See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/337400899

Deep Learning Applications in Agriculture: A Short Review

Chapter · January 2020

DOI: 10.1007/978-3-030-35990-4_12

CITATIONS READS
0 414

4 authors:

Luís Santos Filipe Neves Dos Santos

Institute for Systems and Computer Engineering, Technology and Science (INESC TEC) INESC TEC - Technology and Science
12 PUBLICATIONS   13 CITATIONS    45 PUBLICATIONS   131 CITATIONS   

SEE PROFILE SEE PROFILE

Paulo Moura Oliveira Pranjali Shinde

Universidade de Trás-os-Montes e Alto Douro Institute for Systems and Computer Engineering, Technology and Science (INESC TEC)
135 PUBLICATIONS   996 CITATIONS    8 PUBLICATIONS   1 CITATION   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Robot-to-robot interaction View project

AGRISENSUS: Integrated Monitoring and Control of Water, Nutrients and Plant Protection Products Towards a Sustainable Agricultural Sector (WATER JPI ERANET H2020
Cofund) View project

All content following this page was uploaded by Luís Santos on 09 January 2020.

The user has requested enhancement of the downloaded file.

 Deep Learning applications in agriculture: a short review

Luı́s Santos1,2 , Filipe N. Santos1 , Paulo Moura Oliveira1,2 , and Pranjali Shinde1
1
INESC TEC - INESC Technology and Science
email{luis.c.santos, fbnsantos, pranjali.shinde}@inesctec.pt,
2
UTAD - University of Trás-os-Montes e Alto Douro, Vila Real, Portugal
oliveira@utad.pt

Abstract. Deep learning (DL) incorporates a modern technique for image processing and
big data analysis with large potential. Deep learning is a recent tool in the agricultural
domain, being already successfully applied to other domains. This article performs a survey
of different deep learning techniques applied to various agricultural problems, such as disease
detection/identification, fruit/plants classification and fruit counting among other domains.
The paper analyses the specific employed models, the source of the data, the performance
of each study, the employed hardware and the possibility of real-time application to study
eventual integration with autonomous robotic platforms. The conclusions indicate that deep
learning provides high accuracy results, surpassing, with occasional exceptions, alternative
traditional image processing techniques in terms of accuracy.

Keywords: deep learning · agriculture · image processing · survey,

1 Introduction
Agriculture is a relevant activity for the global economy. Along the years, this sector suffered several
changes to fulfil the world’s growing population, which has doubled in the last 50 years [20]. There
are several predictions for the continuous growth of the world population, is expected to have 9
billion people on earth in the year 2050, a 60% increase. Besides, the predictions indicate an increase
of people living in urban areas and a decrease in the ratio between working people and retired
people [34][31]. This means that the world’s agriculture productivity has to increase sustainably,
and more independent of human work. Technology was introduced to agriculture more than one
century ago, being the first tractor presented in 1913. Nowadays, mechanical technology has an
incredible evolution with a huge amount of commercial technology available [43]. This increased
the productivity resulting in minimal human labour. However, it may not be enough to sustain the
world’s demand along the future years. To improve the production efficiency, several studies were
performed since the 1990s, originating the concept of ”precision agriculture”, a farm management
notion based on the observation, measurement and actuation to the variability in the crops, with the
goal of optimizing the returns while preserving resources [30]. More recently, existing technologies
common in the industry have been applied to agriculture, such as remote sensing [1], Internet of
Things (IoT) [33], and robotics platforms [39] [38] leading to the concept of ”smart farming” [48].
Smart farming is important to face the challenges of agricultural production in terms of productivity,
environmental impact, food security and sustainability. To address these challenges it is necessary
to analyze and understand the agricultural ecosystems which imply constant monitoring of different
variables. This creates huge amounts of data that needs to be stored and processed in real time for
some operations [18]. This data can be constituted by images, which can be processed with different
2 Luı́s Santos , Filipe N. Santos , Paulo Moura Oliveira, and Pranjali Shinde

image analysis techniques for the identification of plants, diseases, etc., in different agricultural
contexts. Some of the image processing techniques are based on machine learning processes such as
K-means, support vector machines (SVM) and artificial neural networks (ANN) [41]. Deep learning
(DL) [42], is a modern approach that has been successfully employed in various domains. DL is
inserted in machine learning techniques, being similar to ANN, with better learning capability, and
thus has a higher classification accuracy [17]. There are some approaches with specific hardware
like Field-programmable Gate Array (FPGA)[49] or Graphics Processing Unit (GPU) [5] used in
order to accelerate the processing time of complex DL models. Several DL techniques have been
applied to different agricultural methods with a rise of popularity in recent years. Kamilaris et
al.[17][19] present a review work of the DL applications in agriculture. Other wider review works
also include some DL techniques [41][25]. However, these works do not check parameters like time
or hardware restrictions imposed by the complexity of deep learning models. With the exponential
growth of this area, it is possible to find in literature a big number of new research works applying
DL to agriculture. So, this paper focuses on the review of recent applications of DL techniques to
several agricultural domains considering the employed hardware to run the application. To the best
of our knowledge, the works mentioned in this paper are not covered by others review documents.
Section 2 of this article presents the methodology used for this review work. Section 3 presents a
brief explanation of the Deep Learning concept. In section 4 all the related work with deep learning
applied to agriculture is presented and analyzed. Section 5 presents the conclusions of the work.

2 Methodology

Two main steps are on the basis of this review: search for recent related work with no more than
four years; and review and analysis of the collected work. The collection of the related work was
performed between April and May 2019 mainly resorting to the scientific search engine Google
Scholar. The search was open to any deep learning method and 29 papers were selected from 14
agricultural areas. The analysis of the related work intends to answer the following issues: i) problem
identification; ii) type of data used; iii) deep learning architecture; iv) deep learning model; v) overall
accuracy; vi) comparison to alternative machine learning techniques as well as similar works existing
in the literature; vii) employed hardware and real-time application. Most of the works present the
Top-1 and Top-5 accuracy, and we present only the Top-1 accuracy. Top-1 is the conventional
accuracy: the model answer must be exactly the expected answer. Top-5 accuracy means that any
of the model 5 highest probability answers must match the expected answer. Some works use F1
score as a quality metrics, which is adequate for binary classification problems. Usually, the authors
give a detailed explanation about their quality metrics, for example, Smith et al [44] present an
explanation of the metric F1 score.

3 Deep Learning

DL is part of machine learning methods based on artificial neural networks (ANN). Deep learning
has varied applications from natural language to image processing [22]. Different deep learning ar-
chitectures such as Deep Neural Networks (DNN), Deep Belief Networks (DBN), Recurrent Neural
Networks (RNN), recursive neural networks, Fully Convolutional Networks (FCN) and Convolu-
tional Neural Networks (CNN) have been successfully applied to diverse research areas, including
agriculture. DL extends the complexity of ANN and represents data hierarchically, through multiple
 Deep Learning applications in agriculture: a short review 3

layers of abstraction. For example, in image processing lower levels can identify edges, while higher
levels will identify items like objects or faces. [19] [6] There are several frameworks, architectures
and data-sets publicly available for researchers to build their own model. Frameworks like CAFFE,
TensorFlow, Theano, DL Matlab toolbox are some of the most popular tools to experiment with DL
and they are used by some of the works reviewed in this paper. AlexNet, VGG, CaffeNet, GoogleNet
and ResNet are popular DL models publicly available for research, with the advantage that most
of them are already pre-trained with open data-sets, which means that the network is ready for
successfully identifying several features. Image-Net3 or COCO4 data-set are common among these
architectures. DL main disadvantage might be the long training time and the necessity of power-
ful hardware suitable for parallel programming (GPU, FPGA), while classic methods like Support
Vector Machine (SVM) or Scale-invariant feature transform (SIFT) have simpler training processes.
However, testing time is faster in DL methods, and generally more accurate.[19]
The following sub-sections present a brief introduction to CNN, FCN, and RNN the principal
architectures found in this literature review.

3.1 Convolutional Neural Networks

CNN’s are a class of deep, feed-forward ANN which has been applied to solve computer vision
problems. This is the most common architecture used almost in every paper approached in this
review. In contrast to ANN, CNN is capable of learning complex problems relatively fast due to
the weight sharing and complex models, which allow parallelization. Fig. 1 shows an example of a
CNN model, where the various convolutions are performed at some layers of the network, creating
different representations of the dataset. The convolutional layers act as feature extractors from
input images whose dimensionality is reduced by the pooling layers. The fully connected layers act
as classifiers, exploiting high-level features to classify input images in its corresponding class.
CNN’s increase the probability of correct classifications, provided that adequate large data-
set is available. A common technique to increase the data-set and improve CNN accuracy is the
data augmentation. As CNN’s are invariant to translation, size or illumination, the dataset can be
enlarged by applying a series of random transformations to the original images, such as rotation
and translation.[17][7].

3.2 Fully Convolutional Networks

FCN is a CNN based architecture that uses down-sampling (convolution) followed by up-sampling
(deconvolution) to produce semantic mask as output [4]. Typically, the input image is downsized to
pass through several convolution layers, and the output is one predicted label for the input image.
FCN networks do not downsize the image, thus the output is not a single label, which allow to
up-sample the output and predict the pixel-wise of the class [27]. Fig. 2 shows an example of a
semantic segmentation with FCN.

