Abstract: The precise extraction of oilseed rape field can realize the separation of road, field ridge and plot, provide accurate operation area and nonoperation area for plant protection UAV, and promote the realization of autonomous operation of plant protection UAV. Based on UAV lowaltitude RGB images of rapeseed field in Minle County, Zhangye City, Gansu Province, this paper constructed a method combining simple linear iterative clustering (SLIC) segmentation and VGG16 classification network to achieve boundary extraction of rapeseed field blocks. Firstly, the image was grayed with green index to distinguish bare ground and vegetation covered area. Secondly, the main part of the field was extracted by histogram analysis and contour detection. Finally, by combining simple linear iterative clustering (SLIC) and VGG16 model, the grid was divided to identify the cropping regions in the segmented region and extract the whole field. The results showed that the average intersection ratio and average accuracy of the proposed model were 95.9% and 96.0%, respectively. The accuracy and integrity of the proposed model were better than those of the traditional algorithm. The model proposed in this paper can eliminate the images of weeds taken at low altitude, which can provide reference for the extraction of farmland boundaries and pave the way for the autonomous operation of plant protection UAV.

Key words: field extraction, UAV, excess green index, super pixel segmentation, VGG16, image recognition

摘要： 油菜地块精准提取可实现道路、田埂和地块三部分的分离，为植保无人机提供准确作业区域和非作业区域，推动植保无人机实现自主作业。基于甘肃省张掖市民乐县油菜地无人机低空RGB影像，构建基于简单线性迭代聚类（SLIC）分割和VGG16分类网络相结合的方法实现油菜田地块边界提取。首先，以过绿指数方式灰度化图像，区分裸露地表与植被覆盖区域，其次，通过直方图分析、轮廓检测提取地块主体部分；最后，通过简单线性迭代聚类（SLIC）和VGG16模型相结合，划分网格，识别过分割区域中的作物种植区域，提取完整地块。对比所提算法与传统边界检测算法地块边界提取效果，结果表明：所提模型的平均交并比和平均准确率分别为95.9%、96.0%，边界提取精度和完整性明显优于传统算法。所提模型能够消除低空拍摄下杂草的影像，可为农田边界提取提供参考，可为植保无人机完全自主作业做好铺垫。

关键词: 地块提取, 无人机, 过绿指数, 超像素分割, VGG16, 图像识别