无人机遥感影像的YOLOv3鼠洞识别技术

doi:10.11707/j.1001-7488.20201022

摘要/Abstract

摘要：

目的: 利用无人机遥感结合机器视觉识别荒漠林大沙鼠鼠洞，以期能够自动、准确和高效地获取鼠洞分布情况，为提高科学治理鼠害效率提供支持。方法: 针对古尔班通古特沙漠南缘荒漠林大沙鼠典型鼠害区，提出一种运用机器视觉结合无人机遥感对鼠害区域鼠洞进行识别定位的方法。该方法包含利用YOLOv3网络及其轻量级解决方案YOLOv3-tiny网络2种识别大沙鼠鼠洞的方案。首先从2次拍摄的低空遥感影像中挑选合适的影像，将其裁剪为416×416像素的图像，利用目标标注工具labelling对图像中鼠洞进行标注，构建荒漠林大沙鼠鼠洞数据集。而后利用k-means聚类算法对大沙鼠鼠洞数据集目标候选框的个数和宽高比维度进行聚类分析，选出适合数据集的先验框。最后调整参数对网络进行训练，采用对大沙鼠鼠洞数据集目标检测的准确率（precision，P）、召回率（recall，R）和平均精确率（average precision，AP作为目标检测算法的评价指标。结果: YOLOv3网络在大沙鼠鼠洞数据集上的平均查准率为92.37%，YOLOv3-tingy的平均查准率为85.86%，YOLOv3网络比YOLOv3-tiny网络AP要高6.51%，但YOLOv3-tiny参数量仅为YOLOv3的1/13；YOLOv3网络在本文硬件环境中检测单张图像的平均时间为0.83 s，YOLOv3-tiny为0.22 s。结论: 以上结果表明无人机遥感结合YOLOv3网络与YOLOv3-tiny网络可以实现对大沙鼠鼠洞的识别定位，高效的监测大沙鼠鼠洞的分布情况，弥补了传统方法监测鼠害时的不足，提高了监测大沙鼠鼠害的实时性、灵活性。

关键词: 机器视觉, 无人机, 监测, 鼠洞识别, 目标检测, YOLOv3

Abstract:

Objective: Unmanned aerial vehicle (UAV) remote sensing combined with machine vision is used to identify large gerbil holes in desert forests,and the technology can automatically,accurately and efficiently acquire the distribution of rat holes and provide support for improving the efficiency of scientific rodent pest management. Method: In this paper,a method for identifying and locating rat holes was proposed by combining machine vision with remote sensing of unmanned aerial vehicle (UAV),which is suitable for the typical haunt of rodents of desert forests in the southern edge of Gurbantunggut Desert. The method includes two schemes for identifying rat holes using the YOLOv3 network and its lightweight solution YOLOv3-tiny network. First,the appropriate images were selected from the two low-altitude remote sensing images,and the images were cropped to 416×416 pixel images. The rat holes in the images were labeled with a labeling tool,and data set of the gerbil holes in desert forest was constructed. Then,k-means clustering algorithm was used to cluster the number and aspect ratio dimensions of target candidate boxes in the gerbil hole data set,so as to select a prior box suitable for the data set. Finally,the adjusted parameters were used to train the network,and the target detection precision (P),recall (R) and average precision (AP) of the rat hole data set were used as the evaluation indexes of the target detection algorithm. Result: The average accuracy of YOLOv3 network on the rat hole data set is 92.37%,the average accuracy of YOLOv3-tiny is 85.86%,and the AP of YOLOv3 network is 6.51% higher than that of YOLOv3-tiny network,but the number of YOLOv3-tiny parameters is only 1/13 of that of YOLOv3. The average time for the YOLOv3 network to detect a single image in this paper's hardware environment is 0.83 s,and that detection time for YOLOv3-tiny is 0.22 s. Conclusion: The above results indicate that UAV remote sensing combined with YOLOv3 network and YOLOv3-tiny network can realize the identification and location of large gerbil holes,effectively monitor the distribution of rat holes,and make up the deficiency of traditional methods in monitoring rat damage and improve the real-time and flexibility of monitoring gerbil damage.

Key words: machine vision, UAV, monitoring, mouseholes to identify, target detection, YOLOv3

中图分类号:

TP753
S764.5

崔博超,郑江华,刘忠军,马涛,沈江龙,赵雪迷. 无人机遥感影像的YOLOv3鼠洞识别技术[J]. 林业科学, 2020, 56(10): 199-208.

