Abstract
Dynamic World provides near-real-time 10 m land-cover data, but scene-level classifications are affected by observation conditions and seasonal changes, causing missing valid data and unstable class labels. Using 2020 Dynamic World data, this study evaluated different temporal windows and compositing methods to identify suitable strategies for mapping stable annual land-cover types in China. We compared mode, cumulative-probability, and maximum-confidence composites at monthly, growing-season (April–October), and annual scales. Spatial coverage was assessed, and classification accuracy was evaluated using 14,042 reference samples. Effective observation rates showed clear spatial and temporal differences. Rates were lower in southwestern China, higher in northwestern and northern China, and generally lower in summer. Longer compositing windows significantly reduced data gaps. The monthly scale spatial coverage was 68.78%, while the growing-season and full-year windows increased spatial coverage to 98.52% and 98.57%, respectively. However, full-year coverage gains over the growing-season window were limited. Growing-season mode and cumulative-probability composites outperformed the other strategies, with the highest macro-F1 (64.96%) and OA (76.19%), respectively. However, performance varied by time window and land-cover type, and no method consistently performed best. These results can guide the selection of time windows and compositing methods for regional Dynamic World mapping.