Spray Quality Assessment on Water-Sensitive Paper Comparing AI and

Classical Computer Vision Methods

es Sim

oes

1,2 a

, Andr

e Baltazar

2 b

, Armando Sousa

1,2 c

and Filipe Neves dos Santos

2 d

FEUP - Faculdade de Engenharia da Universidade do Porto, Universidade do Porto, R. Dr. Roberto Frias, Porto, Portugal

INESC TEC - Instituto de Engenharia de Sistemas e Computadores, Tecnologia e Ci

encia,

R. Dr. Roberto Frias, Porto, Portugal

ﬁ

Keywords:

Precision Agriculture, Water-Sensitive Spray, Classical Computer Vision, Spray Quality Assessment, Instance

Segmentation.

Abstract:

Over recent decades, precision agriculture has revolutionized farming by optimizing crop yields and reducing

resource use through targeted applications. Existing portable spray quality assessors lack precision, espe-

cially in detecting overlapping droplets on water-sensitive paper. This proposal aims to develop a smartphone

application that uses the integrated camera to assess spray quality. Two approaches were implemented for

segmentation and evaluation of both the water-sensitive paper and the individual droplets: classical computer

vision techniques and a pre-trained YOLOv8 deep learning model. Due to the labor-intensive nature of an-

notating real datasets, a synthetic dataset was created for model training through sim-to-real transfer. Results

show YOLOv8 achieves commendable metrics and efﬁcient processing times but struggles with low image

resolution and small droplet sizes, scoring an average Intersection over Union of 97.76% for water-sensitive

spray segmentation and 60.77% for droplet segmentation. Classical computer vision techniques demonstrate

high precision but lower recall with a precision of 36.64% for water-sensitive paper and 90.85% for droplets.

This study highlights the potential of advanced computer vision and deep learning in enhancing spray quality

assessors, emphasizing the need for ongoing reﬁnement to improve precision agriculture tools.

1 INTRODUCTION

Accurate evaluation of agricultural spray applica-

tions is critical for ensuring optimal pesticide usage

and minimizing environmental impact. Although vi-

sual assessment of Water-Sensitive Papers (WSPs) is

the norm for spray application of pesticides, it has

shown to be an unreliable method for correctly assess-

ing the paper’s density and distribution of droplets.

These challenges call for the development of im-

proved methods for analyzing WSP data to enhance

the precision and reliability of agricultural spray as-

sessments.

The use of pesticides in crop management is cru-

cial but can have signiﬁcant economic, environmen-

tal, and health consequences if not applied correctly.

The effectiveness of pesticides relies heavily on the

application method to achieve an optimal droplet pat-

https://orcid.org/0009-0004-2124-4351

https://orcid.org/0000-0003-4047-1395

https://orcid.org/0000-0002-0317-4714

https://orcid.org/0000-0002-8486-6113

tern, reducing drift and ensuring precise product de-

position on the target (Privitera et al., 2023). How-

ever, inconsistent pesticide spray coverage often re-

sults in unequal protection across ﬁelds, leading to nu-

merous issues, such as the need for repeating pesticide

applications, the development of weed resistance, and

potential shifts in pest behavior. These shortcom-

ings impose not only economic burdens but also am-

plify environmental and health risks (Machado et al.,

2018). Moreover, factors such as the variability in

weather conditions, equipment, and applicators fur-

ther contribute to this inconsistency, posing signiﬁ-

cant challenges for effective and sustainable ﬁeld crop

protection (Nansen et al., 2021).

Precision spraying can be optimized by knowing

the values of certain metrics such as Volume Me-

dian Diameter (VMD), which is measured by mul-

tiple image analyzers and stands as a key factor in

reducing wasted spray and thereby minimizing envi-

ronmental damage and cutting operational costs (Priv-

itera et al., 2023). There are already a few meth-

ods available to assess the performance of pesticide

300

Simões, I., Baltazar, A., Sousa, A. and Santos, F.

Spray Quality Assessment on Water-Sensitive Paper Comparing AI and Classical Computer Vision Methods.

