基于空间交叠主动学习的无人机光学影像桉树树冠检测  

UAVs optical image-based Eucalyptus canopy detection using active learning with spatial overlap indicator

作　　者：段炼 梁波 李震 罗天啸 黄超群 黄国斌 DUAN Lian;LIANG Bo;LI Zhen;LUO Tianxiao;HUANG Chaoqun;HUANG Guobin(School of Natural Resources and Surveying,Nanning Normal University,Nanning 530001,Guangxi,China;Guangxi Forest Resources and Environment Monitoring Center,Nanning 530001,Guangxi,China;Management Center of National Dagui Mountain Alligator Lizard Nature Reserve in Guangxi Province,Hezhou 542800,Guangxi,China)

机构地区：[1]南宁师范大学自然资源与测绘学院,广西南宁530001 [2]广西壮族自治区森林资源与生态环境监测中心,广西南宁530001 [3]广西大桂山鳄蜥国家级自然保护区管理中心,广西贺州542800

出　　处：《中南林业科技大学学报》2025年第3期10-19,共10页Journal of Central South University of Forestry & Technology

基　　金：广西林业科技推广示范项目(2023GXLK05);南宁市武鸣区科学研究与技术开发计划项目(20220107);广西大学生创新创业训练计划项目(S202410603111)。

摘　　要：【目的】基于监督深度学习的遥感影像树冠检测逐渐成为森林资源清查和监测的重要技术手段。然而,已有方法需要大量标注样例,导致了高昂的成本和较低的通用性。在该挑战下,本研究设计了一种新的主动学习方法开展了UAVs光学影像下端到端的桉树树冠识别。【方法】采用“Teacher-student”交互学习模式,在每轮学习中通过Teacher模型生成候选伪样本,基于伪样本筛选策略得到高价值的目标伪样本,再将其与已有样本放入Student模型学习,之后将Student模型的参数迁移给Teacher模型。经过以上多轮交互学习,Teacher模型即为所求。特别地,该方法在模型中引入了梯度均衡机制损失以降低对易样本的过度训练,设计了新颖的空间交叠度以加强模型对树冠遮挡严重和多树种共存等难伪样本的学习比重,采用了多尺寸网格掩码等数据增强方法提升模型对桉树空间分布模式、多光照场景和异常拍摄视角的适应性,显著减少标注工作量并提高模型性能。【结果】利用大疆Phantom4 Pro V2.0无人机对广西国有高峰林进行数据采集,形成了8 000张256 pixel×256 pixel影像构成的包含幼龄林和近成过熟林标注的样本集。利用该样本集对所提方法和其他监督学习方法及主动学习方法进行对比分析。结果表明,所提的主动学习方法在使用少量样本的情况下具有比监督学习方法和最新主动学习方法更优异的性能:在使用26%的数据作为样本时,该方法的F1值为0.8,满足了树冠识别实用性要求;而当样本增加到34%时,该方法更是取得了0.9的F1值,与全监督学习性能相当。【结论】提出的主动学习方法在小样本约束下能自动获取准确的树冠范围信息,可节省大量数据处理和样本制作时间,具有高效、便捷和低成本的优势,对提升森林监测效率和自动化水平具有重要意义。【Objective】Deep learning-based canopy detection from remote sensing imagery is gradually becoming an important technique for forest inventory and monitoring.However,existing methodologies usually require a large number of labeled examples,resulting in high costs for both annotation and sample acquisition,thus limiting their applicability.To address this challenge,a novel active learning method was designed for comprehensive Eucalyptus canopy recognition utilizing UAV optical imagery.【Method】The proposed method employs a “Teacher-student” interactive learning mode.In each learning stage,the Teacher model generates candidate pseudo-samples,and high-value target pseudo-samples are obtained based on a pseudo-sample selecting strategy,then combined with existing labeled samples and input into the Student model for training.Subsequently,the parameters of the Student model are transferred to the Teacher model.After multiple rounds of interactive learning,the teacher model becomes the final model for tree canopy detection applications.Specifically,the method introduces a gradient Harmonized mechanism loss(GHM loss) in the Student model to reduce over-training on easy samples.It also designs a novel spatial overlap indicator to strengthen the model's learning emphasis on difficult pseudo-samples with severe canopy occlusions and coexistence of multiple tree species.Moreover,the method adopts multi-size grid mask and other data augmentation methods to enhance the model's adaptability to the spatial distribution patterns of trees,diverse lighting conditions and unconventional photographic angles.These enhancements collectively lead to a significant decrease in labeling workload and a improvement in model performance.【Result】Data collection was conducted on Guangxi Gaofeng forests using the DJI Phantom 4 Pro V2.0 UAVs,resulting in a dataset consisting of 8 000 annotated samples with 256 pixel×256 pixel images of both young and nearly mature forests.This dataset was used to compare our proposed method with ot

关 键 词：树冠识别 主动学习 伪样本筛选 目标检测 

分 类 号：S771.8[农业科学—森林工程] S792.39[农业科学—林学]