诺木洪地区是青海省重要的枸杞种植基地，针对主要植被类型进行准确和快速提取对于种植业的可持续发展具有关键意义。然而，诺木洪地区所属的干旱区具有稀疏的植被覆盖和土壤背景影响显著的特点，仅使用少数遥感源或部分特征无法满足干旱区植被提取的要求。因此，整合多种遥感数据源，挖掘植被分类显著特征，并尝试不同的分类优化方法，在提高干旱区植被分类精度和可靠性方面具有重要意义。本研究基于谷歌地球引擎（Google Earth Engine，GEE）平台，使用Sentinel-1合成孔径雷达（Synthetic Aperture Radar，SAR）数据、Sentinel-2光学数据，探讨了红边光谱、纹理以及雷达特征对干旱区植被类型提取的重要性，验证了利用基尼指数（Gini Index，Gini）寻找最优特征组合的可行性，结合支持向量机算法对2021年青海诺木洪地区地物类型进行提取，并对最终的分类结果优化处理。研究表明：（1）Sentinel-2红边指数、纹理信息和Sentinel-1雷达波段有利于植被信息提取，分类总体精度和Kappa系数分别达到了95.51%和0.9406；（2）根据Gini指数得到特征重要性，将分类特征由29个压缩至17个，且表明雷达极化特征、光谱特征和纹理特征对于分类的重要性依此递减；（3）使用简单非迭代聚类算法以及邻域滤波投票决策融合方法，不仅最优总体精度和Kappa系数达到了96.06%和0.9479，且针对孤点类型的噪声也有较好的抑制效果。本研究利用GEE遥感云平台和多源遥感数据以及机器学习算法，能够准确、快速、高效地提取大尺度范围干旱区地物信息，具有较大的应用潜力。

Nuomuhong region is an important Wolfberry cultivation base in Qinghai Province, China. Accurate and rapid extraction of the primary vegetation types is of critical significance for the sustainable development of agriculture in this region. However, the arid nature of the Nuomuhong area, characterized by sparse vegetation cover and significant soil background effects, presents challenges for vegetation extraction using only a limited number of remote sensing sources or partial features. Therefore, integrating multiple remote sensing data sources, exploring significant features for vegetation classification, and experimenting with different classification and optimization methods are paramount for enhancing the accuracy and reliability of vegetation classification in arid regions. Based on the Google Earth Engine (GEE) platform, this study used Sentinel-1 Synthetic Aperture Radar and Sentinel-2 optical data to explore the importance of red edge, texture, and radar features in extracting vegetation types in arid regions. Additionally, it verifies the feasibility of using the GINI index (GINI) to determine the optimal feature combination. The main geospatial types in Nomu Hong, Qinghai, China, in 2021 were extracted by combining them with the support vector machine algorithm. The classification results were processed using decision fusion methods. The results showed that: (1) Sentinel-2 red edge index, texture data, and Sentinel-1 radar band were beneficial for the extraction of vegetation-related information, with an overall classification accuracy and Kappa coefficient of 95.51% and 0.9406, respectively. (2) Based on the importance obtained by the GINI index, the features involved in the classification were reduced from 29 to 17, and the significance was radar polarization features > spectral features > texture features. (3) Using a simple noniterative clustering algorithm and neighborhood filtering voting decision fusion method not only achieved the optimal overall accuracy and Kappa coefficient but also had an excellent suppression effect on isolated noise. Using the GEE remote sensing cloud platform, multisource remote sensing data, and machine learning algorithms, this study can accurately, quickly, and efficiently extract large-scale arid region geospatial information, which can have great application potential.