Available at www.sciencedirect.

com

INFORMATION PROCESSING IN AGRICULTURE 7 (2020) 566– 574

journal homepage: www.elsevier.com/locate/inpa

Performance analysis of deep learning CNN models

for disease detection in plants using image
segmentation

Parul Sharma a,*, Yash Paul Singh Berwal b, Wiqas Ghai a

a
Department of Computer Science and Engineering, RIMT University, Mandi Gobindgarh 147301, India
b
Additional Director, Department of Technical Education, Haryana, India

A R T I C L E I N F O A B S T R A C T

Article history: Food security for the 7 billion people on earth requires minimizing crop damage by timely
Received 8 July 2019 detection of diseases. Most deep learning models for automated detection of diseases in plants
Received in revised form suffer from the fatal flaw that once tested on independent data, their performance drops sig-
7 November 2019 nificantly. This work investigates a potential solution to this problem by using segmented
Accepted 14 November 2019 image data to train the convolutional neural network (CNN) models. As compared to the
Available online 18 November 2019 F-CNN model trained using full images, S-CNN model trained using segmented images more
than doubles in performance to 98.6% accuracy when tested on independent data previously
Keywords: unseen by the models even with 10 disease classes. Not only this, by using tomato plant and
Machine learning target spot disease type as an example, we show that the confidence of self-classification for
Plant disease detection S-CNN model improves significantly over F-CNN model. This research work brings
Image segmentation applicability of automated methods closer to non-experts for timely detection of diseases.
Ó 2019 China Agricultural University. Production and hosting by Elsevier B.V. on behalf of
KeAi. This is an open access article under the CC BY-NC-ND license (http://creativecommons.
org/licenses/by-nc-nd/4.0/).

1. Introduction The development of computer vision models has fueled

exponential growth in efforts to develop automated systems
With world population expected to increase to 9 billion by for detection of plant diseases using visible symptoms on
2050, developing methods for detection and mitigation of leaves. These systems try to make the involvement of farmers
plant diseases serves the dual purpose of increasing crop as simple as possible and the detection system as foolproof as
yield and reducing pesticide use. Along with development of possible. Before the easy availability of deep learning models,
better crop varieties, disease detection is thus paramount studies mainly focused on using image processing/feature
goal for achieving food security. The traditional method of extraction to develop techniques for disease identification
disease detection has been to use manual examination by with mixed results. See for example Pydipati et al. [1], Sanyal
either farmers or experts, which can be time consuming and Patel [2], Kurniawati et al. [3], Camargo and Smith [4],
and costly, proving infeasible for millions of small and med- Story et al. [5], Pugoy and Mariano [6], Phadikar et al. [7], Kruse
ium sized farms around the world. et al. [8], Clement et al. [9] and Barbedo et al. [10]. The biggest
challenge of this approach, difficulties in describing the
symptoms for identification by computers, has been solved
* Corresponding author. by using deep learning where the models do not need the
E-mail address: [email protected] (P. Sharma).
features to be described, but rather learn the features using
Peer review under responsibility of China Agricultural University.
https://doi.org/10.1016/j.inpa.2019.11.001
2214-3173 Ó 2019 China Agricultural University. Production and hosting by Elsevier B.V. on behalf of KeAi.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
 Information Processing in Agriculture 7 ( 2 0 2 0 ) 5 6 6 –5 7 4 567

