Semantic Image Segmentation with Deep Learning for Vine Leaf Phenotyping Petros N. TamvakisChairi Kiourt

2025-05-03 0 0 5.82MB 6 页 10玖币

侵权投诉

Semantic Image Segmentation with Deep

Learning for Vine Leaf Phenotyping

Petros N. Tamvakis ∗Chairi Kiourt ∗

Alexandra D. Solomou ∗∗ George Ioannakis ∗

Nestoras C. Tsirliganis ∗

∗Athena Research and Innovation Center, Xanthi, 67100 Greece

(e-mail: [petros.tamvakis,chairiq,gioannak,tnestor]@athenarc.gr)

∗∗ Institute of Mediterranean & Forest Ecosystems, Hellenic

Agricultural Organization ”DEMETER”, Athens, 11528 Greece

(e-mail: solomou@fria.gr)

Abstract:

Plant phenotyping refers to a quantitative description of the plant’s properties, however in

image-based phenotyping analysis, our focus is primarily on the plant’s anatomical, ontogenet-

ical and physiological properties. This technique reinforced by the success of Deep Learning

in the ﬁeld of image based analysis is applicable to a wide range of research areas making

high-throughput screens of plants possible, reducing the time and eﬀort needed for phenotypic

characterization. In this study, we use Deep Learning methods (supervised and unsupervised

learning based approaches) to semantically segment grapevine leaves images in order to develop

an automated object detection (through segmentation) system for leaf phenotyping which

will yield information regarding their structure and function. In these directions we studied

several deep learning approaches with promising results as well as we reported some future

challenging tasks in the area of precision agriculture. Our work contributes to plant lifecycle

monitoring through which dynamic traits such as growth and development can be captured

and quantiﬁed, targeted intervention and selective application of agrochemicals and grapevine

variety identiﬁcation which are key prerequisites in sustainable agriculture.

Keywords: Semantic segmentation, Grapevines, Phenotyping, Pattern recognition

1. INTRODUCTION

Recently, in the wine industry, there is growing inter-

est in exploring the role and eﬀect that grapevine vari-

eties play in adaptation strategies against global warming

(Guti´errez-Gamboa et al., 2021b). In light of this, there

is a need to ensure the genuiness of the plants in order

to avoid planting the wrong material, which can result

in considerable losses to viticulturists. The ﬁeld of am-

pelography is concerned with identifying and classifying

grapevine varieties using several parameters or descrip-

tors (Soldavini et al., 2009; Garcia-Mu˜noz et al., 2011).

It provides relevant morphological and agronomical in-

formation for varietal characterization studies, breeding

programs and conservation purposes (Khalil et al., 2017).

It is possible to distinguish grapevine varieties by the

phenotypic features of their leaves, which are unique to

each variety (Galet, 1998). More speciﬁcally, the features

of the leaves are the following: geometrical shape of the leaf

surface, perimeter of the leaf surface (Guti´errez-Gamboa

et al., 2021a), number of lobes, size of teeth, length of

teeth, ratio length/width of teeth, shape of blade (Fig. 1),

leaf length (L) and width (W) measurement (Fig. 2(a))

(Eftekhari and Kamkar, 2011), color of the upper side

of blade, undulation of blade between main and lateral

veins, general shape of petiole sinus, tooth at petiole sinus,

petiole sinus limited by veins, shape of upper lateral sinus,

Fig. 1. Diﬀerent leaf blade shapes.

depth of upper lateral sinus, shape of base, length of petiole

compared to middle vein (IPGRI et al., 1997). Grapevine

mature leaf’s parts are illustrated in Fig. 2(b).

Quantifying the dimensions of leaf veins and their struc-

ture has primarily relied on optical observation and empir-

ical experience of domain experts i.e. agronomists, agricul-

turists, botanologists etc. However, due to the signiﬁcantly

high degree of complexity of vein networks, their intense

color similarity with the overall leaf and the large number

of various cultivated plants and their problems, even well

trained experts of the area often fail to annotate success-

fully the over structure of the lead and are consequently

led to mistaken conclusions.

Unfortunately, there exists no suitable empirical method

to quantify physical vein network geometry with suﬃcient

scope and resolution which makes the evaluation of the

predicted vein network structure nearly impossible (Price,

2012). The development of an automated computational

arXiv:2210.13296v1 [cs.CV] 24 Oct 2022

(a) Leaf dimensions (b) Leaf parts

Fig. 2. Vine leaf characteristics.

system that detects leaf vein network structure, in real-

time, would oﬀer a valuable tool in the hands of the

agronomist who wishes to analyze such networks and

formulate/evaluate hypotheses regarding their structure

and function. Furthermore it can aid agricultural robots:

unmanned ground and aerial vehicles will identify genetic

traits, crop growth and plant diseases. Agribots equipped

with such systems will observe and measure crop growth

and inform on the ﬂy about a plant’s performance against

its predicted growth plan. In addition, the process of

the automated leaf samples collection, plant harvesting

and pruning through soft robotic arms (ﬁngers), will be

strengthened with more accurate and targeted actions.

This need spurred the development of a number of theoret-

ical Machine Learning (ML) models that predict optimal

vein network structures across a broad array of taxonomic

groups, from mammals to plants. In particular, a number

of empirical methods to quantify the vein network struc-

ture of plants, from roots, to xylem networks in shoots and

within leaves have been developed.

