Fine-scale population spatialization with accumulable features from street view imagery and points of interest

Population spatialization servers as a pivotal technique for delineating the geographical distribution patterns of human populations, holding profound implications for urban planning and the optimal allocation of public resources. Most existing population spatialization studies construct population indicators based on the quantity or density of points of interest (POI), but overlook the heterogeneity in population attraction among POIs of the same category. To address this gap, this paper proposes a population spatialization method integrating POI features with accumulable street view imagery (SVI) features. Firstly, it extracts the counts of dynamic objects from SVIs, substituting pixel proportion-based features with accumulable object count features to enhance the spatial heterogeneity. Second, a virtual street combination-based data augmentation (VSCDA) strategy is proposed to boost the model's generalization capability across multi-scale grids. Finally, a random forest model is employed to estimate the grid-scale population distribution by integrating these multi-source features. Experiments conducted in Wuhan demonstrate that the proposed method outperforms prevailing benchmark models and public datasets (e.g., WorldPop, GPW) across different grid scales. Ablation experiments further corroborate that the semantic complementarity between POI and SVI features, coupled with the VSCDA strategy, collectively contribute to the enhancement of prediction accuracy at small and medium scales. The proposed accumulable visual features and VSCDA strategy provide a novel technical paradigm for advancing population spatialization research.