Detection of Sugarcane Crop Rows From UAV

Images Using Semantic Segmentation and

Radon Transform

Renato Rodrigues da Silva

Universidade Federal de Uberlândia

Faculdade de Computação
Programa de Pós-Graduação em Ciência da Computação

Uberlândia
2020
 Renato Rodrigues da Silva

Detection of Sugarcane Crop Rows From UAV

Images Using Semantic Segmentation and
Radon Transform

Dissertação de mestrado apresentada ao Programa

de Pós-graduação da Faculdade de Computação da
Universidade Federal de Uberlândia como parte dos
requisitos para a obtenção do título de Mestre em
Ciência da Computação.

Área de concentração: Ciência da Computação

Orientador: Prof. Dr. André Ricardo Backes

Coorientador: Prof. Dr. Mauricio Cunha Escarpinati

Uberlândia
2020
I dedicate this work to all type 1 diabetics who, like a tightrope walker, have to keep the
balance between managing their lives while handling all the other responsibilities this
world demands.
A special feeling of gratitude to my love Pâmela, whose words of encouragement and
push for tenacity were crucial, and my sister Susana who had always believed in me.
This work is for you guys, too.
 Acknowledgements

First of all, I thank God, who have lit my way during this path. I thank all the
professors of the course, who were very important in my academic life, especially those
who have been true mentors to me: Mauricio Cunha Escarpinati, my co-advisor, for his
companionship, friendship and all the advice given along these years; and André Ricardo
Backes, who, in addition to being my advisor, is a great friend and has always been a key
player in this journey. I also thank my friends and all the people who are important to me
and who have always been by my side, especially my parents and wife. Thanks, Pâmela
for the patience and for giving me strength at all times. Finally, I thank all those who,
in some way, were and are close to me, making this life more and more worthwhile.
ŞLife is like riding a bicycle. To keep your balance, you must keep moving.Ť
(Albert Einstein)
 Resumo

Nos últimos anos, os VANTs (Veículos Aéreos Não Tripulados) têm se tornado cada
vez mais populares no setor agrícola, promovendo e possibilitando o monitoramento de
imagens aéreas tanto no contexto cientíĄco, quanto no de negócios. Imagens capturadas
por VANTs são fundamentais para práticas de agricultura de precisão, pois permitem a
realização de atividades que lidam com imagens de baixa ou média altitude. O cenário
da área plantada pode mudar drasticamente ao longo do tempo devido ao aparecimento
de erosões, falhas de plantio, morte e ressecamento de parte da cultura, intervenções de
animais, etc. Assim, o processo de detecção das linhas de plantio é de grande importância
para o planejamento da colheita, controle de custos de produção, contagem de plantas,
correção de falhas de semeadura, irrigação eĄciente, entre outros. Além disso, a infor-
mação de geolocalização das linhas detectadas permite o uso de maquinários autônomos e
um melhor planejamento de aplicação de insumos, reduzindo custos e a agressão ao meio
ambiente. Neste trabalho, abordamos o problema de segmentação e detecção de linhas de
plantio de cana-de-açúcar em imagens de VANTs. Primeiro, experimentamos uma abor-
dagem baseada em Algoritmo Genético (AG) e Otsu para produzir imagens binarizadas.
Posteriormente, devido a alguns motivos, incluindo a relevância recente da Segmentação
Semântica, seus níveis de abstração e os resultados inviáveis obtidos com AG, estudamos
e propusemos uma nova abordagem baseada em Semantic Segmentation Network (SSN)
em duas etapas. Primeiro, usamos uma SSN para segmentar as imagens, classiĄcando
suas regiões como linhas de plantio ou como solo não plantado. Em seguida, utilizamos a
transformada de Radon para reconstruir e melhorar as linhas já segmentadas, tornando-
as mais uniformes ou agrupando fragmentos de linhas e plantas soltas. Comparamos
nossos resultados com segmentações feitas manualmente por especialistas e os resultados
demonstram a eĄciência e a viabilidade de nossa abordagem para a tarefa proposta.

Palavras-chave: Linhas de Plantio, Cana-de-açúcar, Segmentação, CNN, VANT, Trans-

formada de Radon.
Detection of Sugarcane Crop Rows From UAV
Images Using Semantic Segmentation and
Radon Transform

Renato Rodrigues da Silva

Universidade Federal de Uberlândia

Faculdade de Computação
Programa de Pós-Graduação em Ciência da Computação

Uberlândia
2020
 Abstract

In recent years, UAVs (Unmanned Aerial Vehicles) have become increasingly popular in
the agricultural sector, promoting and enabling the application of aerial image monitoring
in both scientiĄc and business contexts. Images captured by UAVs are fundamental
for precision farming practices, as they allow activities that deal with low and medium
altitude images. After the effective sowing, the scenario of the planted area may change
drastically over time due to the appearance of erosion, gaps, death and drying of part
of the crop, animal interventions, etc. Thus, the process of detecting the crop rows
is strongly important for planning the harvest, estimating the use of inputs, control of
costs of production, plant stand counts, early correction of sowing failures, more-efficient
watering, etc. In addition, the geolocation information of the detected lines allows the
use of autonomous machinery and a better application of inputs, reducing Ąnancial costs
and the aggression to the environment. In this work we address the problem of detection
and segmentation of sugarcane crop lines using UAV imagery. First, we experimented
an approach based on Genetic Algorithm (GA) associated with Otsu method to produce
binarized images. Then, due to some reasons including the recent relevance of Semantic
Segmentation in the literature, its levels of abstraction, and the non-feasible results of
Otsu associated with GA, we proposed a new approach based on SSN divided in two
steps. First, we use a Convolutional Neural Network (CNN) to automatically segment
the images, classifying their regions as crop lines or as non-planted soil. Then, we use
the Radon transform to reconstruct and improve the already segmented lines, making
them more uniform or grouping fragments of lines and loose plants belonging to the same
planting line. We compare our results with segmentation performed manually by experts
and the results demonstrate the efficiency and feasibility of our approach to the proposed
task.

Keywords: Crop-row, Sugarcane, Segmentation, CNN, UAV, Radon Transform.

 List of Figures

Figure 1 Ű Example of crop-row identiĄcation performed manually by an expert. . 18

Figure 2 Ű Example of an autonomous tractor equipped with a device showing the
crop-rows which it is being guided by. CommandCenter™ Premium
produced bu John Deer, extracted from https://www.agriexpo.online/prod/john-
deere/product-169419-2710.html . . . . . . . . . . . . . . . . . . . . . . 19
Figure 3 Ű Example of precision agriculture equipment developed for farm man-
agement and tasks such as high precision positioning systems, laser land
levelling, and precision seeding/fertilizer/irrigation/harvesting, extracted
from (LI et al., 2020). . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
Figure 4 Ű Examples of rotary-wing UAVs, a more common type popularly known
as ŚdronesŠ, extracted from (RADOGLOU-GRAMMATIKIS et al., 2020). 26
Figure 5 Ű Examples of Ąxed-wing UAVs, extracted from (RADOGLOU-GRAMMATIKIS
et al., 2020). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Figure 6 Ű Ąxed-wing UAV SX2 made by Sensix Innovations and responsible for
capturing the imagery used in this work. . . . . . . . . . . . . . . . . . 27
Figure 7 Ű Example of sugarcane crop seen from above. This image was taken by
the unmanned aerial vehicle (UAV) shown in Figure 6. . . . . . . . . . 28
Figure 8 Ű An illustration of convolution and pooling operation in CNN, extracted
from (DEWA; AFIAHAYATI, 2018). . . . . . . . . . . . . . . . . . . . 29
Figure 9 Ű Example of architecture of CNN having 4 convolution followed by pool-
ing and one fully-connected layer, extracted from (ZADA; ULLAH, 2020). 30
Figure 10 Ű ReLu activation function. . . . . . . . . . . . . . . . . . . . . . . . . . 31
Figure 11 Ű Example of a semantic segmentation performed in some images, their
results, as well their classiĄcations and respective percentage score per
segment/label. Extracted from (NAGATA et al., 2020) . . . . . . . . . 34
Figure 12 Ű Basic Principle of HT. (a) shows representation on equation of straight
line, (b) illustrates intersection of many lines to a point, (c) transfor-
mation of point in image space to polar space, extracted from (VARUN
et al., 2015). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
Figure 13 Ű Working of HT as feature extractor. (a) shows a set of points in image
space, (b) illustrates convergence of points in image space to sinusoidal
waves in polar space, (c) represents the accumulator space, extracted
from (VARUN et al., 2015). . . . . . . . . . . . . . . . . . . . . . . . . 38
Figure 14 Ű A hand-operated vehicle carrying the camera, computer and a battery
for power supply, extracted from (SØGAARD; OLSEN, 2003). . . . . . 42
Figure 15 Ű A tractor carrying a camera to capture crop-row images. The camera
can be seen in the upper part of the image pointing forward-downward
to the left of the tractorŠs left rear wheel, extracted from (SØGAARD;
OLSEN, 2003). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
Figure 16 Ű Examples of images captured by cameras attached to tractors. The
Ąrst two are raw images used in (LEEMANS; DESTAIN, 2006). The
last one is an image captured by the camera shown in Figure 15 and
changed to its grey-scale representation for typographical reasons, ex-
tracted from (SØGAARD; OLSEN, 2003). . . . . . . . . . . . . . . . . 44
Figure 17 Ű Examples of crop images used for RHT line detection technique. Ex-
tracted from (JI; QI, 2011). . . . . . . . . . . . . . . . . . . . . . . . . 45
Figure 18 Ű Framework architecture from (GUERRERO; RUZ; PAJARES, 2017). . 46
Figure 19 Ű Example of an image processed by the framework in Figure 18: (a)
Original image; (b) Greenness Index Combination index from (a); (c)
Binary image after Otsu thresholding. Extracted from (GUERRERO;
RUZ; PAJARES, 2017). . . . . . . . . . . . . . . . . . . . . . . . . . . 46
Figure 20 Ű Proposed image processing method architecture. Extracted from (GARCÍA-
SANTILLÁN et al., 2018). . . . . . . . . . . . . . . . . . . . . . . . . . 47
Figure 21 Ű Early-season maize stand count determination system proposed by Pang
et al.. Extracted from (PANG et al., 2020). . . . . . . . . . . . . . . . 48
Figure 22 Ű Encoder-decoder architecture of SegNet. Extracted from (BAH; HAFI-
ANE; CANALS, 2020). . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
Figure 23 Ű Architecture of the HoughCNet network. Extracted from (BAH; HAFI-
ANE; CANALS, 2020). . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
Figure 24 Ű Test images used to evaluate our approach and their respective sizes:
(a) 11180×8449; (b) 19833×30255; (c) 17497×10771; (d) 16677×24181. 53
Figure 25 Ű Examples of crop lines and the segmentation provided by an expert. . . 54
Figure 26 Ű Example of sugar cane crops in different phases: (a) Ratoon phase; (b)
Plant Phase. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
Figure 27 Ű Flow chart of the Ąrst approach based on Genetic Algorithm and Radon
transform. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
Figure 28 Ű Architectures used for semantic segmentation. Adapted from (YAKUBOVSKIY,
2019). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
Figure 29 Ű Example of problems encountered after the segmentation step: (a)
Original image; (b) Planting lines provided by an expert; (c) Image
after segmentation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
Figure 30 Ű Proposed scheme for crop line reconstruction using Radon transform:
(a) Input image; (b) Matrix obtained with the Radon transform. The
red dot represents the location of the maximum point and the orien-
tation angle of the input image; (c) Radon transform obtained for the
image orientation angle (red line in (b)). Each peak of the curve cor-
responds to the center of a line in the input image; (d) Reconstruction
of the lines using the orientation angle and the peaks of the Radon
transform for that angle. . . . . . . . . . . . . . . . . . . . . . . . . . . 60
Figure 31 Ű Average Dice coefficient and standard deviation for different images for
5 different GA kernel masks. . . . . . . . . . . . . . . . . . . . . . . . . 62
Figure 32 Ű Results for different sections of the map: (a) Original image; (b) Ex-
pertŠs segmentation; (c) Manual threshold (𝑡 = 0.8); (d) Global Otsu;
(e) Local Otsu (𝑊 = 50 and 𝑆 = 25). . . . . . . . . . . . . . . . . . . 63
Figure 33 Ű Dice coefficient for various global threshold values. . . . . . . . . . . . . 64
Figure 34 Ű Dice coefficient obtained using Global Otsu and Local Otsu for different
combinations of Window 𝑊 and Stride 𝑆. . . . . . . . . . . . . . . . . 64
Figure 35 Ű Dice coefficient obtained for the line reconstruction for different com-
binations of Window 𝑊 and Stride 𝑆. . . . . . . . . . . . . . . . . . . 64
Figure 36 Ű Errors detected during the line reconstruction. . . . . . . . . . . . . . . 65
Figure 37 Ű Curved crop lines: (a) ExpertŠs segmentation; (b) Line reconstruction
using Radon transform in a tiling scheme. . . . . . . . . . . . . . . . . 65
Figure 38 Ű Results obtained for each segmentation networks. Left column shows
the loss function, while the right column shows the Dice Segmentation
Coefficient: (a) LinkNet (b) PSPNet and (c) U-net. . . . . . . . . . . . 67
Figure 39 Ű Average Dice coefficient obtained for different selection approaches dur-
ing the crop line reconstruction. . . . . . . . . . . . . . . . . . . . . . . 68
Figure 40 Ű Examples of images where there was an improvement in the Dice coeffi-
cients after line reconstruction using the Radon transform. (a) Original
image; (b) Segmentation provided by the expert; (c) Segmentation ob-
tained using LinkNet; (d) Line reconstructed. . . . . . . . . . . . . . . 69
Figure 41 Ű Examples of images where there was a decrease in the Dice coefficients
after line reconstruction using the Radon transform. (a) Original image;
(b) Segmentation provided by the expert; (c) Segmentation obtained
using LinkNet; (d) Line reconstructed. . . . . . . . . . . . . . . . . . . 70
 List of Tables

Table 1 Ű Segmentation results obtained with the application of the segmentation

networks in Dataset A. . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
Table 2 Ű Result obtained with the application of the LinkNet network trained in
dataset A to segment other datasets. . . . . . . . . . . . . . . . . . . . . 66
 Acronyms list

AUVSI Association for Unmanned Vehicle Systems International

CNN Convolutional Neural Network

DIP Digital Image Processing

DNN Deep Neural Networks

DSC Dice Similarity Coefficient

FCN Fully Convolutional Network

GA Genetic Algorithm

GNSS Global Navigation Satellite Systems

HSV Hue, Saturation, Value

HT Hough Transform

ICT Information and Communications Technology

IMU Inertial Measurements Unit

IoT Internet of Things

JSC Jaccard Similarity Coefficient

PA Precision Agriculture

RHT Random Hough Transform

RGB Red, Green, Blue

SHT Standard Hough Transform

SSN Semantic Segmentation Network

UAV Unmanned Aerial Vehicle

VI Vegetation Index
 Contents

1 INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.2 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.3 Hypothesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.4 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.5 Thesis Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

2 FUNDAMENTALS . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.1 Precision Agriculture . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2 Unmanned Aerial Vehicles . . . . . . . . . . . . . . . . . . . . . . . 25
2.3 Convolutional neural network . . . . . . . . . . . . . . . . . . . . . 28
2.4 Image Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.4.1 Binarization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.4.2 Semantic Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.5 Genetic Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.6 Hough Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.7 Radon Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

3 RELATED-WORK . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.1 Hough Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.2 Otsu Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.3 Convolutional Neural Networks . . . . . . . . . . . . . . . . . . . . 48
3.4 Other Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.5 Final Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

4 METHODOLOGY . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.1 Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.1.1 Evaluation metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.2 Segmentation using Genetic Algorithm . . . . . . . . . . . . . . . 55
4.3 Semantic Segmentation Networks . . . . . . . . . . . . . . . . . . . 56
4.3.1 Proposed Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.4 Line Reconstruction and Refinement . . . . . . . . . . . . . . . . . 59

5 EXPERIMENTAL RESULTS . . . . . . . . . . . . . . . . . . . 61
5.1 Segmentation using Generic Algorithm . . . . . . . . . . . . . . . 61
5.2 Semantic Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . 66
5.3 Comparison of approaches . . . . . . . . . . . . . . . . . . . . . . . 70

6 CONCLUSION . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
6.1 Main Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
6.2 Contributions in Bibliographic Production . . . . . . . . . . . . . 74
6.3 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

BIBLIOGRAPHY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
 I hereby certify that I have obtained all legal permissions from the owner(s) of each
third-party copyrighted matter included in my dissertation, and that their permissions
allow availability such as being deposited in public digital libraries.

Renato Rodrigues da Silva

14 LIST OF TABLES
 15

Chapter 1
Introduction

Sugarcane is one of the most planted cultures in the planet. Its planting practice and
mechanization are the development tendency of the modern agro-industry (UCHIMIYA;
SPAUNHORST, 2020). Brazil is the largest producer of sugarcane in the world. The
country registered an area around 10,123.5 million hectares (Mha) of land planted with
the crop in the 2018/2019 harvest. This area includes Ąelds meant for both sugar and
ethanol production (LIMA et al., 2020).
The main destination for ethanol is the biofuel industry supplying the Brazilian vehicle
Ćeet with the mixture of anhydrous ethanol for gasoline and for the engines with Ćex
fuel technology which make up an increasingly emerging market in Brazil and worldwide
(LIMA et al., 2020). As expected, Brazil is also the largest producer of sugarcane ethanol
worldwide and this production is expected to have a substantial increase in the coming
years as the sugarcane ethanol sector contributes signiĄcantly to the national economy
(BRINKMAN et al., 2018).
Despite all the economic beneĄts that sugarcane and other crop cultures bring, the
massive expansion of agriculture also leads to some social and ecological issues. Some of
them due to drastic deforestation, social conĆicts, land disputes and the degradation of the
environment caused by the spread of pesticides and agricultural inputs. A great ally for
helping solving some of these problems is the Precision Agriculture (PA) (MCBRATNEY
et al., 2005; MILELLA; REINA; NIELSEN, 2018).
PA is a modern concept for managing agricultural activities and it has been increas-
ingly adopted by producers in several countries. This concept is associated with research,
information gathering, and the use of various technologies to analyze and monitor the
conditions of planted areas in a more precise and efficient way. It is based on observation,
measurement, monitoring, and rapid decision making in the face of the variability that
planting crops can present (LINDBLOM et al., 2017).
The main objective of the research in the Ąeld of PA is to deĄne a support system
for the necessary decisions aiming at a better management of the crop to optimize the
use of resources and inputs, while increasing the Ąnancial return (MCBRATNEY et al.,
16 Chapter 1. Introduction

2005; MILELLA; REINA; NIELSEN, 2018). More recently, research in this area has
had a major positive impact on the growth of agricultural production (REN et al., 2020).
Methodologies for improving seed quality, more efficient irrigation systems and soil quality
control are just a few examples of the techniques that beneĄt from this research (JR;
DAUGHTRY, 2018).
Technological advances in the use of unmanned aerial vehicles (UAVs) have also opened
up new opportunities in the PA sector. This type of equipment allows for more effective
monitoring and greater agility in cultivation. Sensors coupled to an UAV are able to collect
large amounts of information about the plantation. In addition, they enable more frequent
data collections and less cloud interference due to their lower Ćight altitude (SILVA et al.,
2017; SOARES; ABDALA; ESCARPINATI, 2018; SOUZA; ESCARPINATI; ABDALA,
2018; FUENTES-PEÑAILILLO et al., 2018). The use of UAVs has also fostered the
development of new and more efficient digital image processing techniques to analyze im-
ages acquired by their sensors (SILVA et al., 2017; SOARES; ABDALA; ESCARPINATI,
2018; SOUZA; ESCARPINATI; ABDALA, 2018). Most of these techniques aim to es-
timate the growth of the crop or to identify other important agronomic characteristics,
such as nitrogen stress, water stress, new diseases, known pests and vegetation indexes.
After the initial planning and effective sowing, the scenario of the planted area may
change over time due to the appearance of failures, erosion, death, and drying of part
of the plantation, tipping of plants, animal interventions, among others. This makes
the identiĄcation of crop lines, and how they are arranged in a region, an important
task within PA. With this information, it is possible, for example, to better plan the
application of inputs, thus reducing Ąnancial costs and the aggression to the environment.
Recently, Convolutional Neural Networks (CNNs) have emerged as a powerful ap-
proach to computer vision tasks. Its use has been widespread in the most diverse areas
of research and it has presented relevant results in applications of classiĄcation, object
detection, and facial recognition (LIU et al., 2020; SIMONYAN; ZISSERMAN, 2014;
KANG et al., 2014). They have, for example, been used with great success in identifying
pests in agricultural environments of complex soil structures (CHENG et al., 2017), in the
detection of weeds (FERREIRA et al., 2017), detection of plant diseases (FERENTINOS,
2018) and even in the detection of Ćowers (DIAS; TABB; MEDEIROS, 2018).
In this work we address the problem of crop line detection and segmentation in aerial
images of sugarcane plantations obtained by UAVs. First, we experimented an approach
based on Genetic Algorithm associated with Otsu method to produce binarized images
that were then reconstructed using a Radon transform. Then, due to some reasons includ-
ing the recent relevance of Semantic Segmentation in the literature, its levels of abstrac-
tion, and the non-feasible results of Otsu associated with GA, we studied and proposed
a new automatic segmentation approach based on SSN consisting in two steps. First, we
use a Convolutional Neural Network (CNN) to segment the planted area into regions of
1.1. Motivation 17

crop lines (region of interest) and unplanted soil (background). Then, we use a reĄne-
ment process which aims to reconstruct and to improve the previous detected lines. This
is performed in order to make the detected crop lines more uniform and to connect line
fragments and isolated plants that originally belonged to the same crop line.
The remainder of this dissertation is organized as follows: In the chapter 2 we describe
the fundamental concepts relevant to the understanding of the working. In the chapter
3 we present a review of the related architecture and the state of the art in detecting
crop lines. In chapter 4 we describe our methodology, showing the datasets as well as the
techniques used in the proposed approach. In the chapter 5 we show our experiments and
obtained results and Ąnally, the chapter 6 concludes this dissertation.

1.1 Motivation
The identiĄcation of crop lines, and how they are arranged in the planted area, in
low and medium altitude images obtained by Unmanned Aerial Vehicle (UAV) is an
important problem within PA. The lower cost of obtaining images by the UAV also allows
farmers to monitor them more frequently. This is important because after the initial
planning and effective sowing, the scenario of the planted area may change over time,
such as the appearance of failures, erosion, death and drying of part of the plantation,
tipping of plants, animal interventions, among others. Thus, this process of detecting
the lines is important for planning the harvest, estimating the use of inputs, controlling
costs, estimating production, counting plants and early correction of sowing failures. In
addition, the geolocation information of the crop allows a better planning of application of
inputs, thus reducing Ąnancial costs and less aggression to the environment as sugarcane
represents a great percentage of all plantation crop worldwide.
Another important point is the fact that the harvest may come from autonomous
vehicles and machinery. The geolocation of the crop-rows are crucial for these machines
to drive themselves through the Ąeld. Plus, additional information such as which parts of
the crop-row have gaps help the machinery to know which parts of the row do not need
to receive inputs, suppressing their spread and thus saving money and, more importantly,
lessening the degradation of the environment as some of theses substances can be harmful.
Figure 1 shows an example of sugarcane crop where the crop lines were detected by an
expert. Also, the exact location of the line can minimize the stump trampling and soil
compaction in the seedling zone done by the own machinery. In the beneath layer we can
see the image captured by the UAV. In the above layer we can see the segments of the
crop lines. The green segments represent the part of the line where there is plants and the
gaps of the line are described in different colors depending on the extension of the gap.
The red color segments represent small gaps. Orange represents the medium segments.
The yellow, large gaps.
20 Chapter 1. Introduction

1.3 Hypothesis
This research explores basically two hypotheses:

1. An effective binarization can be obtained from a two-class trained CNN. This

segmentation is effective in sugarcane crops producing a result of quality when
compared with binarization performed manually by experts.

2. Using the binarized output images from the CNN, it is possible to perform a Radon
transform to achieve a reĄnement process that targets the reconstruction (gap Ąlling)
and enhancement of the previous detected lines. This is performed in order to make
the detected crop lines more uniform and to link row fragments and isolated plant
areas that originally belonged together.

