Evaluating the Impact of Cellular Tower Density on Intercity Travel OD Identification Using Mobile Signalling Data: An Empirical Comparison of ST-DBSCAN and Bi-LSTM Algorithms

Abstract

Significant differences in cellular tower density across cities pose a major challenge for identifying intercity travel origin-destination (OD) pairs from mobile phone signalling data. Many existing OD identification algorithms apply uniform parameters across cities, which can undermine detection accuracy in heterogeneous networks, and their performance remains underexplored under varied tower density conditions. To address this gap, we conducted a field experiment collecting mobile signalling data from intercity trips in two metropolitan regions with different tower densities, while recording GPS trajectories and travel diaries as ground truth. We compared the unsupervised spatiotemporal clustering method (ST-DBSCAN) and the supervised deep learning model (Bi-LSTM) for OD identification. Furthermore, we introduced a novel genetic algorithm adaptive parameter selection mechanism to enhance performance under different density conditions by dynamically adjusting ST-DBSCAN's clustering radius, time threshold and minimum cluster size, as well as tuning Bi-LSTM's input features and time window length. Results show that this adaptive approach significantly improved OD identification accuracy, with optimised ST-DBSCAN achieving 84% accuracy and Bi-LSTM 91%. These findings highlight the importance of adaptive algorithm calibration and offer theoretical insights and practical guidance for more reliable intercity travel modelling in metropolitan areas with heterogeneous cellular infrastructure.