Scholarlink

<p>Air pollution caused by fine particulate matter (PM2.5) poses a serious threat to public health. This is particularly evident in urban areas with complex topography and recurring haze, such as Chiang Mai, Thailand. Accurate short-term forecasting of PM2.5 concentrations is essential for early-warning systems and proactive public health interventions. This study proposes a semi-supervised multimodal deep learning framework that integrates spatial and temporal data sources to predict next-day PM2.5 levels. The spatial input consists of RGB composite images derived from Sentinel-2 satellite data (bands B4, B3, B2), while the temporal input comprises daily PM2.5 values from six Air4Thai monitoring stations across Chiang Mai Province. A Convolutional Neural Network (CNN) based on ResNet50 is used to extract spatial features from satellite imagery, and a Long Short-Term Memory (LSTM) network models recent temporal trends in pollutant concentrations. The outputs from both branches are fused and passed through a regression head to generate the final prediction. To enhance generalization under limited labeled data, a consistency regularization strategy is applied, enabling the model to benefit from unlabeled satellite imagery via semi-supervised learning. The framework is trained and evaluated on real-world haze-season data from 2018 to 2025, achieving a Mean Absolute Error (MAE) of 11.34, Root Mean Squared Error (RMSE) of 17.32, and a coefficient of determination (R²) of 0.57. These results demonstrate the model’s ability to capture complex spatiotemporal patterns and highlight its potential as a lightweight, reliable forecasting tool for PM2.5 in data-scarce, pollution-prone urban environments.</p> Graphical Abstract <p></p> <p>This graphical abstract presents a semi-supervised multimodal deep learning framework for forecasting daily PM2.5 concentrations in Chiang Mai, Thailand, a region prone to recurring haze events. The framework integrates two types of data: Sentinel-2 RGB satellite imagery, filtered to exclude cloud coverage above 20%, and ground-based PM2.5 time-series values obtained from Air4Thai monitoring stations. Both labeled and unlabeled satellite images are used, with pseudo-labels generated to expand the training set under data-scarce conditions. For analysis, the system applies a Convolutional Neural Network (CNN) based ResNet50 to extract spatial features from satellite images and a Long Short-Term Memory (LSTM) network to capture temporal dependencies in pollutant concentration trends. A dedicated fusion layer concatenates spatial and temporal features, which are then passed through a fully connected regression head to produce next-day PM2.5 forecasts. Results from haze-season data (2018 to 2025) demonstrate that this semi-supervised approach achieves a Mean Absolute Error of 11.34, Root Mean Squared Error of 17.32, and R² of 0.57, showing improvements in generalization. Overall, the study concludes that leveraging multimodal data with a lightweight CNN-LSTM architecture and semi-supervised consistency regularization provides a scalable, data-efficient, and reliable tool for air quality forecasting in urban environments with limited monitoring infrastructure.</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

Multimodal PM2.5 Forecasting Using Satellite Imagery and Sensor Data with Semi-supervised Deep Learning

  • 期刊论文

  • 发表日期：

    2026年4月6日

查看全文
  • Tanittha Srimanee,
  • Benchawan Kiatphibun,
  • Chumpol Thanomngern,
  • Nattawan Mhuaksa,
  • Chanya Banju-ngam,
  • Thitirat Siriborvornratanakul

摘要

Air pollution caused by fine particulate matter (PM2.5) poses a serious threat to public health. This is particularly evident in urban areas with complex topography and recurring haze, such as Chiang Mai, Thailand. Accurate short-term forecasting of PM2.5 concentrations is essential for early-warning systems and proactive public health interventions. This study proposes a semi-supervised multimodal deep learning framework that integrates spatial and temporal data sources to predict next-day PM2.5 levels. The spatial input consists of RGB composite images derived from Sentinel-2 satellite data (bands B4, B3, B2), while the temporal input comprises daily PM2.5 values from six Air4Thai monitoring stations across Chiang Mai Province. A Convolutional Neural Network (CNN) based on ResNet50 is used to extract spatial features from satellite imagery, and a Long Short-Term Memory (LSTM) network models recent temporal trends in pollutant concentrations. The outputs from both branches are fused and passed through a regression head to generate the final prediction. To enhance generalization under limited labeled data, a consistency regularization strategy is applied, enabling the model to benefit from unlabeled satellite imagery via semi-supervised learning. The framework is trained and evaluated on real-world haze-season data from 2018 to 2025, achieving a Mean Absolute Error (MAE) of 11.34, Root Mean Squared Error (RMSE) of 17.32, and a coefficient of determination (R²) of 0.57. These results demonstrate the model’s ability to capture complex spatiotemporal patterns and highlight its potential as a lightweight, reliable forecasting tool for PM2.5 in data-scarce, pollution-prone urban environments.

Graphical Abstract

This graphical abstract presents a semi-supervised multimodal deep learning framework for forecasting daily PM2.5 concentrations in Chiang Mai, Thailand, a region prone to recurring haze events. The framework integrates two types of data: Sentinel-2 RGB satellite imagery, filtered to exclude cloud coverage above 20%, and ground-based PM2.5 time-series values obtained from Air4Thai monitoring stations. Both labeled and unlabeled satellite images are used, with pseudo-labels generated to expand the training set under data-scarce conditions. For analysis, the system applies a Convolutional Neural Network (CNN) based ResNet50 to extract spatial features from satellite images and a Long Short-Term Memory (LSTM) network to capture temporal dependencies in pollutant concentration trends. A dedicated fusion layer concatenates spatial and temporal features, which are then passed through a fully connected regression head to produce next-day PM2.5 forecasts. Results from haze-season data (2018 to 2025) demonstrate that this semi-supervised approach achieves a Mean Absolute Error of 11.34, Root Mean Squared Error of 17.32, and R² of 0.57, showing improvements in generalization. Overall, the study concludes that leveraging multimodal data with a lightweight CNN-LSTM architecture and semi-supervised consistency regularization provides a scalable, data-efficient, and reliable tool for air quality forecasting in urban environments with limited monitoring infrastructure.