Tea is a high value-added economic crop with extremely high economic value. It is the main starting point for rural revitalization in mountainous areas of China. However, due to destructive behaviors such as deforestation and planting tea, forest resources are destroyed, and ecological and environmental problems such as soil erosion are caused. Acquiring the spatial distribution of tea plantations quickly and accurately is very important for government supervision and the planning and development of the tea industry. However, due to the rainy weather in the study area and the scattered distribution of tea plantations, which are close to the spectrum of vegetation such as forests, the extraction based on satellite imagery has become a problem. Tea plantations are challenging. In order to find out the spatial distribution of tea plantations in Yingde, this paper systematically analyzes the application potential of medium and high-resolution multispectral Sentinel-2 image data combined with multi-time-series and multi-feature information in tea garden extraction. Taking the whole territory of Yingde as the research area, this paper selects 9 phases of Sentinel-2 image data from 2019 to 2021 to analyze the phenological characteristics of tea tree growth in detail and further explore the characteristics changes of tea plantations and other land types in multiple time series, using the Relief algorithm to sort the importance of all features. According to the result of feature sorting, the feature factors weighted by 90% of the feature weight value are selected, namely 7 vegetation index features and 2 texture features, and 9 kinds of tea garden classification scenes are constructed through different combination rankings, and the RF algorithm is used to evaluate the accuracy of all classification scenes. To select the best classification scene and further discuss the feasibility of the RF classification algorithm and SVM classification algorithm for tea garden extraction. The results show that: (1) When extracting tea plantations in Yingde, February and October are the best combinations to construct multiple characteristics of tea plantations using multi-temporal phases. (2) Compared with the SVM classification method, the RF classification method has high accuracy. Its overall accuracy reaches 91.56%, the Kappa coefficient is 0.89, and the producer accuracy and user accuracy are 80.22% and 84.56%, respectively. This study provides an efficient method for quickly and efficiently obtaining the spatial distribution information of tea plantations in Yingde and provides data support for the government to plan and manage the tea industry.