optimization. A number of studies in recent years have uti- involves application of a mask to each pixel and then select-
lized the capabilities of deep learning models to successfully ing a single value (eg. maximum) from with the mask. Mixing
attain various accuracy levels on laboratory/field images steps in Fig. 1 involve multiple convolution and pooling steps.
(see, for example, Mohanty et al., [11], Fuentes et al. [12], By feeding the final output into a SoftMax function, a gradient
Wang et al. [13], Johannes et al. [14], Barbedo [15], Kamilaris optimizer is used to train the model weights. By introducing
[16], Plantix [17] and Sharma et al. [18]). These accurate clas- non-linearities using a Rectified Linear Unit (ReLU) and ran-
sifiers, when tested on data similar to the one used during dom dropouts during learning, the model is able to automat-
training, achieve high accuracy but fail miserably when tested ically learn the decision boundaries required to classify
on different data. Examples of this include Mohanty et al. [11] images into one class or another. As the data proceeds
and Ferentinos [19], where training models attained accuracy through the network, the size of the filters and outputs
and confidence more than 99%, but when tested against changes (getting smaller progressing through the network),
images taken from trusted online resources, confidence which allows for training for and detection of similar features
dropped to less than 40%. Barbedo [15] outlined a number of with different scaling. The convolution filters applied over the
factors affecting the performance of deep learning models whole images for training data mean that the CNNs are
applied to plant leaf disease detection and concluded that invariant to transformations of features such as rotation
despite the significantly high success rate of developed sys- and translation. Furthermore, the pooling layers effectively
tems, including their own, there are various reasons that make the CNNs tolerant to distortions in the features.
make it still quite far from being a generic tool that could The model is implemented using Python and Ten-
be used in real conditions. sorFlowTM [27]. Training and validation runs were carried
Transfer learning, tweaking the final output of a trained out on a hosted server at Google CloudTM using Nvidia GPUs.
neural network to a new dataset, has been a popular tool of In order to analyze the working of the model in detail, we pick
choice for researchers given the relatively small sizes of data- out tomato plant, which includes 9 disease types and healthy
sets currently available (tens of thousands of images as com- leaves, the symptoms are shown in Fig. 2. We have trained
pared to millions of images required for training models from two CNN models: the first was trained using a dataset consist-
scratch). Examples of using this technique and images from ing of full leaf images with varying background and disease
publicly available database such as Plant Village [20] include progressions. This model is hereafter called F-CNN. The sec-
Durmusß et al. [21], Zhang et al. [22], Too et al. [23], Suryawati ond model, hereafter called S-CNN, was trained using a data-
et al. [24] and Zhang et al. [25]. set generated from the same images used in training F-CNN,
Most efforts utilizing deep learning for disease detection but segmented in order to only include regions-of-interest
have trained and tested their efforts using the Plant Village with disease symptoms. The regions-of-interest were chosen
dataset, which consists of images with low variability and to contain multiple spots/lesions of similar disease. We devel-
similar backgrounds. Recently, Barbedo [26] investigated the oped a user interface, which allows for rapid image segmen-
performance of deep learning models when trained using tation capable of processing thousands of images per hour.
individual lesions and spots, using image segmentation and Although the models presented here are generally applica-
augmentation to increase the size of the dataset from a rela- ble to any plant type/disease with visible symptoms, to test
tively small number of images. They concluded that the per- the performance of the CNN models, we select the tomato
formance of the models improved (12% higher on average) plant type with 10 classes. Images for training and classifica-
in every plant type under investigation with accuracies con- tion (total 17929) were locally sourced from farms (637), taken
sistently above 75% in the most complex case with 10 dis- from the Plant Village database ([20]) (16579) and crawled
eases. This research work follows a similar approach: the from the Internet (713). The Plant Village database by itself
image datasets are segmented to use maximum information is not optimum for training deep learning models because
about the disease symptoms by extracting affected parts of of similar backgrounds and lighting conditions. We used mul-
the leaves instead of whole leaf images. By quantifying the tiple approaches to offset this problem: image segmentation
results of the two models and comparing the self- to extract relevant features, applying random transforms to
classification confidence for each disease type, it is shown the extracted images to vary brightness/contrast/blur and
that we can train using more meaningful data and very good the farm-(and internet-)sourced images contained many dif-
results can be achieved even in real world conditions. ferent lighting conditions and backgrounds. Images sourced
from internet and farms were not included in the training/-
2. Method validation data used during model-training but were used
only during performance evaluation to provide an indepen-
In our previous work ([18]), we developed a CNN model, which dent source of data. For training deep learning models from
achieved accuracy greater than 93% for 15 different plant scratch, 17000 images for 10 classes can be insufficient
types. This study will investigate the model in more detail. and previous studies have used data augmentation to multi-
The CNN model trained in this study is shown in Fig. 1. Once ply the number of samples multifold (Zhang et al. [22], Bar-
images are read, 256x256 pixel random parts of the images bedo [26] and Liu et al. [28]). However, as we shall
are extracted and noise, distortion, flip or rotation transforms demonstrate, our model can achieve very high levels of accu-
are applied. By controlling stride lengths (spacing interval for racy with the dataset at hand.
placement of the filters/masks), dimensions of masks, multi- The images downloaded from the internet, 713 in total,
ple convolution and pooling steps are applied. Pooling were manually annotated and labeled using the image
568 Information Processing in Agriculture 7 (2 0 2 0) 5 6 6–57 4

Fig. 1 – Convolutional neural network model overview. The different steps are described in the text.

Fig. 2 – Example of disease symptoms for the nine tomato disease and healthy leaf. (a) Bacterial Spot. (b) Early Blight. (c)
Healthy. (d) Late Blight. (e) Leaf Mold. (f) Septoria Leaf Spot. (g) Spider Mite. (h) Target Spot. (i) Tomato Mosaic Virus. (j) Yellow
Leaf Curl Virus.

segmentation user interface we developed. Segmentation, detected class. This is because a number of images for Spider
annotation and labeling involved selection of only the areas Mite had Target Spot symptoms as well, confusing the model
of interest (lesions), application of random stretch, blur, during training. Example of this is shown in Fig. 4. The aver-
brightness and contrast adjustments and assigning a single age accuracy of 96.3% for the model is similar to what other
lable to each. Further 637 images were locally taken from farms researchers have found [15,26,11,14,19]. This high accuracy
in North India covering 8 of the 10 classes under investigation. can partly be attributed to the similar nature of images in
Tomato Mosaic Virus and Septoria Leaf Spot were not found in the Plant Village data, images, which are used for both train-
farms. The dataset was set up in such a way to have roughly ing and validation. However, the model suffers from specific
the same number of images for each class and covering the problem that when tested with the same images but cropped
widest spectrum of disease symptoms. Before testing on the to contain partial leaves or with independent images from the
CNN models, random transformations were applied on each internet, the average self classification accuracy drops to only
segmented image (stretch/rotation/brightness/contrast blur). 42.3%. This is consistent with other studies which considered
images from the Plant Village dataset for training and tested
3. Results and discussion against independent images [11,19].
The confusion matrix for deep learning model trained with
Fig. 3(a) shows the confusion matrix for tomato using full the segmented images (S-CNN) is shown in Fig. 3(b). It can be
images (F-CNN). Confusion matrix shows the corresponding seen that, except for Bacterial Spot and Tomato Mosaic Virus,
percentage of detected class for each true class. The diagonal the self classification accuracy for each disease type
in Fig. 3(a) shows the self classification case, with the text increases. For Bacterial Spot, accuracy remains the same
showing the percentage of classification result for corre- whereas for Tomato Mosaic Virus, accuracy decreases by
sponding class. It can be seen that the self classification accu- 2%. The average accuracy increases from 96.3% in case of
racy is always greater than 92% in each case. The worst F-CNN to 98.0% in case of S-CNN. The problem of Spider Mite
performing class is Target Spot (TS) with 92% accuracy. About detection for Target Spot images, even when the symptoms
3% of images with Target Spot gave Spider Mite as the are quite different, is resolved using segmented images.
 Information Processing in Agriculture 7 ( 2 0 2 0 ) 5 6 6 –5 7 4 569

Fig. 3 – Confusion matrix for the tomato plant. (a) F-CNN model. (b) S-CNN model. Text depicts corresponding result in %
rounded to the nearest integer. Results rounding to 0% are not shown for clarity. Abbreviations: Early Blight (EB), Bacterial
Spot (BS), Healthy (H), Septoria Leaf Spot (SLS), Yellow Leaf Curl Virus (YLCV), Late Blight (LB), Target Spot (TS), Leaf Mold (LM),
Spider Mite (SM) and Tomato Mosaic Virus (TMV).

