Cross-sectional accuracy does not imply the reliability of population change in gridded population datasets of China

Abstract Time-series gridded population datasets are foundational resources for diverse fields, from public health to sustainable development. Current datasets’ evaluations primarily focus on their cross-sectional accuracy, and their applications often carry an implicit assumption that datasets with higher cross-sectional accuracy are more effective in representing population dynamics. This study evaluated six time-series gridded population datasets (CnPop, GHS-POP, GlobPop, GPWv4, LandScan, and WorldPop) using high-resolution Chinese township-level census data (2010 and 2020). The findings challenge this assumption. Although most time-series datasets exhibited high cross-sectional accuracy (Pearson’s r with census data exceeding 0.8), their ability to accurately represent decadal population change was severely limited. Specifically, these datasets showed substantial inaccuracies in identifying the population decline trends and weak performance in capturing the magnitude of change (Pearson’s r ≤ 0.27). Notably, despite excelling in cross-sectional accuracy, both GlobPop and WorldPop performed below the average rate (53.25%) in capturing the direction of decadal population change. Conversely, CnPop, which exhibited the lowest cross-sectional accuracy, achieved the highest directional accuracy rate (60.42%). These results underscore the significant limitations of existing gridded datasets in capturing population changes during China’s rapid urbanization and highlight the need for careful evaluation before application. Improvements in future gridded population data production can be achieved by enhancing input data quality and advancing spatiotemporal modeling techniques, enabling better representation of population dynamics, rather than focusing solely on cross-sectional precision.