DOI: 10.5220/0013027700003822

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 21st International Conference on Informatics in Control, Automation and Robotics (ICINCO 2024) - Volume 2, pages 300-307

ISBN: 978-989-758-717-7; ISSN: 2184-2809

applications with sophisticated imaging analyses and

water-sensitive spray cards, which are able to quan-

tify spray coverage and evaluate the efﬁciency of the

pesticide distribution (Nansen et al., 2021). Tech-

niques have evolved to calculate these metrics in the

last few years, from manual processes to automated

solutions such as DropletScan (Machado et al., 2018)

and SmartSpray (Nansen et al., 2021). This signiﬁes

a growing need for more precise and efﬁcient assess-

ment tools (Xun and Gil, 2024). The integration of

artiﬁcial intelligence and machine learning in agricul-

ture can also make signiﬁcant advancements in as-

sessing spray quality (Privitera et al., 2023).

This research addresses the challenges described

by developing a dual-method framework integrating

both classical computer vision and advanced machine

learning solutions for accurate WSP and droplet seg-

mentation. To allow real-time image processing and

statistical analysis of WSP in a practical tool for ﬁeld

use, an Android application was developed that is

ready to be used for photography and analysis of the

WSP. Additionally, a synthetic dataset for images of

WSP with droplets was created to augment the train-

ing data as a means to improve the accuracy and gen-

eralization of the machine-learning models, given the

limited number of annotated WSP images available.

This paper reviews existing methodologies for

Water Spray Pattern (WSP) analysis, identifying their

limitations through a gap analysis. It presents the pri-

mary metrics for evaluating spray quality and intro-

duces an innovative approach to enhance model train-

ing using synthetic datasets. The methodology de-

scribes both classical computer vision and machine

learning techniques for WSP and droplet segmenta-

tion, alongside the development of an Android appli-

cation and web service. The results demonstrate the

efﬁcacy of the proposed solutions, comparing perfor-

mance metrics across real and synthetic datasets. The

study concludes with a summary of ﬁndings, implica-

tions for future research, and potential directions for

further work.

2 LITERATURE REVIEW ON

WATER-SENSITIVE SPRAY

ASSESSMENT

Water Sensitive Papers (WSPs) have been a funda-

mental tool for agricultural spray evaluation for over

40 years, used in both aerial and ground applica-

tions (Cerruto et al., 2019; Marc¸al and Cunha, 2008).

WSPs are semi-rigid papers coated with bromoethyl

blue on one side and are available in various sizes.

The coating appears yellow when dry but turns differ-

ent shades of brown, blue, and purple when it comes

into contact with water droplets. These stains cre-

ate a signiﬁcant contrast with the dry yellow back-

ground, making it easier to assess the dispersion of

the droplets (Privitera et al., 2023; Machado et al.,

2018; Hoffmann and Hewitt, 2005). However, this

method faces limitations such as the inability to mea-

sure droplets smaller than 50 µm Machado et al.

(2018); Cerruto et al. (2019), sensitivity to high hu-

midity (Nansen et al., 2021; Hoffmann and Hewitt,

2005), and inaccuracies with high coverage (Marc¸al

and Cunha, 2008). These limitations call for alterna-

tive methods for accurate spray analysis.

Rosin-Rammler distribution is a widely used func-

tion expressing drop size distribution with two pa-

rameters, representative diameter and a measure of

drop size dispersion, and is useful for single-peaked

results. Its simplicity and ability to extrapolate into

difﬁcult-to-measure ranges make it popular (Lefebvre

and McDonell, 2017; D

echelette et al., 2011).

The evaluation of droplet sizes commonly uses

Volume Median Diameter (VMD), which expresses

drop size based on the volume of liquid sprayed.

The spray is divided into two equal parts based on

the sprayed volume, meaning that 50% of the total

volume is made up of drops with diameters larger

than the VMD value, and the other 50% by droplets

with diameters smaller than the VMD value (Privit-

era et al., 2023; Schick, 2008). Relative Span Fac-

tor (RSF) is another metric commonly used. It is a

dimensionless parameter indicative of the uniformity

of the drop size distribution. RSF is a practical way

of comparing various drop size distributions. High

RSF values indicate wider drop size distributions and

lower RSF values indicate less variation among drop

sizes (Lefebvre and McDonell, 2017; Privitera et al.,

2023; Machado et al., 2018; Schick, 2008).

