A database and model for the PM_2.5 concentration measurement with visible and infrared imaging

With the escalating severity of environmental pollution, monitoring PM_2.5 concentration has become paramount. Among diverse monitoring techniques, the vision-based methods stand out due to their cost efficiency, high efficacy, and ease of integration. However, these methods are currently constrained by their sole reliance on visible light (RGB) imaging, which limits their ability to capture nuanced variations in particulate matter. To address this, we present the first RGB and far-infrared (FIR) multi-source image dataset for PM_2.5 measurement, comprising 4909 RGB-FIR image pairs. We further propose a prior knowledge-guided RGB-FIR multi-source fusion network, which integrates RGB-FIR multi-source features via attention mechanisms to enhance sensitivity to PM_2.5 variations. Moreover, a correction module leverages existing RGB data as prior knowledge to refine predictions. Experiments demonstrate that our method outperforms state-of-the-art vision-based algorithms, reducing RMSE by 7.859 and MAE by 5.483 while improving R² by 12.77%. The self-built dataset and code will be released at https://github.com/StarDJJ/RGB-FIR-dataset.