Bochao Cui,Jianghua Zheng,Zhongjun Liu,Tao Ma,Jianglong Shen,Xuemi Zhao. YOLOv3 Mouse Hole Recognition Based on Remote Sensing Images from Technology for Unmanned Aerial Vehicle[J]. Scientia Silvae Sinicae, 2020, 56(10): 199-208.

图/表 10

图1

表1

图2

图3

图4

图5

图6

表2

图7

图8

参考文献 0

	陈梦, 杨永刚, 刘忠军, 等. 新疆昌吉州荒漠林害鼠种类调查初报. 防护林科技, 2014. (7): 37- 39.
	Chen M , Yang Y G , Liu Z J , et al. Investigation on rodent species in desert forest of Changji prefecture in Xinjiang. Protection Forest Science and Technology, 2014. (7): 37- 39.
	冯文武, 阿地力·沙塔尔. 昌吉市荒漠林鼠害种类调查及其防治技术研究. 防护林科技, 2014. (2): 8- 11, 13.
	Feng W W , Adili Shatar . Investigation on rodent species in desert forest of changji city and its control technology. Protection Forest Science and Technology, 2014. (2): 8- 11, 13.
	黄建文, 鞠洪波, 特木钦, 等. 阿拉善左旗天然梭梭林鼠害防治的遥感监测. 林业科学, 2004. 40 (3): 107- 110.
	Huang J W , Ju H B , Te M Q , et al. Monitoring on controlling effect of rhombomys opimus in haloxylon ammodendron bunge stands using remote sensing TM imagery in Alashan. Scientia Silvae Sinicae, 2004. 40 (3): 107- 110.
	康淑红. 新疆草原鼠害的综合防治技术分析. 当代畜牧, 2015. (32): 41- 42.
	Kang S H . Comprehensive control technique analysis of rodent pests in grassland in Xinjiang. Modern Animal Husbandry, 2015. (32): 41- 42.
	梁倩玲, 刘萍, 陈梦, 等. 林业有害生物灾害损失评估研究进展. 林业资源管理, 2015. (2): 139- 144.
	Liang Q L , Liu P , Chen M , et al. Research review on forest pest disaster losse assessment. Forest Resources Management, 2015. (2): 139- 144.
	马涛, 郑江华, 温阿敏, 等. 基于UAV低空遥感的荒漠林大沙鼠洞群覆盖率及分布特征研究——以新疆古尔班通古特沙漠南缘局部为例. 生态学报, 2018a. 38 (3): 953- 963.
	Ma T , Zheng J H , Wen A M , et al. Group coverage of burrow entrances and distribution characteristics of desert forest-dwelling Rhombomys opimus based on unmanned aerial vehicle(UAV) low-altitude remote sensing:A case study at the southern margin of the Gurbantunggut Desert in Xinjiang. Acta Ecologica Sinica, 2018a. 38 (3): 953- 963.
	马涛, 郑江华, 温阿敏, 等. 基于无人机低空遥感的荒漠林大沙鼠鼠洞分布与地形的关系——以新疆古尔班通古特沙漠南缘局部为例. 林业科学, 2018b. 54 (10): 180- 188.
	Ma T , Zheng J H , Wen A M , et al. Relationship between the distribution of Rhombomys opimus holes and the topography in desert forests based on low-altitude remote sensing with the unmanned aerial vehicle(UAV):a case study at the southern margin of the Gurbantunggut desert in Xinjiang, China. Scientia Silvae Sinicae, 2018b. 54 (10): 180- 188.
	盛兆湖, 陈梦, 刘忠军, 等. 新疆昌吉州荒漠林鼠(兔)害灾害损失评估指标重要性研究. 中国森林病虫, 2015. 34 (1): 14- 17, 22.
	Sheng Z H , Chen M , Liu Z J , et al. Importance of loss evaluation indexes of rodent disaster in the desert forests in Changji, Xinjiang Uygur Autonomous Region. Forest Pest and Disease, 2015. 34 (1): 14- 17, 22.
	孙迪, 倪亦非, 陈吉军, 等. 应用无人机(UAV)低空影像监测黄兔尾鼠鼠洞初探. 中国植保导刊, 2019. 39 (4): 35- 43.
	Sun D , Ni Y F , Chen J J , et al. Application of UAV low-altitude image on rathole monitoring of Eolagurus luteus. China Plant Protection, 2019. 39 (4): 35- 43.
	孙钰, 周焱, 袁明帅, 等. 基于深度学习的森林虫害无人机实时监测方法. 农业工程学报, 2018. 34 (21): 74- 81.
	Sun Y , Zhou Y , Yuan M S , et al. UAV real-time monitoring for forest pest based on deep learning. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2018. 34 (21): 74- 81.
