Comparative Analysis of Statistical Filtering and Deep Learning for GPS Trajectory Denoising in Personal Exposure Assessment

Background Accurate assessment of personal air pollution exposure is essential in environmental epidemiological studies. Indirect approaches often estimate exposures using measurements from the nearest monitoring station based on an individual’s location. Although these locations are typically extracted using devices equipped with global positioning systems (GPS), raw GPS signals frequently lack indicators of positional accuracy. This deficiency makes it difficult to reliably identify positional errors within continuous GPS tracking data across diverse time–activity patterns. Objective This study aims to develop and systematically evaluate GPS correction models designed to restore accurate personal movement trajectories Methods We used a dataset from the Korean Air pollutant EXposure (KAPEX) model project, which includes time-location diaries labeling participants’ locations at one-minute intervals and positional coordinates simultaneously tracked with two GPS devices. A GPS trajectory from one device providing only raw signals were used to correction target and verified with the other reference trajectory, preprocessed with GPS signal quality information. The correction models were developed employing both statistical state-space and deep-learning based algorithms. Results Kalman Filter consistently demonstrated robust GPS correction performance in terms of denoising accuracy, computational efficiency, and trajectory smoothness when compared with deep learning–based models. Conversely, deep learning approaches exhibited reasonable denoising capability primarily in indoor settings characterized by frequent GPS signal degradation.