胡杨是荒漠河岸地区的建群树种，是绿洲的生态屏障，对于保护绿洲生态系统的稳定至关重要。塔里木河流域是世界上天然胡杨林分布最为密集的区域，获取塔里木河干流区域的天然胡杨林的精确分布数据将为天然胡杨林的保护与恢复提供重要支持。本研究基于Google earth engine云平台和机器学习方法自动提取天然胡杨林，主要研究结论如下：

（1）对比S-G、HANTS、Whittaker和SWCF平滑方法重构初始NDVI时间序列的效果，发现SWCF方法重构效果最佳。利用SWCF方法重构NDVI时间序列后，构建了包括SoS、EoS、LoS、Max value、DoM和AoS等6个物候参数特征；与Sentinel-2 Band2、Band3、Band4、EVI、NDPI、IRECI、GCVI和PSRI等8个光谱指数特征；与CON、ASM、CORR等3个纹理特征；以及Sentinel-1的VV、VH波段等2个后向散射系数特征的特征集合。将该特征集合输入随机森林模型，采用主动学习方法不断优化模型，获得了塔里木河干流区域的10米分辨率天然胡杨林分布数据，总体精度为0.9274，生产者精度为0.9269，用户精度为0.9189，kappa系数为0.8543。

（2）比较仅输入物候特征（P）和光谱指数特征（S），添加后向散射系数特征（B），添加纹理特征（T），以及同时添加后向散射系数特征（B）和纹理特征（T）提取天然胡杨林的结果，发现物候特征和光谱指数特征是提取天然胡杨林的基础特征，加入后向散射系数特征可以改善胡杨林提取精度，单独加入纹理特征会导致胡杨林提取效果变差，同时加入后向散射系数特征和纹理特征则会进一步提升胡杨林的识别精度。

（3）天然胡杨林具有沿着地表河流分布的特点，呈现出带状分支特征。50%的天然胡杨林分布在距离塔里木河1.2km的范围内，90%的天然胡杨林分布在距离塔里木河6.5km的范围内，99%的天然胡杨林生长在距离河流12.9km的范围内。

P.euphratica is a group of trees in desert riparian areas, which is an ecological barrier for oases and is essential for protecting the stability of oasis ecosystems. The Tarim River Basin is the most densely distributed area of natural P.euphratica forests in the world, and obtaining accurate distribution data of natural P.euphratica forests in the main stream of the Tarim River will provide important support for the protection and restoration of natural P.euphratica forests. In this study, natural P.euphratica forests were automatically extracted based on Google earth engine cloud platform and machine learning method, and the main conclusions are as follows:

(1) Compared with the effects of S-G, HANTS, Whittaker and SWCF smoothing methods in reconstructing the raw NDVI time series, it was found that the SWCF method has the best reconstruction effect. After using the SWCF method to reconstruct the NDVI time series, a feature collection consists of six phenological parameter features, including SoS, EoS, LoS, Max value, DoM and AoS, eight spectral index features including Sentinel-2 Band2, Band3, Band4, EVI, NDPI, IRECI, GCVI and PSRI, three texture features including CON, ASM and CORR, and two backscatter coefficient features includingVV,VH band from Sentinel-1, was created. The feature set was input into the random forest model, and the active learning method was used to continuously optimize the model, and the distribution data of natural P.euphratica forest with a resolution of 10 meters in the main stream of the Tarim River were obtained, with an overall accuracy of 0.9274, a producer accuracy of 0.9269, a user accuracy of 0.9189, and a kappa coefficient of 0.8543.

(2) The results of extracting natural P.euphratica forest by only inputting phenological features (P) and spectral index features (S), adding backscatter coefficient features (B), adding texture features (T), and adding backscatter coefficient features (B) and texture features (T) at the same time were compared. It was found that phenological features and spectral index features were the basic features for extracting natural P.euphratica forests, adding backscattering coefficient features could improve the extraction accuracy of P.euphratica forests, adding texture features alone would make the extraction effect of P.euphratica forest worse, and adding backscattering coefficient features and texture features at the same time would further improve the identification accuracy of P.euphratica forests.

(3) The natural P.euphratica forest has the characteristics of distribution along the surface river, showing the characteristics of band-like branches. 50% of the natural P.euphratica forests are distributed within 1.2 km of the Tarim River, 90% of the natural poplar forests are distributed within 6.5 km of the Tarim River, and 99% of the natural poplar forests are grown within 12.9 km of the river.