3.3 Recurrent Neural Networks

RNN is a multi-temporal network model which presents connections between nodes with direct
graphs along a temporal sequence. This allows temporal dynamic behaviour. Unlike mono-temporal
3
Image-net data-set: http://image-net.org/
4
COCO data-set: http://cocodataset.org/
4 Luı́s Santos , Filipe N. Santos , Paulo Moura Oliveira, and Pranjali Shinde

Fig. 1: AlexNet: an example CNN architecture model [14]

Fig. 2: FCN semantic segmentation example [27]

models (CNN’s), RNN has access to information in its own internal memory, which means that it
can consult the previous observation for the classification of the current observation. This makes it
suitable for temporal sequence tasks as speech recognition [24][37].

4 Deep Learning applications in agriculture

DL applications in agriculture are spread over several areas, being that 14 areas were identified
in a total of 43 recent papers. For this review, only 29 papers got selected, however, the list of
the remaining papers is publicly available 5 . The most popular areas are the disease identification
(6 papers), plants recognition (4 papers) and land cover and weed identification (3 papers each).
CNN’s have been the most popular architecture used in 24 papers while the few alternatives are
based in FCN or RNN.
The oldest article referred in this review was published in 2016 (one article). There are 3 articles
published in 2017, 17 published in 2018 and 8 articles from 2019. Tables 1, 2 and 3 present a list of
the selected papers and short answers for the questions presented in section 2.
5
Review List: https://bit.ly/2ZtS8tA
 Deep Learning applications in agriculture: a short review 5

The authors from the several reviewed papers used different DL architectures such as CNN,
FCN, RNN, Single Shot MultiBox Detector (SSD) and Bidirectional Long Short-Term Memory
(BLSTM). However, the current analysis shows that CNN has been clearly the most popular DL
architecture in the past years. From the 29 papers approached in this review, only 7 resorted
to others architectures, and even in this small group, CNN is just excluded in 2 works [26][37].
Some authors resorted to the combination of CNN/FCN [45] or CNN/RNN [54], while others
make use of Region-Based Convolutional Neural Network (R-CNN) and/or Region Based Fully
Convolutional Neural Networks (R-FCN). As the name suggests, R-CNN bypasses the problem of
selecting a huge number of regions to classify, instead, only some regions are classified, improving the
classification time. Fast R-CNN and Faster R-CNN are improved versions of R-CNN built in order
to achieve real-time object identification, being that Faster R-CNN is used in 3 papers [12][4][15].
Following a similar logic, R-FCN improves speed classification by reducing the amount of work in
the classification process and it is used is 2 papers [12][15]. Most of the authors resorted to known
architectures models publicly available such as VGG, AlexNet, MatConvnet, DenseNet, Darknet
(Yolo) using known frameworks (FW). The authors developed their own DL model in about 12
articles, however, it is not always clear in the paper if the author built its own model from scratch,
or modified one of the existing models. Most of the authors created their own data-set in various
ways such as manually-captured images, images from large public data-sets like ImageNet or COCO
[45], Unmanned Ground Vehicles (UGV) [45] or Unmanned Aerial Vehicles (UAV) [36] equipped
with sensors to capture the desired data. Data augmentation was a common practice among the
majority of the authors to enlarge the dataset and improve the results of the neural network. The
majority of the papers compares their DL approach with alternative machine learning methods such
as: SVM, SIFT, Random Forest (RF), backpropagation (BP), Particle Swarm Optimization (PSO),
Multilayer Perceptron (MLP), AdaBoost ([40]), Principal Component Analysis (PCA) plus Kernel
SVM (kSVM), Random Forests Uncorrelated Plant (RFUP) and Random Forests Correlated Field
(RFCF). The overall accuracy of the reviewed papers is generally good and surpasses alternative
machine learning methods with occasional exceptions where the traditional methods have the same
or higher accuracy [40]. The hardware restriction imposed by the complexity of DL is clear in
this review, as only 4 papers worked just with regular Central Processing Units (CPU). Most of
the authors resort to high rated GPUs, and Lammie et al [22] used a set of GPU and FPGA to
accelerate their CNN model. Even with this power-full hardware, only 11 papers claim to be capable
of performing the test operation in real time. The rest of the authors do not provide information
about the testing time of their networks.
This analysis leads to the conclusion that DL techniques in agriculture are rising at an incredibly
high rate with lots of possible applications and combination with aerial our ground platforms such
as UAVs or UGVs, as DL can be used in the process of automatizing these machines, by collecting
and processing the data to give an order to the autonomous vehicle. However, in this review, the
authors only used unmanned platforms for collecting data. To fully automatize a robotic platform
using deep learning might have some restrictions due to the hardware demand. GPUs, the most
used hardware platform to process DL models, are subject to large power requirements[22] and
this could represent a limitation to a robotic platform that is typically electric and powered by
batteries. To run the necessary DL tools in a remote computer, and thus save the battery life of
unmanned platforms is not always feasible for remote harsh agricultural terrains. FPGAs could
represent a low power alternative but the research with these platforms in agriculture is not yet
popular with only one article [22] found for this review. As the review shows, DL surpasses most
of the traditional machine learning methods in terms of accuracy, and real-time application, which
6 Luı́s Santos , Filipe N. Santos , Paulo Moura Oliveira, and Pranjali Shinde

is crucial for autonomous vehicles is hard to accomplish without DL approaches, as the testing
time of other methods is longer. As example, Mendes et al [32] present a work for Autonomous
Ground Robot (AGV) localization in a steep slope vineyard. They use an SVM classifier to detect
natural features, which requires an accurate and real-time detection, something that this work is not
capable to provide. Even with its GPU implementation [2], the real-time detection is not assured.
With other DL approaches, as seen on the works covered by this review, it could be possible to
perform a fast and more accurate detection, solving this localization problem. To conclude, DL is in
a good path to collaborate with diverse agricultural areas, and the focus of research should proceed
to low power demanding hardware and integration with autonomous platforms.

5 Conclusion

The current paper presented a review of DL based research efforts applied to agricultural domains.
It examined the agricultural area and described the problem they focus on, listed technical details
such as DL architecture and model, described the data source, reported the overall accuracy of
each work compared to alternative methods, verified the employed hardware and possible real-time
application. The findings indigent that DL reached high accuracy in the majority of the reviewed
works, scoring higher precision than other traditional technique. The mains advantage to non DL
techniques is the relatively low time for the classification process, which allows an easier execution in
real time. The hardware restriction might be the main disadvantage, as they require mostly powerful
GPUs, which are subject to large power demands. The alternative lower power demand systems
(FPGAs) are still with low research work in the agricultural domain. For a proper integration in
real time with autonomous robotic platforms (AGVs) the research should focus on decreasing the
complexity and hardware demand for the existing effective DL techniques.
 Table 1: Different deep learning applications in agriculture

v) Overall vi) Comparison to

Agricultural i) Problem ii) Data iii) DL vii) Hardware /
iv) DL Model Accuracy other methods Ref
Area Description Used Architecture Real Time?
(Top1 accuracy) accuracy
Find suitable DL
5000 images dataset
architecture for VGG GPU
created by the Faster R-CNN, 86 %
real-time tomato ResNet N/A / [12]
authors in several SDD and R-FCN (R-FCN)
diseases and ResNeXt Yes
farms from Korea
pest recognition
GPU
Recognize cucumber 14208 images AlexNet: 94%
Developed ( Nvidia Quadro)
diseases using leaf dataset created CNN 93.4 % SVM: 81.9% [29]
by authors /
symptons images by the authors RF: 84.8%
N/A
6000 images dataset GPU
Plant Detection of ESCA SIFT
created by (Nvidia GTX)
Disease disease in Bordeaux CNN MobileNet 90.7% Encoded : [35]
the authors in two /
Identification Vineyards 87.9%
Bordeaux Vineyards Yes
Identification of
Open database GPU
plant/disease AlexNet, VGG,
with 87484 photos 99.48% (Nvidia GTX)
combination in 25 CNN GoogleNet N/A [11]
of healthy and (VGG) /
different species and Overfeat
infected leaves Yes
with 58 classes
500 images of

Deep Learning applications in agriculture: a short review

Identify 10 healthy and diseased BP: 92% N/A
Developed by
common rice rice leaves from CNN 95.48% SVM: 91% / [28]
authors
diseases in leaves public datasets PSO: 88% N/A
and books
500 images collected
Identification of Modified GoogleNet GPU
from public sources 98.9%
8 kinds of maize CNN and Cifar10 N/A / [52]
(Plant Village dataset (GoogleNet)
leaves diseases (CAFFE FW) N/A
and Google Websites)
DataSet collected GPU
Identification of 16 by governmental Developed by (Nvidia Quadro)
CNN 97.47% SVM : 89.94% [50]
plant species project over authors /
1200 agro-stations N/A
LeafSnap dataset, SVM accuracy: GPU
LeafNet LeafSnap: 86.3%
Identify until 184 FOLIAGE dataset LeafSnap: 79.66% (Nvidia Gtx)
CNN (Developed by FOLIAGE: 95.8% [3]
different species and FLAVIA FOLIAGE: 98.75% /
authors) FLAVIA: 97.9%
dataset FLAVIA: 98.69% N/A
Images of 1000 GPU
Species and trait Modified Alternative DL
Plant species on GBIF 82.4% (Nvidia Titan)
recognition from CNN ResNet model approach: 90.3% [51]
Recognition dataset (subset (96.3% Top5) /
herbarium scans (TensorFlow FW) (Top5 accuracy)
of 170 species) N/A
AlexNet, GPU
LifeClef 2015
Identification of 1000 GoogleNet (Nvidia Tesla)
dataset CNN 80% N/A [13]
species and VGGNet /
(90 000 images)

7
(CAFFE FW) N/A
 8
Table 2: Different deep learning applications in agriculture (continuation)

v) Overall
Agricultural i) Problem ii) Data iii) DL vi) Comparison to vii) Hardware /

Luı́s Santos , Filipe N. Santos , Paulo Moura Oliveira, and Pranjali Shinde
iv) DL Model Accuracy Ref
Area Description Used Architecture other methods Real Time?
(Top-1 accuracy)
MLP accuracy: 83.81% GPU
Multi-temporal
Identify 13 different RNN LSTM and 85.54% XGBoost: 84.17% (Nvidia Quadro)
LandSat enchaced [54]
crop types and CNN Conv1D (Conv1D) RF: 84.09% /
Vegetation Index
SVM: 83.09% N/A
CPU
Multi-temporal
Land Cover Crop mapping of 14 Keras (Intel Xeon)
Sentinel-1 SAR CNN 91% N/A [23]
Classification different species (TensorFlow FW) /
images
N/A
Other RNN: 83.4% N/A
Sentinel-2A
Identify 19 crop types RNN LSTM 84.4% CNN: 76.8% / [37]
observations
SVM: 40.9% N/A
400 crop images SVM: 98% GPU
Weed detection and
captured by the CaffeNet AdaBoost: 98.2 % (Nvidia TITAN)
classification in CNN 98% [40]
authors with (CAFFE FW) Random Forest: 96 /
soybean crops
UAV % N/A
Weed detection CPU
200 Hyperspectral
Weed and classification (intel i7)
images with 61 CNN MatConvnet 94.72 % HoG: 74.34 % [10]
classification by spectral /
bands
band analysis N/A
Accelerate a DL 18 000 weed GPU and FPGA
Other work
approach with FPGA images from VGG-16, 90.08% (intel DE1-SoC)
CNN approach: [22]
for weed classification DeepWeedX DenseNet-128-10, (DenseNet) /
ResNet: 95.7 %
with 8 classes dataset Yes
Pods photography GPU
Estimate number of Developed
Seed over ligthbox. (Nvidia GTX)
seeds into soybeans CNN by authors 82.7% SVM: 50.4 % [47]
classification Dataset created /
on pods** (Theano FW)
by authors N/A
Pictures of hundreds 4 CPU and 1
Classify and sort YOLO9000: 86%
of thousands of seeds CNN ResNet-18 99 % GPU [16]
6 kinds of seeds SSD: 94 %
(created by authors) /Yes (500 fps)
Detect mango fruits Darknet R-CNN: 95.3%
1300 images GPU
Fruit in trees canopies (MangoYolo - SSD: 98.3 %
acquired by authors CNN 98.3 % (Nvidia GTX) [21]
counting and estimate YOLO modified YOLOv3: 96.7%
at 5 mango orchards /Yes (14 fps)
fruit load by authors) YOLOv2: 95.9 %
Orange
Combine DL, tracking
count: 99 % N/A
Fruit and SfM for robust Developed
N/A FCN N/A / [26]
classification visible fruit counting on by authors
Apple N/A
orange and apple orchards
count: 97 %
3600 images acquired GPU
PCA + kSVM: 89.11%
Classification of by authors and Developed (Nvidia GTX)
CNN 94.94 % WE+BBO: 89.47% [53]
18 types of fruit downloaded from by authors /
FRFE + BPNN: 88.99%
public websites Yes
 Table 3: Different deep learning applications in agriculture (continuation)

v) Overall
Agricultural i) Problem ii) Data iii) DL vi) Comparison to vii) Hardware /
iv) DL Model Accuracy Ref
Area Description Used Architecture other methods Real Time?
(Top-1 accuracy)
Segment Soil/ CPU
0.57
/root in X-ray (intel Xeon)
Soil/Root Imagenet Dataset CNN N/A (Specific N/A [8]
tomography and GPU(AMD)
segmentation quality metric)
with CNN+SVM / N/A
50 annotated GPU
Detect and
Chicory root (Nvidia Titan)
quantify CNN U-Net 99.7 % FrangiNet: 99.6% [44]
images created /
chicory roots
by authors N/A
Detect and count 13520 images CPU (Intel i7)
Cattle Developed
cattle in UAV captured during CNN 95.5 % N/A / [36]
detection by authors

