Multichannel Analysis in Weed Detection

Hericles Felipe Ferraz

1 a

, Jocival Dantas Dias Junior

, Andr

e Ricardo Backes

Daniel Duarte Abdala

3 b

and Mauricio Cunha Escarpinati

3 c

Faculty of Mechanical Engeneering, Uberl

andia Federal University, Uberl

andia, Brazil

Department of Computing, Federal University of S

ao Carlos, S

ao Carlos-SP, 13565-905, Brazil

Faculty of Computing, Federal University of Uberl

andia, Uberl

andia, Brazil

Keywords:

Precision Agriculture, Weed Detection, Color Spaces, Deep Learning, Remote Sensing.

Abstract:

In this paper a new classiﬁcation scheme is investigated aiming to improve the current classiﬁcation models

used in weed detection based on UAV imaging data. The premise is that the investigation regarding the

relevance of a given color space channel regarding its classiﬁcation power of important features could lead

to a better selection of training data. Consequently it could culminate on a superior classiﬁcation result. An

hybrid image is constructed using only the channels which least overlapping regarding their contribution to

represent the weed and soil data. It is then fed to a deep neural net in which a process of transfer learning

takes place incorporating the previously trained knowledge with the new data provided by the hybrid images.

Three publicly available datasets were used both in training and testing. Preliminary results seems to indicate

the feasibility of the proposed methodology.

1 INTRODUCTION

Agriculture represents by all accounts a relevant por-

tion of Brazilian economy. According to CEPEA

2020 data it represented 26.6% of the country’s GDP.

It is also a growing economical activity showing 6.1%

growth from 2019 to 2020. Brazil ﬁgures as the third

largest country in food production been the ﬁrst in

soy bean coffee and sugar cane production (Nachiluk,

2022). There are a number of reasons for those ﬁgures

such as the farming land available, climate and soil

fertility. Another contributing factor for the Brazilian

aggro business success could be the early and ever in-

creasing adoption of Precision Agriculture - PA. Ofﬁ-

cial data regarding the usage of PA technologies in

Brazil are scarce despite efforts in regulating such

as the creation of ”Comiss

ao Brasileira de Agricul-

tura de Precis

ao - CBAP” or Brazilian Commission

for Precision Agriculture a consulting group reporting

directly to the Federal Agricultural Ministry and the

Brazilian Association of Precision Agriculture Ser-

vice Providers (Molin, 2017). However, most data re-

garding the adaption of PA in Brazil still stems from

https://orcid.org/0000-0003-4299-8553

https://orcid.org/0000-0002-4647-8689

https://orcid.org/0000-0002-3792-056X

AP equipment companies.

The most recent data available points out to vari-

able adoption levels for distinct PA methods. Geo-

referencing in sowing and harvesting are among the

most widespread technology been quickly caught up

by soil testing for automated variable soil fertilization

based on grid maps. PA techniques focusing of weed

and invasive plants detection are still largely uncov-

ered. According to (de Carvalho, 2013) weed and in-

vasive plants are regarded as ”any plant that grows

spontaneously in a crop ﬁeld causing losses to the

farming activity.”. They can be very prejudicial to the

overall crop yield since they would naturally compete

for the soil nutrients, water and sunlight. Addition-

ally they can be vectors for invading plagues as well

as disease spreading agents. Weed invasion usually

are prejudicial to harvesting operations negatively im-

pacting the crop quality an yield. Weed and invasive

plants are usually managed by means of herbicides

since they are relatively inexpensive and presents a

speedy application process compared to the manual

weed removal. In fact, in large ﬁelds the manual ap-

proach is virtually impossible (Bucci et al., 2018). Al-

though cheap and quick, the indiscriminate herbicide

application presents by itself a number of troubling

environmental problems. It can damage the intended

crop, can pose risks for pollinating agents, inﬁltrate

Ferraz, H., Dias Junior, J., Backes, A., Abdala, D. and Escarpinati, M.