YOLOv8, a state-of-the-art model in object detec-

tion, excels in efﬁciency and speed, making it suitable

for real-time applications like precision agriculture.

There are an array of software solutions available

to characterize spray quality. Zhu et al. (2011) cre-

ated DepositScan, which is a portable scanner com-

prised of a handheld card scanner and a deposit col-

lector. Nansen et al. (2015) presented SnapCard as

an app available for both iOS and Android smart-

phones that offers two core functions: prediction of

spray coverage based on pre-application weather data

and spray settings, and post-application measurement

for quality control. Machado et al. (2018) intro-

duced DropLeaf, a smartphone app tailored to esti-

mating pesticide amounts on WSPs, assisting farmers

and agronomists in measuring spray coverage. Addi-

Spray Quality Assessment on Water-Sensitive Paper Comparing AI and Classical Computer Vision Methods

301

tionally, there are multiple other methodology solu-

tions that although are not readily available and com-

mercialized, they serve important contributions to the

ﬁeld.

Omer Barıs¸

Ozl

uoymak and Bolat (2020) de-

veloped a novel image processing software within

Vision Acquisition Software (VAS) by National In-

struments that aims to assess accurately spray cov-

erage rates and droplet counts on WSP, including

overlapped droplets. Xun and Gil (2024) developed

a novel methodology that focuses on precisely seg-

menting overlapping droplets by utilizing concave

point detection and ellipse ﬁtting, ensuring optimal

accuracy when the coverage is below 25%. Liu et al.

(2024) presented a novel Optical Droplet Edge Imag-

ing method that uses a device to acquire images from

the top and bottom side of the droplet deposition sur-

face to obtain the correct size measurement of the

droplets.

3 DESIGN AND

IMPLEMENTATION OF

PROPOSED SOLUTION

The main goal of the system is to create an An-

droid application tool for real-time assessment of

spray quality on water-sensitive paper. The An-

droid application, client-side, captures an image of

the water-sensitive paper and sends it to the server

side. The server analyzes the image, provides droplet

statistics, and sends the results back to the client.

The server-side application uses a dual-method al-

gorithm that combines classical computer vision and

machine learning techniques to accurately segment

water-sensitive paper and droplets. The primary ob-

jective of both algorithms is to detect and separate

overlapped droplets, ensuring that droplets that over-

lap are counted as separate entities for statistical pur-

poses. Both methods developed follow an equiva-

lent logic: detect the water-sensitive paper, remove

distortion, segment individual droplets, and calculate

the WSP statistics. This dual-algorithm approach al-

lows to directly compare distinct methods of analyz-

ing water-sensitive papers.

The standard WSP image contains over 1000

droplets that require manual labeling to establish ac-

curate ground truth. This is challenging due to pixel-

level segmentation requirements and the risk of incor-

rect labels. To overcome this, an algorithm was cre-

ated to generate a simulation-based dataset, produc-

ing synthetic data that replicates real-world processes

and automatically generates ground truth annotations.

This approach saves time, ensures consistency, and

improves annotation accuracy, which is vital for ef-

fective machine learning model training.

Two distinct sets of images of water-sensitive pa-

per were provided that form the basis for the analy-

sis and study of the visual aspect of real spraying on

a WSP. The images were also used to test the appli-

cation’s algorithms implemented. Given the task of

manually labeling each droplet, only two images were

properly annotated for testing purposes.

Both datasets will be made available further into

the future.

3.1 Synthetic Dataset

The synthetic dataset addresses the data acquisition

and annotation challenges and was designed for trans-

fer learning during the training of CNN models. Ele-

ments from real datasets, such as droplet colors, pat-

terns, and shapes, were combined to enhance accu-

racy and simulate realistic conditions. The dataset

represents a range of spraying scenarios with over-

lapped droplets. The Rosin-Rammler distribution was

used to calculate droplet sizes, a method commonly

used to describe particle sizes in sprays. The synthetic

datasets were created with three different image res-

olutions to test model generalization across varying

resolutions, as well as multiple different droplet size

distributions.

