High-Resolution Population Distribution Prediction Considering the Spatiotemporal Heterogeneity of Geolocated Behavior

High spatiotemporal-resolution population distribution prediction (HST-PDP) is critical for urban management and emergency response. Existing methods often use geolocated digital footprints (GDF) as a direct population proxy, ignoring the representation bias of GDF for population counts that stem from the spatiotemporal heterogeneity of human behavior. To address this, we propose a High SpatioTemporal-resolution Population prediction (HSTP) method. Drawing from behavioral geography, HSTP explicitly models the GDF-population relationship by introducing a latent variable, the per capita Trigger Frequency of Digital Footprints (TFDF), to quantify behavioral heterogeneity, thus decomposing GDF as the product of the true population and TFDF. The HSTP architecture uses a Transformer encoder to learn contextual representations from multi-source geospatial data. A dual-decoder then uses these contexts to predict the latent population and TFDF variables. This decomposition acts as a training constraint, enabling the model to leverage abundant GDF data to optimize population estimates even with sparse population labels, thus enhancing generalization. Experiments on hourly, 200-meter-grid data from Wuhan show HSTP outperforms the state-of-the-art baselines like NHITS and TFT by over 35.5% in sMAPE. Thus, HSTP serves as both a high-precision prediction tool and a novel framework for modeling and analyzing behavioral heterogeneity in urban computing.