1.4 Contributions
The main contributions of this work are :

❏ Development of a CNN training capable of classifying crop images in two classes

(crop rows and background) generating binarized images. The efficient binarization
can reduce drastically the cost of the post-processing step.

❏ Development of a framework that receives sugarcane crop images, binarize them

using the CNN already trained and then perform a post-processing step based on
Radon transform to perform a reĄnement process reconstructing and enhancing the
lines making them more uniform and to linking row fragments that are supposed to
belong together at the same line/row.

1.5 Thesis Organization

The remainder of this dissertation is organized as follows:

❏ In Chapter 2 we present the concepts of PA, image segmentation, remote sensing

as well as some techniques important to the understanding of the work;

❏ In chapter 3 we describe some of relevant state of the art works;

❏ Chapter 4 states the methodology of this project, including the image acquisition
process and the datasets used, a description of our Ąrst approach based on genetic
algorithm, and Ąnally our prime approach based on semantic segmentation to bina-
rize images, as well as the post-processing step to reconstruct lines and the metrics
used for the work evaluation;
1.5. Thesis Organization 21

❏ Chapter 5 describes the experimental results obtained for each dataset. Discussions
and comparisons between the results obtained by GA and CNN as well as the results
obtained by the Radon transform.

❏ Finally, in Chapter 6 we present the conclusions of this research work, its results,
contributions and future work.
22 Chapter 1. Introduction
 23

Chapter 2
Fundamentals

2.1 Precision Agriculture

It is indisputable that the positive economic impacts that agriculture cause locally
and globally are enormous. However the massive expansion of agriculture also has led
to some social conĆicts, land disputes, as well as some serious ecological issues. Drastic
deforestation, such as soil erosion and pollution, water scarcity, are usually caused by
the overuse of pesticides and other chemical inputs (LI et al., 2020). Consequently,
technological advances in the agricultural context have been crucial to promote sustainable
solutions in Ąeld productivity, economical farm incomes, more security in the food section
and general economic growth (LI et al., 2020), as well as reducing agricultural harmful
impacts in the environment.
One solution for helping solving some of these problems is to implement advanced
agricultural technologies, and the precision agriculture technologies are some of the most
important of them (MCBRATNEY et al., 2005; MILELLA; REINA; NIELSEN, 2018)
as they enable the more precise use of agricultural inputs (LI et al., 2020). Hence, it
is clear that PA is a great ally for increasing efficiency, productivity and proĄtability
in Ąeld operations, enhancing food security, and minimizing the unintended impacts of
overuse of inputs on agricultural production systems and affecting positively the envi-
ronment (LI et al., 2020). One example of PA equipment developed for farm man-
agement and tasks such as high precision positioning systems, levelling, and precision
seeding/fertilizer/irrigation/harvesting can bee seen in Figure 3.
The main objective of the researching in the Ąeld of PA is to deĄne a support system
for the necessary decisions aiming at a better management of the crop to optimize the
use of resources and inputs, while increasing the Ąnancial return (MCBRATNEY et al.,
2005; MILELLA; REINA; NIELSEN, 2018). More recently, research in this area has
had a major positive impact on the growth of agricultural production (REN et al., 2020).
Methodologies for improving seed quality, more efficient irrigation systems and soil quality
control are just a few more examples of the techniques that beneĄt from research in
2.2. Unmanned Aerial Vehicles 25

their lower Ćight altitude (SILVA et al., 2017; SOARES; ABDALA; ESCARPINATI,
2018; SOUZA; ESCARPINATI; ABDALA, 2018; FUENTES-PEÑAILILLO et al., 2018).
The use of UAVs has also fostered the development of new and more efficient digital
image processing techniques to analyze images acquired by their sensors (SILVA et al.,
2017; SOARES; ABDALA; ESCARPINATI, 2018; SOUZA; ESCARPINATI; ABDALA,
2018). In addition, most of these techniques aim to estimate the growth of the crop
or to identify other important agronomic characteristics, such as nitrogen stress, water
stress, new diseases, known pests and VI. In the next section the UAV technology is more
explored and explained.

2.2 Unmanned Aerial Vehicles

The unmanned aerial vehicles, popularly known as UAVs, are vehicles capable of Ćying
medium and low altitudes with no need of a on-board pilot. Commonly, the Ćight mission
of UAV is pre-deĄned; however a pilot can also control it through remote teleoperation
commands from the ground (RADOGLOU-GRAMMATIKIS et al., 2020). There are
many kinds of UAVs and they can be categorized in many different ways. A way to
classify them is by their type of wing: rotary and Ąxed (RADOGLOU-GRAMMATIKIS
et al., 2020). Some examples of rotary-wing UAVs, a more common type commonly known
as ŚdronesŠ, can be seen in Figure 4, while some examples of Ąxed-wing UAVs are shown
in Figure 5.
Even though the Ąrst UAV equipments were developed for primarily military opera-
tions, the fast evolution of technologies such as imaging sensors, Inertial Measurements
Unit (IMU) (MIRZAEI; ROUMELIOTIS, 2008), synthetic aperture radar (CHAN; KOO,
2008) and Global Navigation Satellite Systems (GNSS) (DOW; NEILAN; RIZOS, 2009)
resulted in the development of nonmilitary UAVs fostering the development of multi-
ple areas, such as PA, geomatics, logistics and infrastructure monitoring (RADOGLOU-
GRAMMATIKIS et al., 2020).
UAVs are important for understanding crop and soil variability as it is one of the
oldest challenges faced by agriculture and researches in PA (LAGACHERIE; MCBRAT-
NEY, 2006). The use of this technology is strongly increasing not only in the agricultural
but in all as they are able to perform some air operations that manned airborne may
struggle to do or are not feasible such as regular crop and soil analysis (RADOGLOU-
GRAMMATIKIS et al., 2020). Their use even leads to economic savings and environ-
mental welfare, also reducing the need of a human pilot in the air, thus reducing the risk
of human lives according to (OUTAY; MENGASH; ADNAN, 2020). The authors also
highlight that the drones market (considering the commercial and the civilian ones) is
growing at a compound rate of 19%.
The Association for Unmanned Vehicle Systems International (AUVSI) in its economic
28 Chapter 2. Fundamentals

Figure 7 Ű Example of sugarcane crop seen from above. This image was taken by the
unmanned aerial vehicle (UAV) shown in Figure 6.

2.3 Convolutional neural network

Convolutional neural networks have been showing relevant results in computer vision
applications in the last years (LIU et al., 2020). Their use has been widespread in the
most diverse areas of research and it has presented relevant results in applications of clas-
siĄcation, object detection, image recognition, target classiĄcation and facial recognition
(LIU et al., 2020; LIN; SHEN, 2018). They have, for example, been used with great suc-
cess to identify pests in agricultural environments of complex soil structures (CHENG et
al., 2017), in the detection of weeds (FERREIRA et al., 2017), detection of plant diseases
(FERENTINOS, 2018) and even in the detection of Ćowers (DIAS; TABB; MEDEIROS,
2018).
This machine learning method is a trainable multi-layer network structure composed
by stacks of other convolution neural networks. The basic structure of a CNN normally
includes two layers. The Ąrst one, the convolutional layer, is the feature-extraction layer
where the input of each neuron connects to the local acceptance domain of the previous
layer and extract the local feature. The second common layer is the feature-mapping or
fully connected layer that assumes that each computing layer of the network represents
multiple feature maps, each of these maps is considered a plane, and the weights of all
the neurons in each plane are equal (LIN; SHEN, 2018). Each of the these layers can be
32 Chapter 2. Fundamentals

that are normally black and white (VERMA; PARIHAR, 2017). These two values labels
the regions of the image as regions of interest and background (SEZGIN; SANKUR, 2004).
There are several ways to perform the binarization, but the simplest technique is the
use of a threshold value to classify the pixels of the source image based on that value.
All pixels with values greater than this threshold are deĄned as white or intensity 255,
and all other pixels will receive an intensity value of 0 or black. The threshold value in a
binarization can be set based on the region of the image or globally (VERMA; PARIHAR,
2017). One of the most common approaches to calculate a threshold is through the use of
Otsu method. This algorithm basically assumes that the image has two classes of pixels
following a bi-modal histogram (background and foreground). Then, the ideal threshold
is reached by separating the two classes so that the combined intra-class variation is
minimized (OTSU, 1979). Usually, Ąnding a single correct threshold that satisfactorily
represents the entire image can be quite difficult or even impossible. Thus, applying local
thresholds to smaller regions of the image may be the best option (GONZALEZ; WOODS,
2000).

2.4.2 Semantic Segmentation

Another segmentation technique that has shown great efficacy is the semantic seg-
mentation method (NEMOTO et al., 2020). It can be deĄned as the act of classifying
each pixel of an image into a class (BRAS et al., 2020). Semantic segmentation is one
of the most important tasks in machine learning and computer vision. This technique
has been used with success in many applications, namely: autonomous driving, medi-
cal diagnosing, image editing, among others (FENG et al., 2020; STAN; THOMPSON;
VOORHEES, 2020; BRAS et al., 2020). Its relevance has been increasingly growing in
the last years due the resumption of convolution neural networks and its fast development
(KRIZHEVSKY; SUTSKEVER; HINTON, 2012; HE et al., 2016).
Convolutional neural networks (CNNs) have received great attention, among other
machine learning and semantic segmentation techniques. Nonetheless, unlike traditional
machine learning methods, a CNN does not demand manual feature extraction on an
image, as it automatically incorporates feature extraction engineering into the training
steps. Features are extracted in convolutional layers, while the classiĄcation process is
performed in the fully connected layers based on the feature values. CNNs were initially
applied for object recognition (KRIZHEVSKY; SUTSKEVER; HINTON, 2012); lately, a
range of semantic segmentation works in the literature are based on Fully Convolutional
Network (FCN), which, according to (LONG; SHELHAMER; DARRELL, 2015).
An FCN is a network composed of only convolutional layers and it can be applied to
semantic segmentation training (NEMOTO et al., 2020). It arose from the CNN con-
cept (LONG; SHELHAMER; DARRELL, 2014) and replaces the fully connected layer
with convolution layer to support arbitrary input sizes (SANG; ZHOU; ZHAO, 2020).
2.4. Image Segmentation 33

Despite the impressive advances these networks have made, there are still two main chal-
lenges associated with this task to be worked out. The Ąrst one is how to correctly capture
rich contextual information and features to determining confusing classes. The second
one is related to how to accurately recover feature map resolution to improve spatial
performance (SANG; ZHOU; ZHAO, 2020). Figure 11 shows an example of a semantic
segmentation process performed in some images, their results, as well as their classiĄ-
cations and respective percentage score per segment/label. In the sequence we explain
three of the most used semantic segmentation networks (U-net, PSPNet and LinkNet).
We chose these networks due to the great results they have been showing in literature.
Further explanation regarding CNNs is provided in section 2.3

2.4.2.1 U-net

U-net was initially proposed by (RONNEBERGER; FISCHER; BROX, 2015) for tasks
that needed precise segmentation, but had few examples available for training, such as
medical images. It consists of two main segments: contracting and expanding paths,
giving to the network the shape of an ŞUŤ, which justiĄes its name.
The contracting path consists of the repeated application of blocks of two 3 × 3 con-
volution layers (each followed by a ReLU unit) and a 2 × 2 max-polling layer. After
each block, the number of Ąlters doubles so that the network can learn the more complex
structures.
In the expanding path, each block consists of two 3 × 3 convolution layers (each
followed by a ReLU unit) and a 2 × 2 up convolution layer. It also concatenates the
high-resolution features maps from the respective step of the contracting path in order
to ensure the proper reconstruction of the image. After each block, the number of Ąlters
halves. This is necessary due to the loss of the border region in each convolution.
Both contracting and expanding path present the same number of blocks. After ex-
panding path, the resulting feature map passes through a 3 × 3 convolution layer where
the number of feature maps is equal to the number of classes in the segmented image.

2.4.2.2 PSPNet

The Pyramid Scene Parsing Network (PSPNet) (ZHAO et al., 2017) has as its fun-
damental principle the use of global information from the image by extracting context
information in each scene. Its architecture consists of a fully convolutional network, be-
ing the ResNet (HE et al., 2016) used in (ZHAO et al., 2017). From the feature map,
four pooling layers of different sizes are applied, thus generating four feature sub-maps.
Subsequently, the network applies a 1 × 1 convolution to reduce the mapsŠ dimensionality,
which are enlarged through a bilinear interpolation to return to the size of the original
feature map. Finally, the feature maps are concatenated and a convolution is applied to
obtain the prediction map.
2.5. Genetic Algorithm 35

2.5 Genetic Algorithm

Introduced by John Holland in 1960, a genetic algorithm (GA) is a heuristic inspired by
Charles DarwinŠs theory of natural evolution. This algorithm reĆect the process of natural
selection where the Ąttest individuals are selected for reproduction in order to produce
offspring of the next generation. Genetic Algorithms are commonly used to generate high-
quality solutions to optimization and search problems by relying on bio-inspired operators
such as mutation, crossover and selection (MITCHELL, 1996).
A genetic algorithm starts with an initial population of individuals, where each one is
assumed to be a solution to the problem to be solved. Each possible solution is represented
as a chromosome 𝑔(𝑖)j , where 𝑖 = 1, . . . , 𝑁 is the 𝑖-th gene of an individual 𝑗, 𝑗 = 1, . . . , 𝑀 ,
of a population of 𝑀 individuals.
Then, a Ątness function 𝑓 is applied to determine how an individual 𝑗 Ąts in the
relation. Each individual will have a Ątness score 𝑓 (𝑗) in the end and this score will be
used to determine the probability of that individual being selected for reproduction, where
individuals containing high Ątness have more chance to be selected. At the selection stage,
the Ąttest individuals pass their genes (parameters) to the next generation. Two pairs of
parent-individuals, 𝐴 and 𝐵, are selected based on their Ątness scores, 𝑓 (𝐴) and 𝑓 (𝐵),
and have their parameters combined through a process called Crossover. This process is
one of the most signiĄcant phase in a genetic algorithm. A crossover point 𝑝, 1 < 𝑝 < 𝑁 ,
is randomly picked from within the parameters for each pair of parents to be mated. The
genes of the parents, 𝐴 and 𝐵, are combined in order to produce two new individuals, 𝐶
and 𝐷, called offspring:

𝑔C = [𝑔(1)A , . . . , 𝑔(𝑝)A , 𝑔(𝑝 + 1)B , . . . , 𝑔(𝑁 )B ] (1)

𝑔D = [𝑔(1)B , . . . , 𝑔(𝑝)B , 𝑔(𝑝 + 1)A , . . . , 𝑔(𝑁 )A ] (2)

In certain individuals of offspring, some of their genes can be subjected to a mutation with
a low random probability, 𝑟. This implies that some of the genes in the chromosome 𝑔(𝑖)j
that represents the individual can have its value modiĄed. A mutation occurs to maintain
diversity within the population and to prevent premature convergence of the population.
The mutation depends on how the gene is used to represent the data. For binary genes,
selected genes may have their bits Ćipped. For non-binary genes, for example, a unit
Gaussian distributed random value can be added to the gene.
Finally, the algorithm terminates if the population converges to a solution, i.e., the
population does not produce offspring signiĄcantly different from the previous generation.
When it occurs, it can be said that the genetic algorithm has provided a set of solutions
to the problem (EIBEN; SMITH, 2003).
In our work, we propose to use a genetic algorithm (GA) to estimate the best pa-
rameters of a kernel mask used to segment crop lines. We opted for this algorithm as
it presents a large range of application in several areas. For example, GA is used to
36 Chapter 2. Fundamentals

select optimal parameter values for image defogging algorithms, a technique commonly
used to correct image degradation produced by many outdoor working systems (GUO;
PENG; TANG, 2016). A modiĄed genetic algorithm (HEMANTH; ANITHA, 2019) is
proposed to minimize the random nature of conventional GA. The authors proposed this
modiĄcation aiming to improve medical image classiĄcation, more speciĄcally, the classi-
Ącation of abnormal brain images from four different classes. This modiĄcation presented
promising results and achieve 98% accuracy in the given problem. In (GHOSH et al.,
2016), GA is used for prostate automatic segmentation on pelvic images. The authors
propose a framework where GA evaluates candidate contours by combining representa-
tions of learned information (e.g., known shapes and local properties). Visual analysis of
the three dimensional segmentation indicates that GA is a feasible approach for pelvic
CT and MRI image segmentation.

2.6 Hough Transform

The Hough Transform (HT) was introduced and patented by P.V.C. Hough in 1962
(HOUGH, 1962). The primary original application of this technique was to detect lines
and arcs in the area of physics, more speciĄcally in particle detectors photographs. How-
ever, along the years various approaches have been used not only to improve this tech-
nique, but also to propose new applications for it (BELTRAMETTI; ROBBIANO, 2012).
Despite its initial use, in the last years, this technique has been extended to identifying
positions of arbitrary shapes being mostly used for DIP and image analysis, especially
for the detection of simple curves such as lines, circles (BELTRAMETTI; ROBBIANO,
2012) and ellipses (KHADANGA; JAIN, 2020).
As said, this technique is capable of detecting arbitrary images shapes. However, the
parametrized description of the shape in question must be given (BELTRAMETTI; ROB-
BIANO, 2012). Despite the great results in performing linear detection, it has been stated
that this technique requires a huge memory use and it has a considerable computational
complexity (EL HAJJOUJI et al., 2020).
In the last years, in particular in problems of recognition of special shapes in medical
and astronomical images, much effort has been made to apply the above described pro-
cedure to the detection of more complicated objects, in particular special algebraic plane
and space curves (BELTRAMETTI; ROBBIANO, 2012).
HT is a technique that aims feature extraction (DUDA; HART, 1972) and the mech-
anism behind this method uses a ŞvotingŤ process to Ąnd imperfect objects contours
encountered in the image that represent the class of shapes described in the parametrized
description (KHADANGA; JAIN, 2020). The voting scheme is carried out to obtain
the values in the parameter space that represent the line with more probability to exist.
Concurrently, when the shape aimed to be detected is a straight line, the process, also
2.7. Radon Transform 39

𝑥 = 𝑥′ cos ã ⊗ 𝑦 ′ sin ã (8)

𝑦 = 𝑥′ sin ã + 𝑦 ′ cos ã (9)

Denoted by the operator ℜ, the Radon transform is mathematically deĄned as:

𝑝φ (𝑥′ ) = ℜ[𝑓 (𝑥, 𝑦)] (10)

∫︁ ∞ ∫︁ ∞
= 𝑓 (𝑥, 𝑦)Ó(𝑥 cos ã + 𝑦 sin ã ⊗ 𝑥′ )𝑑𝑥𝑑𝑦 (11)
⊗∞ ⊗∞
∫︁ inf
= 𝑓 (𝑥′ cos ã ⊗ 𝑦 ′ sin ã, 𝑥′ sin ã + 𝑦 ′ cos ã)𝑑𝑦 ′ (12)
⊗ inf

Since the one-dimensional projection of the 𝑓 (𝑥, 𝑦) function at the ã angle is deĄned
as 𝑝φ (𝑥′ ), the Radon transform calculates the integral of a two-dimensional image on the
axis 𝑦 ′ .
40 Chapter 2. Fundamentals
 41

Chapter 3
Related Work

As already mentioned in this dissertation, the popularization of UAVs has been en-
abling authors to address many existing problems in agriculture. One of these problems
is to accurately identify existing crop lines in a region and, consequently, its geolocation,
arrangement, as well as how failures and gaps are distributed in the Ąeld. Literature
presents some existing approaches to address this problem, many of them are commonly
based on the Hough transform (HT) (HOUGH, 1962), which is widely used in problems
involving the detection of known objects, such as straight lines and circles (ILLING-
WORTH; KITTLER, 1988; HASSANEIN et al., 2015). Nonetheless, other techniques in
the state of the art have also been used to study problems similar to the one exposed
in this work. Hence, in the remainder of this chapter, we will discuss some of classical
approaches that have been taken mostly based on HT technique as well as some new and
very interesting approaches applied to similar problems that, by the way, also use CNN
as our work.

3.1 Hough Transform

Considering the imaging processing techniques available, the Hough transform (HOUGH,
1962) Ągures as an initial clear choice to detect crop lines. As presented in (ILLING-
WORTH; KITTLER, 1988) this technique is widely used to identify Ąxed parameterized
shapes formed by points on images. The basic Hough transform works well for straight
lines and circles, but it can also be used to Ąnd arbitrary shapes (BALLARD, 1981). How-
ever it has the disadvantage of requiring that the object shape to be known in advance,
which limits its application. Nonetheless, a number of solutions were proposed using the
Hough transform as basis.
Some authors along the years have presented adaptations of the Hough transform
to deal with images captured by cameras attached to tractors, such as (LEEMANS;
DESTAIN, 2006). In this work a row localization method was proposed using an adap-
tation of the HT. This approach works on video sequences captured on ground level by
3.2. Otsu Method 47

Figure 20 Ű Proposed image processing method architecture. Extracted from (GARCÍA-

SANTILLÁN et al., 2018).
50 Chapter 3. related-work

root rot that, despite being related to another culture crop, it gives us a great overview
of the state of the art in terms of modern technology and techniques applied in the Ąeld
and alternative approaches.

The work in (VIDOVIC; CUPEC; HOCENSKI, 2016) also uses ExG index to segment
images under perspective projection. It combines image evidence and prior knowledge
of the geometric structure of the crop using a dynamic programming technique. This is
performed to detect regular patterns related to the appearance of crop rows, both straight
and curved. This method is used as base algorithm in (BASSO; FREITAS, 2020), where
an entire guiding system for spraying UAVs is proposed. The idea is to identify the crop
rows during the UAV Ćight and to use this information to generate the driving parameters
sent to the Ćight controller. The authors claim that their approach is able to deal with
curved crop rows by dividing the curves into segments of straight line.

In (MONTALVO et al., 2012) yet another method of crop row identiĄcation is pre-
sented. It was devised to work on crops with high incidence of weeds and with camera
mounted on ground vehicles. According to the authors the image processing of this work
consists of three main stages: image segmentation, double thresholding based on the
OtsuŠs method, and crop row detection. The image segmentation is based on the applica-
tion of a VI, the double thresholding handles the separation between weeds and crops and
the crop row detection applies least squares linear regression for line adjustment. Also,
in (MONTALVO et al., 2013), the authors explore another technique based on image
segmentation procedures that works independently of the loss of greenness. First, they
perform a combination of vegetation indices and apply a Ąrst Otsu thresholding. Then,
they select black pixels and apply a second Otsu thresholding. Lastly, the histogram
obtained from pixels belonging to the background and masked plants is thresholded by
applying a last stage of Otsu. However, the main focus of these works falls on the image
processing task, since the crop row identiĄcation became trivial due the fact that only
straight lines are expected.

A different approach (SOUZA; ESCARPINATI; ABDALA, 2018) was used for images
of sugar cane and coffee taken on low altitude (≡100m). They assumed the images
are well segmented and, from a cloud of points representing the plants in a Ąeld they
subdivided such points into lines representing the true plantation rows. The proposed
algorithm is a two-fold process. First the points were subdivided into preliminary lines
by a procedure inspired in the formulation of hierarchical clustering. Afterward the lines
are pruned to correct for imprecisions introduced by Ąeld speciĄcities and the image pre-
processing. They achieved good results, but the total dependence of the segmentation
process represents a real problem for the algorithm.
3.5. Final Considerations 51

3.5 Final Considerations

In general, most of the works in the current literature still focus on straight lines
based on the HT technique (HASSANEIN et al., 2015). One of the problems with the
use of the Hough transform is that the shape of the detected object must be previously
known and described, which limits its application. In general, the crop lines can have
irregular shapes as they follow the terrain, thus creating curves and non-uniform shapes.
To circumvent this deĄciency, other possibilities have been explored. However, most of
these approaches either require processing of a crop line at a time, or focusing on plant
counting, or even focus on different crop cultures. In sugarcane crops, for example, its is
common that plants of different lines intersect to each-other, which impairs the detection
and correct reconstruction of the line.
52 Chapter 3. related-work
 53

Chapter 4
Methodology

4.1 Datasets
To evaluate the proposed methodology we used four test images of different sizes.
These images are mosaics of aerial images that represent areas of sugar cane cultivation
and that contain planting lines of different ages and widths. These images were acquired
using an eBee SenseFly mapping drone. We used a senseFly S.O.D.A. camera of 1in
Sensor, 5472 × 3648 pixel resolution and RGB lens F/2.8-11, 10.6 mm. Each pixel in the
image represents 5𝑐𝑚 of ground (Ground sample distance - GSD - of 0.053 meters).
Figure 24 shows a preview of the four test images used in the experiments (named
Dataset A, B, C, and D, respectively). It is important to notice that each image has the
crop lines of its entire region segmented by an expert, as illustrated in Figure 25. Note
that the rows can vary in width depending on the age of the crop and the level of success
of the planting process.

Figure 24 Ű Test images used to evaluate our approach and their respective sizes: (a)
11180 × 8449; (b) 19833 × 30255; (c) 17497 × 10771; (d) 16677 × 24181.
4.2. Segmentation using Genetic Algorithm 55

4.1.1 Evaluation metrics

For the evaluation step, we compared the binary images resulting from our segmen-
tation process with the crop line markings provided by an expert. For this comparison
we used the Dice Similarity Coefficient (DSC), a common index of segmentation accuracy
that measures how similar two binary images are.
Given two images, 𝐴 and 𝐵, Dice coefficient measures the intersection of the objects
present in the images, as described in Equation 13. The result from this equation is the
similarity 𝐷, 0 ⊘ 𝐷 ⊘ 1, where the more the value 𝐷 is close to 1, the more similar the
images 𝐴 and 𝐵 are (González Sánchez et al., 2020).

♣𝐴 ∩ 𝐵♣
𝐷=2 (13)
♣𝐴♣ + ♣𝐵♣
Also, for the SSN approach, the Jaccard Similarity Coefficient (JSC) (also known as
Intersection over Union) was used as a loss function in the training process. This measure
is highly recommended for segmentation problems where there are unbalanced classes.
The Jaccard Loss between two images 𝐴 and 𝐵 is deĄned as:

♣𝐴 ∩ 𝐵♣ ♣𝐴 ∩ 𝐵♣
𝐽(𝐴, 𝐵) = 1 ⊗ =1⊗ (14)
♣𝐴 ∪ 𝐵♣ ♣𝐴♣ + ♣𝐵♣ ⊗ ♣𝐴 ∩ 𝐵♣
In addition, DSC is quite similar to JSC. In fact, it is possible to make a conversion
from DSC to JSC as described in Equation 15 and vice versa (Equation 16).

𝐷
𝐽= (15)
2⊗𝐷

2𝐽
𝐷= (16)
𝐽 +1

4.2 Segmentation using Genetic Algorithm

Crop lines may vary in width depending of the age of the crop and how successful the
planting process was. Nevertheless, crop lines frequently show a greenish appearance in
contrast to the reddish presence of the soil.
Our Ąrst approach was creating and experimenting a method based on GA. Instead
of using color segmentation methods and different color space, such as Hue, Saturation,
Value (HSV), our Ąrst hypothesis was that it is possible to optimize a 3 × 3 × 3 kernel
mask so that its convolution combines the local characteristics of the RGB channels of the
image into a single gray scale image 𝑆, which could latter be segmented using a simple
and automatic threshold selection method, such as Otsu. To accomplish that we used a
genetic algorithm to optimize the 27 Ćoating point values for the 3 × 3 × 3 kernel mask.
The genetic algorithm ran for 2700 generations, using a population size of 200 individuals.
We used mutation rate of 0.05 and crossover rate of 0.8.
4.3. Semantic Segmentation Networks 57

learning and computer vision and it has been used with success in many applications,
including autonomous driving. Its relevance has been increasingly growing in the last years
due the resumption of convolution neural networks and its fast development, producing
the named Semantic Segmentation Networks (SSN). As better stated in Section 2.4.2,
these networks aim to assign semantic labels accurately to each pixel of an image. Thus,
in the proposed approach, the binarization is performed by using a network trained with
a dataset referenced by two classes (crop rows and background). An unclassiĄed image is
introduced into the CNN in order to generate an output binarized image, where each of
the two color values represents the classiĄcation of pixels based on these two classes.

In addition, it is important to state that while the Genetic Algorithm technique was
able to work over only one type of feature: reddish and greenish color tones to produce
a kernel Ąlter, Semantic Segmentation manages to extract several other different levels of
abstraction, each of these levels focusing on a different type of feature, such as border,
texture, etc. This is a very important aspect as depending on the stage of the after-
cut, dry leaves and ratoon is present in the soil between the crop-rows, confounding the
contrast between plants and the soil interfering in the computational analysis process.
Thus, as this color contrast may be compromised, a GA based method, in this case,
has a disadvantage and that is one of the reasons why we decided to go for the SSN
approach, even though the two methods being trained with both cane plant and cane
ratoon datasets.

Yet, as previously stated, Otsu global binarization method does not perform well to
segment crop lines as they present different local features due to the age and width of
the crop line. Also, the Otsu method picking a local threshold to perform binarization is
not a feasible option. As the CNN approach does not depend on Otsu as the GA based
approach did, this is another reason why we chose to follow the Semantic Segmentation
path in our research.

For this approach, we used the same datasets previewed in Figure 24. For each image,
we cropped the mosaic into pieces of 256 × 256 pixels size, with a stride of 256 pixels.
If the cropped area does not contain at least 80% of useful information (i.e., pixels with
values other than zero), the sample is discarded. After cropping, datasets A, B, C, and
D contained a total of, respectively, 678, 3291, 1550 and 2162 images.

In this work, we evaluated U-net, LinkNet, and PSPNet semantic neural networks
for the segmentation in aerial crop images. We replaced the encoder of each network
by the VGG16 pre-trained with ImageNet weights (SIMONYAN; ZISSERMAN, 2014)
due to its high performance in feature extraction in precision agriculture applications
(FAWAKHERJI et al., 2019). The basics of these networks was explored in section 2.4.2
and the use of this networks in our new approach is described in the next section (4.3.1).
58 Chapter 4. methodology

4.3.1 Proposed Approach

Our approach to perform automatic segmentation of crop lines consists of two steps.
First, we use a Convolutional Neural Network to segment the image into crop lines (region
of interest) and soil (background). To accomplish that we evaluated three segmentation
networks: Unet, LinkNet and PSPNet. For all of them we used VGG16 structure (pre-
trained with the ImageNet) to extract the feature maps. Figure 28 shows the architecture
of the segmentation networks used. Each network was trained with all the images ex-
tracted from the dataset. Each image has a corresponding binarized image provided by
an expert extracted from the binarized mosaic using the same pixel coordinates, so that
the networks could learn which part of the image corresponded to the Ąrst class (crop-
row) and which part corresponded to the other class that represents the background (soil,
weed or ratton depending on the phase of the crop, region and image).

Figure 28 Ű Architectures used for semantic segmentation. Adapted from

(YAKUBOVSKIY, 2019).

In an ideal crop, only the segmentation step should be sufficient to obtain the crop
lines accurately. However, in the vast majority of cases, there are external factors that
may affect the results. Among these factors are sowing failures (i.e., absence of plants
in a section of the line), weed plants (which in the segmentation process will be treated
as plants) and plants that are outside the crop row. An example of these problems are
shown in Figure 29.
Thus, to improve the segmentation obtained, we used a reĄnement step, which aims to
reconstruct and to improve the previous segmented lines by making them more uniform
or linking line fragments and loose plants that belong to the same crop line. For this new
approach we also used the Radon transform (Section 4.4) to reconstruct the line obtained
during the segmentation step to coincide with the one marked by the expert, as shown in
Figure 30.
4.4. Line Reconstruction and Refinement 59

Figure 29 Ű Example of problems encountered after the segmentation step: (a) Original
image; (b) Planting lines provided by an expert; (c) Image after segmentation.

4.4 Line Reconstruction and Refinement

After the binarization process, we needed to reconstruct and improve the already
segmented lines. Our aim was to reconnect fragments of crop lines and loose plants that
were supposed to belong together and have been separated by failures in the seeding
process or in the segmentation step. Also, we aimed to make the crop lines more uniform
and smooth, cutting false connection between parallel crop rows.
For this step, we used the Radon transform method. First, we compute the Radon
transform from a given input image in order to obtain the orientation angle of the image
(Figure 30b). Using this orientation angle it is possible to analyse how the pixels which
compose the detected areas are distributed along the image and to detected the center
of each crop line in the image (Figure 30c). Then, by using the orientation angle of the
image and the center of each crop line we were able to reconstruct the crop lines with the
same constant width as used by the expert, as shown in Figure 30d.
60 Chapter 4. methodology

Figure 30 Ű Proposed scheme for crop line reconstruction using Radon transform: (a) In-
put image; (b) Matrix obtained with the Radon transform. The red dot
represents the location of the maximum point and the orientation angle of
the input image; (c) Radon transform obtained for the image orientation an-
gle (red line in (b)). Each peak of the curve corresponds to the center of a
line in the input image; (d) Reconstruction of the lines using the orientation
angle and the peaks of the Radon transform for that angle.
 61

Chapter 5
Experimental Results

As previously mentioned, in the Ąrst stages of this research we experimented with a

Genetic Algorithm based approach. Then, due to some reasons including the relevance
of Semantic Segmentation in the recent literature, its levels of abstraction, and the non-
feasible results of Otsu associated with GA, we decided to study and proposed a new
approach based on CNN to accomplish the binarizartion. In this chapter, we better
describe experimental results for each of the two approaches including the crop line re-
Ąnement and reconstruction performed in both approaches by the Radon transform. In
the end, we synthesize and discuss the results and differences of our test approaches.

5.1 Segmentation using Generic Algorithm

Since Genetic Algorithm is stochastic, results may be different at independent runs. To
address this important issue, we Ąrst investigated the convergence of the genetic algorithm
through a cross-validation scheme. To accomplish that, we split all 35 training images
into 5-folds and we used the images in the training folds as input for the genetic algorithm
to compute the kernel mask. As for the Ątness function, the GA aimed to search for the
kernel mask that maximized the average Dice coefficient of the images in the current
training folds.
Figure 31 shows the average Dice coefficient obtained by each image using the 5 kernel
mask computed by GA. We must emphasize that, instead of using the images in their
respective test fold, we evaluated each kernel in all images in order to have a better
estimation of the result. Different kernel mask results in similar segmentation in terms of
Dice coefficient. However, it does not mean that the 5 kernel mask present the same set
of values, but that they produce similar segmentation results.
Figure 32 presents some examples of segmented areas obtained at different sections
of the evaluated crop maps. It is important to notice that the crop lines segmented by
experts did not follow exactly the width of each crop line (Figure 32b). Instead, these
images contain uniform lines marking the location of each crop line. Moreover, the expert
62 Chapter 5. Experimental Results

Figure 31 Ű Average Dice coefficient and standard deviation for different images for 5
different GA kernel masks.

also marked regions where the crop line should exist, even though there is not any plant
there. Since these markings do not follow the natural width of the crop lines, an error
is expected when comparing the segmentation provided by an expert with the results
obtained by our approach.
We noticed that the application of the convolution kernel resulted in an image that
is mostly black and white. Although other gray levels are present in the image, their
frequency are not signiĄcant and could compromise the use of a Global Otsu threshold.
This explain the poor segmentation obtained in Figure 32d. The same is true when using
a manually deĄned threshold. Although faster than using the Otsu method, it is not
possible to set a user deĄned global threshold that achieves good results for all evaluated
images, as shown in Figure 33. As a result, depending on the threshold value used, some
crop lines may be missed while we improve the detection of other crop lines (Figure 32c).
One must consider that global binarization may not be the best approach to segment
crop line as they present different local features due to the age and width of the crop
line. In order to investigate such matter we applied Otsu algorithm locally on the images.
To accomplish that we used a square window of 𝑊 × 𝑊 pixels size and we moved this
window along the image using different values of stride, 𝑆. We used ŞOR operatorŤ to
combine all local binarizations into a single one binary image. Figure 34 shows that local
analysis improves the Dice coefficient obtained by Otsu binarizaton in all conĄgurations
evaluated. In general, small windows present better results as they enable Otsu method
to capture the local features with more precision, thus providing a better binarization and
a higher Dice coefficient, as shown in Figure 32e.
Although the use of local Otsu improves the detection of crop lines (Figure 32e),
we noticed that crop lines which are parallels to each other are sometimes connected by
regions incorrectly detected as crop lines. Moreover, the detected line presents an irregular
width which compromises the comparison with the expertŠs segmentation.
Since the orientation angle of the image is a local feature, we evaluated this approach
64 Chapter 5. Experimental Results

Figure 33 Ű Dice coefficient for various global threshold values.

Figure 34 Ű Dice coefficient obtained using Global Otsu and Local Otsu for different com-
binations of Window 𝑊 and Stride 𝑆.

Figure 35 Ű Dice coefficient obtained for the line reconstruction for different combinations
of Window 𝑊 and Stride 𝑆.
66 Chapter 5. Experimental Results

5.2 Semantic Segmentation

Initially, we applied the semantic segmentation networks to the images obtained in each
dataset. To accomplish that we adjusted the input size of the networks to 256 × 256 × 3
(YAKUBOVSKIY, 2019) and each network has been trained for 50 epochs using the Adam
optimizer with a 0.001 learning rate. As loss function, we used the Jaccard-loss (BEERS
et al., 2019). We also used simple data augmentation methods, such as rotation (up to
∘180◇ ), translation (up to ∘20𝑝𝑥), scaling (up to ∘7%), and shearing, during the training
step to prevent overĄtting. To guarantee a correct validation process, a validation step
was implemented using K-fold, with 𝐾 = 10. For the training, we used only dataset A to
train and validate the network. Figure 38 and Table 1 show the average results obtained
by each network.

Segmentation Network Dice Coefficient

VGG16 - LinkNet 0.90 ± 0.0062
VGG16 - PSPNet 0.88 ± 0.0075
VGG16 - Unet 0.87 ± 0.0113
Table 1 Ű Segmentation results obtained with the application of the segmentation net-
works in Dataset A.

Results show that LinkNet obtained the best result for the Dice coefficient when
segmenting crop lines in the images. Besides, this network was also the one that showed
the least variation among test folds. Due to its consistent result in the plant detection
process, we chose LinkNet as the standard network for the segmentation stage. We restore
the weights that generated the best result during the training of the LinkNet and we used
them to evaluate other datasets (B, C, and D). Table 2 shows the results obtained for
datasets B, C, and D.

Dataset Dice Coefficient

A 0.90 ± 0.0062
B 0.80 ± 0.0702
C 0.84 ± 0.0724
D 0.86 ± 0.0588
Table 2 Ű Result obtained with the application of the LinkNet network trained in dataset
A to segment other datasets.

We noticed a slight worsening of the LinkNet result when applied to other datasets.
The decrease in the Dice coefficient depends on the dataset evaluated. However, it is
possible to notice that the average Dice coefficient is superior to 0.80 in all datasets. A
possible explanation for this behavior lies in the fact that crop lines can vary in width
depending on the age of the crop and how the planting process was successful, as illustrated
in Figure 25.
5.3. Comparison of approaches 71

constant Dice coefficient when compared to Genetic Algorithm results. Certainly, this is
due to the fact that in our approach GA was able to work over only one type of feature
(reddish and greenish color tones to produce a kernel Ąlter). Depending on the crop, dry
leaves, weed and ratoon may be present between the crop-rows, compromising the reddish
contrast between plants and the background. Thus, as the Semantic Segmentation based
technique manages to extract several other different levels of abstraction it tends to be
more capable of operating well in different stages of the crop regardless of color contrast.
Yet, the experiments with Genetic Algorithm method did not show good segmentation
results according to the Dice evaluation metric when associated with a global Otsu tech-
nique. It is undeniable that a local Otsu analysis improves the Dice coefficient obtained.
Small windows present better results as they enable Otsu method to better capture the
features with more precision. However, the general results did not reach even 0.78, while
LinkNet shows average Dice results from 0.80 to 0.86 in tested datasets that were not
part of the training, reaching 0.90 when tested with the training dataset.
72 Chapter 5. Experimental Results
 73

Chapter 6
Conclusion

In this work, we presented a methodology to segment crop lines from UAV images.
First, we experimented an approach based on Genetic Algorithm associated with Otsu
method to produce binarized images that were then reconstructed using Radon transform.
Then, due to some reasons including the recent relevance of Semantic Segmentation in the
literature, its levels of abstraction, and the non-feasible results of Otsu associated with
GA, we studied and proposed a new approach based on SSN. This new approach uses a
Convolutional Neural Network to perform the segmentation step. Among the networks
evaluated, the one that stands is LinkNet presenting the best results to segment crop lines,
obtaining a higher and much more consistent Dice coefficient for the datasets evaluated.
Which is extremely positive despite the fact that this method requires a larger number of
images for the training process. We also proposed a line reconstruction approach based
on the Radon transform for this technique testing some variation Ąlters. Although the
crop row reconstruction sometimes producing a slight decrease in the Dice coefficient, it
enables us to improve the segmentation results by connecting fragments of crop lines and
by Ąlling segmentation errors caused by missing plants, thus indicating that our approach
is a feasible solution to segment crop lines in images.

6.1 Main Contributions

The main contributions of this work are: the development of a CNN training capable
of classifying crop images in two classes (crop-rows and background) generating binarized
images, as this efficient segmentation reduces drastically the cost of the post-processing
step; and also the development of a framework that receives sugarcane crop images,
binarize them using the CNN already trained and then perform a post-processing step
based on Radon transform to achieve a reĄnement process reconstructing and enhancing
the lines, making them more uniform and linking row fragments that are supposed to
belong together at the same line/row.
In addition, the process of detecting the crop lines in an effective and concise way is
74 Chapter 6. Conclusion

strongly for important for PA, specially planning the harvests, inputs usage, estimating
of production, counting plants, early correction of sowing failures. In addition, the geolo-
cation information of the detected crop rows lessens the waste of inputs, the harm to the
environment and Ąnancial costs. Also, it allows autonomous machinery guidance through
the crop,
Finally, sugarcane represents a great percentage of all crops worldwide and also it is
a semi-perennial crop, which means that it can be harvested annually for several years
without replanting. Thus, the correct and effective detection and frequent maintenance
over the years can bring a huge economic impact for producers and, consequently, their
countries.

6.2 Contributions in Bibliographic Production

❏ The following papers are strongly connected to the research presented in this dis-
sertation. They were submitted for their respective journals and are currently in
phase of analysis:

– SILVA, R. R.; ESCARPINATI, M. C. and BACKES, A. R. Sugarcane Crop

Line Detection From UAV Images Using Genetic Algorithm and Radon Trans-
form. Submitted to Signal, Image and Video Processing manuscript.
– SILVA, R. R.; DIAS JR., J. D.; ESCARPINATI, M. C. and BACKES, A.
R. Detection of sugarcane crop line from UAV images using Semantic Seg-
mentation and Radon Transform. Submitted to Computers and Electronics in
Agriculture.

❏ The study of CNNs done for this research also allowed us to work in other problems
related to computer vision. Hence, our following paper was recently published:

– SILVA, R. R.; BRITO, L. F. A.; ALBERTINI, M. K.; NASCIMENTO, M. Z.

and BACKES, A. R. Using CNNs for Quality Assessment of No-Reference and
Full-Reference Compressed-Video Frames. In: XVI WORKSHOP DE VISÃO
COMPUTACIONAL, 2020, Uberlândia. Anais do 16◇ Workshop de Visão
Computacional, 2020.

❏ Furthermore, it is worth mentioning that this work is currently running for the
Prêmio Mercosul de Ciência e Tecnologia (Mercosur Science and Technology Award):

– SILVA, R. R.; DIAS JR., J. D.; ESCARPINATI, M. C. and BACKES, A.

R. Detecção de linha de plantio de cana de açúcar a partir de imagens de
VANT usando Segmentação Semântica e Transformada de Radon. Submitted
to Prêmio Mercosul de Ciência e Tecnologia Ű edição 2020.
6.3. Future Work 75

6.3 Future Work

The results obtained by this work demonstrate the good performance obtained by the
proposed approach and motivate new lines of investigation, such as: evaluation of datasets
of different cultures besides sugar cane; explore how mosaic alignment techniques interfere
in the result; explore the use of other sensors in association with the images to produce
better results; study new methods to enhance crop reconstruction of regions with highly-
curved lines.
76 Chapter 6. Conclusion
 77

Bibliography

ANTHIMOPOULOS, M. et al. Lung pattern classiĄcation for interstitial lung diseases

using a deep convolutional neural network. IEEE Trans. Medical Imaging, v. 35,
n. 5, p. 1207Ű1216, 2016. Disponível em: <https://doi.org/10.1109/TMI.2016.2535865>.

BAH, M. D.; HAFIANE, A.; CANALS, R. Crownet: Deep network for crop row
detection in uav images. IEEE Access, v. 8, p. 5189Ű5200, 2020. Disponível em:
<https://doi.org/10.1109/ACCESS.2019.2960873>.

BALLARD, D. H. Generalizing the hough transform to detect arbitrary shapes.

Pattern recognition, Elsevier, v. 13, n. 2, p. 111Ű122, 1981. Disponível em:
<https://doi.org/10.1016/0031-3203(81)90009-1>.

BARRETT, H.; SWINDELL, W. Radiological Imaging: The Theory of Image

Formation, Detection, and Processing. [S.l.]: Elsevier Science, 1996. ISBN
9780120796038.

BASSO, M.; FREITAS, E. P. de. A uav guidance system using crop row detection
and line follower algorithms. J. Intell. Robotic Syst, v. 97, n. 3, p. 605Ű621, 2020.
Disponível em: <https://doi.org/10.1007/s10846-019-01006-0>.

BEERS, F. van et al. Deep neural networks with intersection over union loss for binary
image segmentation. In: ICPRAM. [S.l.: s.n.], 2019. p. 438Ű445.

BELTRAMETTI, M. C.; ROBBIANO, L. An algebraic approach to hough transforms.

Journal of Algebra, v. 371, p. 669 Ű 681, 2012. ISSN 0021-8693. Disponível em:
<http://www.sciencedirect.com/science/article/pii/S002186931200453X>.

BOSILJ, P.; DUCKETT, T.; CIELNIAK, G. Connected attribute morphology

for uniĄed vegetation segmentation and classiĄcation in precision agriculture.
Computers in Industry, v. 98, p. 226 Ű 240, 2018. ISSN 0166-3615. Disponível em:
<http://www.sciencedirect.com/science/article/pii/S0166361517306139>.

BRAS, G. et al. Transfer learning method evaluation for automatic pediatric chest
x-ray image segmentation. In: 2020 International Conference on Systems,
Signals and Image Processing (IWSSIP). [s.n.], 2020. p. 128Ű133. Disponível em:
<https://doi.org/10.1109/IWSSIP48289.2020.9145401>.
78 Bibliography

BRINKMAN, M. L. et al. Interregional assessment of socio-economic effects of

sugarcane ethanol production in brazil. Renewable and Sustainable Energy
Reviews, v. 88, p. 347 Ű 362, 2018. ISSN 1364-0321. Disponível em: <http:
//www.sciencedirect.com/science/article/pii/S1364032118300340>.

CHAMOLA, V. et al. A comprehensive review of the covid-19 pandemic and the role
of iot, drones, ai, blockchain, and 5g in managing its impact. IEEE Access, v. 8, p.
90225Ű90265, 2020. Disponível em: <https://doi.org/10.1109/ACCESS.2020.2992341>.

CHAN, Y. K.; KOO, V. C. An introduction to synthetic aperture radar (sar).

Progress In Electromagnetics Research B, v. 2, p. 27Ű60, 2008. Disponível em:
<https://doi.org/10.2528/PIERB07110101>.

CHAURASIA, A.; CULURCIELLO, E. Linknet: Exploiting encoder representations for

efficient semantic segmentation. In: IEEE. 2017 IEEE Visual Communications and
Image Processing (VCIP). 2017. p. 1Ű4. Disponível em: <https://doi.org/10.1109/
VCIP.2017.8305148>.

CHENG, X. et al. Pest identiĄcation via deep residual learning in complex

background. Comput. Electron. Agric, v. 141, p. 351Ű356, 2017. Disponível em:
<https://doi.org/10.1016/j.compag.2017.08.005>.

CLEVERT, D.-A.; UNTERTHINER, T.; HOCHREITER, S. Fast and accurate deep

network learning by exponential linear units (elus). In: BENGIO, Y.; LECUN, Y. (Ed.).
4th International Conference on Learning Representations, ICLR 2016, San
Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings. [s.n.], 2016.
Disponível em: <http://arxiv.org/abs/1511.07289>.

DEWA, C. K.; AFIAHAYATI. Suitable cnn weight initialization and activation function
for javanese vowels classiĄcation. Procedia Computer Science, v. 144, p. 124 Ű 132,
2018. ISSN 1877-0509. INNS Conference on Big Data and Deep Learning. Disponível
em: <http://www.sciencedirect.com/science/article/pii/S187705091832221X>.

DIAS, P. A.; TABB, A.; MEDEIROS, H. Apple Ćower detection using deep
convolutional networks. Comput. Ind, v. 99, p. 17Ű28, 2018. Disponível em:
<https://doi.org/10.1016/j.compind.2018.03.010>.

DOW, J. M.; NEILAN, R. E.; RIZOS, C. The international gnss service in a changing
landscape of global navigation satellite systems. Journal of Geodesy, v. 83, n. 3, p.
191Ű198, 2009. Disponível em: <https://doi.org/10.1007/s00190-008-0300-3>.

DUDA, R. O.; HART, P. E. Use of the hough transformation to detect lines and
curves in pictures. Commun. ACM, v. 15, n. 1, p. 11Ű15, 1972. Disponível em:
<https://doi.org/10.1145/361237.361242>.

ECKLE, K.; SCHMIDT-HIEBER, J. A comparison of deep networks with relu activation

function and linear spline-type methods. Neural Networks, v. 110, p. 232 Ű 242, 2019.
ISSN 0893-6080. Disponível em: <http://www.sciencedirect.com/science/article/pii/
S0893608018303277>.

EIBEN, A. E.; SMITH, J. E. Introduction to evolutionary computation. 2. ed.

[S.l.]: Springer-Verlag, 2003. (Natural computing series). ISBN 978-3-662-05094-1.
Bibliography 79

EL HAJJOUJI, I. et al. A novel fpga implementation of hough transform for

straight lane detection. Engineering Science and Technology, an International
Journal, v. 23, n. 2, p. 274 Ű 280, 2020. ISSN 2215-0986. Disponível em:
<http://www.sciencedirect.com/science/article/pii/S2215098618314782>.

ELFWING, S.; UCHIBE, E.; DOYA, K. Sigmoid-weighted linear units for neural
network function approximation in reinforcement learning. Neural Networks, v. 107,
p. 3 Ű 11, 2018. ISSN 0893-6080. Special issue on deep reinforcement learning. Disponível
em: <http://www.sciencedirect.com/science/article/pii/S0893608017302976>.

FAWAKHERJI, M. et al. Crop and weeds classiĄcation for precision agriculture

using context-independent pixel-wise segmentation. In: IEEE. 2019 Third IEEE
International Conference on Robotic Computing (IRC). 2019. p. 146Ű152.
Disponível em: <https://doi.org/10.1109/IRC.2019.00029>.

FENG, D. et al. Deep multi-modal object detection and semantic segmentation

for autonomous driving: Datasets, methods, and challenges. IEEE Transactions
on Intelligent Transportation Systems, IEEE, 2020. Disponível em: <https:
//doi.org/10.1109/TITS.2020.2972974>.

FERENTINOS, K. P. Deep learning models for plant disease detection and

diagnosis. Comput. Electron. Agric, v. 145, p. 311Ű318, 2018. Disponível em:
<https://doi.org/10.1016/j.compag.2018.01.009>.

FERREIRA, A. dos S. et al. Weed detection in soybean crops using convnets.

Comput. Electron. Agric, v. 143, p. 314Ű324, 2017. Disponível em: <https:
//doi.org/10.1016/j.compag.2017.10.027>.

FUENTES-PEÑAILILLO, F. et al. Using clustering algorithms to segment uav-based rgb

images. In: IEEE. 2018 IEEE International Conference on Automation/XXIII
Congress of the Chilean Association of Automatic Control (ICA-ACCA).
2018. p. 1Ű5. Disponível em: <https://doi.org/10.1109/ICA-ACCA.2018.8609822>.

GARCÍA-SANTILLÁN, I. et al. Curved and straight crop row detection by accumulation

of green pixels from images in maize Ąelds. Precision Agriculture, v. 19, n. 1, p. 18 Ű
41, 2018. Disponível em: <https://doi.org/10.1007/s11119-016-9494-1>.

GHOSH, P. et al. Incorporating priors for medical image segmentation using a genetic
algorithm. Neurocomputing, v. 195, p. 181 Ű 194, 2016. ISSN 0925-2312. Learning for
Medical Imaging. Disponível em: <https://doi.org/10.1016/j.neucom.2015.09.123>.

GLOROT, X.; BENGIO, Y. Understanding the difficulty of training deep feedforward

neural networks. In: TEH, Y. W.; TITTERINGTON, D. M. (Ed.). Proceedings
of the Thirteenth International Conference on Artificial Intelligence and
Statistics, AISTATS 2010, Chia Laguna Resort, Sardinia, Italy, May
13-15, 2010. JMLR.org, 2010. (JMLR Proceedings, v. 9), p. 249Ű256. Disponível em:
<http://proceedings.mlr.press/v9/>.

GONZALEZ, R.; WOODS, R. Processamento Digital De Imagens. ADDISON

WESLEY BRA, 2000. ISBN 9788576054016. Disponível em: <https://books.google.
com.br/books?id=r5f0RgAACAAJ>.
80 Bibliography

González Sánchez, J. C. et al. Segmentation of bones in medical dual-energy computed

tomography volumes using the 3d u-net. Physica Medica, v. 69, p. 241 Ű 247, 2020.
ISSN 1120-1797. Disponível em: <http://www.sciencedirect.com/science/article/pii/
S1120179719305356>.
GUERRERO, J. M.; RUZ, J. J.; PAJARES, G. Crop rows and weeds detection in
maize Ąelds applying a computer vision system based on geometry. Computers and
Electronics in Agriculture, v. 142, p. 461 Ű 472, 2017. ISSN 0168-1699. Disponível
em: <https://doi.org/10.1016/j.compag.2017.09.028>.
GUO, F.; PENG, H.; TANG, J. Genetic algorithm-based parameter selection approach
to single image defogging. Information Processing Letters, v. 116, n. 10, p. 595 Ű
602, 2016. ISSN 0020-0190. Disponível em: <https://doi.org/10.1016/j.ipl.2016.04.013>.
HASEGAWA, M.; TABBONE, S. Amplitude-only log radon transform for geometric
invariant shape descriptor. Pattern Recognition, v. 47, n. 2, p. 643Ű658, 2014.
Disponível em: <https://doi.org/10.1016/j.patcog.2013.07.024>.
HASEGAWA, M.; TABBONE, S. Histogram of radon transform with angle correlation
matrix for distortion invariant shape descriptor. Neurocomputing, v. 173, p. 24Ű35,
2016. Disponível em: <https://doi.org/10.1016/j.neucom.2015.04.100>.
HASSANEIN, A. S. et al. A survey on hough transform, theory, techniques and
applications. CoRR, abs/1502.02160, 2015. Disponível em: <http://arxiv.org/abs/
1502.02160>.
HE, K. et al. Delving deep into rectiĄers: Surpassing human-level performance on
imagenet classiĄcation. In: ICCV. IEEE Computer Society, 2015. p. 1026Ű1034. ISBN
978-1-4673-8391-2. Disponível em: <http://doi.ieeecomputersociety.org/10.1109/ICCV.
2015.123>.
HE, K. et al. Deep residual learning for image recognition. In: Proceedings of the
IEEE conference on computer vision and pattern recognition. [s.n.], 2016. p.
770Ű778. Disponível em: <https://doi.org/10.1109/CVPR.2016.90>.
HEMANTH, D. J.; ANITHA, J. ModiĄed genetic algorithm approaches for
classiĄcation of abnormal magnetic resonance brain tumour images. Applied
Soft Computing, v. 75, p. 21 Ű 28, 2019. ISSN 1568-4946. Disponível em:
<https://doi.org/10.1016/j.asoc.2018.10.054>.
HOUGH, P. V. Method and means for recognizing complex patterns. [S.l.]:
Google Patents, 1962. US Patent 3,069,654.
ILIEV, A. I.; KYURKCHIEV, N.; MARKOV, S. On the approximation of the step
function by some sigmoid functions. Math. Comput. Simul, v. 133, p. 223Ű234, 2017.
Disponível em: <https://doi.org/10.1016/j.matcom.2015.11.005>.
ILLINGWORTH, J.; KITTLER, J. A survey of the hough transform. Computer
vision, graphics, and image processing, Elsevier, v. 44, n. 1, p. 87Ű116, 1988.
Disponível em: <https://doi.org/10.1016/S0734-189X(88)80033-1>.
JI, R.; QI, L. Crop-row detection algorithm based on random hough transformation.
Mathematical and Computer Modelling, Elsevier, v. 54, n. 3-4, p. 1016Ű1020, 2011.
Disponível em: <https://doi.org/10.1016/j.mcm.2010.11.030>.
Bibliography 81

JR, E. R. H.; DAUGHTRY, C. S. What good are unmanned aircraft systems for
agricultural remote sensing and precision agriculture? International journal of
remote sensing, Taylor & Francis, v. 39, n. 15-16, p. 5345Ű5376, 2018. Disponível em:
<https://doi.org/10.1080/01431161.2017.1410300>.

KANG, L. et al. Convolutional neural networks for no-reference image quality assessment.
In: CVPR. IEEE Computer Society, 2014. p. 1733Ű1740. ISBN 978-1-4799-5118-5.
Disponível em: <http://ieeexplore.ieee.org/xpl/mostRecentIssue.jsp?punumber=
6909096>.

KHADANGA, G.; JAIN, K. Tree census using circular hough transform and grvi.
Procedia Computer Science, v. 171, p. 389 Ű 394, 2020. ISSN 1877-0509.
Third International Conference on Computing and Network Communications
(CoCoNetŠ19). Disponível em: <http://www.sciencedirect.com/science/article/pii/
S1877050920310061>.

KREUZER, M.; ROBBIANO, L. Computational Commutative Algebra 1.

Springer, 2000. Disponível em: <https://doi.org/10.1007/978-3-540-70628-1>.

KRIZHEVSKY, A.; SUTSKEVER, I.; HINTON, G. E. Imagenet classiĄcation with deep

convolutional neural networks. In: NIPS. [S.l.: s.n.], 2012.

LAGACHERIE, P.; MCBRATNEY, A. Chapter 1 spatial soil information systems

and spatial soil inference systems: Perspectives for digital soil mapping. In:
LAGACHERIE, P.; MCBRATNEY, A.; VOLTZ, M. (Ed.). Digital Soil Mapping.
Elsevier, 2006, (Developments in Soil Science, v. 31). p. 3 Ű 22. Disponível em:
<http://www.sciencedirect.com/science/article/pii/S016624810631001X>.

LEEMANS, V.; DESTAIN, M.-F. Line cluster detection using a variant of the hough
transform for culture row localisation. Image and Vision Computing, Elsevier, v. 24,
n. 5, p. 541Ű550, 2006. Disponível em: <https://doi.org/10.1016/j.imavis.2006.02.004>.

LI, W. et al. A hybrid modelling approach to understanding adoption of precision

agriculture technologies in chinese cropping systems. Computers and Electronics
in Agriculture, v. 172, p. 105305, 2020. ISSN 0168-1699. Disponível em:
<http://www.sciencedirect.com/science/article/pii/S0168169919323191>.

LIMA, M. et al. Sugarcane: Brazilian public policies threaten the amazon and
pantanal biomes. Perspectives in Ecology and Conservation, 2020. ISSN
2530-0644. Disponível em: <http://www.sciencedirect.com/science/article/pii/
S2530064420300262>.