The background of a WSP was generated with a

radial gradient and random imperfections to mimic

real-world scenarios. Droplet sizes were determined

using the Rosin-Rammler distribution. Both colors

for the background and the droplets were sourced

directly from real datasets. The algorithm gener-

ates the placement of each droplet randomly. Then,

it identiﬁed the overlaps between the droplets and

applied a ”shape burst” coloring technique for real-

ism. Droplets were positioned using a sliding window

technique to manage the computational load. Once

positioned, droplets were colored based on the dis-

tance to the nearest edge, ensuring smooth color tran-

sitions.

(a) Syn-

thetic

droplets.

(b) Syn-

thetic

droplets.

(d) Real

droplets.

Figure 1: Comparison between droplets from the synthetic

dataset (image (a) and (b)) and the real dataset (image (c)

and (d)).

ICINCO 2024 - 21st International Conference on Informatics in Control, Automation and Robotics

302

In the end, for testing purposes, background com-

positing was applied to seamlessly integrate synthetic

droplets with realistic backgrounds, using the base

images from real datasets. This process involved

scaling and layering synthetic elements to maintain

authenticity, ensuring the resulting composite image

looked natural.

(a) Syn-

thetic WSP.

(b) Syn-

thetic WSP.

WSP.

(d) Real

WSP.

Figure 2: Comparison between droplets from the synthetic

dataset (image (a) and (b) and the real dataset (image (c)

and (d)).

3.2 Water-Sensitive Paper Statistics

To accurately assess the quality of a spray in a water-

sensitive paper (WSP) analysis, several statistical

measures must be considered. The most representa-

tive and widely used statistical measurements are the

Volume Median Diameter (VMD) and Relative Span

Factor (RSF), which both require the diameter of each

droplet.

Since the measurements are conducted using the

pixels of a digital image, droplets are initially mea-

sured in pixels. The software detects the area of each

droplet rather than the diameter. Therefore, to cal-

culate the real-world measurements of each droplet,

the server-side application must obtain the size of the

paper to be analyzed in real-world units, speciﬁcally

in centimeters. This value is then used to establish

the true ratio between the measurements in pixels and

centimeters. With the area of each droplet measured

in pixels, the diameter can be calculated by approxi-

mating the droplet shape to a perfect circle and using

the relation between area and diameter. The following

equation is used throughout the algorithms developed

to calculate the statistics of a WSP:

µm

= 2 ×

width

µm

width

(1)

µm

is the diameter in micrometers, A

is the area

of the droplet in pixels, width

is the measurement

of the width of the paper in pixels and width

µm

is the

measurement of the width of the paper in microme-

ters.

When calculating real-world statistics such as

VMD or RSF, the measurements of the droplets,

which are all in pixels, are converted to real-world

measurements in centimeters using a pre-deﬁned

scale. This scale factor accounts for the resolution

of the image and the physical dimensions of the WSP.

Additionally, a spread factor is applied, assuming that

all droplets have the same physical properties and im-

pact on the WSP under similar conditions. This en-

sures consistency in the interpretation of droplet mea-

surements

The values of VMD and RSF are calculated using

the representative diameters of the droplets. There-

fore, for an accurate analysis of a WSP, it is crucial to

know the correct measurement of the diameter of each

droplet. If droplets overlap, it is important to sepa-

rate them, as the calculation of the diameter based on

the area assumes a perfect circle. Additionally, if a

droplet is considered an ellipse, it introduces further

errors in the calculation.

Moreover, the coverage percentage, which indi-

cates the proportion of the WSP covered by droplets,

along with the quantity of droplets, is computed. This

information is used to assess the precision of the al-

gorithms developed and is displayed as a statistical

result in the Android application to offer additional

details about the paper.

3.3 Water-Sensitive Paper and Droplet

Segmentation

The server-side application focuses on the instance

segmentation of droplets on WSP, meaning that the

algorithm must identify and delineate the individual

each individual droplet within the image, including

the overlapped droplets. This is crucial for accurately

measuring and assessing droplet size and distribution.

These measurements are used to calculate the sta-

tistical values mentioned before like average droplet

size, distribution uniformity, and coverage percent-

age. Two segmentation types are required: segmen-

tation of the WSP and segmentation of the droplets

on the WSP.

