欢迎您访问《智慧农业(中英文)》官方网站! English

Smart Agriculture ›› 2022, Vol. 4 ›› Issue (3): 95-107.doi: 10.12133/j.smartag.SA202206001

• 专刊--智慧果园关键技术与装备 • 上一篇    下一篇

基于改进Linknet网络的黄土高原苹果园精准提取

张志博1(), 赵西宁2(), 高晓东2, 张利1, 杨孟豪2   

  1. 1.西北农林科技大学 水利与建筑工程学院,陕西杨凌 712100
    2.中国科学院水利部水土保持研究所,陕西杨凌 712100
  • 收稿日期:2022-06-01 出版日期:2022-09-30 发布日期:2022-11-23
  • 基金资助:
    国家重点研发计划项目(2021YFD1900700);国家杰出青年科学基金项目(42125705)
  • 作者简介:张志博(1995-),博士研究生,研究方向为农业遥感技术与图像分割。E-mail:zhang_zhibo@nwafu.edu.cn
  • 通讯作者: 赵西宁 E-mail:zhang_zhibo@nwafu.edu.cn;zxn@nwafu.edu.cn

Accurate Extraction of Apple Orchard on the Loess Plateau Based on Improved Linknet Network

ZHANG Zhibo1(), ZHAO Xining2(), GAO Xiaodong2, ZHANG Li1, YANG Menghao2   

  1. 1.College of Water Resources and Architectural Engineering, Northwest A&F University, Yangling 712100, China
    2.Institute of Soil and Water Conservation, CAS & MWR, Yangling 712100, China
  • Received:2022-06-01 Online:2022-09-30 Published:2022-11-23
  • corresponding author: ZHAO Xining E-mail:zhang_zhibo@nwafu.edu.cn;zxn@nwafu.edu.cn

摘要:

黄土高原近20年来苹果栽植面积迅猛增加,对区域生态水文和社会经济发展均产生了重要影响。但该区域果园地块小且场景复杂,仅有县/市尺度统计数据,尚无苹果园实际的空间分布信息。为此,本研究建立了无人机低空遥感影像专业数据集。融合迁移学习与深度学习方法,将残差神经网络ResNet34网络迁移到Linknet网络,得到R_34_Linknet网络。将R_34_Linknet网络与5种常用的深度学习语义分割模型SegNet、FCN_8s、DeeplabV3+、UNet和Linknet应用于黄土高原苹果园空间分布提取,表现最好的模型为R_34_Linknet,其在测试集上的调和平均值F1为87.1%,像素准确度PA为92.3%,均交并比MIoU为81.2%,频权交并比FWIoU为85.7%,平均像素准确度MPA为89.6%。将空间金字塔池化结构(Atrous Spatial Pyramid Pooling,ASPP)与R_34_Linknet网络相结合,扩大网络的感受野,得到R_34_Linknet_ASPP网络;然后对ASPP结构进行改进,得到R_34_Linknet_ASPP+网络。对比三种网络性能,表现最优的为R_34_Linknet_ASPP+,在测试集上F1为86.3%,PA为94.7%,MIoU为82.7%,FWIoU为89.0%,MPA为92.3%。使用R_34_Linknet_ASPP+在长武县王东沟和白水县通积村提取苹果园面积精度分别为94.22%和95.66%。本研究提出的R_34_Linknet_ASPP+方法提取到的苹果园更加准确,苹果园地块边缘处效果更好,可作为黄土高原苹果园空间分布制图等研究的技术支撑和理论依据。

关键词: 无人机遥感, 苹果园提取, 深度学习, 黄土高原, 迁移学习, 残差神经网络, 语义分割

Abstract:

The rapid increasing of apple planting area on the Loess Plateau has exerted an important influence on the regional eco-hydrology and socio-economic development. However, the orchards in this area are small and complex, and there are only county or city scale statistical data, lack of actual spatial distribution information. To this end, for the extraction of apple orchards on the Loess Plateau, in this study, a professional dataset of low-altitude remote sensing images acquired by unmanned aerial vehicle was firstly established. The R_34_Linknet network and other five commonly used deep learning semantic segmentation models SegNet, FCN_8s, DeeplabV3+, UNet and Linknet were applied to the spatial distribution extraction of apple orchards on the Loess Plateau, and the best-performing model was R_34_Linknet, with a F1 score of 87.1%, a pixel accuracy (PA) of 92.3%, an mean intersection over union (MioU) of 81.2%, a frequency weighted intersection over union (FWIoU) of 85.7%, and the mean pixel accuracy (MPA) was 89.6%. The spatial pyramid pool structure (ASPP) and R_34_Linknet network was combined to expand the receptive field of the network and get R_34_Linknet_ASPP network, and then ASPP structure was improved. Combining the spatial pyramid pooling (ASPP) with the R_34_Linknet network to expand the receptive field of the network and obtain a R_34_Linknet_ASPP network; Then the ASPP structure was improved to get a R_34_Linknet_ASPP+ network. The performance of the three networks were compared. R_34_Linknet_ASPP+ got the best performance, with 86.3% for F1, 94.7% for PA, 82.7% for MIoU, 89.0% for FWIoU, and 92.3% for MPA on the test set. The accuracy of apple orchard extraction in Wangdonggou, Changwu County and Tongji Village, Baishui County using R_34_Linknet_ASPP+ were 94.22% and 95.66%, respectively. In Wangdonggou, it was 1.21% and 0.58% higher than R_34_Linknet and R_34_Linknet_ASPP, respectively. In Tongji village, it was 1.70% and 0.90% higher than R_34_Linknet and R_34_Linknet_ASPP, respectively. The results show that the proposed R_34_Linknet_ASPP+ method can extract apple orchards accurately, the edge treatment of apple orchard plots is better, the method can be used as the technical support and theoretical basis for research on the spatial distribution mapping of apple orchards on the Loess Plateau.

Key words: UAV remote sensing, apple orchard extraction, deep learning, Loess Plateau, transfer learning, residual neural network, semantic segmentation

中图分类号: