Abstract: To reveal the spatio-temporal pattern and influencing factors of vegetation changes in the Sanjiangyuan region from 2000 to 2023， and to forecast the possible change trend of vegetation under climate change， in this study NDVI data and climte data including temperature， precipitation， wind speed and barometric pressure from four regions within Sanjiangyuan-Dari， Mado， Yushu， and Qumalai were utilized to analyze the normalized Difference Vegetation Index at the same period. The NDVI prediction model was established by using machine learning algorithms-Gradient Boosting Regressor， AdaBoost Regressor， Random Forest， and Neural Networks. On this bisis， all models were fine-tuned and validated to enhance performance and reliability. Finally， an optimal model of simulation accuracy was selected to simulate vegetation change under multiple scenarios. The results showed that temperature was the most significant meteorological factor influencing NDVI， explaining up to 67.29% of the variability. The Gradient Boosting Regressor showed better performance than other models in all the study areas. This model achieved a Mean Squared Error （MSE） ranging from 0.000 45 to 0.001 04 and an R2 value exceeding 0.90. It showed strong fitting ability.The Gradient Boosting Regressor proved to be highly accurate and stable in predicting NDVI， which provides a robust approach for forecasting vegetation changes and is instrumental for early warning of vegetation degradation in response to climate change. The research findings provide a robust scientific basis for ecological conservation initiatives， facilitating the formulation of strategies to alleviate the effects of climate change on the vegetation within the Sanjiangyuan area.

Key words: NDVI, Machine learning, Gradient boosted regression tree, Sanjiangyuan region

摘要： 为了揭示三江源地区2000—2023年植被时空变化格局及影响因素，并预测气候变化条件下三江源地区植被可能的变化趋势，本研究基于三江源达日、玛多、玉树、曲麻莱四个地区2000—2023年归一化植被指数（Normalized difference vegetation index， NDVI）数据，以及温度、降水、风速和气压等气候数据进行分析。研究采用了梯度提升回归树、自适应增强回归、随机森林以及神经网络等机器学习算法建立NDVI预测模型。在此基础上，对所有模型参数进行了精细调优和验证，以提升模型性能和可靠性。最终，筛选出了模拟精度最优模型，进行多情景下植被变化模拟。研究结果表明，温度对NDVI的气象特征值占比最高，达0.6486。梯度提升回归模型在所有研究区综合表现优于其他模型，平均均方误差（Mean squared error，MSE）在0.000 45~0.001 04之间，拟合系数（Coefficient of determination，R2）均超过0.90，显示出强大的拟合能力。梯度提升回归树在预测三江源地区NDVI方面具有较高的准确性和稳定性，并对NDVI数据具有良好拟合效果，为三江源地区NDVI预测提供了科学方法。研究结果有助于预警气候变化条件下植被退化的潜能，为气候变化背景下该区域植被生态保护提供科学依据。

关键词: NDVI, 机器学习, 梯度提升回归树, 三江源地区