3.3.1 Segmentation of the Water-Sensitive Paper

The classical computer vision method for segmenta-

tion of the WSP involves a multi-step process to ac-

curately delineate the paper from its background. Ini-

tially, the image is processed with a Gaussian blur

to minimize noise and a 3D color histogram is em-

Spray Quality Assessment on Water-Sensitive Paper Comparing AI and Classical Computer Vision Methods

303

ployed to identify and isolate the predominant color,

masked to generate a binary image. This binary im-

age is produced by converting the original image to

grayscale and applying a binary threshold to accentu-

ate the edges. OpenCV’s findContours function is

used to detect contours, with the largest contour pre-

sumed to outline the paper’s boundary. The convex

hull of this contour is computed to establish a more

precise boundary. Finally, the corner points of the de-

tected shape are identiﬁed, and a perspective trans-

formation matrix is applied to correct any distortion,

resulting in a properly aligned image of the WSP.

For the machine learning approach, a pre-trained

model from YOLOv8 is utilized due to its efﬁciency

and rapid performance in object detection tasks. The

model, speciﬁcally a small-sized version optimized

for segmentation, was retrained using a dataset com-

prising 591 images for training, 54 images for val-

idation, and 27 images for testing. The model was

trained with 50 epochs and AdamW optimizer for

30 minutes. This dataset originally included 278

annotated real images of WSP with diverse back-

grounds. The annotations were done manually us-

ing Roboﬂow’s smart polygon tool. The augmenta-

tion techniques were applied on Roboﬂow as well,

expanding the dataset to a total of 672 images.

3.3.2 Droplet Segmentation on Water-Sensitive

Paper

The developed algorithm utilizes classical computer

vision techniques to segment and classify WSP

droplets, thereby providing meaningful distribution

statistics. The process begins with Otsu threshold-

ing (Mugele and Evans, 1951) from the OpenCV li-

brary, which effectively distinguishes droplets from

the background and outperforms adaptive threshold-

ing that often results in fragmented and inconsis-

tent segmentations. As cited by various researchers

(Lipi

nski and Lipi

nski, 2020; Wen et al., 2022), this

thresholding method stands out as particularly effec-

tive in the domain of this research

To minimize noise and enhance droplet separa-

tion, the image is ﬁrst blurred before applying thresh-

olding. The thresholding mask is inverted, and the

findContours function in OpenCV is employed to

identify contours. Contours are classiﬁed into single

circles, single ellipses, and overlapped shapes, with

additional notation for edge shapes. Circularity and

ellipse ﬁtting are determined using OpenCV func-

tions, while convexity analysis helps identify poten-

tial individual droplets within complex shapes.

For complex shapes, the Hough Circle Transform

combined with KMeans clustering is used to detect

circles. A mask isolates the region of interest, and

Original

Thresholding

Contour Classiﬁcation

Droplet Detection

Figure 3: Image processing algorithm using classical com-

puter vision methods.

parameters are ﬁnely tuned to maximize circle detec-

tion. Circles are then reﬁned through clustering and

Intersection over Union (IoU) score improvements,

ensuring accurate droplet identiﬁcation and analy-

sis. Figure 4 illustrates the steps taken for individual

droplet segmentation.

ROI

Hough Circle

KMeans

IoU

Figure 4: Separation of individual droplets process given a

non-circular shape.

The machine learning approach uses a pre-trained

medium segmentation model from YOLOv8. The

model is trained with 5 000 images of the synthetic

dataset. Given the small size of the droplets to be

detected, the images are cut into squares of 320 by

320 pixels to minimize the number of droplets per im-

age and maximize the visibility of each droplet for the

model. The model was trained with 300 epochs and

SGD optimizer for 7 hours and 30 minutes.

YOLOv8 was trained using images of 320 by 320

pixels, which was chosen to balance the need for de-

tail and the computational efﬁciency required for real-

time processing. This resolution provides enough de-

tail for identifying and segmenting droplets while en-

suring the processing time remains practical for appli-

cations that demand quick responses.

3.4 Android Application

The client-side is an Android application designed to

meet the system’s real-time requirements and was pri-

marily developed using the programming languages

Kotlin and Java in the Android Studio IDE. The ap-

plication GUI is illustrated in Figure 5

When using the app, the user can select to take a

ICINCO 2024 - 21st International Conference on Informatics in Control, Automation and Robotics

304

picture of the WSP using the smartphone-integrated

camera or choose an image from the gallery. After

that, they need to choose the segmentation algorithm

they intend to use and will be able to view the statis-

tics obtained through the computer vision analysis of

the server-side application. The communication be-

tween the two sides is done by sending a ﬁle in JSON

format.

