A Prototype Feature Representation-Guided Multitask Learning Framework for Image-Based PM_2.5Concentration Monitoring

With rapid industrialization, air pollution has emerged as a critical societal problem. Particulate matter (PM2.5) poses severe threats to public health and ecosystems, making it paramount to establish accurate and scalable monitoring solutions to safeguard the health of humans and animals as well as the ecological environment. While conventional instruments provide precise measurements, their widespread deployment is hindered by high costs and limited coverage. The image-based methods that utilize widely distributed traffic cameras or drones to capture PM2.5-related visual features (e.g., contrast and saturation) offer a promising alternative. However, existing approaches, including both traditional hand-crafted feature-based methods and deep learning-based methods, face challenges in complex environments and suffer from performance limitations due to the long-tailed distribution of pollution data. To this aim, we propose a prototype feature representation-guided multitask learning (PFRML) framework, which integrates classification and regression for PM2.5 concentration estimation. The framework comprises two main components. First, a convolutional neural network (CNN)-Transformer ensemble backbone combines CNN's strength in local spatial feature extraction with Transformer's capability in modeling global contextual relationships, enabling robust feature representation under complex conditions. Second, a classification-regression multitask paradigm employs prototype learning (effective in few-shot scenarios) to explicitly construct a PM2.5 pollution level representation space (i.e., prototype features), thereby enhancing sparse sample recognition through clustering constraints. Meanwhile, the regression branch implicitly learns continuous concentrations via a multilayer perception while leveraging the prototype features from the classification branch for auxiliary optimization. Experiments conducted on multiple public benchmarks demonstrate that our method achieves superior prediction accuracy and robustness compared to state-of-the-art (SOTA) PM2.5 estimation methods.