Phy-APMR: A physics-informed air pollution map reconstruction approach with mobile crowd-sensing for fine-grained measurement

Fine-grained air pollution map reconstruction is critical for urban pollution management and healthy building construction. Recent advancements in air pollution measurement, such as mobile sensing platforms that equip sensors on urban vehicles (e.g., taxis, buses), have greatly enhanced data collection for pollution analysis. However, two major challenges remain: (1) the uncontrolled mobility of vehicles leads to data sparsity in certain areas and times, reducing the effectiveness of data-driven reconstruction methods, and (2) training these methods is often time-consuming, hindering frequent updates needed for improved accuracy. In this paper, we propose Phy-APMR, i.e., a fine-grained phy sics-informed a ir p ollution m ap r econstruction approach, which combines the advantages of a physical air pollution propagation model and a deep-learning model to mitigate the data sparsity issue. In addition, to make our method feasible for a high-frequency updating strategy to further improve the reconstruction accuracy, we propose ASUS ( a daptive s hort-time u pdate s ampling), a novel collocation point sampling algorithm, speeding up the convergence of the training for Phy-APMR. Experiments are conducted in three cities, showing that Phy-APMR surpasses the state-of-the-art by 15% in reconstruction accuracy and 84% in convergence efficiency. • Propose Phy-APMR for fine-grained air pollution map reconstruction to mitigate the data sparsity issue in certain time and space. To the best of our knowledge, Phy-APMR is the first hybrid inference method leveraging the Physics-Informed Neural Network (PINN) framework for the air pollution map reconstruction problem. • Propose Phy-APMR+W/S, an enhanced variant of Phy-APMR, to fuse additional environmental information of wind speed and air pollution sources. Phy-APMR+W/S can make an efficient use of such additional environmental information to further improve the air pollution map reconstruction accuracy. • Propose Adaptive Short-time Update Sampling (ASUS) algorithm to adaptively accelerate the training of Phy-APMR without deteriorating the accuracy, ensuring the feasibility of Phy-APMR in terms of speed, accuracy, and efficiency. • Design and conduct evaluation experiments in three cities using a total of 75 air sensing devices. Experimental results show the significantly improved performance of the Phy-APMR over the state-of-the-art.