Figure 5: Preview of the Android application.

Given the computationally intensive nature of the

image segmentation algorithms, processing on the

smartphone’s processor was deemed impractical. To

address this, a web service server was developed to

handle the image-processing tasks. This server re-

ceives requests to process images, performs the nec-

essary computations, and returns the processed image

along with relevant statistics.

The web service was developed using Python and

Flask, a lightweight web framework. The service fea-

tures a single REST API endpoint designed to receive

and process images and send back the evaluation of

the WSP.

4 EXPERIMENTAL RESULTS

4.1 Segmentation of Water Sensitive

Paper Results

The validation of the segmentation of the water-

sensitive paper is done by calculating the IoU met-

ric using both methods on a collection of 20 images,

which were put aside when the training of YOLOv8

took place. These images include a multitude of back-

grounds with the WSP being not always centered or

the main focus of the image, assuming the worst type

of user for the images.

Table 1: Results of the two algorithms developed for detect-

ing water-sensitive paper.

Method IoU Time (ms)

YOLOv8 0.9776 0.2801

CCV 0.3664 0.0357

The results presented in Table 6 indicate that the

YOLOv8 method signiﬁcantly outperforms the CCV

approach in terms of IoU. This outcome aligns with

expectations, as CCV techniques rely on predeﬁned

ﬁxed values and are often less adaptable to image

variations. The CCV method struggles with gener-

alization, particularly when the dataset exhibits diver-

sity in the shapes, colors, and orientations of the WSP.

The detection of the WSP using the CCV method

implements a histogram of the colors in the image.

Therefore, when the main focus of the WSP image is

not the paper itself, this method is not the most effec-

tive. The algorithm performs with much more preci-

sion when the WSP is centered and occupies most of

the pixels in the image.

Figure 6 depicts two examples of the images used

to validate the algorithms. The image on the left

achieved an IoU value of 97.34%, whereas the image

on the right attained a value of 8.59%. This illustrates

that while segmentation with classical computer vi-

sion can be effective in certain cases, it has notable

failure points. On the other hand, the YOLOv8 model

obtained values of 97.03% and 93.78% respectively.

Figure 6: Two examples of images used for validating the

algorithms of WSP segmentation.

4.2 Segmentation of Droplets

For the evaluation of the segmentation of droplets, the

metrics presented in Table 2 were chosen for the task

of validating the different algorithms across various

datasets. The primary goal was to assess the perfor-

mance of each algorithm in terms of precision (P),

recall (R), F1-score (F1), mean average precision at

50% intersection over union (mAP50), and process-

ing time (Time).

Three different datasets were used. The synthetic

dataset (SD) comprises 30 images, which were put

aside during the training of YOLOv8, representing a

comprehensive sample of the synthetic dataset cre-

ated, including multiple image resolutions, colors,

and droplet distributions available. From the real

dataset, two images were annotated and veriﬁed, RD1

and RD2. Although this task was performed meticu-

lously, it is still susceptible to human error.

From the results, it can be inferred that YOLOv8

Spray Quality Assessment on Water-Sensitive Paper Comparing AI and Classical Computer Vision Methods

305

Table 2: Results of algorithms for calculating statistics for water-sensitive paper.

Method Dataset P R F1 mAP50 Time (ms)

YOLOv8 RD 0.1516 0.0859 0.1096 0.01378 10.6307

YOLOv8 SD 0.6077 0.5725 0.5896 0.3796 8.2397

CCV RD 0.6551 0.8485 0.7393 0.5545 5.3724

CCV SD 0.9085 0.6304 0.7443 0.6010 5.3204

does not outperform CCV in segmentation. The seg-

mentation by YOLOv8 is imperfect, with many rough

edges due to the low resolution of images and the

small size of objects. Nevertheless, YOLOv8 exhibits

acceptable metrics, especially given the constraints of

the task. The performance metrics highlight speciﬁc

strengths and weaknesses of each method. For ex-

ample, while CCV shows higher precision on RD, its

recall is considerably lower, indicating that it misses

more droplets than YOLOv8.

