欢迎访问林业科学,今天是

林业科学 ›› 2021, Vol. 57 ›› Issue (3): 98-107.doi: 10.11707/j.1001-7488.20210310

• 论文与研究报告 • 上一篇    下一篇

基于深度学习的小目标受灾树木检测方法

周焱1,刘文萍1,*,骆有庆2,宗世祥2   

  1. 1. 北京林业大学信息学院 北京 100083
    2. 北京林业大学林学院 北京 100083
  • 收稿日期:2019-10-17 出版日期:2021-03-01 发布日期:2021-04-07
  • 通讯作者: 刘文萍
  • 基金资助:
    十三五国家科技重大专项和国家研发计划项目"人工林重大灾害防控关键技术研究"(2018YFD0600200);北京市科技计划项目"北京生态公益林重大有害生物防空关键技术"(Z191100008519004)

Small Object Detection for Infected Trees Based on the Deep Learning Method

Yan Zhou1,Wenping Liu1,*,Youqing Luo2,Shixiang Zong2   

  1. 1. College of Information, Beijing Forestry University Beijing 100083
    2. College of Forestry, Beijing Forestry University Beijing 100083
  • Received:2019-10-17 Online:2021-03-01 Published:2021-04-07
  • Contact: Wenping Liu

摘要:

目的: 针对无人机森林图像中树木尺度小、生长密集以及分布不规律等问题,提出一种基于深度学习的小目标受灾树木检测方法,以准确识别和定位高分辨率无人机森林图像中的小尺度受灾树木。方法: 以辽宁省凌源县受红脂大小蠹侵害的油松林无人机图像为数据源,利用LabelImg开源软件标注拍摄高度为180~240 m的图像,构建无人机森林图像数据集。设计小目标受灾树木检测框架,该框架在轻量级目标检测框架(SSD)基础上,首先从conv3_3开始构建预测模块,并根据图像中目标树木的尺寸删减预测模块,同时优化特征图上默认框生成方式;然后,通过特征增强模块将基础特征图转化为增强特征图,生成基础和增强预测模块;最后,利用基于默认框的双层损失函数训练检测模型以促进特征学习。结果: 基于深度学习的小目标受灾树木检测方法可实现无人机森林虫害远程监测,较准确检测无人机图像中细小密集的受灾树木。采用5组不同结构的SITD模型以及SSD、FSSD和RFBNet目标检测框架对无人机森林图像数据集进行训练和测试,以平均查准率(AP)和Precision-Recall曲线作为评价指标,试验选出最优SITD模型在测试集上的AP为92.62%,相比原始SSD300框架提升69.71%,且优于其他3种目标检测框架。结论: 基于深度学习的小目标受灾树木检测方法可实现对森林中受灾树木的自动化检测,能够简化林业有害生物监测流程,提升对森林虫害的预警能力。

关键词: 无人机图像, 森林病虫害, 深度学习, 目标检测

Abstract:

Objective: The recognition of forest images taken with the unmanned aerial vehicle (UAV) It is an the important technology using the unmanned aerial vehicle (UAV) images for pest automatic monitoring and early warning tree recognition for the automatic pest-control warning system in a forestry. An approach on object detection based on deep learning was used to accurately recognize and locates small-scale infected trees in high-resolution forest UAV images are accurately recognized and located through deep learning object detection technology. Method: The images of pine forest infected by Dendroctonus valens from Liaoning Province were chosen used as the experimental data. The UAV forest image dataset was constructed by images taken from the a height of 180-240 m and annotated by open source software LabelImg. Because the instance size of UAV forest images dataset in this paper is far smaller than general datasets, a small infected trees detector (SITD) is designed to improve the ability of recognition and localization for small object, since the instance size of UAV forest images dataset in this paper is far smaller than general datasets. Firstly, based on the framework of single shot multi-box detector (SSD), the prediction module is built from conv3_3 and tailored according to the size of the target tree in the image, then the generation method of default boxes is optimized. Secondly, the base feature map is transformed into the enhanced feature map by the feature enhance module, thus generating the base and enhanced prediction module are generated by using base feature maps and enhanced feature maps which is transferred by the feature enhance module. Finally, dual loss functions are used to facilitate the feature learning. Result: The method of small object detection based on deep learning can realize UAV remote monitoring of for forest pests, and accurately detect small and dense infected trees in UAV images accurately. In this paper, five SITD models with different structures and other object detection frameworks as SSD, FSSD and RFBNet are used to trained and tested with on UAV forest image dataset, were compared by and average precision (AP) and Precision-Recall curve are selected as evaluation indexes. The proposed method achieves the average precision of 92.62% using the UAV images captured in our study, which gets a gain of 69.71% compared to the original SSD300, and higher than the other three frameworks. Conclusion: Experimental results demonstrate that the method of small object detection based on deep learning can realize automatic detection of infected trees, and simplify the monitoring process for forest pests, and which can improve the warning ability of forest pests.

Key words: unmanned aerial vehicle imagery, forest pest, deep learning, object detection

中图分类号: