A smart agricultural system to detect Rust and Powdery Mildew

disease using deep learning

Oussama Bakkali Yedri *, Mohamed benahmed, Mohamed Bouhorma and Lotfi Elaachak
Computer Science, Systems and Telecommunication Laboratory (LIST),
Faculty of Sciences and Technologies,
University Abdelmalek Essaadi, Tangier, Morocco
bakkali.yedri.oussama@gmail.com, m.benahmed@gmail.com, bouhorma@gmail.com,
lotfi1002@gmail.com

Abstract. This paper proposes a new smart agriculture architecture using the Internet of
Things (IoT) and Deep Learning algorithms (DL). Starting from the field, controlling soil and
weather conditions, based on each plant needs, in parallel to that, treating field images,
plant by plant, in order to detect diseases and heal them at early stages using another way
for plant species classification from digital leaf image using Convolutional Neural Network
algorithm (CNN). This algorithm helps us to match different plant features in order to classify
them based on pre-learned and pre-processed features.
Keywords: Smart Agriculture, IoT, Machine Learning, Deep Learning, classification, CNN.

I. Introduction
After the plant breeding and genetics revolutions, we are opening the doors to a new era
of agriculture. Combining agriculture with a set of tools, concepts and practices that
promote the economic growth of farms. Moving from the tractor to the Internet of Things,
big data, artificial intelligence and other innovative practices to become a "smart"
ecosystem: intelligent, collaborative and sustainable.
As in other economic sectors, the greatest innovation and the main tool of the modern
farmer will be the data, collected and analysed to make the right decisions in order to have a
reduced soil erosion, control of soil temperature, better soil’s water retention, reduce
pesticide intake, preserve soil quality, increase soil organic matter accumulation (etc.)
To do so, in our research, we will develop an important problem of agriculture, which is
plant diseases, using Deep Learning algorithms in order to classify different plants species,
and also detect if this plant is affected by the disease or not.
In this paper, first, we will develop a CNN algorithm that helps us to classify tree plants:
Mint, verbena and cannabis from their leaves, after that, we will go for the recognition of
Powdery Mildew and Puccinia on a mint plant, and finally, we will conclude this paper with
future prospects.
 II. Theoretical Background

a. Smart Agriculture
Smart Agriculture is an important dimension of the smart cities concept, in
consideration of its impact on the economical side of any country. considering the value of
the agricultural economical source of income of the country and improving the existing
sources rather than spending money on new investments.
It is a combination of practices that are able to adapt to climate change and control
the necessary supplements while having the potential to increase food production. and high-
quality management of natural resources, such as land, water, soil and genetic resources to
enable more productive and resilient agriculture.

b. The Problem of Rust and Powdery Mildew disease:

Rust is a fungal disease. In fact, many different fungi can be responsible for what is
called "rust": behind this term are hiding many diseases whose symptoms and damage are
similar. The fungus - and therefore the disease - develops between spring and autumn. It is
carried by the wind, or transmitted by contact (contamination of one plant by another,
gardening tool badly cleaned ...). He spends the winter in the debris of plants, where he
waits for the return of spring to resume its development and contaminate new hosts.
Powdery Mildew is a fungus of the family Erysiphe quite common in some plants, is a
fungal disease familiar to gardeners. This generic name corresponds in fact to a set of similar
diseases, but due to different parasitic fungi, and more or less specific to each affected plant
species. It is characterized by the appearance of a kind of mould or a light white down.

c. Convolutional Neural Network (CNN):

A convolutional neuronal network (with the acronyms CNNs or ConvNets) is a
concrete case of Deep Learning neural networks, which were already used at the end of the
90s but which in recent years have become enormously popular[1] when achieving very
impressive results in the recognition of image, deeply impacting the area of computer vision.
CNN is a type of deep learning model for processing data that has a grid pattern, such
as images, which is inspired by the organization of animal visual cortex [2]
In recent years, deep learning has made promising achievements in the machine
learning field [3], [4], [5]. Among many deep learning algorithms, we find that Convolutional
neural network (CNN) make great success in image classification, Hu [6] proposed a CNN that
contains different layers, which are input, convolutional, max pooling, fully connected, and
an output layer ,for image classification. Where Makantasis[7] presented a deep learning-
based classification method that exploits a CNN to encode the image information and a
multilayer perceptron to conduct the classification task. Unlike Chen [8] who proposed a
regularized 3-D CNN-based feature extraction model to extract efficient spectral-spatial
features for hyperspectral image classification.

III. Related works

Deep learning is being used in multiple fields for image classification and recognition,
we find multiple authors talk about classification using multiple algorithms such as SVM,
CNN, RNN etc. by proving it efficiency for image recognition, most of its technics have been
replaced by deep learning for many domains[9], such as pathology, medical image
analysis[10], [11], [12] and agriculture[13], [14].
In our research, we find several works focused on smart agriculture in order to classify
leaf crops, using several Machine Learning algorithms, each of these algorithms uses its own
way to predict this classification.
In this paper, we present a comprehensive review of the application of ML in agriculture.
A number of relevant papers are presented that emphasise key and unique features of
popular ML models.
Typically, Machine Learning methodologies involve a learning process with the objective
to learn from “experience” (training data) to perform a task. Data in ML consists of a set of
examples. Usually, an individual example is described by a set of attributes, also known as
features or variables. A feature can be nominal (enumeration), binary (i.e., 0 or 1), ordinal
(e.g., A+ or B−), or numeric (integer, real number, etc.).
The performance of the ML model in a specific task is measured by a performance metric
that is improved with experience over time. To calculate the performance of ML models and
algorithms, various statistical and mathematical models are used. After the end of the
learning process, the trained model can be used to classify, predict, or cluster new examples
(testing data) using the experience obtained during the training process. Recently, many
authors used deep learning for crop leaf classification and disease detection.
First, we find the paper entitled Leaf Classification Based on GLCM Texture and SVM[15]
the authors used classification of leaves using Grey level Co-Occurrence Matrix GLCM, by
extracting leave’s features and Support vector machines SVM algorithm, the author
describes how Artificial Neural Network is used to identify plant by inputting leaf image,
using image processing techniques to extract leaf shape and features.
The paper has as a focus to:
1. Introduce an approach of plant classification based on the characterization of texture
properties using combined “classifier learning vector quantization” and “radial basis
function”, the main idea is to recognize a leaf from only parts of it and its texture
with no need for full leaf picture.
2. Propose a method that incorporates shape, vein, colour and texture features based
on Probabilistic neural network PNN
 None of the above has used a colour parameter having algorithm parameters doesn’t
accept colour.
However, the main idea of the project is to classify different leaves species based on
their texture using Grey Level Co-Occurrence Matrix (GLCM) for texture feature extraction,
and Support Vector Machines (SVM) for classification.
Moreover, the second study entitled Architecture-Based Classification of Plant Leaf
Images, The author used as a method three main steps to classify any leaf based on its shape
and boundaries[16] using The Otsu, polygon decimation simplification and centroid contour
distances algorithms:
1. Leaf image pre-processing:
By taking only the leaf shape and ignoring the colour by turning it to black and white
image
2. Architecture-based features extraction
3. Classification and species identification
The third study, entitled Multi-class K-support Vector Nearest Neighbour for Mango Leaf
Classification, the author[17] research is to classify mango varieties using K-Support vector
nearest neighbour algorithm (K-SVNN) that uses data reduction while preserving the
accuracy. while the K-SVNN works only for binary class, the author also proposed the
entropy to calculate the Significant Degree (SD) to solve the multi-class problem[17].
Entropy can measure the impurity of data class distribution so that the selection of the
SD can be conducted based on the high entropy. The data with same class distribution has
zero impurity, whereas data with uniform class distribution has the highest impurity.
Other studies used many other algorithms, such in this study CNN-RNN: a large-scale
hierarchical image classification framework[18] we find that the author talked about his own
data set before proposing the deep learning algorithm. Traditionally, input dataset for
classifiers uses small datasets, by encoding hand-crafted features [19].
Mosin Hasan and his co-authors[13] go for tomato leaf disease detection using Transfer
Learning in convolutional neural network, transfer learning, as the author says, is an
approach which reduces the number of parameters, using a neural network which has
already been trained. Unlike Jihen Amara [14] used convolutional neural network for banana
diseases classification, the author compared the results of two types of images, RBG and
grey scale image in order to identify which type of images gives us the best accuracy. While
other authors [20], [21] and [22] went for hyperspectral analysis with supervised
classification using decision tree random forest which develops multiple trees from samples
and combines the results via voting or using maximum and support vector machines for
multi class classification who has achieved great success in various applications and is
considered a stable and efficient algorithm for image classification.
The following Table 1table summarize algorithms used by each article discussed above.
Article Algorithm
GLCM SVM PNN K-SVNN CNN RNN decision tree
random forest
Leaf classification based on X X X
GLCM texture and SVM[15]
Multi-class K-support Vector X
Nearest Neighbour for Mango
Leaf Classification [17]
CNN-RNN: a large-scale X X
hierarchical image classification
framework [18]
Deep Learning Precision X
Farming: Tomato Leaf Disease
Detection by Transfer Learning
[13]
A Deep Learning-based X
Approach for Banana Leaf
Diseases Classification [14]
Heathland conservation status X X
mapping through integration of
hyperspectral mixture analysis
and decision tree classifiers [20]
Investigation of the random X X
forest framework for
classification of hyperspectral
data [21]
The random subspace method X X
for constructing decision forests
[22]
Deep Convolutional Neural X
Networks for Hyperspectral
Image Classification [6]
Table 2: algorithms used in each discussed article

Deep learning methods have proven to be successful for image classification. In recent
years, researchers have built various deep structures[7], [8], and have achieved quite
accurate predictions on small datasets [18].
The first contribution of this paper is a framework capable of generating hierarchical
labels, by integrating the powerful Convolutional Neural Networks (CNN) and Recurrent
Neural Networks (RNN). CNN is used to generate discriminative features, and RNN is used to
generate sequential labels.
In recent years, deep learning methods have attracted significant attention [25] and have
achieved revolutionary successes in various applications[23]. Two important structures for
 deep learning are CNN and RNN. CNN has proven to be successful in processing imagelike
data, while RNN is more appropriate in modelling sequential data.
From the studies, we show that ML and DL models have been applied in multiple
applications for crop management; mostly yield prediction and disease detection,
distribution of ML applications in livestock management (19%), water management (10%),
and soil management (10%). This trend in the distribution of the application reflects the
data-intense applications within the crop and high use of images. Data analysis, as a
scientific field, provides the basics for the development of numerous applications related to
this field because ML / DL predictions can be extracted without the need for third-party
resources. It is also evident from the analysis that most of the studies used multiple models.
More specifically, ANNs were used mostly for implementations in the crop, water, and soil
management, while SVMs were used mostly for livestock management.
By applying machine learning to crop image analysis, crop growth rate and crop disease
detection at early stages provide an ultimate crop production improvement, a real increase
of production levels and bio-products quality.

IV. Proposed architecture and solution:

Our project is to develop smart agriculture practices that are able to adapt to climate
change and control the necessary supplements while having the potential to increase food
production. Also, to ensure high-quality management of natural resources, such as land,
water, soil and genetic resources to enable more productive and resilient agriculture.
However, as in modern economic sectors, the greatest key of improvement is Data,
collected, filtered and analysed to make the right decisions in order to have a reduced soil
erosion, control of soil temperature, better water retention in soil, less pesticide intake,
preserve soil quality, increase soil organic matter accumulation and many other things.
Weather pressure, atmospheric pressure, Wind vane, pluviometry, air quality, leaf witness,
humidity Temperature, luminosity Soil moisture, soil temperature, Solar radiation,
Ultraviolet radiation and other sensors are useful to gather data, send it to the cloud servers,
perform treatments and analysis to help us understand field’s needs.
To meet the growing demand of the sector and democratize access to innovation,
our proposed architecture matches well with many other solutions proposed by
technological players.

Article Architect Layers

ure
Perceptio MAC Aggregation Decisi Application Security
n layer layer layer on layer layer
layer
Machine to M2M X X X X
machine
communication
[26]
Smart monitoring Cyber- X X X X
of potato crop [27] physical
systems
(CPS)

Future Internet X X X X
The Internet of
Things
Architecture [28]
Internet of things X X X X X
[29]
Multidimensional X X X X
model for smart
agriculture using
internet of things
[30]
Research and X X X X
application of fire
power cloud
platform [31]
Smart agriculture X X X X X
based on cloud
computing and IoT
[32]
Table 3: Layers used in reviewed articles

As described in Table 4:, multiple authors proposed many architectures for an IoT
application. Such as [26], [27], [28], [30] go for Physical layer, networking layer, decision
layer and application layer, with some changes at decision layer level, where they differ only
in the decision layer details. [27] defines decision layer as a, analysis layer, where he makes
all the needed data analysis at this stage, then transfers the knowledge to the application
layer to define actions to proceed. Where [30] add also information processing to take good
actions on the field. [26] says that the decision layer should contain three sub-layers in order
to have good results, which are analysing sub-layer, information processing sub-layer and
also decision making sub-layer, which is responsible of taking the right decisions, then send
the result to application layer to notify and take actions on the field.
Unlike [31] goes for Physical layer, device monitoring layer, networking layer and
application layer. Where [29] and [32] add the security layer in order to secure transferred
data from the physical layer to application one.
 IoT in Smart Agriculture:
Advanced electronic devices, computing languages and telecommunication
technologies allow IoT to be developed in many fields of application. Which made architects
think of multiple layers of architecture in order to have an efficient IoT system. The basic
model of IoT has a 3 layers architecture which is of Perception, Network and Application
Layers[33]. Where Perception, Network and Application are the main layers for [34], [35].
However, there are many other authors who proposed their own architecture, where [36]
goes for Sensors, Edge, Network and Application, and [8] goes for Business, Application,
Service, Object abstraction and Objects.

Data analysis:
IoT brings multiple objects such as Sensors and cameras, every connected object
brings us a massive amount of data that need both a storage and analysis system.
Cloud computing offers many services which can scale to storage and processing
requirement. However, there are many limitations[33], there are applications such as sensor
monitoring, control and analysis response that require low latency, therefore, the delay
caused by transferring data to the cloud and then back to the application can seriously
impact their performances. Therefore[33], proposes a Fog and edge computing, the main
idea is to bring the analysis layer to the field.
The key difference between the two architectures is exactly where that intelligence
and computing power is placed:

• Fog computing pushes intelligence down to the local area network level,
processing data in a fog node or IoT gateway
• Edge computing pushes the intelligence, processing power and communication
capabilities of an edge gateway or appliance directly into devices like
programmable automation controllers
To propose a proper architecture, there is not only one criterion to analyse, but we
must also, study the need of the field, and ask these questions and many others:

• Do we need local or cloud storage?

• Do we need Offline analysis?
• Do we need Real-Time feedback?
These questions will allow us to know which architecture we need to apply, [37] say
that we need real-time analytics where we can analyse large amounts of generated data. Do
we need a real-time in local or memory level, or on large computing platforms etc.

Smart Agriculture Architecture:

From all the above, we find that many authors elaborate on the idea of a smart
agriculture by developing its components, and the different layers on which each smart city
should stand. Some of the authors keep only 3 layers, while others say that an extended
architecture must be applied in order to develop a powerful architecture:

• Perception / physical / Sensing layer:

Used for acquisition process, Responsible for using sensors at ground level in
order to collect data and measurements. All of the collected data will be
georeferenced using GPS.
• MAC layer (device monitoring, energy, etc.):
Mac / device monitoring: Used to send data collected by the Device Layer to the
information processing system using wireless communication technology.
• Aggregation / Networking layer (access and transmit information)

• Decision Layer:
Based on machine learning algorithms, this layer performs information processing
and takes an automatic decision based on learned process.
• Application layer / SOA:
Which provides global management of the application based on processed
information, also provide solutions to incoming problems based on information
from decision layer.
• Security layer:
Trust and privacy are likely to be the major hurdles in IoT uptake. The main
purpose is to identify management, security, and access control standardization
[8].
• Business Layer:
That is responsible for the management of the overall IoT system.
In our smart agriculture project, we will focus on three principal cases, Smart
Resource Monitoring, Crop Leaf Analysis and climate conditioning.
Smart agriculture architecture has many goals as [38] and [27] define these goals as:

• Increase the agricultural management with major economic impact

• Protect the environment and to preserve the natural resources in a more efficient
manner
• Optimizing the use of water resources by careful monitoring the crop vegetation
indices regarding the hydric stress
• Obtain large and quality productions due to the integrated management system
• Optimizing the chemical inputs (fertilizers and pesticides) by knowing the right doses
and moments when to be used
• Increase the sustainability of agricultural systems
• Provide a low-cost solution for the farmers with limited financial possibilities that can
be easily implemented in practice
At this purpose our proposed architecture is based on many layers, starting from
device or acquisition layer for collecting data, then transferring all these data to the
middleware layer using network layer. The middleware layer is responsible of performing
information processing and taking automatic decisions. After that, we use the application
layer which contains global management of the application based on information processed
at early stages. Ending by the business layer which is responsible for the management of the
overall IoT systems.

Smart Resource management:

Smart resource management focuses on optimizing the use of resources, such as
water irrigation control and effective use of fertilizers, by ensuring large and quality
production.
With the SA framework, we will minimize farmers costs and increase their benefits,
doing more with less,

Figure 1: Smart Agriculture Architecture, Resource management.

In Figure 2 , we use real-time monitoring process, starting with collecting raw data
from sensors, underground sensors for soil analysis and aboveground for atmospheric
condition analysis, sending it to distant cloud servers for storage, processing, analysis,
monitoring and decision making.
The most important part of our architecture is data analysis of collected data using
machine learning algorithms, considering pre-learned soil nature and crop needs.
The analysis part sends the necessary information to Application layer, in order to
make good decisions and to quantify our field needs from both water and fertilizers. This
application layer also contains a deep learning algorithm that analyses captured images of
the field, classify each plant species, and detect plants disease, we will talk about it in details
in the next chapter.
As project perspectives, it will be data visualization using Augmented Reality
technology applied to an agricultural field in order to display our captured and analyzed
data.
 V. The proposed deep learning model:

Figure 3: Deep Learning process

The important idea of convolutional networks is that each neuron in each

intermediate layer is exposed is only exposed to a particular receptor field, and the analysis
of this receptor field is the same as analysis performed by another neuron of the same layer
with its own receiver field.
We multiply the value of each pixel by a synaptic weight and calculate the sum of the
weighted luminosities, and hold us only the positive part of the result, which corresponds to
the activation function.
The synaptic weights then form a convolution matrix (Filter/kernel), which will be
applied to all regions of the image and other neurons of the layer in question, in particular,
these synapses have always the synaptic weight mm, one talk about sharing synaptic
weights (weight sharing). This allows us to drastically reduce the parameters of our neural
network.
CNN's have a pre-programmed architecture for translational invariance of images. To
reduce the size of the images, we use a dimensionality reduction function which amounts to
summarizing the information of several neighbouring neurons into a single piece of
information, in order to do so, the average value of the excitations of these neurons is then
calculated, we then talk about the extraction of the average (Average Pooling) or in other
cases, the extraction of the maximum (Max pooling) using the same convolution matrix for
all the receptive fields that automatically guarantees our network of convolutional neurons,
they have a pre-programmed architecture for translation invariance of the images.
After pooling, the different sub-regions of the image were then described using the
resumé composed of a few canes. then we start again any debt of similar operations conv-
pool etc.
Generally, after a few lines, we obtain a vector of much smaller dimension and which
has an abstract representation of the contents of the different regions of the image.
For our datasets, we used 267 images of all plants, we took 80% of them as train
datasets and the other 20% as test datasets. After that, we take a random selection of
images as validation datasets in order to have predictions results as an output of the
algorithm.

VI. Results:
In this part, we will talk about our developed CNN algorithms and discuss its results.
In our algorithm, we used two convolutional layers in order to process our data source,
which is composed of 267 images, we can see a sample of it in figure 3, devised on three
types, verbena, mint and cannabis. We split this data source aleatory on 80% train sets, and
20% test tests. To process our train sets, we apply the filter on the first convolutional layer,
we can see it in figure 4.

Figure 4: Sample of the data set

 Figure 5: Filter of the first convolutional layer

In order to well understand the filter, we applied it on a real image, so we can see the
added value of the filter:

Figure 6:Filtered of the first convolutional layer applied to a real image

Then we apply the filter of the second convolutional layer:

Figure 7: Filter of the second convolutional layer

As before we display the results of applied filter on a real image:

 Figure 8: Filtered of the second convolutional layer applied to a real image

In order to have better results of the prediction, it is recommended to apply

performance optimization iterations, we can see that if we apply one iteration, we have
these learning performances:

Figure 9: Performance of the training after one iteration

But after applying 99 iterations we get better results:

Figure 10: Performance of the training after 99 iteration

Performance after multiple iterations:

 Figure 11: Performance of the training after multiple iterations

Results of leaf type classification prediction:

Figure 12: Leaf type classification data set

Figure 13: Results of leaf type classification

After applying the same process on the affected plants, we have this classification as an
output of the algorithm:
 Figure 14: Affected leaf data set

Results of the predicted classification:

Figure 15: Results of disease classification

As we can see in figure 14, we see that the images 14, 18, 19 and 22 are affected of ‘rouille’
disease, which is the rust disease, with a certainty of 78%, 53%, 59% and 60% consecutively,
and the image 2 is affected with the Powdery Mildew with a certainty of 52% .
But we have in figure 18, a classification error, the disease is a Powdery Mildew and not the
Rust, therefor, our algorithm needs more certainty, to do so, we need to insert more
learning data in order to have better results.
 VII. Conclusion:

In this article, we have talked about Smart agriculture architecture, we started from
its definition by other players and authors to end up by proposing our own architecture and
model. After that, we have developed a deep learning algorithm to classify and detect plants
diseases at early stages in order to ensure a qualitative production with maximum
rentability.
In order to compare our results with others, we talked in the comparative study part,
about the vision of other authors and their algorithms. Each algorithm has its own pros and
cons, but what we can say after this work, Is that our CNN does not only take leaf shape, or
leaf texture or its colour, it takes all the criteria and specifications of each leaf to give us the
best results we can have.
We can see in training performance figures, that our algorithm performance
increases from 46.9% in Training accuracy, 40.6% in validation accuracy and 1.086 in
validation loss to 100% in Training accuracy, 75.0% in validation accuracy and 0.584 in
validation loss.
In our results, we see that the disease classification certainty that goes up to 78%,
and for some cases, we find only 52 or 53%, these values need to be higher to consider them
as reliable results.
Unlike leaf crop classification, we find that the certainty varies from 65% to 91%
which we consider them as reliable results.
As this project perspective, we will increase diseases classification certainty to have
more reliable results to put them into production.
Table of figures:

Figure 1: Smart Agriculture Architecture, Resource management. ...................................................... 10

Figure 2: Deep Learning process ........................................................................................................... 12
Figure 3: Sample of the data set............................................................................................................ 13
Figure 4: Filter of the first convolutional layer ...................................................................................... 14
Figure 5:Filtered of the first convolutional layer applied to a real image ............................................. 14
Figure 6: Filter of the second convolutional layer ................................................................................. 14
Figure 7: Filtered of the second convolutional layer applied to a real image ....................................... 15
Figure 8: Performance of the training after one iteration .................................................................... 15
Figure 9: Performance of the training after 99 iteration ...................................................................... 15
Figure 10: Performance of the training after multiple iterations ......................................................... 16
Figure 11: Leaf type classification data set ........................................................................................... 16
Figure 12: Results of leaf type classification ......................................................................................... 16
Figure 13: Affected leaf data set ........................................................................................................... 17
Figure 14: Results of disease classification............................................................................................ 17
 Bibliography:

[1] A. McCoss, « Quantum Deep Learning Triuniverse », JQIS, vol. 06, no 04, p. 223‑248,
2016, doi: 10.4236/jqis.2016.64015.
[2] R. Yamashita, M. Nishio, R. K. G. Do, et K. Togashi, « Convolutional neural networks: an
overview and application in radiology », Insights Imaging, vol. 9, no 4, p. 611‑629, août
2018, doi: 10.1007/s13244-018-0639-9.
[3] K. Fragkiadaki, S. Levine, P. Felsen, et J. Malik, « Recurrent Network Models for Human
Dynamics », in 2015 IEEE International Conference on Computer Vision (ICCV), Santiago,
Chile, 2015, p. 4346‑4354, doi: 10.1109/ICCV.2015.494.
[4] Y. Yuan, L. Mou, et X. Lu, « Scene Recognition by Manifold Regularized Deep Learning
Architecture », IEEE Trans. Neural Netw. Learning Syst., vol. 26, no 10, p. 2222‑2233,
oct. 2015, doi: 10.1109/TNNLS.2014.2359471.
[5] R. Girshick, J. Donahue, T. Darrell, et J. Malik, « Region-Based Convolutional Networks
for Accurate Object Detection and Segmentation », IEEE Trans. Pattern Anal. Mach.
Intell., vol. 38, no 1, p. 142‑158, janv. 2016, doi: 10.1109/TPAMI.2015.2437384.
[6] W. Hu, Y. Huang, L. Wei, F. Zhang, et H. Li, « Deep Convolutional Neural Networks for
Hyperspectral Image Classification », Journal of Sensors, vol. 2015, p. 1‑12, 2015, doi:
10.1155/2015/258619.
[7] K. Makantasis, K. Karantzalos, A. Doulamis, et N. Doulamis, « Deep supervised learning
for hyperspectral data classification through convolutional neural networks », in 2015
IEEE International Geoscience and Remote Sensing Symposium (IGARSS), Milan, 2015, p.
4959‑4962, doi: 10.1109/IGARSS.2015.7326945.
[8] Y. Chen, H. Jiang, C. Li, X. Jia, et P. Ghamisi, « Deep Feature Extraction and Classification
of Hyperspectral Images Based on Convolutional Neural Networks », IEEE Trans. Geosci.
Remote Sensing, vol. 54, no 10, p. 6232‑6251, oct. 2016, doi:
10.1109/TGRS.2016.2584107.
[9] D. Komura et S. Ishikawa, « Machine Learning Methods for Histopathological Image
Analysis », Computational and Structural Biotechnology Journal, vol. 16, p. 34‑42, 2018,
doi: 10.1016/j.csbj.2018.01.001.
[10] G. Litjens et al., « A survey on deep learning in medical image analysis », Medical Image
Analysis, vol. 42, p. 60‑88, déc. 2017, doi: 10.1016/j.media.2017.07.005.
[11] D. Shen, G. Wu, et H.-I. Suk, « Deep Learning in Medical Image Analysis », Annu. Rev.
Biomed. Eng., vol. 19, no 1, p. 221‑248, juin 2017, doi: 10.1146/annurev-bioeng-071516-
044442.
[12] F. Xing et L. Yang, « Robust Nucleus/Cell Detection and Segmentation in Digital
Pathology and Microscopy Images: A Comprehensive Review », IEEE Rev. Biomed. Eng.,
vol. 9, p. 234‑263, 2016, doi: 10.1109/RBME.2016.2515127.
[13] M. Hasan, B. Tanawala, et K. J. Patel, « Deep Learning Precision Farming: Tomato Leaf
Disease Detection by Transfer Learning », SSRN Journal, 2019, doi:
10.2139/ssrn.3349597.
[14] J. Amara, B. Bouaziz, et A. Algergawy, « A Deep Learning-based Approach for Banana
Leaf Diseases Classification », p. 10.
[15] « Leaf Classification Based on GLCM Texture and SVM », vol. 4, no 3, p. 4.
[16] M. Sadeghi, A. Zakerolhosseini, et A. Sonboli, « Architecture-Based Classification of
Plant Leaf Images », p. 28.
[17] E. Prasetyo, R. D. Adityo, N. Suciati, et C. Fatichah, « Multi-class K-support Vector
Nearest Neighbor for Mango Leaf Classification », TELKOMNIKA, vol. 16, no 4, p. 1826,
août 2018, doi: 10.12928/telkomnika.v16i4.8482.
[18] Y. Guo, Y. Liu, E. M. Bakker, Y. Guo, et M. S. Lew, « CNN-RNN: a large-scale hierarchical
image classification framework », Multimed Tools Appl, vol. 77, no 8, p. 10251‑10271,
avr. 2018, doi: 10.1007/s11042-017-5443-x.
[19] S. Zhang, F. Sun, et H. Liu, « Locality-Constrained Linear Coding with Spatial Pyramid
Matching for SAR Image Classification », in Foundations and Practical Applications of
Cognitive Systems and Information Processing, vol. 215, F. Sun, D. Hu, et H. Liu, Éd.
Berlin, Heidelberg: Springer Berlin Heidelberg, 2014, p. 867‑876.
