摘要
【目的】针对物候期识别传统方法特征提取不充分、未对关键特征进行区分,导致方法泛化能力较差、迁移应用识别精度低的问题,本研究将注意力机制引入残差神经网络,结合基于数字照相的物候观测方式,提出具有较强细粒度特征识别能力且实用性较强的林木物候期识别方法,从而为林木的长期实时物候监测提供技术支撑。【方法】以PhenoCam中的栎林及槭林像片为研究材料,选取2017—2018年的数据作为训练集,以2019年的数据评价模型的泛化能力。研究结合实地观测数据对研究区的林木物候期进行划分,设计数据裁剪公式,在增强数据的同时均衡各个物候期数据的数量。研究基于ResNet50残差神经网络设计了深度学习模型,针对林木物候期差异较小的特性引入了注意力机制,注意力模块在通道及空间维度对神经网络提取的特征进行再处理,提升了模型对细粒度图像差异的识别能力。【结果】注意力机制的引入有效提升了模型的泛化能力,增强了模型对易混淆物候期的识别能力,在未参与训练的19年数据集的栎林物候期识别取得了90.58%的准确率,槭林物候期识别准确率为89.27%,较引入前模型准确率在两个研究区分别提升21.86%与13.15%,优于传统AlexNet和HOG-SVM方法。【结论】基于残差注意力网络的林木物候期识别方法较好解决了原有方法泛化能力较差的问题,该方法具有准确度高、迁移应用能力强等优势,能对易混淆的林木物候期进行精准区分,适用于林木物候的长期观测。
【Objective】In view of the insufficient feature extraction of traditional methods for phenological phase recognition and the indistinguishment of key features,resulting in poor generalization ability of the method and low recognition accuracy of migration applications.The attention mechanism was introduced into the residual neural network,combined with the phenological observation method based on digital photography,proposed a forest tree phenological period recognition method with strong fine-grained feature recognition ability and strong practicability,thereby provided technical support for long-term real-time phenological monitoring of forest.【Method】This article uses the Quercus and Acer photos in PhenoCam as the research material,selects the data from 2017 to 2018 as the training set.The data from 2019 is used to evaluate the generalization ability of the model.The study combines the field observation data to divide the forest trees in the study area,and designs the data cutting formula to balance the number of data in each phenological period while enhancing the data.The study designed a deep learning model based on the ResNet50 residual neural network,introduced an attention mechanism for the characteristics of small differences in forest phenology.The attention module reprocessed the features extracted by the neural network in the channel and spatial dimensions,improved the model The ability to recognize differences in fine-grained images.【Result】The introduction of the attention mechanism has effectively improved the generalization ability of the model and enhanced the model’s ability to recognize the easily confused phenology period.The Quercus phenological period recognition of the 19-year data set which did not participate in the training achieved an accuracy of 90.58%.The accuracy of Acer L.phenology recognition is 89.27%,compared with the model before the introduction,the accuracy rate in the two study areas is increased by 13.15%and 21.86%,which is better than the traditional AlexNet a
作者
崔晓晖
陈民
陈志泊
许福
王新阳
CUI Xiaohui;CHEN Min;CHEN Zhibo;XU Fu;WANG Xinyang(College of Information Science and Technology,Beijing Forestry University,Beijing 100083,China;Engineering Research Center for Forestry-oriented Intelligent Information Processing,Beijing Forestry University,Beijing 100083,China)
出处
《中南林业科技大学学报》
CAS
CSCD
北大核心
2021年第7期11-19,共9页
Journal of Central South University of Forestry & Technology
基金
国家自然科学基金项目(61772078,32071775)。
关键词
物候期识别
深度学习
注意力机制
精准林业
phenology recognition
deep learning
attention mechanism
precision forestry
