Abstract: This study focused on Pinus densata in Shangri-La City and systematically depicted the spatiotemporal dynamic characteristics of carbon density in high-altitude coniferous forests based on long time-series remote sensing data by adopting a dynamic indicator of total above- and root system carbon density, multi-time scale analysis and multi-model comparison. We compared the Random Forest (RF) and Gradient Boosting Regression Tree (GBRT) models to analyze the differences in their estimation accuracy for 5- and 10-year carbon density changes and their rates, thus identifying the optimal estimation indicators. A total of 573 remote sensing variables and environmental factors were analyzed to identify the dominant drivers of carbon density dynamics. Dynamic spatial distribution maps were further produced to reveal temporal and spatial patterns of regional carbon density.Results showed that the RF model performed best for the 5-year carbon density change rate, with R²=0.88 and RMSE=3.24 tC·hm^-2. Feature importance analysis of the four change indicators showed that texture factors contributed relatively highly, and the modeling variable R3T4SK had the highest feature importance (49.12%). Incorporating temperature improved the RF model’s estimation of the 5-year carbon density change rate (R²=0.91, RMSE=1.96 tC·hm^-2). From 1987 to 2017, Pinus densata forests achieved a cumulative net carbon storage increment of 12.01 × 10⁴ tC, with root carbon storage contributing 1.71 × 10⁴ tC (14.24% of the total net increment).