Prediction of particulate matter (PM) in rural built environments based on Generative Adversarial Network (GAN)

Abstract

Particulate matter (PM) is a key parameter for characterizing outdoor air quality. PM concentration is closely related to features of built environment. Rural built environment elements at block scale, such as building massing, impermeable surfaces, and farmlands, significantly impact the PM concentration. However, current research has focused on large-scale and broad-spectrum forecasting models, which are difficult for guide designers to apply because they lack rapid, detailed forecasting and specific visualization. This study proposes an automated design procedure using the Generative Adversarial Network (GAN) model to perform spatial planning oriented by environmental performance in rural blocks. This study collected and obtained data, including satellite land cover maps and PM concentrations, to construct a prediction model. Then, the model was used to quickly and accurately predict the concentrations of three kinds of particulate matter, PM₁, PM_2.5, and PM₁₀ under different design scenarios. This study found that first, the ratio between industrial and residential buildings (IB:RB) was positively correlated with PM concentration. The buildings with a short-strip configuration exhibited the lowest PM concentration in their environment compared to clusters of buildings in long strips or block-form structures. Second, the ratio between roads and small squares (R:SS) showed a positive correlation with PM concentration. The impervious surfaces characterized by large block configurations demonstrated the lowest PM concentration among the five planar forms evaluated. Third, farmland coverage exhibited a weak negative correlation with PM concentration. Farmlands with small blocks had the lowest PM₁₀ levels among five different planar forms, and small dotted farmlands had the lowest PM₁ and PM_2.5 levels. Finally, the model was used to simulate PM concentration under different design scenarios and suggested interactive strategies for future rural spatial planning design.