LIN, G.; SHEN, W. Research on convolutional neural network based on improved

relu piecewise activation function. Procedia Computer Science, v. 131, p. 977 Ű
984, 2018. ISSN 1877-0509. Recent Advancement in Information and Communication
Technology:. Disponível em: <http://www.sciencedirect.com/science/article/pii/
S1877050918306197>.

LINDBLOM, J. et al. Promoting sustainable intensiĄcation in precision agriculture :

review of decision support systems development and strategies. Högskolan i Skövde,
Institutionen för informationsteknologi; Högskolan i Skövde, Forskningscentrum för
Informationsteknologi; Department of Urban and Rural development, Swedish University
82 Bibliography

of Agricultural Sciences, Skara, Sweden; National Competence Centre for Advisory

Services, Swedish University of Agricultural Sciences, Skara, Sweden; Department of Soil
and Environment, Swedish University of Agricultural Sciences, Skara, Sweden, 2017. ISSN
1385-2256. Disponível em: <http://urn.kb.se/resolve?urn=urn:nbn:se:his:diva-13271>.

LIU, Z. et al. A heterogeneous processor design for cnn-based ai applications on iot

devices. Procedia Computer Science, v. 174, p. 2 Ű 8, 2020. ISSN 1877-0509.
2019 International Conference on IdentiĄcation, Information and Knowledge in the
Internet of Things. Disponível em: <http://www.sciencedirect.com/science/article/pii/
S1877050920315611>.

LONG, J.; SHELHAMER, E.; DARRELL, T. Fully convolutional networks for

semantic segmentation. CoRR, abs/1411.4038, 2014. Disponível em: <http:
//arxiv.org/abs/1411.4038>.

LONG, J.; SHELHAMER, E.; DARRELL, T. Fully convolutional networks for

semantic segmentation. In: 2015 IEEE Conference on Computer Vision
and Pattern Recognition (CVPR). [s.n.], 2015. p. 3431Ű3440. Disponível em:
<https://doi.org/10.1109/CVPR.2015.7298965>.

MCBRATNEY, A. et al. Future directions of precision agriculture. Precision

agriculture, Springer, v. 6, n. 1, p. 7Ű23, 2005. Disponível em: <https://doi.org/10.
1007/s11119-005-0681-8>.

MILELLA, A.; REINA, G.; NIELSEN, M. A multi-sensor robotic platform for ground
mapping and estimation beyond the visible spectrum. Precision Agriculture, Springer,
p. 1Ű22, 2018. Disponível em: <https://doi.org/10.1007/s11119-018-9605-2>.

MIRZAEI, F. M.; ROUMELIOTIS, S. I. A kalman Ąlter-based algorithm for

imu-camera calibration: Observability analysis and performance evaluation. IEEE
Transactions on Robotics, v. 24, n. 5, p. 1143Ű1156, 2008. Disponível em:
<https://doi.org/10.1109/TRO.2008.2004486>.

MITCHELL, M. An Introduction to Genetic Algorithms. [S.l.]: The MIT Press,

1996.

MONTALVO, M. et al. Automatic expert system for weeds/crops identiĄcation in images

from maize Ąelds. Expert Systems with Applications, Elsevier, v. 40, n. 1, p. 75Ű82,
2013. Disponível em: <https://doi.org/10.1016/j.eswa.2012.07.034>.

MONTALVO, M. et al. Automatic detection of crop rows in maize Ąelds with high weeds
pressure. Expert Systems with Applications, Elsevier, v. 39, n. 15, p. 11889Ű11897,
2012. Disponível em: <https://doi.org/10.1016/j.eswa.2012.02.117>.

NADIPALLY, M. Chapter 2 - optimization of methods for image-texture segmentation

using ant colony optimization. In: HEMANTH, D. J.; GUPTA, D.; Emilia Balas, V.
(Ed.). Intelligent Data Analysis for Biomedical Applications. Academic Press,
2019, (Intelligent Data-Centric Systems). p. 21 Ű 47. ISBN 978-0-12-815553-0. Disponível
em: <http://www.sciencedirect.com/science/article/pii/B9780128155530000021>.

NAGATA, S. et al. Objective scoring of streetscape walkability related to leisure

walking: Statistical modeling approach with semantic segmentation of google street
Bibliography 83

view images. Health & Place, v. 66, p. 102428, 2020. ISSN 1353-8292. Disponível em:
<http://www.sciencedirect.com/science/article/pii/S1353829220302720>.

NAIR, V.; HINTON, G. E. RectiĄed linear units improve restricted boltzmann machines.
In: FÜRNKRANZ, J.; JOACHIMS, T. (Ed.). Proceedings of the 27th International
Conference on Machine Learning (ICML-10), June 21-24, 2010, Haifa, Israel.
[S.l.]: Omnipress, 2010. p. 807Ű814.

NEMOTO, T. et al. Simple low-cost approaches to semantic segmentation in radiation

therapy planning for prostate cancer using deep learning with non-contrast planning ct
images. Physica Medica, v. 78, p. 93 Ű 100, 2020. ISSN 1120-1797. Disponível em:
<http://www.sciencedirect.com/science/article/pii/S1120179720302209>.

OSPINA, R.; NOGUCHI, N. Simultaneous mapping and crop row detection by

fusing data from wide angle and telephoto images. Computers and Electronics
in Agriculture, v. 162, p. 602 Ű 612, 2019. ISSN 0168-1699. Disponível em:
<http://www.sciencedirect.com/science/article/pii/S0168169919301565>.

OTSU, N. A threshold selection method from gray-level histograms. IEEE Transactions

on Systems, Man, and Cybernetics, v. 9, n. 1, p. 62Ű66, 1979. Disponível em:
<https://doi.org/10.1109/TSMC.1979.4310076>.

OUTAY, F.; MENGASH, H. A.; ADNAN, M. Applications of unmanned aerial

vehicle (uav) in road safety, traffic and highway infrastructure management:
Recent advances and challenges. Transportation Research Part A: Policy
and Practice, v. 141, p. 116 Ű 129, 2020. ISSN 0965-8564. Disponível em:
<http://www.sciencedirect.com/science/article/pii/S096585642030728X>.

PANG, Y. et al. Improved crop row detection with deep neural network for
early-season maize stand count in uav imagery. Computers and Electronics
in Agriculture, v. 178, p. 105766, 2020. ISSN 0168-1699. Disponível em:
<http://www.sciencedirect.com/science/article/pii/S0168169920311376>.

RADOGLOU-GRAMMATIKIS, P. et al. A compilation of uav applications for precision

agriculture. Computer Networks, v. 172, p. 107148, 2020. ISSN 1389-1286. Disponível
em: <http://www.sciencedirect.com/science/article/pii/S138912862030116X>.

REN, G. et al. Agricultural robotics research applicable to poultry production:

A review. Comput. Electron. Agric, v. 169, p. 105216, 2020. Disponível em:
<https://doi.org/10.1016/j.compag.2020.105216>.

RONNEBERGER, O.; FISCHER, P.; BROX, T. U-net: Convolutional networks for

biomedical image segmentation. In: SPRINGER. International Conference on
Medical image computing and computer-assisted intervention. 2015. p. 234Ű241.
Disponível em: <https://doi.org/10.1007/978-3-319-24574-4_28>.

SANG, H.; ZHOU, Q.; ZHAO, Y. Pcanet: Pyramid convolutional attention network
for semantic segmentation. Image and Vision Computing, v. 103, p. 103997, 2020.
ISSN 0262-8856. Disponível em: <http://www.sciencedirect.com/science/article/pii/
S0262885620301293>.
84 Bibliography

SENIOR, A. W.; LEI, X. Fine context, low-rank, softplus deep neural networks for
mobile speech recognition. In: ICASSP. IEEE, 2014. p. 7644Ű7648. Disponível em:
<https://doi.org/10.1109/ICASSP.2014.6855087>.

SEZGIN, M.; SANKUR, B. Survey over image thresholding techniques and quantitative
performance evaluation. Journal of Electronic Imaging, v. 13, p. 13 Ű 13 Ű 20, 2004.
Disponível em: <https://doi.org/10.1117/1.1631315>.

SILVA, G. R. et al. DeĄnition of management zones through image processing for

precision agriculture. In: IEEE. 2017 Workshop of Computer Vision (WVC).
[S.l.], 2017. p. 150Ű154.

SIMONYAN, K.; ZISSERMAN, A. Very Deep Convolutional Networks for Large-

Scale Image Recognition. 2014. Disponível em: <http://arxiv.org/abs/1409.1556>.

SOARES, G. A.; ABDALA, D. D.; ESCARPINATI, M. Plantation rows identiĄcation by

means of image tiling and hough transform. In: VISIGRAPP (4: VISAPP). [s.n.],
2018. p. 453Ű459. Disponível em: <https://doi.org/10.5220/0006657704530459>.

SØGAARD, H. T.; OLSEN, H. J. Determination of crop rows by image analysis without

segmentation. Computers and electronics in agriculture, Elsevier, v. 38, n. 2, p.
141Ű158, 2003. Disponível em: <https://doi.org/10.1016/S0168-1699(02)00140-0>.

SOUZA, I. R.; ESCARPINATI, M. C.; ABDALA, D. D. A curve completion

algorithm for agricultural planning. In: ACM. Proceedings of the 33rd Annual
ACM Symposium on Applied Computing. 2018. p. 284Ű291. Disponível em:
<https://doi.org/10.1145/3167132.3167158>.

STAN, T.; THOMPSON, Z. T.; VOORHEES, P. W. Optimizing convolutional neural

networks to perform semantic segmentation on large materials imaging datasets: X-ray
tomography and serial sectioning. Materials Characterization, Elsevier, v. 160, p.
110119, 2020. Disponível em: <https://doi.org/10.1016/j.matchar.2020.110119>.

SYLVESTER, G. E-agriculture in action: Drones for agriculture. Food and

Agriculture Organization of the United Nations and International, 2018. ISBN 978-92-5-
130246-0. Disponível em: <http://www.fao.org/documents/card/en/c/I8494EN/>.

UCHIMIYA, M.; SPAUNHORST, D. J. InĆuence of summer fallow on aromatic

secondary products in sugarcane (saccharum spp. hybrids). Journal of Agriculture
and Food Research, v. 2, p. 100064, 2020. ISSN 2666-1543. Disponível em:
<http://www.sciencedirect.com/science/article/pii/S2666154320300454>.

VARUN, R. et al. Face recognition using hough transform based feature extraction.
Procedia Computer Science, v. 46, p. 1491 Ű 1500, 2015. ISSN 1877-0509. Proceedings
of the International Conference on Information and Communication Technologies, ICICT
2014, 3-5 December 2014 at Bolgatty Palace & Island Resort, Kochi, India. Disponível
em: <http://www.sciencedirect.com/science/article/pii/S1877050915001337>.

VERMA, O. P.; PARIHAR, A. S. An optimal fuzzy system for edge detection in color
images using bacterial foraging algorithm. IEEE Trans. Fuzzy Syst, v. 25, n. 1, p.
114Ű127, 2017. Disponível em: <https://doi.org/10.1109/TFUZZ.2016.2551289>.
Bibliography 85

VIDOVIC, I.; CUPEC, R.; HOCENSKI, Z. Crop row detection by global

energy minimization. Pattern Recognit, v. 55, p. 68Ű86, 2016. Disponível em:
<https://doi.org/10.1016/j.patcog.2016.01.013>.

XU, B. et al. Empirical evaluation of rectiĄed activations in convolutional network.

CoRR, abs/1505.00853, 2015. Disponível em: <http://arxiv.org/abs/1505.00853>.

YAKUBOVSKIY, P. Segmentation Models. [S.l.]: GitHub, 2019. <https:

//github.com/qubvel/segmentation_models>.

ZADA, B.; ULLAH, R. Pashto isolated digits recognition using deep convolutional
neural network. Heliyon, v. 6, n. 2, p. e03372, 2020. ISSN 2405-8440. Disponível em:
<http://www.sciencedirect.com/science/article/pii/S2405844020302176>.

ZHAO, H. et al. Pyramid scene parsing network. In: Proceedings of the IEEE
conference on computer vision and pattern recognition. [s.n.], 2017. p. 2881Ű2890.
Disponível em: <https://doi.org/10.1109/CVPR.2017.660>.