[20] S. Delalieux, B. Somers, B. Haest, T. Spanhove, J. Vanden Borre, et C. A. Mücher,
« Heathland conservation status mapping through integration of hyperspectral mixture
analysis and decision tree classifiers », Remote Sensing of Environment, vol. 126, p.
222‑231, nov. 2012, doi: 10.1016/j.rse.2012.08.029.
[21] J. Ham, Yangchi Chen, M. M. Crawford, et J. Ghosh, « Investigation of the random forest
framework for classification of hyperspectral data », IEEE Trans. Geosci. Remote
Sensing, vol. 43, no 3, p. 492‑501, mars 2005, doi: 10.1109/TGRS.2004.842481.
[22] Tin Kam Ho, « The random subspace method for constructing decision forests », IEEE
Trans. Pattern Anal. Machine Intell., vol. 20, no 8, p. 832‑844, août 1998, doi:
10.1109/34.709601.
[23] K. He, X. Zhang, S. Ren, et J. Sun, « Deep Residual Learning for Image Recognition »,
arXiv:1512.03385 [cs], déc. 2015.
[24] K. Simonyan et A. Zisserman, « Very Deep Convolutional Networks for Large-Scale
Image Recognition », arXiv:1409.1556 [cs], avr. 2015.
[25] D. E. Fisher et B. C. Bastian, Melanoma. 2019.
[26] V. B. Misic et J. Misic, Éd., Machine-to-Machine Communications: Architectures,
Technology, Standards, and Applications, 0 éd. CRC Press, 2014.
[27] C.-R. Rad, O. Hancu, I.-A. Takacs, et G. Olteanu, « Smart Monitoring of Potato Crop: A
Cyber-Physical System Architecture Model in the Field of Precision Agriculture »,
Agriculture and Agricultural Science Procedia, vol. 6, p. 73‑79, 2015, doi:
10.1016/j.aaspro.2015.08.041.
[28] R. Khan, S. U. Khan, R. Zaheer, et S. Khan, « Future Internet: The Internet of Things
Architecture, Possible Applications and Key Challenges », in 2012 10th International
Conference on Frontiers of Information Technology, Islamabad, Pakistan, 2012, p.
257‑260, doi: 10.1109/FIT.2012.53.
[29] I. B. Thingom, « Internet of Things. », IOSR Journal of Computer Engineering, p. 5.
[30] H. Channe, S. Kothari, et D. Kadam, « Multidisciplinary Model for Smart Agriculture
using Internet-of-Things (IoT), Sensors, Cloud-Computing, Mobile-Computing & Big-
Data Analysis », vol. 6, p. 9, 2015.
[31] J. Deng, W. Yang, Q. Li, J. Liu, Y. Zhang, et L. Bai, « Research and Application of Fire
Power Cloud Platform », Procedia Engineering, vol. 211, p. 911‑916, 2018, doi:
10.1016/j.proeng.2017.12.092.
[32] F. T. -, « Smart Agriculture Based on Cloud Computing and IOT », JCIT, vol. 8, no 2, p.
210‑216, janv. 2013, doi: 10.4156/jcit.vol8.issue2.26.
[33] F. Ferrández-Pastor, J. García-Chamizo, M. Nieto-Hidalgo, et J. Mora-Martínez,
« Precision Agriculture Design Method Using a Distributed Computing Architecture on
 Internet of Things Context », Sensors, vol. 18, no 6, p. 1731, mai 2018, doi:
10.3390/s18061731.
[34] T. Ara, P. Gajkumar Shah, et M. Prabhakar, « Internet of Things Architecture and
Applications: A Survey », Indian Journal of Science and Technology, vol. 9, no 45, déc.
2016, doi: 10.17485/ijst/2016/v9i45/106507.
[35] I. Ungurean, N.-C. Gaitan, et V. G. Gaitan, « An IoT architecture for things from
industrial environment », in 2014 10th International Conference on Communications
(COMM), Bucharest, Romania, 2014, p. 1‑4, doi: 10.1109/ICComm.2014.6866713.
[36] F. Ferrández-Pastor, J. García-Chamizo, M. Nieto-Hidalgo, J. Mora-Pascual, et J. Mora-
Martínez, « Developing Ubiquitous Sensor Network Platform Using Internet of Things:
Application in Precision Agriculture », Sensors, vol. 16, no 7, p. 1141, juill. 2016, doi:
10.3390/s16071141.
[37] M. W. Pfaffl, « A new mathematical model for relative quantification in real-time RT-
PCR », Nucleic Acids Research, vol. 29, no 9, p. 45e‑445, mai 2001, doi:
10.1093/nar/29.9.e45.
[38] Y.-R. Chen, K. Chao, et M. S. Kim, « Machine vision technology for agricultural
applications », Computers and Electronics in Agriculture, vol. 36, no 2‑3, p. 173‑191,
nov. 2002, doi: 10.1016/S0168-1699(02)00100-X.