Abstract: In order to quickly extract spatial information of citrus orchards in districts and counties by using highresolution remote sensing images and convolutional neural network models, we selected Pujiang County, a key citrus production area in Sichuan Province, as the research area, downloaded highresolution images on Google earth as the data source and constructed 3 types of citrus orchard sample data sets of different tree age stages. On this basis, we trained U-net and DeepLabv3+ semantic segmentation models to extract citrus orchard spatial information and verified its classification accuracy. The results showed that the U-net and DeepLabv3+ models with different neural network structures performed well, with the close citrus information extraction accuracy. The overall accuracy was 88.30% and 86.79% and the Kappa coefficient was 0.75 and 0.72, respectively. In addition, the identification accuracy of citrus orchards in small plots were analyzed, their minimum identification plot area was about 120 m2, and the average accuracy was above 85% when the plots area exceeds this number. This research could provide reference for farmer or local agricultural departments to use highresolution remote sensing and open source deep learning classification methods to quickly and automatically extract orchard spatial information.

Key words: orchard, citrus, remote sensing, semantic segmentation, spatial information

摘要： 为应用高分辨率遥感影像和卷积神经网络模型快速提取柑橘园空间信息，选择四川省柑橘重点产区蒲江县为研究区，以高分辨率Google earth图像为数据源，构建3类不同树龄的柑橘园样本数据集，训练U-net和DeepLabv3+语义分割模型，提取柑橘园空间信息。通过验证，具有不同神经网络结构的U-net和DeepLabv3+模型提取柑橘园信息总体精度分别为88.30%和86.79%，Kappa系数为0.75和0.72，二者精度相当；通过分析小地块的果园遥感识别精度，测试区最小识别图斑面积约为120 m2，大于该面积的果园遥感面积平均精度在85%以上。该研究可为经营者、农业部门使用高分辨率遥感影像和开源的深度学习分类工具快速获取果园空间信息提供参考。

关键词: 果园, 柑橘, 遥感, 语义分割, 空间信息