Enhancing PM2.5 Detection via Multispectral Image-Based Double-Channel Ensemble Learning for Drones

Due to the rapid development of modern cities over the past decade, air pollution has been continuously rising. In areas with severe air pollution, it threatens human health and significantly exacerbates climate change, energy, and food production. Air quality monitoring focuses mainly on detecting particulate matter ${\mathrm{PM}}_{{2}.{5}}$ in the air. However, traditional monitoring stations are expensive and can only provide localized data. On the other hand, RGB images captured by drones often struggle to distinguish between ${\mathrm{PM}}_{{2}.{5}}$ and clouds or fog in the sky. Therefore, based on the above reasons, we developed a unique method in this study, named double-channel ensemble learning (DCEL- ${\mathrm{PM}}_{{2}.{5}}$ ), to be applied to the ${\mathrm{PM}}_{{2}.{5}}$ detection tasks of drones. Unlike traditional deep learning models that use RGB images for ${\mathrm{PM}}_{{2}.{5}}$ detection, the DCEL method utilizes input data from aero multispectral cameras and builds a dual-channel neural network, which processes raw RGB images and multispectral images [near infrared $\text {(}\text {NIR}\text {)}$ , red edge $\text {(}\text {RE}\text {)}$ , red $\text {(}\text {R}\text {)}$ , green $\text {(}\text {G}\text {)}$ , and blue $\text {(}\text {B}\text {)}$ ]. Each channel independently trains the multispectral and RGB images to extract attention features and mask information, which are fed into the global weighted average pooling (GWAP). The dual-channel GWAP features are concatenated into a total GWAP feature to detect ( ${\mathrm{PM}}_{{2}.{5}}$ ). Compared with other existing methods, the experimental results show that our proposed DCEL- ${\mathrm{PM}}_{{2}.{5}}$ method provides more accurate ${\mathrm{PM}}_{{2}.{5}}$ predictions while maintaining a low parameter count and high frames/s (FPS), making it highly suitable for drone missions.