耦合卷积神经网络与注意力机制的无人机摄影测量果树树冠分割方法  

Fruit Tree Canopy Segmentation by Unmanned Aerial Vehicle Photogrammetry Coupled on Convolutional Neural Network and Attention Mechanism

作　　者：何海清[1,2] 周福阳 陈敏 陈婷 官云兰[1,2] 曾怀恩[5] 魏燕 HE Haiqing;ZHOU Fuyang;CHEN Min;CHEN Ting;GUAN Yunlan;ZENG Huaien;WEI Yan(School of Surveying and Geoinformation Engineering,East China University of Technology,Nanchang 330013,China;Key Laboratory of Mine Environmental Monitoring and Improving around Poyang Lake of Ministry of Natural Resources,East China University of Technology,Nanchang 330013,China;Faculty of Geosciences and Environmental Engineering,Southwest Jiaotong University,Chengdu 611756,China;School of Water Resources and Environmental Engineering,East China University of Technology,Nanchang 330013,China;National Field Observation and Research Station of Landslides in the Three Gorges Reservoir Area of Yangtze River,China Three Gorges University,Yichang 443002,China)

机构地区：[1]东华理工大学测绘与空间信息工程学院,南昌330013 [2]东华理工大学自然资源部环鄱阳湖区域矿山环境监测与治理重点实验室,南昌330013 [3]西南交通大学地球科学与环境工程学院,成都611756 [4]东华理工大学水资源与环境工程学院,南昌330013 [5]三峡大学中国长江三峡库区滑坡国家野外观测研究站,宜昌443002

出　　处：《地球信息科学学报》2023年第12期2387-2401,共15页Journal of Geo-information Science

基　　金：国家自然科学基金项目(42261075、41861062);江西省自然科学基金资助项目(20224ACB212003);地理信息工程国家重点实验室、自然资源部测绘科学与地球空间信息技术重点实验室联合资助基金项目(2022-02-04)。

摘　　要：基于无人机可见光影像的果树树冠分割易受地形起伏、灌木及杂草等复杂背景影响,尽管现有深度神经网络能在一定程度上提高树冠分割的鲁棒性,但因受感受野和信息交互限制而忽略了树冠全局上下文和局部细节信息,制约了树冠分割精度进一步提高。针对此,本文引入果树高度模型与深度学习算法,提出一种耦合卷积神经网络与注意力机制的无人机摄影测量果树树冠分割方法。该方法首先通过迁移学习构建基于卷积神经网络与注意力机制的耦合深度网络模型来提取果树树冠局部和全局上下文高级语义特征;同时,顾及深度语义特征与果树树冠位置关联性,设计了局部与全局特征融合模块来实现属性与空间位置协同树冠分割。以柑橘果树树冠分割为例,实验结果表明,引入树冠高度模型能有效抑制地形起伏影响,提出的方法总体精度、F1评分和均交并比最高分别达到97.57%、95.49%和94.05%,能显著削弱低矮杂草或灌木对树冠提取的干扰。此外,与SegFormer、SETR_PUP、TransUNet、TransFuse和CCTNet等先进网络模型相比,均交并比分别提升了1.79%、8.83%、1.16%、1.43%和1.85%。提出的方法可实现复杂背景下果树树冠高精度分割,对于掌握果树生长状况和果园精细化管理具有重要的实用价值。The segmentation of fruit tree canopy based on Unmanned Aerial Vehicle(UAV)visible spectral images is greatly influenced by complex background information such as topographic relief,shrubs,and weeds.Although existing deep neural networks can improve the robustness of canopy segmentation to a certain extent,they ignore the global context and local detailed information of the canopy due to limited receptive field and information interaction,which restricts the improvement of canopy segmentation accuracy.To address these issues,this paper introduces the Canopy Height Model(CHM)and deep learning algorithms,and proposes a fruit tree canopy segmentation method that couples Convolutional Neural Networks(CNN)and Attention Mechanisms(AM)based on UAV photogrammetry.This method first constructs a coupled deep neural network based on CNN and AM through transfer learning to extract both the local and global high-level contextual features of fruit tree canopies.Meanwhile,considering the correlation between deep semantic features and the position information of fruit tree canopies,a local and global feature fusion module is designed to achieve collaborative tree canopy segmentation of attributes and spatial positions.Taking the citrus tree canopy segmentation as an example,the experimental results demonstrate that the use of the CHM can effectively suppress the influence of topographic relief.Our proposed method can also significantly reduce the interference of underlying weeds or shrubs on canopy segmentation,and achieves the highest Overall Accuracy(OA),F1 score,and mean Intersection over Union(mIoU)of 97.57%,95.49%,and 94.05%,respectively.Compared with other state-of-the-art networks such as SegFormer,SETR_PUP,TransUNet,TransFuse,and CCTNet,the mIoU obtained by the proposed method increases by 1.79%,8.83%,1.16%,1.43%,and 1.85%,respectively.The proposed method can achieve high-precision segmentation of fruit tree canopies with complex background information,which has important practical value for understanding the growth stat

关 键 词：无人机摄影测量 树冠分割 深度学习 迁移学习 卷积神经网络 注意力机制 感受野 语义特征 

分 类 号：S66[农业科学—果树学] TP18[农业科学—园艺学] TP391.41[自动化与计算机技术—控制理论与控制工程]