Forest disturbances and forestation profoundly affect terrestrial ecosystem dynamics and carbon cycling. However, the lack of long-term national reference samples has greatly restricted the development and application of forest disturbance and forestation monitoring across China. In this study, we constructed a 30 m resolution Reference sample dataset of Stable, Disturbed and Forestated forests (RSDF), covering China from 1986 to 2025 with 12,991 geographically representative samples obtained via stratified random sampling. Across the full dataset, 79.61 % of the samples represent stable forest during 1986–2025, 15.99 % record a single disturbance or forestation event, and 4.40 % undergo two or more events. All samples were annotated through hierarchical visual interpretation by integrating Landsat time-series data, multi-source high-resolution satellite imagery, and available forest inventory records. A strict multi-expert blind review framework combined with a sequential consensus mechanism was further implemented for multi-stage quality control. All samples with forest disturbance or forestation were cross-validated using standardized visual diagnostic evidence, including time-series trajectories of six spectral bands, two vegetation indices, and matched pre- and post-change RGB composite imagery. Spatial-temporal pattern analysis revealed distinct regional disparities in forest change activities: forest disturbances were most frequent in East and South China and relatively weak in Northwest, North and Northeast China; harvest was the dominant forest disturbance type nationwide, while fire, other disturbance and forestation events accounted for a small proportion. Quality assessments showed over 95 % high-confidence samples, with a mean inter-expert pairwise agreement of 0.83 ± 0.07 and an average Cohen’s Kappa coefficient of 0.65 ± 0.13, demonstrating high reliability and interpretative consistency of the RSDF dataset. This 40-year nationwide sample dataset can be widely applied to algorithm calibration, model training, and product validation. It also provides a credible benchmark for large-scale forest dynamic monitoring and carbon budget assessment in China. The dataset is publicly available at <a href="https://doi.org/10.6084/m9.figshare.33313359" target="_blank" rel="noopener">https://doi.org/10.6084/m9.figshare.33313359</a> (Shang et al., 2026).