	王术波, 韩宇, 陈建, 等. 基于深度学习的无人机遥感生态灌区杂草分类. 排灌机械工程学报, 2018. 36 (11): 1137- 1141.
	Wang S B , Han Y , Chen J , et al. Weed classification of remote sensing by UAV in ecological irrigation areas based on deep learning. Journal of drainage and irrigation machinery engineering(JDIME), 2018. 36 (11): 1137- 1141.
	温阿敏, 郑江华, 陈梦, 等. 荒漠生态林区大沙鼠鼠洞密度的无人机遥感监测技术初探. 林业科学, 2018. 54 (4): 186- 192.
	Wen A M , Zheng J H , Chen M , et al. Monitoring mouse-hole density by Rhombomys opimus in desert forests with UAV remote sensing technology. Scientia Silvae Sinicae, 2018. 54 (4): 186- 192.
	徐正刚, 赵运林, 李波, 等. 洞庭湖区东方田鼠灾害预警分析. 生态学杂志, 2013. 32 (10): 2830- 2836.
	Xu Z G , Zhao Y L , Li B , et al. Forecast of Microtus fortis disaster in Dongting Lake region of China. Chinese Journal of Ecology, 2013. 32 (10): 2830- 2836.
	轩俊伟, 郑江华, 倪亦非, 等. 基于动力三角翼平台的草原鼠害遥感监测研究. 中国植保导刊, 2015. 35 (2): 52- 55.
	Xuan J W , Zheng J H , Ni Y F , et al. Remote sensing monitoring of rodent infestation in grassland based on dynamic delta-wing platform. China Plant Protection, 2015. 35 (2): 52- 55.
	杨红艳, 杜健民, 王圆, 等. 基于无人机遥感与卷积神经网络的草原物种分类方法. 农业机械学报, 2019. 50 (4): 188- 195.
	Yang H Y , Du J M , Wang Y , et al. Classification method of grassland species based on Unmanned Aerial Vehicle remote sensing and Convolutional Neural Network. Transactions of the Chinese Society for Agricultural Machinery, 2019. 50 (4): 188- 195.
	郑二功, 田迎芳, 陈涛, 等. 基于深度学习的无人机影像玉米倒伏区域提取. 河南农业科学, 2018. 47 (8): 155- 160.
	Zheng E G , Tian Y F , Chen T , et al. Region extraction of corn lodging in UAV images based on deep learning. Journal of Hennan Agricultural Sciences, 2018. 47 (8): 155- 160.
	周晓琳, 安如, 陈跃红, 等. 三江源典型区鼠洞无人机遥感识别研究. 亚热带资源与环境学报, 2018. 13 (4): 85- 92.
	Zhou X L , An R , Chen Y H , et al. Identification of rat holes in the typical area of"Three-River Headwaters"region By UAV remote sensing. Journal of Subtropical Resources and Environment, 2018. 13 (4): 85- 92.
	Addink E A , De Jong S M , Davis S A , et al. The use of high-resolution remote sensing for plague surveillance in Kazakhstan. Remote Sens Environ, 2010. 114, 674- 681. doi: 10.1016/j.rse.2009.11.015
	Everingham M , Eslami S M A , Gool L V , et al. The pascalvisual object classes challenge:A retrospective. International Journal of Computer Vision, 2015. 111 (1): 98- 136.
	Kausrud K L , Viljugrein H , Frigessi A , et al. Climatically driven synchrony of gerbil populations allows large-scale plague outbreaks. Proc R Soc B Biol Sci, 2007. 274, 1963- 1969. doi: 10.1098/rspb.2007.0568
	Redmon J , Divvala S , Girshick R , et al. You only look once:unified, real-time object detection. IEEE Conference on Computer Vision and Pattern Recognition, 2016. 779- 788.
	Ren S , He K , Girshick R B , et al. Faster R-CNN:towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 39 (6): 1137- 1149. doi: 10.1109/TPAMI.2016.2577031
	Redmon J , Farhadi A . YOLO9000:better, faster, stronger. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 6517- 6525.
	Redmon J, Farhadi A. 2018. YOLOv3: an incremental improvement. IEEE Conference on Computer Vision and Pattern Recognition, arXiv preprint arXiv: 1804.02767.