Multichannel Analysis in Weed Detection.

DOI: 10.5220/0011780900003417

In Proceedings of the 18th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2023) - Volume 4: VISAPP, pages

419-426

ISBN: 978-989-758-634-7; ISSN: 2184-4321

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

419

water tables and even poison fauna and humans.

A feasible way to mitigate the herbicide conse-

quences is to restrict its application to speciﬁc ﬁeld

regions in which the are prevalence of weed instead

of a uniform approach aiming the entire ﬁeld. Smart

herbicide usage is obviously preferable since it mit-

igates its harmful impact. A secondary advantage

would be the cost savings deriving of the smaller her-

bicides quantities that would have to be acquired and

managed. In order to enable the smart herbicide man-

agement a number of aiding PA technologies are re-

quired.

This work aims to investigate the inﬂuence of dis-

tinct color spaces used in various imaging sensors as

raw data to be fed into intelligent weed detection sys-

tems. Ten color spaces (RGB, HED, HSV, LAB, RG-

BCIE, XYP, YCBCR, YDBCRm YIQ, YPBPR and

YUV) are investigated. Distinct color spaces not nec-

essarily encode the same bands in the color spec-

tra. Consequently there is the possibility that spe-

ciﬁc color bands could highlight important features

pertaining the problem in question. The identiﬁca-

tion of an adequate color space can impact consider-

ably the success to be achieved by a given classiﬁca-

tion method. It is believed that such inﬂuence merits

a more thorough investigation. The premise is that

the investigation regarding the relevance of a given

color space channel regarding its classiﬁcation power

of important features could lead to a better selection

of training data. Consequently it could culminate on

a superior classiﬁcation result.

Preliminary evaluation had shown distinct chan-

nels deriving from different color spaces contribute

asymmetrically to the overall classiﬁcation process.

Consequently this study focus on combining channels

from distinct color spaces in order to maximize the

discriminating power of the weed classiﬁer. The ex-

periments were carried out using two publicly avail-

able annotated datasets.

The remainder of this document is organized as

follows: Section 2 Presents the theoretical back-

ground covering a brief review of the color spaces,

and the classiﬁcation model devised to process the in-

put data. Section 3 discusses the methods in which

the general methods workﬂow is presented followed

by a detailed description of the color channel selec-

tion process. This section closes with the presentation

of the experimental dataset. Section 4 presents the de-

vised experiments and discusses the obtained results.

Finally this work closes with conclusions and a pos-

sible description of the road ahead.

2 THEORETICAL BACKGROUND

One of the current PA research interest is to ﬁnd meth-

ods to accurately detect weed and infesting plants.

There are a number of different approaches been in-

vestigated.

Survey invasive plants in crop ﬁelds has increas-

ingly been carried out be means of imaging using

UAV - Unmanned Aerial Vehicles. It is a low cost

high resolution technique allowing the usage of dis-

tinct imaging sensors. UAVs can capture high res-

olution images using virtually any imaging sensor

able to be mounted on the vehicles frame. The most

common sensors used refers to the visible spectra us-

ing RBG sensors, infrared, multi-spectral and hyper-

spectral sensors (Yao et al., 2019). Sensors based

of refractance spectroscopy work based of the elec-

tromagnetic radiation reﬂection on different surfaces.

Those are the sensors which capture most information

about the crop since they sense the physicochemical

properties of the ﬁeld been imaged including chemi-

cal and biochemical properties. Unfortunately multi

and hyper-spectral sensors are out of reach for a con-

siderable portion of small and medium size farmers

given their elevated costs compared to simpler visi-

ble spectra sensors (Jafarbiglu and Pourreza, 2022).

Therefore most UAV crop surveys are carried out us-

ing cheaper readily available RGB band sensors.

With the emergence of deep learning by late

2010’s new decision models are been tested in PA.

Current work has shown satisfactory an above aver-

age. In special, weed and invasive plants detection has

shown considerable performance. Most work done so

far focus on UAV RGB sensed data given the reason

aforementioned (Hasan et al., 2021). Early efforts uti-

lize vegetation index computation as a pre-processing

step in order to enhance some visual features as pre-

sented in (Osorio et al., 2020) in which the NDVI in-

dex was chosen. In (Milioto et al., 2018) various veg-

etation indexes computed using RGB data are com-

pared in order to assess their inﬂuence in predictive

models. To the best of our knowledge there are no

work investigating the possibility of combining dis-

tinct sensors bands as input data to deep learning base

classiﬁcation models.

2.1 Color Spaces

Image pixel color perceived by the human eye are

nothing more than a speciﬁc spectral potency distri-

bution. Color is a uniquely useful feature to discrimi-

nate image data (Packyanathan et al., 2015). A color

space is a numerical system usually represented as a

3D or 4D matrix. According to (Hastings and Rubin,

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

420

2012) there are a number of color spaces available

tailored for speciﬁc applications. YCbCr and HSV

were devised for skin detection (Shaik et al., 2015),

CIELab and CIEluv (Mahy et al., 1991) were effec-

tive in image segmentation, and HED and LAB in

medical imaging.

It is possible to map a color space to another (Bi

and Cao, 2021). However the mapping is not always

perfect since a given color spaces focus on distinct

ranges of the electromagnetic spectra. This is the

main reason imaging sensors usually choose a given

color space that best represents the imaging range

provided by the sensor. There are additionally cur-

rent work that seems to demonstrate that a color space

change can impact positively the classiﬁcation model

performance. As an example (Fu et al., 2019) shows

improved results in retina image classiﬁcation by sim-

ply changing the color space used.

2.2 Classiﬁcation Model

The classiﬁcation model used in this work was chosen

given the actual success of deep neural nets in classi-

ﬁcation problems in special on computer vision appli-

cations. The proposed net can be seen in Figure 1. By

following the diagram the proposed net can be easily

reproduced. This net was realized using the Tensor

Flow framework.

This net is a composition of LinkNet (Chaurasia

and Culurciello, 2017) and vgg16 (Elharrouss et al.,

2022). The net model is composed as a pre-trained

vgg16 as encoder layer (nets upper half) in place of

LinkNets downsampling. This choice was made seek-

ing to take advantage of vgg16 good convolutional ar-

chitecture in order to emphasize the input image fea-

tures by means of their convolutional layers. As it

is usual, after the convolution a maxpulling layer is

necessary as shown in Figure 1. The lines positioned

along the diagram depict the information ﬂow along

the net architecture. Each layers is linked in sequence

to the next layer (from up to bottom). There are also

direct connections among the autoencoders meaning

that encoders and decoders are all connected. Such

connection confer to this net the status of a FCNN -

Fully Convolutional Neural Network. The diagrams

bottom half represents the UpSampling meaning that

for each convolution layer there are a corresponding

batch normalization followed by a RELU layer. The

encoder is responsible for condense the input layers.

The attained effect is to compact incrementally the in-

put data. The decoder does exactly the opposite in-

ﬂating the data in order to produce an output image of

desirable dimensions. This step is necessary in order

to confer to this net the ability of output data similar

Figure 1: Deep Neural Net architecture. The upper half

(downSampling) is basically vgg16 merged inside LinkNet.

to the input data.

3 METHODOLOGY

The proposed method is depicted in Figure 2. Ini-

tially the selected datasets is pre-processed in order

to convert the input data to the appropriated color

spaces (Ansari and Singh, 2022). The dataset images

were fed to step 1, in which the color space conver-

sion takes place. The conversion is necessary because

not all color spaces considered are provided directly

by the selected datasets. Additionally SC and SC2

provides only RGB data. SC provides some addi-

tional channels but still insufﬁcient to compose di-

rectly all color spaces. Once the conversion is over the

overlapping analysis takes place in step 2. It utilizes

the groundtruth to create two images. One composed

only by the pixels pertaining to weed data, and a sec-

ond one representing only soil data. The color distri-

Multichannel Analysis in Weed Detection

421

bution of both images is compared in order to mea-

sure the percentage of overlapping seeking to identify

which channels should be selected. The overlapping

analysis is somehow the most complex process in all

model, therefore it will be discussed in details in sub-

section 3.1. once the desirable channels are selected

they are combined in a new hybrid image on step 3

in order to produce the training data to be fed to the

classiﬁcation model. The training phase takes place

in step 4 in which a pre-trained deep neural net com-

posed by linknet and vgg16 is further trained to incor-

porate the new knowledge. In step 5 the trained net is

used in order to evaluate its classiﬁcation power. Step

6 represented the trained net that can be used in step

7 to predict which portions of the input data are weed

or soil.

Figure 2: General Workﬂow: Source Images and ﬁrstly con-

verted to distinct color spaces followed by a superposition

channel analysis. Hybrid images are generated using the

most signiﬁcant channels and fed to the training phase. Dur-

ing the training phase a process of transfer learning takes

place culminating in a new net of weights adequate to the

weed detection problem.

Figure 3 depicts the process of hybrid image gen-

eration. In order to provide a suitable input to the

training and classiﬁcation the selected channels must

be merged. The selected method is somehow simple.

An RGB template is used as a placeholder for the se-

lected channels from various color spaces. Each of

the three selected channels are mapped to the R, G

and B channels respectively. This is necessary given

the fact the classiﬁer expects a three channel image as

input as well as it provides a suitable way to visualize

the composed image.

3.1 Channel Selection

The selection of the most representative channels in

different color spaces was performed based on the fol-

lowing premise: ”Good channels to feed the classi-

ﬁcation model are the ones which present the least

overlapping of weed and soil”. Those two objects

of interest present distinct colors. Additionally the

dataset groundtruth provide pixelwise annotation on

them. Figure 3 highlights the overall process.

The ﬁrst step comprises the input images segmen-

tation into soil and weed data.The segmentation is

straightforward. Guided by the groundtruth the input

image is separated in two new images. One pertain-

ing only pixels annotated as weed data and the second

composed by the remainder soil.

For each original input image twenty new images

are produced since ten color spaces were considered

and for each color spaces two new images are pro-

duced. The color spaces considered in this work are

the following: RGB, HED, HSV, LAB, RGBCIE,

XYZ, YCBCR, YDBCR, YIQ, YPBPR, and YUV.

For each image the range ([min, max]) is computed

on each channel on both the weed and soil images.

Based on the range values the interval overlapping is

computed and expressed in percentage values. Table 1

presents the overlapping data computed for the entire

Beet dataset. Considering e.g. the lab color space. It

shows that channel 1 overlaps in 54.13%. This means

that channel 1 was used to represent both soil and

weed data and therefore could not be considered good

information to discriminate them. The desirable case

would be to ﬁnd a channel that was used to represent

weed but not soil and vice versa.

A hard threshold α is set to 5% meaning that a

given channel is selected if its overlapping score is α

or less. All channels stemming from all color spaces

are ranked and the three with smallest scores are se-

lected to compose the hybrid image which will be the

input data to be fed to the classiﬁer.

3.2 Experimental Dataset

In order to evaluate the methodology proposed in this

work a number of publicly available datasets were

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

422

Figure 3: Channel selection process. An input image is pro-

cessed based on groudtruth data in order to separate soil and

weed data. Afterwards those images are processed in or-

der to convert them in ten distinct color spaces. The color

spaces channels are decomposed and compared in order to

verify their soil/weed overlapping. Then three least over-

lapped channels among all color spaces are selected as in-

put data to be fed into the classiﬁer.

surveyed to ensure the reproducibility of the pro-

posed method. They were surveyed seeking annotated

data showing weed and invasive plants. The selected

datasets are the following:

(B) Beet - A publicly available dataset provided

by (Sa et al., 2018) was chosen since it presents inva-

sive plants. The predominant culture is beet and the

data was collected by an UAV using a multi-spectral

camera (RedEdge) in Germany.

(SC and SC2) Sugar-cane - In (Monteiro and

von Wangenheim, 2019) and (Pereira Junior and von

Table 1: Beet dataset overlapping rate considering all ten

color spaces.

Color Space Overlapping (%)

ch 1 ch 2 ch 3

rgb 5,06 100,00 33,18

hed 0,00 0,00 56,52

hsv 100,00 51,58 35,23

lab 54,13 0,00 13,17

rgbcie 9,90 100,00 35,85

xyz 30,48 100,00 36,57

ycbcr 42,62 31,22 0,00

ydbdr 42,62 31,25 0,00

yiq 42,62 0,00 0,00

ypbpr 42,62 31,22 0,00

yuv 42,62 31,22 0,00

Wangenheim, 2019) two datasets were made avail-

able depicting sugar-cane cultures along with invasive

plants. Those images were captured using an standard

RGB camera mounted on an UAV. The images were

manually annotated providing a suitable groundtruth.

4 EXPERIMENTS AND RESULTS

Section 3 describes how the hybrid images were cre-

ated. The entirety of the datasets were processed in

such way. It is then divided in training and testing

sets. A process of transfer learning takes place in or-

der to incorporate the knowledge provided by the new

data in the neural net also described in Section 3. Af-

terwards the testing set is fed to the classiﬁer and pre-

diction images are achieved.

Figure 4 shows a representative sample set of 30

testing images drawn from the Beet dataset. Presented

results are given in terms of VI - Variation of Informa-

tion (Meil

a, 2003) and f-measure or dice (Pandit et al.,

2011) indexes. Regarding the VI values close to zero

should be regarded as good scores indicating little to

no variation in comparison to the groundtruth data.

Values close to 1.0 represent high variation and con-

sequently an undesirable performance. The f-measure

is also set in the [0, 1.0] interval but with reverse inter-

pretation. The ﬁrst measure is information theoretical

and the second one is pair counting based. As one can

see by inspecting the data the overall performance is

good. For most data the VI index scores around and

bellow 0.2 and the f-measure index above and beyond

0.8. Notable exceptions do occur, in special in the

cases of sample data 14, 15, and 28. A closer in-

spection of the processed data is in order to try and

understand such anomalies.

Figure 5 present a sample set comprised of data

drawn from cases 14, 15, 11, 12, 20 and 21 respec-

Multichannel Analysis in Weed Detection

423

Figure 4: VI and F-measure results for a sample of 30 im-

ages drawn from the Beet dataset.

tively representing an undesirable, intermediate and

usual or desirable results. They are organized in ﬁve

columns. The ﬁrst column a RGB representation of

the original image. Column 2 is the processed hy-

brid image. Column 3 is an image processed using

the vegetation index (Erunova et al., 2021). Column

4 gives the dataset provided groundtruth and ﬁnally

column 5 depicts the prediction provided by the pro-

posed method.

The ﬁrst two rows representing cases 14 and 15

are samples in which the prediction diverged consid-

erably from the provided groundtruth. By direct vi-

sual inspection one can see that the groundtruth which

was generated by human experts does not necessar-

ily highlights all samples of weed present in the im-

age. This assertion is corroborated by the fact the

much of the weed data present in this image is con-

siderably small and hard to be accurately captured

by mid altitude images such as the ones generated

by a UAV mounted camera. However if one directs

his attention to the image provided in the third col-

umn representing the vegetation index it is possible

to see that the hybrid and prediction images tend to

agree much more with the prediction in comparison

to the provided groundtruth. This is due the fact the

ground truth was generated by visual inspection and

annotated manually by a human expect. It stands to

reason that very small patches of weed which would

be represented by sometimes one or even less pixels

(in which case the weed data would be spread frac-

Figure 5: Sample dataset highlighting images generated

during the process described in this method.

tionally among neighboring pixels) would be neglect

to be annotated. Another factor that could have im-

pacted negatively the ground truth generation process

would be the fact in those particular images the soil is

covered by dead matter usually stems from previous

crops which will partially cover the weeds and further

impending its visual identiﬁcation.

Rows three and four refers to cases 11 and 12

in which the overall performance of the proposed

method achieved a median result. By direct visual in-

spection one can conclude that hybrid, vegetation in-

dex and groundtruth tend to agree among themselves

considerably more. However the effects of weed cov-

ering by dead matter and small weed patches could

still be present.

Finally rows 5 and 6 referring to cases 20 and 21

which achieved above 0.8 f-measure and bellow 0.2

VI represent the most recurrent cases in which the

weed patches are well deﬁned and therefore could be

annotated very precisely on the groundtruth. Vegeta-

tion index seems to agree almost perfectly with the

annotated data which further corroborates the predic-

tion achieved.

5 CONCLUSIONS

This paper presented a work in progress pertaining to

a broader project aiming to devise a feasible method

to identify and highlight patches of week in crop

ﬁelds. Although very encouraging the results pre-

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

424

sented are only preliminary. The idea of cherry pick-

ing the most adequate channels among different color

spaces seems to indicate that this approach can lead to

superior results. However a number of further steps

needs to be taken in order to ensure the reproducibil-

ity and levels of accuracy could be ensured. The most

immediate problem is regarding the data. As one can

see in the results section it was identiﬁed that there

is a good change the annotation process available in

the used datasets are not necessarily precise leading to

miss-classiﬁcation. In order to address this issue we

are currently working on a huge multi-spectral dataset

featuring invasive plants. The aforementioned dataset

is been manually annotated by a pool of specialists

and it will be made publicly available as soon as it

is complete and validated. A further development to

ensure the accuracy of the produced groundtruth is to

provide the expects with vegetation index images in

order to aid and further guide them during the anno-

tation process.

ACKNOWLEDGEMENTS

The authors would like to thank the company Sen-

six Inovac¸

oes em Drones Ltda (http://sensix.com.br)

for providing the images used in the tests. Andr

e R.

Backes gratefully acknowledges the ﬁnancial support

of CNPq (National Council for Scientiﬁc and Techno-

logical Development, Brazil) (Grant #307100/2021-

9). This study was ﬁnanced in part by the

Coordenac¸

ao de Aperfeic¸oamento de Pessoal de

ıvel Superior - Brazil (CAPES) - Finance Code 001.

REFERENCES

Ansari, A. M. and Singh, D. K. (2022). Signiﬁcance of

color spaces and their selection for image processing:

A survey. Recent Advances in Computer Science and

Communications (Formerly: Recent Patents on Com-

puter Science), 15(7):946–956.

Bi, Z. and Cao, P. (2021). Color space conversion algorithm

and comparison study. Journal of Physics: Confer-

ence Series, 1976(1):012008.

Bucci, G., Bentivoglio, D., and Finco, A. (2018). Pre-

cision agriculture as a driver for sustainable farming

systems: state of art in literature and research. Cali-

tatea, 19(S1):114–121.

Chaurasia, A. and Culurciello, E. (2017). Linknet: Exploit-

ing encoder representations for efﬁcient semantic seg-

mentation. In 2017 IEEE Visual Communications and

Image Processing (VCIP), pages 1–4.

de Carvalho, L. B. (2013). Plantas Daninhas, volume 1. 1st

edition.

Elharrouss, O., Akbari, Y., Almaadeed, N., and Al-

Maadeed, S. (2022). Backbones-review: Feature

extraction networks for deep learning and deep re-

inforcement learning approaches. arXiv preprint

arXiv:2206.08016.

Erunova, M. G., Pisman, T. I., and Shevyrnogov, A. P.

(2021). The technology for detecting weeds in agri-

cultural crops based on vegetation index vari (plan-

etscope). Journal of Siberian Federal University. En-

gineering & Technologies, 14(3):347–353.

Fu, H., Wang, B., Shen, J., Cui, S., Xu, Y., Liu, J., and Shao,

L. (2019). Evaluation of Retinal Image Quality As-

sessment Networks in Different Color-Spaces, pages

48–56.

Hasan, A. S. M. M., Sohel, F., Diepeveen, D., Laga, H., and

Jones, M. G. (2021). A survey of deep learning tech-

niques for weed detection from images. Computers

and Electronics in Agriculture, 184:106067.

Hastings, G. D. and Rubin, A. (2012). Color spaces - a

review of historic and modern colour models. African

Vision and Eye Health, 71(3).

Jafarbiglu, H. and Pourreza, A. (2022). A comprehensive

review of remote sensing platforms, sensors, and ap-

plications in nut crops. Computers and Electronics in

Agriculture, 197:106844.

Mahy, M., Van Mellaert, B., Van Eycken, L., and Oost-

erlinck, A. (1991). The inﬂuence of uniform color

spaces on color image-processing-a comparative-

study of cielab, cieluv, and atd. Journal of imaging

technology, 17(5):232–243.

Meil

a, M. (2003). Comparing clusterings by the variation of

information. In Learning theory and kernel machines,

pages 173–187. Springer.

Milioto, A., Lottes, P., and Stachniss, C. (2018). Real-time

semantic segmentation of crop and weed for precision

agriculture robots leveraging background knowledge

in cnns. In 2018 IEEE International Conference on

Robotics and Automation (ICRA), pages 2229–2235.

Molin, J. P. (2017). Agricultura de precis

ao: N

umeros do

mercado brasileiro. Technical report, USP.

Monteiro, A. and von Wangenheim, A. (2019). Orthomo-

saic dataset of rgb aerial images for weed mapping.

http://www.lapix.ufsc.br/weed-mapping-sugar-cane.

Nachiluk, K. (2022). Alta na prodduc¸

ao e exportac¸

oes de

ac¸ucar marcam a safra 2020/21 de cana. an

alises e in-

dicadores o agroneg

ocio. Electronically. v. 16 n. 16.

Osorio, K., Puerto, A., Pedraza, C., Jamaica, D., and

Rodr

ıguez, L. (2020). A deep learning approach for

weed detection in lettuce crops using multispectral

images. AgriEngineering, 2(3):471–488.

Packyanathan, G., Veeraraghavalu, R., B.S, S., Kalist, V.,

and Basha, S. (2015). Satellite image segmentation

based on ycbcr color space. Indian Journal of Science

and Technology, 8:35.

Pandit, S., Gupta, S., et al. (2011). A comparative study

on distance measuring approaches for clustering. In-

ternational journal of research in computer science,

2(1):29–31.

Pereira Junior, P. and von Wangenheim, A. (2019). Or-

thomosaic dataset of rgb aerial images for crop rows

Multichannel Analysis in Weed Detection

425

detection. http://www.lapix.ufsc.br/weed-mapping-

sugar-cane.

Sa, I., Popovi

c, M., Khanna, R., Chen, Z., Lottes, P.,

Liebisch, F., Nieto, J., Stachniss, C., Walter, A., and

Siegwart, R. (2018). Weedmap: A large-scale seman-

tic weed mapping framework using aerial multispec-

tral imaging and deep neural network for precision

farming. Remote Sensing, 10(9).

Shaik, K. B., Ganesan, P., Kalist, V., Sathish, B., and

Jenitha, J. M. M. (2015). Comparative study of skin

color detection and segmentation in hsv and ycbcr

color space. Procedia Computer Science, 57:41–

48. 3rd International Conference on Recent Trends

in Computing 2015 (ICRTC-2015).

Yao, H., Qin, R., and Chen, X. (2019). Unmanned aerial ve-

hicle for remote sensing applications—a review. Re-

mote Sensing, 11(12).

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

426