijing, images from 2018-2021 were used.
  • For Tianjin, images from May 1, 2014, to May 1, 2015, were used.
  • Images were converted from NetCDF to TIFF format, geographically cropped, normalized, and redundant channels were removed.

Key Insights and Ideas

  • The Power of Multi-Modality: The most significant insight is that combining remote-sensing imagery with traditional time-series data leads to substantial improvements in air quality prediction accuracy. Remote-sensing images provide crucial spatial context and fine-grained details about pollutant dispersion that are not captured by ground-based sensors alone.
  • Dynamic Graphs for Dynamic Phenomena: The use of Dynamic Time Warping (DTW) to construct dynamic graphs in DSTGCN is a key innovation. Air pollution is a dynamic phenomenon, and static graph representations can be insufficient. DTW allows the model to better capture the evolving spatial relationships between monitoring stations over time.
  • ResNet's Role in Spatial Context: Even with a limited geographical scope, the ResNet component effectively extracts valuable visual features from remote-sensing images, demonstrating its utility in providing spatial context for air quality prediction.
  • DSTGCN's Adaptability: The DSTGCN's ability to capture both spatial and temporal dependencies, coupled with its dynamic graph construction, makes it well-suited for the complex and non-stationary nature of air quality data.

Key Results

  • Superior Performance: The Res-GCN model consistently outperformed all baseline models (ARIMA, SVR, DNN, LSTM, TCN, CNN-LSTM, STGCN, Informer, STSGCN) across various prediction horizons (1, 8, 16, and 24 hours) on both Beijing and Tianjin datasets.
  • Significant Improvement: Compared to the best-performing baseline model (Informer/STSGCN), Res-GCN achieved performance improvements of approximately 6-7% in MAE and RMSE.
  • Ablation Study Findings: Removing the ResNet component (w/o ResNet) led to a noticeable decline in performance, highlighting the importance of remote-sensing image data. Similarly, removing DSTGCN (w/o DSTGCN) resulted in a significant deterioration, underscoring the critical role of time-series data and spatio-temporal feature extraction.
  • Graph Construction Effectiveness: DTW distance proved to be a superior method for graph construction compared to Euclidean distance, Pearson correlation, and Spearman correlation, leading to better predictive performance.
  • Visualizations: The predicted PM2.5 concentration curves for multiple stations showed that Res-GCN accurately captured trends and maintained errors within a reasonable range, demonstrating its generalization capabilities.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that while remote-sensing images provide valuable auxiliary information, they are not a complete substitute for ground-based monitoring data.
  • The study focused on PM2.5; extending the model to predict other pollutants could be beneficial.
  • Future Work:
    • The authors suggest exploring the potential of Res-GCN in other domains, such as traffic prediction and urban planning.
  • They also propose investigating the integration of additional data sources, such as points of interest, to further enhance predictive performance.

Summary updated

BibTeX
@article{xia2024multi,
  author = {Hanzhong Xia and Xiaoxia Chen and Zhen Wang and Xinyi Chen and Fangyan Dong},
  title = {{A Multi-Modal Deep-Learning Air Quality Prediction Method Based on Multi-Station Time-Series Data and Remote-Sensing Images: Case Study of Beijing and Tianjin}},
  journal = {Entropy},
  year = {2024},
  doi = {10.3390/e26010091},
}
Figure