The conﬁdence threshold for the prediction of

YOLOv8 was set to 0.1 and 0.5, respectively. Dur-

ing the segmentation task, the YOLOv8 solution was

found to be conservative. It often failed to detect

droplets even when there was a clear distinction be-

tween the background and the droplet. Upon evaluat-

ing the results, it was apparent that assigning the low-

est conﬁdence score during the prediction generated

much better results for the task at hand. To assess the

precision and recall of the predictions, the IoU thresh-

old was set to 0.5 (or 50%). This means that a predic-

tion was considered accurate when the IoU relation-

ship between the predicted mask and the ground-truth

mask of the object was over 0.5.

Another important aspect to consider to validate

the algorithms developed is to evaluate the values ob-

tained for the most relevant statitstics for spray qual-

ity assessment. For this, it was used an external soft-

ware developed by Machado et al. (2018), DropLeaf.

Multiple applications are available on the Google Play

Store, but only DropLeaf allows image uploads from

the gallery. This process was done manually for both

the datasets and the results were annotated for later

evaluation and metric calculation. The datasets used

are the same datasets used for the before table. To

utilize DropLeaf, a virtual machine for Android 7.3.1

was utilized since the software stopped being compat-

ible with newer Android versions.

DropLeaf consistently evaluates the coverage per-

centage as high, approximately 90%. However, ac-

cording to Machado et al. (2018), the density should

be lower and more accurate. When recalibrating the

value to 100% minus the Coverage Percentage, the ac-

curacy of DropLeaf improved signiﬁcantly. This sug-

gests a potential error in the application rather than

the calculation of the coverage percentage in the al-

gorithm, as no other cause was identiﬁed.

The ground truth statistics of the datasets were cal-

culated using the same algorithm as the segmentation

algorithm developed. Therefore, it is crucial to con-

sider that signiﬁcant discrepancies in DropLeaf values

might stem from differences in the calculation mod-

ule, as it is the only varying factor.

5 CONCLUSIONS

This study analyzes WSP methodologies, identiﬁes

limitations, and proposes innovative solutions. A syn-

thetic dataset is generated to improve model training

and generalizability. The development of a partial an-

notation algorithm for real datasets reduces manual

effort while maintaining accuracy. However, further

improvements to the synthetic dataset are needed to

enhance both human perception and transfer learning

results.

The methodology compares classical computer vi-

sion techniques with advanced machine learning for

WSP and droplet segmentation. An Android app and

web service are also developed for real-time data col-

lection, demonstrating practical applications.

The evaluation of droplet segmentation algorithms

on both synthetic and real datasets revealed distinct

performance differences between the methods. While

YOLOv8 demonstrated reasonable results, it was out-

performed by CCV in terms of segmentation accu-

racy, particularly with real-world data. CCV achieved

superior precision and recall, making it more effec-

tive at detecting droplets in diverse conditions. Nev-

ertheless, YOLOv8 showed promising results in syn-

thetic datasets, especially when dealing with reason-

ably sized objects and higher-resolution images.

Further analysis using metrics such as droplet

count, VMD, RSF, and coverage percentage high-

lighted additional performance disparities between

the algorithms. These ﬁndings emphasize the criti-

cal need to select algorithms tailored to the speciﬁc

characteristics of a dataset and segmentation task, and

they underline the importance of thorough validation

to ensure precise droplet segmentation and accurate

statistical assessment.

In summary, this research contributes to the ﬁeld

by addressing current limitations in WSP analysis

by providing a smartphone application and instance

segmentation methodology for future studies. The

ICINCO 2024 - 21st International Conference on Informatics in Control, Automation and Robotics

306

Table 3: Percentage error of each one of the metrics used to evaluate the spray quality of a water-sensitive paper.

Method Dataset No Droplets VMD RSF Coverage Percentage

YOLOv8 RD 0.5561 0.2233 2.7338 0.6095

YOLOv8 SD 0.2621 0.1423 0.2729 0.4593

CCV RD 0.4189 0.1952 1.9945 0.1533

CCV SD 0.3144 2.5103 0.5108 0.0842

DropLeaf RD 0.6096 0.7362 0.0916 0.5223

DropLeaf SD 0.9667 0.3738 8.5897 0.2639

ﬁndings emphasize the potential of using synthetic

datasets to train machine learning models to enhance

accuracy and efﬁciency. Future research could fo-

cus on reﬁning the process of generating synthetic

datasets of WSP and developing more advanced

machine-learning models to further advance its anal-

ysis.