Advances in automated plant handling and digital imaging

along with the recent increases in computational power

and the expansion of system memory capabilities now

make it possible to use image-based analysis for plant phe-

notyping that allows for high-throughput characterization

of the extracted plant traits and improved understanding

of the adaptive and ecological signiﬁcance of vessel bundle

(veins) network structure.

2. RELATED WORK

The introduction of Deep Learning (DL) techniques into

agriculture (Carranza-Rojas et al., 2017) began to take

place in the context of precision agriculture. Traditional

ML approaches that have been extensively adopted in the

agricultural ﬁeld, such as plant disease investigation and

pest detection are gradually being replaced by more so-

phisticated techniques.The majority of which rely on some

kind of image recognition/classiﬁcation i.e. DL image pro-

cessing, image-based phenotyping, including weed detec-

tion (Milioto et al., 2017), crop disease diagnosis (Mohanty

et al., 2016; Bresilla et al., 2019), fruit detection (Ghosal

et al., 2018), and many other applications as listed in the

recent review (Kamilaris and Prenafeta-Bold´u, 2018).

The notion that leaves are a major hydraulic bottleneck

in plants (Sack and Holbrook, 2006), motivated attempts

to model patterns of conductance (Cochard et al., 2004;

Price et al., 2011) by measuring the geometry of veins.

Vein network structure can inﬂuence photosynthesis via

hydraulic eﬃciency, with recent work implicating vein den-

sity as a good predictor of photosynthetic rates (Brodribb

et al., 2007; Sack and Frole, 2006). Leaf vein patterning

is also associated with leaf shape in general, suggesting

shared developmental pathways (Dengler and Kang, 2001).

In Price et al. (2010), the authors developed a tool which

produces descriptive statistics about the dimensions and

positions of leaf veins and areoles by utilizing a series

of thresholding, cleaning, and segmentation algorithms

applied to images of leaf veins. An all-inclusive software

tool for mathematical and statistical calculations in plant

growth analysis that calculates up to six of the most

fundamental growth parameters according to a purely clas-

sical approach across one harvest-interval was developed

in (Hunt et al., 2002). Today the ﬁeld has broadened its

range from the initial characterization of single-plant traits

in controlled conditions towards real-life applications of

robust ﬁeld techniques in plant plots and canopies (Walter

et al., 2015).

Convolutional Neural Networks (CNN) (LeCun, 1989)

have been used in image recognition with remarkable suc-

cess and constitute one of the most powerful DL tech-

niques for modeling complex processes such as pattern

recognition in image based applications. Naturally, the ma-

jority of image-based approaches in precision agriculture

are based on popular CNN architectures. In (Lee et al.,

2015) the authors developed CNN architectures for auto-

matic leaf-based plant recognition. Pawara et al. (2017)

compared the performance of some conventional pattern

recognition techniques with that of CNN models, in plant

identiﬁcation, using three diﬀerent image-databases of ei-

ther entire plants and fruits, or plant leaves, concluding

that CNNs drastically outperform conventional methods.

Grinblat et al. (2016) presented a simple, yet powerful

Neural Network (NN) for the successful identiﬁcation of

three diﬀerent legume species based on the morphological

patterns of leave nerves. In their work, Dyrmann et al.

(2016) presented a method that can recognize weeds and

plant species using colored images. They used CNNs and

tested a total 10.413 images of 22 weeds and crop species.

Their model was able to achieve a classiﬁcation accuracy

of 86.2%. Mohanty et al. (2016) developed a smartphone-

assisted disease diagnosis system by employing two state-

of-the-art architectures: AlexNet (Krizhevsky et al., 2012)

and GoogLeNet (Szegedy et al., 2015) and trained their

model to identify 14 crop species and 26 diseases.

3. METHOD

Semantic image segmentation (Ronneberger et al., 2015;

He et al., 2020; Long et al., 2015) is a popular image

analyzing technique where each pixel is assigned to one

from a set of predeﬁned classes. It is a complex method

that entails the description, categorization, and visualiza-

tion of the regions of interest in an image contributing

to complete scene understanding. Semantic segmentation

models ﬁrst determine the presence or not of the objects

of interest in the picture and then classify object’s pixels

to their corresponding class.

3.1 Dataset

At this part it should be highlighted that the data ac-

quisition and dataset organization approach is targeted

on small number of images. Our dataset consists of 24

images with dimensions 7,952x5,304 pixels of vine leaves

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

SemanticImageSegmentationwithDeepLearningforVineLeafPhenotypingPetrosN.TamvakisChairiKiourtAlexandraD.SolomouGeorgeIoannakisNestorasC.TsirliganisAthenaResearchandInnovationCenter,Xanthi,67100Greece(e-mail:[petros.tamvakis,chairiq,gioannak,tnestor]@athenarc.gr)InstituteofMediterranean&Forest...

展开>> 收起<<

Semantic Image Segmentation with Deep Learning for Vine Leaf Phenotyping Petros N. TamvakisChairi Kiourt.pdf

共6页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Semantic Image Segmentation with Deep Learning for Vine Leaf Phenotyping Petros N. TamvakisChairi Kiourt

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: