Abstract: To accurately map kiwifruit cultivation zones，Songbai Town，Yongshun County，Hunan Province was selected as the experimental area. Based on the domestically developed GF‐6 sensor，an object‐oriented random forest classification model integrating multi‐temporal and textural features was constructed.The classification performance was systematically evaluated across 12 feature combinations：single‐temporal（July），single‐temporal+textural，and multi‐temporal（May，July，September）+textural.The results showed that the multi‐temporal+texture feature combination classification model performed better in the extraction of kiwifruit orchards，which could achieve high‐precision identification of kiwifruit orchards in the study area.The multi‐temporal feature was the key factor to improve the classification accuracy.Compared with the single‐temporal+texture feature combination（RF2—RF6），after introducing multi‐temporal features（RF7—RF12），the F value increased from 89.35%—91.09% to 91.59%—93.48%，the overall accuracy（OA）increased from 83.65%—87.33% to 91.00%—93.18%，and the Kappa coefficient increased from 0.78—0.80 to 0.86—0.90.Compared with the single‐phase feature（RF1），the single‐phase+texture feature combination showed limited improvement in classification accuracy.The F value was only increased by 1.03 percentage points，and the Kappa coefficient was only 0.80.In the combination of different texture+multi‐temporal features，the wavelet texture（RF9）extracted based on energy statistics performed best.The F value and OA of RF9 reached 93.48% and 93.18%，respectively，which were 1.89 and 2.04 percentage points higher than those of GLCM texture（RF8），and 1.59 and 2.18 percentage points higher than those of multi‐temporal features（RF7）.These indicate that the object‐oriented random forest classification method utilizing multi‐temporal+wavelet texture feature combination is an effective approach for high‐precision extraction of kiwifruit orchards in complex mountainous terrain.

Key words: Kiwifruit orchard, Object‐oriented, Temporal features, Wavelet texture, Random forest

摘要： 为准确掌握猕猴桃种植区分布信息，选取湖南省永顺县松柏镇为试验区，基于国产GF-6，构建了融合多时相特征与纹理特征的面向对象随机森林分类模型，系统比较了单时相特征（7月）、单时相+纹理特征、多时相（5、7、9月）+纹理特征12种特征组合下的分类性能。结果表明，多时相+纹理特征组合分类模型在猕猴桃园提取中表现较优，可实现研究区猕猴桃园的高精度识别。多时相特征是提升分类精度的关键因素，相较于单时相+纹理特征组合（RF2—RF6），引入多时相特征后（RF7—RF12），F值从89.35%~91.09%提升至91.59%~93.48%，总体精度（OA）从83.65%~87.33%提升至91.00%~93.18%，Kappa系数从0.78~0.80提升至0.86~0.90。相较于单时相特征（RF1），单时相+纹理特征组合对分类精度提升有限，F值最高仅提升1.03百分点，Kappa系数最大仅为0.80。在不同纹理+多时相特征组合中，基于Energy统计量提取的小波纹理（RF9）表现最优，RF9的F值和OA分别达到93.48%和93.18%，相对于GLCM纹理（RF8）分别提高1.89、2.04百分点，相对于多时相特征（RF7）分别提高1.59、2.18百分点。表明利用多时相+小波纹理特征组合的面向对象随机森林分类方法是复杂山区猕猴桃园高精度提取的有效途径。

关键词: 猕猴桃园, 面向对象, 时序特征, 小波纹理, 随机森林