ACKNOWLEDGEMENTS

This work is co-ﬁnanced by Component 5 - Capital-

ization and Business Innovation, integrated in the Re-

silience Dimension of the Recovery and Resilience

Plan within the scope of the Recovery and Resilience

Mechanism (MRR) of the European Union (EU),

framed in the Next Generation EU, for the period

2021 - 2026, within project Vine&Wine PT, with ref-

erence 67.

REFERENCES

Cerruto, E., Manetto, G., Longo, D., Failla, S., and Papa, R.

(2019). A model to estimate the spray deposit by sim-

ulated water sensitive papers. Crop Protection, 124.

echelette, A., Babinsky, E., and Sojka, P. (2011). Hand-

book of atomization and sprays.

Hoffmann, W. and Hewitt, A. (2005). Comparison of three

imaging systems for water-sensitive papers. Applied

Engineering in Agriculture, 21:961–964. USDA-

ARS.

Lefebvre, A. and McDonell, V. (2017). Atomization and

sprays. CRC Press.

Lipi

nski, A. and Lipi

nski, S. (2020). Binarizing water sen-

sitive papers – how to assess the coverage area prop-

erly? Crop Protection, 127.

Liu, J., Yu, S., Liu, X., Wang, Q., Cui, H., Zhu, Y., and

Yuan, J. (2024). A novel optical shadow edge imaging

method based fast in-situ measuring portable device

for droplet deposition. Computers and Electronics in

Agriculture, 217:108632.

Machado, B. B., Spadon, G., Arruda, M. S., Goncalves,

W. N., Carvalho, A. C., and Rodrigues-Jr, J. F. (2018).

A smartphone application to measure the quality of

pest control spraying machines via image analysis.

Proceedings of the ACM Symposium on Applied Com-

puting.

Marc¸al, A. and Cunha, M. (2008). Image processing of arti-

ﬁcial targets for automatic evaluation of spray quality.

Transactions of the ASABE, 51:811–821.

Mugele, R. A. and Evans, H. D. (1951). Droplet size distri-

bution in sprays. Industrial & Engineering Chemistry,

43:1317–1324.

Nansen, C., Ferguson, J., Moore, J., Groves, L., Emery, R.,

Garel, N., and Hewitt, A. (2015). Optimizing pes-

ticide spray coverage using a novel web and smart-

phone tool, snapcard. Agronomy for Sustainable De-

velopment, 35:1075–1085. SnapCard.

Nansen, C., Villar, G., Recalde, A., Alvarado, E., and Chen-

napragada, K. (2021). Phone app to perform quality

control of pesticide spray applications in ﬁeld crops.

Agriculture (Switzerland), 11. SmartSpray.

Privitera, S., Manetto, G., Pascuzzi, S., Pessina, D., and

Cerruto, E. (2023). Drop size measurement tech-

niques for agricultural sprays:a state-of-the-art review.

Agronomy.

Schick, R. J. (2008). Spray technology reference guide:

Understanding drop size. Spray Systems Co.

Wen, T., Tong, B., Liu, Y., Pan, T., Du, Y., Chen, Y., and

Zhang, S. (2022). Review of research on the instance

segmentation of cell images. Computer Methods and

Programs in Biomedicine, 227:107211.

Xun, L. and Gil, E. (2024). A novel methodology for water-

sensitive papers analysis focusing on the segmentation

of overlapping droplets to better characterize deposi-

tion pattern. Crop Protection, 176.

Zhu, H., Salyani, M., and Fox, R. (2011). A portable

scanning system for evaluation of spray deposit dis-

tribution. Computers and Electronics in Agriculture,

76:38–43. DepositScan.

Omer Barıs¸

Ozl

uoymak and Bolat, A. (2020). Develop-

ment and assessment of a novel imaging software for

optimizing the spray parameters on water-sensitive

papers. Computers and Electronics in Agriculture,

168:105104.

Spray Quality Assessment on Water-Sensitive Paper Comparing AI and Classical Computer Vision Methods

307