Machine learning-based methods for dust concentration distribution prediction by utilizing back scattering images

Dust is a significant environmental pollutant that poses serious risks to human health, so its proper monitoring is quite significant. However, traditional light scattering single-point measurements are limited to reflecting the dust concentration distribution in a large-scale environment. In this paper, we introduced the light intensity distance and obtained a large-scale dust concentration distribution prediction model by analyzing the relationship between dust concentration and the attenuation of light intensity caused by dust scattering and absorption in the optical path. Fifteen classical machine learning algorithms were applied, which proved the importance of light intensity distance in predicting dust concentration. When only fitting the relationship between light intensity and concentration, each algorithm gave the result with R2 of about 0.9000, while Kolmogorov–Arnold Networks (KAN) had the best prediction result 0.9076. When light intensity distance was added, the prediction accuracy of each algorithm was improved correspondingly. Light Gradient Boosting Machine (LightGBM) methods have the best performance (R2: 0.9500), and KAN followed (R2: 0.9472). We added another set of ${\mathrm{SiO}}_2$ experiments and sensitivity analyses for light intensity distance to demonstrate the applicability of the model. Finally, LightGBM was used to predict the dust concentration distribution in the whole process.