名称 Name	参数 Parameter	数值 Value
DOPSV360无人机 DOPSV360 UAV	最大起飞质量Maximum take-off mass/kg	5.5
	机身翼展长×宽×高Fuselage wingspan× width× height	2 000 mm× 1 300 mm× 400 mm
	最大飞行高度Maximum flight altitude/m	5 000
	最大水平飞行速度Maximum horizontal speed/(m·s^-1)	19.4
	最大飞行时间Maximum flight time/min	90
	通讯半径Communication radius/km	30
	正常工作风力允许等级Normal working wind	4
索尼NEX7相机 Sony NEX7 camera	质量Mass/g	291
	传感器尺寸Sensor size	35.9 mm×24 mm
	有效像素Effective pixels	24 300 000
	最高分辨率Maximum resolution(pixels)	6 000×4 000
索尼ilce-7m相机 Sony ilce-7m camera	质量Mass/g	407
	传感器尺寸Sensor size	23.4 mm×15.6 mm
	有效像素Effective pixels	36 400 000
	最高分辨率Maximum resolution(pixels)	7 360×4 912

网络 Network	真正例 TP	假正例 FP	假负例 FN	准确率 P(%)	召回率 R(%)
YOLOv3	589	56	32	91.3	94.8
YOLOv3-tiny	513	86	78	85.6	86.8

[1]	张会儒,雷相东,李凤日. 中国森林经理学研究进展与展望[J]. 林业科学, 2020, 56(9): 130-142.
[2]	谢锦莹,丁丽霞,王志辉,刘丽娟. 基于FCN与面向对象的滨海湿地植被分类[J]. 林业科学, 2020, 56(8): 98-106.
[3]	李亚东,曹明兰,李长青,冯仲科. 基于POS约束的无人机森林航摄影像分区匹配策略[J]. 林业科学, 2020, 56(10): 113-120.
[4]	吴项乾,曹林,申鑫,汪贵斌,曹福亮. 基于无人机激光雷达的银杏人工林有效叶面积指数估测[J]. 林业科学, 2020, 56(1): 74-86.
[5]	王卫, 杨俊杰, 罗晓莹, 周长江, 陈世发, 杨志军, 侯荣丰, 陈再雄, 李永生. 基于Maxent模型的丹霞山国家级自然保护区极小种群植物丹霞梧桐的潜在生境评价[J]. 林业科学, 2019, 55(8): 19-27.
[6]	李丹, 张俊杰, 赵梦溪. 基于FCM和分水岭算法的无人机影像中林分因子提取[J]. 林业科学, 2019, 55(5): 180-187.
[7]	孔维尧, 孙权, 刘鑫鑫, 曲丽, 王福友, 姚明远, 邹红菲. 基于红外相机监测的汪清自然保护区东北豹种群动态[J]. 林业科学, 2019, 55(5): 188-196.
[8]	夏永杰, 庞勇, 刘鲁霞, 陈博伟, 董斌, 黄庆丰. 高精度DEM支持下的多时期航片杉木人工林树高生长监测[J]. 林业科学, 2019, 55(4): 108-121.
[9]	宋以宁, 刘文萍, 骆有庆, 宗世祥. 基于线性谱聚类的林地图像中枯死树监测[J]. 林业科学, 2019, 55(4): 187-195.
[10]	周小成, 何艺, 黄洪宇, 许雪琴. 基于两期无人机影像的针叶林伐区蓄积量估算[J]. 林业科学, 2019, 55(11): 117-125.
[11]	莫锦华,李佳,刘芳,李晓光,李迪强. 利用红外相机调查海南尖峰岭地区兽类和鸟类多样性[J]. 林业科学, 2019, 55(10): 203-210.
[12]	温阿敏, 郑江华, 陈梦, 穆晨, 马涛. 荒漠生态林区大沙鼠鼠洞密度的无人机遥感监测技术初探[J]. 林业科学, 2018, 54(4): 186-192.
[13]	马涛, 郑江华, 温阿敏, 陈梦, 穆晨. 基于无人机低空遥感的荒漠林大沙鼠鼠洞分布与地形的关系——以新疆古尔通古特沙漠南缘局部为例[J]. 林业科学, 2018, 54(10): 180-188.
[14]	刘奇, 吴怀栋, 谭运洪, 张教林. 西双版纳热带季雨林木质藤本多样性及其攀援方式[J]. 林业科学, 2017, 53(8): 1-8.
[15]	张煜星, 王雪军, 黄国胜, 党永峰, 陈新云. 森林面积多阶遥感监测方法[J]. 林业科学, 2017, 53(7): 94-104.