Abstract: With the development of facility agriculture, it is important to extract updated and accurate information regarding agricultural greenhouses for fine-tuning the growth mode of agricultural economy and efficiently using agricultural resources. This paper applied an object-based image classification random forest to extract grape greenhouses from a GF-2 remote sensing image in southern hilly areas acquired in May 2017. For the classification, 15 optimal features of image objects were selected based on the scale evaluation tool (ESP) and the Ratio of Mean Diff. to Neighbors (ABS) combining with Standard Deviation (RMAS) methods. The results showed that: ① combining ESP and RMAS methods provided a promising means to select the optimal segmentation scale parameter for image segmentation. ② Feature selection based on random forest  reduced data redundancy and was crucial to improve extraction accuracy. Among the 15 selected optimal features, spectral features had the highest importance, followed by texture features and geometric features. ③ By using the selected optimal features, this method produced the grape greenhouses classification with an overall accuracy of 92.5%, F-value of 0.91, and global error index (GTC) of 0.12. The proposed method was of high potential for grape greenhouses extraction in the southern hilly region, and provided an effective means for extracting agricultural greenhouses in many other areas.

Key words: grape greenhouses, GF-2, random forest, feature selection, object-based image analysis

摘要： 随着设施农业的不断发展，快速准确获取农业大棚的空间分布和种植面积有助于农业经济增长模式调整，实现农业资源的高效利用。以2017年5月的GF-2遥感影像     为数据源，在构建最优特征空间的基础上，采用面向对象随机森林分类方法开展南方丘陵地区葡萄大棚信息提取。结果表明：①利用尺度评价工具ESP和邻域差分绝对值与标准差比RMAS结合的方法可以实现特定地物目标的最优分割尺度选择，分割效果良好；②通过Gini指数进行特征选择能减少数据冗余，提高分类精度，在优选的15个特征变量中，光谱特征占有绝对优势，其次是纹理特征和几何特征；③基于最优特征子空间的随机森林模型能有效提取葡萄大棚的分布信息，总体精度高达92.5%，F值为0.91，其面向对象的精度评价指数GTC为0.12。 结果表明，该方法对基于GF-2影像的南方丘陵区域葡萄大棚信息提取具有较大的应用潜力，并可为其他地区的农业大棚信息提取提供较好的解决思路。

关键词: 葡萄大棚, GF-2, 随机森林, 最优特征空间, 面向对象信息提取