Deep Learning applications in agriculture: a short review

images UAV flight 3.2s/frame
SVM: 0.038 1 CPU (intel i7)
Simulated large
Inferring moisture CNNUP RFUP: 0.060 and
Precision datasets of 1200 0.034
conditions CNN and RFCF: 0.094 3 GPU (Nvidia [46]
Irrigation aerial images (mean error)
from images CNNCF 2 layer NN: 0.086 Titan)
of Vineyards
(mean error) / N/A
Detect humans, Darknet
70.81 %
obstacles and (improved
(accuracy of N/A
Obstacle traversable obstacle ImageNet and CNN and YOLO),
fusion of N/A / [45]
Detection agricultural fields COCO dataset FCN VGG and
different Yes
for safe machinery DeepAnomal
approaches)
operation** (AlexNet)
10% error for
400 Stereo camera StalkNet GPU
Measure Stalk count and Faster RCNN count;
Images collected (created by N/A (Nvidia GTX)/ [4]
Features stalk width of crops and FCN 2.76 mm error
with robotic platform authors) 2 fps
for measure
Yeld prognosis 50 GB of Images Faster R-CNN, 99.6% N/A
Crop yield
in vineyards by from 2 vineyards at R-FCN, TensorFlow FW (F.R-CNN for N/A / [15]
estimation
object counting different crop stages SSD object detection) Yes
Automatic Translation of GPU
labeling of phytosanitary CNN and Developed 88.3% (Nvidia Quadro)
N/A N/A [9]
agricultural regulations into BLSTM by authors (F1 score) /
regulations formal rules N/A

9
10 Luı́s Santos , Filipe N. Santos , Paulo Moura Oliveira, and Pranjali Shinde

Acknowledgements
This work is co-financed by the European Regional Development Fund (ERDF) through the Interreg
V-A Espanha-Portugal Programme (POCTEP) 2014-2020 within project 0095 BIOTECFOR 1 P.
This work also was co-financed by the ERDF European Regional Development Fund through the
Operational Programme for Competitiveness and Internationalisation - COMPETE 2020 under
the PORTUGAL 2020 Partnership Agreement, and through the Portuguese National Innovation
Agency (ANI) as a part of project ROMOVI: POCI-01-0247-FEDER-017945 The opinions included
in this paper shall be the sole responsibility of their authors. The European Commission and the
Authorities of the Programme aren’t responsible for the use of information contained therein.

References
1. Atzberger, C.: Advances in remote sensing of agriculture: Context description, existing operational
monitoring systems and major information needs. Remote sensing 5(2), 949–981 (2013)
2. Azevedo, F., Shinde, P., Santos, L., Mendes, J., Santos, F.N., Mendonça, H.: Parallelization of a vine
trunk detection algorithm for a real time robot localization system. In: 2019 IEEE International Con-
ference on Autonomous Robot Systems and Competitions (ICARSC). pp. 1–6. IEEE (2019)
3. Barré, P., Stöver, B.C., Müller, K.F., Steinhage, V.: Leafnet: A computer vision system for automatic
plant species identification. Ecological Informatics 40, 50–56 (2017)
4. Baweja, H.S., Parhar, T., Mirbod, O., Nuske, S.: Stalknet: A deep learning pipeline for high-throughput
measurement of plant stalk count and stalk width. In: Field and Service Robotics. pp. 271–284. Springer
(2018)
5. Cui, H., Zhang, H., Ganger, G.R., Gibbons, P.B., Xing, E.P.: Geeps: Scalable deep learning on dis-
tributed gpus with a gpu-specialized parameter server. In: Proceedings of the Eleventh European Con-
ference on Computer Systems. p. 4. ACM (2016)
6. Deng, L., Yu, D., et al.: Deep learning: methods and applications. Foundations and Trends R in Signal
Processing 7(3–4), 197–387 (2014)
7. Ding, J., Chen, B., Liu, H., Huang, M.: Convolutional neural network with data augmentation for sar
target recognition. IEEE Geoscience and remote sensing letters 13(3), 364–368 (2016)
8. Douarre, C., Schielein, R., Frindel, C., Gerth, S., Rousseau, D.: Transfer learning from synthetic data
applied to soil–root segmentation in x-ray tomography images. Journal of Imaging 4(5), 65 (2018)
9. Espejo-Garcia, B., Lopez-Pellicer, F.J., Lacasta, J., Moreno, R.P., Zarazaga-Soria, F.J.: End-to-end
sequence labeling via deep learning for automatic extraction of agricultural regulations. Computers
and Electronics in Agriculture 162, 106–111 (2019)
10. Farooq, A., Hu, J., Jia, X.: Analysis of spectral bands and spatial resolutions for weed classification
via deep convolutional neural network. IEEE Geoscience and Remote Sensing Letters 16(2), 183–187
(2018)
11. Ferentinos, K.P.: Deep learning models for plant disease detection and diagnosis. Computers and Elec-
tronics in Agriculture 145, 311–318 (2018)
12. Fuentes, A., Yoon, S., Kim, S., Park, D.: A robust deep-learning-based detector for real-time tomato
plant diseases and pests recognition. Sensors 17(9), 2022 (2017)
13. Ghazi, M.M., Yanikoglu, B., Aptoula, E.: Plant identification using deep neural networks via optimiza-
tion of transfer learning parameters. Neurocomputing 235, 228–235 (2017)
14. Han, X., Zhong, Y., Cao, L., Zhang, L.: Pre-trained alexnet architecture with pyramid pooling and
supervision for high spatial resolution remote sensing image scene classification. Remote Sensing 9(8),
848 (2017)
15. Heinrich, K., Roth, A., Breithaupt, L., Möller, B., Maresch, J.: Yield prognosis for the agrarian man-
agement of vineyards using deep learning for object counting (2019)
 Deep Learning applications in agriculture: a short review 11

16. Heo, Y.J., Kim, S.J., Kim, D., Lee, K., Chung, W.K.: Super-high-purity seed sorter using low-latency
image-recognition based on deep learning. IEEE Robotics and Automation Letters 3(4), 3035–3042
(2018)
17. Kamilaris, A., Prenafeta-Boldú, F.: A review of the use of convolutional neural networks in agriculture.
The Journal of Agricultural Science 156(3), 312–322 (2018)
18. Kamilaris, A., Kartakoullis, A., Prenafeta-Boldú, F.X.: A review on the practice of big data analysis
in agriculture. Computers and Electronics in Agriculture 143, 23–37 (2017)
19. Kamilaris, A., Prenafeta-Boldú, F.X.: Deep learning in agriculture: A survey. Computers and electronics
in agriculture 147, 70–90 (2018)
20. Kitzes, J., Wackernagel, M., Loh, J., Peller, A., Goldfinger, S., Cheng, D., Tea, K.: Shrink and share:
humanity’s present and future ecological footprint. Philosophical Transactions of the Royal Society B:
Biological Sciences 363(1491), 467–475 (2007)
21. Koirala, A., Walsh, K., Wang, Z., McCarthy, C.: Deep learning for real-time fruit detection and orchard
fruit load estimation: Benchmarking of mangoyolo. Precision Agriculture pp. 1–29 (2019)
22. Lammie, C., Olsen, A., Carrick, T., Azghadi, M.R.: Low-power and high-speed deep fpga inference
engines for weed classification at the edge. IEEE Access 7, 51171–51184 (2019)
23. Lavreniuk, M., Kussul, N., Novikov, A.: Deep learning crop classification approach based on coding
input satellite data into the unified hyperspace. In: 2018 IEEE 38th International Conference on Elec-
tronics and Nanotechnology (ELNANO). pp. 239–244. IEEE (2018)
24. Li, X., Wu, X.: Constructing long short-term memory based deep recurrent neural networks for large
vocabulary speech recognition. In: 2015 IEEE International Conference on Acoustics, Speech and Signal
Processing (ICASSP). pp. 4520–4524. IEEE (2015)
25. Liakos, K., Busato, P., Moshou, D., Pearson, S., Bochtis, D.: Machine learning in agriculture: A review.
Sensors 18(8), 2674 (2018)
26. Liu, X., Chen, S.W., Aditya, S., Sivakumar, N., Dcunha, S., Qu, C., Taylor, C.J., Das, J., Kumar,
V.: Robust fruit counting: Combining deep learning, tracking, and structure from motion. In: 2018
IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1045–1052. IEEE
(2018)
27. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Pro-
ceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431–3440 (2015)
28. Lu, Y., Yi, S., Zeng, N., Liu, Y., Zhang, Y.: Identification of rice diseases using deep convolutional
neural networks. Neurocomputing 267, 378–384 (2017)
29. Ma, J., Du, K., Zheng, F., Zhang, L., Gong, Z., Sun, Z.: A recognition method for cucumber diseases
using leaf symptom images based on deep convolutional neural network. Computers and electronics in
agriculture 154, 18–24 (2018)
30. McBratney, A., Whelan, B., Ancev, T., Bouma, J.: Future directions of precision agriculture. Precision
agriculture 6(1), 7–23 (2005)
31. McNicoll, G.: World population ageing 1950-2050. Population and Development Review 28(4), 814–816
(2002)
32. Mendes, J., Dos Santos, F.N., Ferraz, N., Couto, P., Morais, R.: Vine trunk detector for a reliable robot
localization system. In: 2016 International Conference on Autonomous Robot Systems and Competitions
(ICARSC). pp. 1–6. IEEE (2016)
33. Patil, K., Kale, N.: A model for smart agriculture using iot. In: 2016 International Conference on Global
Trends in Signal Processing, Information Computing and Communication (ICGTSPICC). pp. 543–545.
IEEE (2016)
34. Perry, M.: Science and innovation strategic policy plans for the 2020s (eu, au, uk): Will they prepare
us for the world in 2050? Applied Economics and Finance 2(3), 76–84 (2015)
35. Rançon, F., Bombrun, L., Keresztes, B., Germain, C.: Comparison of sift encoded and deep learning
features for the classification and detection of esca disease in bordeaux vineyards. Remote Sensing
11(1), 1 (2019)
 12 Luı́s Santos , Filipe N. Santos , Paulo Moura Oliveira, and Pranjali Shinde

36. Rivas, A., Chamoso, P., González-Briones, A., Corchado, J.: Detection of cattle using drones and
convolutional neural networks. Sensors 18(7), 2048 (2018)
37. Rußwurm, M., Körner, M.: Multi-temporal land cover classification with long short-term memory neural
networks. The International Archives of Photogrammetry, Remote Sensing and Spatial Information
Sciences 42, 551 (2017)
38. Santos, L., Santos, F.N., Magalhães, S., Costa, P., Reis, R.: Path planning approach with the extraction
of topological maps from occupancy grid maps in steep slope vineyards. In: 2019 IEEE International
Conference on Autonomous Robot Systems and Competitions (ICARSC). pp. 1–7. IEEE (2019)
39. Santos, L., Ferraz, N., dos Santos, F.N., Mendes, J., Morais, R., Costa, P., Reis, R.: Path planning aware
of soil compaction for steep slope vineyards. In: 2018 IEEE International Conference on Autonomous
Robot Systems and Competitions (ICARSC). pp. 250–255. IEEE (2018)
40. dos Santos Ferreira, A., Freitas, D.M., da Silva, G.G., Pistori, H., Folhes, M.T.: Weed detection in
soybean crops using convnets. Computers and Electronics in Agriculture 143, 314–324 (2017)
41. Saxena, L., Armstrong, L.: A survey of image processing techniques for agriculture (2014)
42. Schmidhuber, J.: Deep learning in neural networks: An overview. Neural networks 61, 85–117 (2015)
43. Schmitz, A., Moss, C.B.: Mechanized agriculture: machine adoption, farm size, and labor displacement
(2015)
44. Smith, A.G., Petersen, J., Selvan, R., Rasmussen, C.R.: Segmentation of roots in soil with u-net. arXiv
preprint arXiv:1902.11050 (2019)
45. Tseng, D., Wang, D., Chen, C., Miller, L., Song, W., Viers, J., Vougioukas, S., Carpin, S., Ojea,
J.A., Goldberg, K.: Towards automating precision irrigation: Deep learning to infer local soil moisture
conditions from synthetic aerial agricultural images. In: 2018 IEEE 14th International Conference on
Automation Science and Engineering (CASE). pp. 284–291. IEEE (2018)
46. Tseng, D., Wang, D., Chen, C., Miller, L., Song, W., Viers, J., Vougioukas, S., Carpin, S., Ojea,
J.A., Goldberg, K.: Towards automating precision irrigation: Deep learning to infer local soil moisture
conditions from synthetic aerial agricultural images. In: 2018 IEEE 14th International Conference on
Automation Science and Engineering (CASE). pp. 284–291. IEEE (2018)
47. Uzal, L.C., Grinblat, G.L., Namı́as, R., Larese, M.G., Bianchi, J., Morandi, E., Granitto, P.M.: Seed-
per-pod estimation for plant breeding using deep learning. Computers and electronics in agriculture
150, 196–204 (2018)
48. Walter, A., Finger, R., Huber, R., Buchmann, N.: Opinion: Smart farming is key to developing sustain-
able agriculture. Proceedings of the National Academy of Sciences 114(24), 6148–6150 (2017)
49. Wang, C., Gong, L., Yu, Q., Li, X., Xie, Y., Zhou, X.: Dlau: A scalable deep learning accelerator unit on
fpga. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems 36(3), 513–517
(2016)
50. Yalcin, H., Razavi, S.: Plant classification using convolutional neural networks. In: 2016 Fifth Interna-
tional Conference on Agro-Geoinformatics (Agro-Geoinformatics). pp. 1–5. IEEE (2016)
51. Younis, S., Weiland, C., Hoehndorf, R., Dressler, S., Hickler, T., Seeger, B., Schmidt, M.: Taxon and
trait recognition from digitized herbarium specimens using deep convolutional neural networks. Botany
Letters 165(3-4), 377–383 (2018)
52. Zhang, X., Qiao, Y., Meng, F., Fan, C., Zhang, M.: Identification of maize leaf diseases using improved
deep convolutional neural networks. IEEE Access 6, 30370–30377 (2018)
53. Zhang, Y.D., Dong, Z., Chen, X., Jia, W., Du, S., Muhammad, K., Wang, S.H.: Image based fruit cat-
egory classification by 13-layer deep convolutional neural network and data augmentation. Multimedia
Tools and Applications 78(3), 3613–3632 (2019)
54. Zhong, L., Hu, L., Zhou, H.: Deep learning based multi-temporal crop classification. Remote sensing of
environment 221, 430–443 (2019)

View publication stats