Fig. 5(a and b), had more than two classes with significant
confidence as compared to markers located close to the red
lines, which had only two classes with significant confidence.
The black dashed lines in Fig. 5(a and b) represent the case
where CTrue CSecondBest = 0. Thus, markers close to the black
lines, for example, markers in green circles in Fig. 5(a and
b), had a very small difference between CTrue and CSecondBest.
In such cases, even though the correct class was chosen,
the second best class had similar confidence. Qualitatively,
it can be seen from Fig. 5 that as compared to F-CNN model
results in Fig. 5(a), S-CNN model results in Fig. 5(b) contain
fewer markers located away from the red dashed line. Simi-
larly, more markers are located away from the black dashed
line in Fig. 5(b) (S-CNN) as compared to Fig. 5(a) (F-CNN).
Fig. 6 shows confidence for the predicted (wrong) class
(CBest) as compared to confidence for the true class (CTrue) for
Fig. 4 – Example of tomato leaf classified as spider mite but
images of tomato target spot disease type. Similar to Fig. 5,
showing symptoms similar to target spot as well (in white
the black dashed lines in Fig. 6(a and b) represent the case
circles).
where CBest CTrue = 0 and red dashed lines represent the case
where CBest + CTrue = 1. Markers close to black dashed lines
Most studies of application of deep learning for detection mean the difference in confidence for True class and Best-
of diseases in plants use the final classification accuracy as detected class was very small and markers close to red
the only factor to determine model performance dashed lines signify a binary result, only two classes were
[15,26,11,14,19]. However, once self classification score detected with significant confidence. It can be seen from
reaches greater than 95%, small changes in the score cannot Fig. 6(a and b) that both F-CNN and S-CNN models have a
be the best metric for judging model performance. In our number of images which were mis-classified with high
case, self classification accuracy shows an increase of 1.7%. confidence.
To judge the improvement in quality of classification in more Quantification of results of F-CNN and S-CNN to compare
detail, we look at the results shown in Fig. 5. The two plots model performance can be done by using Fig. 7, which shows
show confidence for the true class (CTrue) as compared to con- histogram of confidence for correctly classified images with
fidence of the second best choice (CSecondBest) output from the tomato target spot disease type. It can be seen that more than
models for tomato target spot disease type. Confidence here 600 images had confidence greater than 0.93 for S-CNN model
refers to the ‘‘confidence” of the model that the image belongs (Fig. 7(a)) as compared to approx. 270 images for F-CNN model
to a certain class. The sum of confidence for all classes is 1. (Fig. 7(b)). The mean and median values for confidence are
Red dashed lines in Fig. 5(a and b) show the case where given in Table 1. An increase of 0.093 in both mean confidence
CTrue + CSecondBest = 1. This means that markers located further and median confidence for S-CNN represent a significant
away from the red lines for example, markers in red circles in improvement over F-CNN. This improvement is also seen in
570 Information Processing in Agriculture 7 (2 0 2 0) 5 6 6–57 4

Fig. 5 – Scatter plot showing confidence for true class (CTrue) or Best class (CBest) as compared to confidence of the second best
choice (CSecondBest) for tomato target spot disease type for correctly classified images. (a) F-CNN model. (b) S-CNN model. Size of
markers is proportional to CTrue CSecondBest. Black dashed line corresponds to CTrue CSecondBest = 0 and red dashed line
corresponds to CTrue + CSecondBest = 1.

Fig. 6 – Scatter plot showing confidence for predicted (wrong) class (CBest) as compared to confidence for true class (CTrue) for
tomato target spot disease type for mis-classified images. (a) F-CNN model. (b) S-CNN model. Size of markers is proportional
to CBest CTrue. Black dashed line corresponds to CBest CTrue = 0 and red dashed line corresponds to CBest + CTrue = 1.

Fig. 7 – Histogram of confidence for correctly classified images with tomato target spot disease type. (a) F-CNN model. (b) S-
CNN model.
 Information Processing in Agriculture 7 ( 2 0 2 0 ) 5 6 6 –5 7 4 571

Table 1 – Result comparison between F-CNN and S-CNN models. QCoD: Quartile Coefficient of Dispersion.
F-CNN S-CNN

Confidence Mean 0.840 0.933

Median 0.889 0.982
Std. Dev. 0.150 0.111
QCoD 0.118 0.036
Confidence Difference Mean 0.745 0.885
Median 0.828 0.969
Std. Dev. 0.244 0.192
QCoD 0.205 0.063

reduction in standard deviation for confidence for S-CNN by The strength of S-CNN over F-CNN can also be gauged
0.039 as compared to F-CNN. using the example leaf shown in Fig. 9, which contains symp-
Fig. 8 shows the difference in confidence between true toms of both leaf mold and target spot (encircled regions).
class (CTrue) and second best class (CSecondBest) for the same When the image was classified using F-CNN model, the model
images in Fig. 7. The trend of improvement in performance returned spider mite as the detected disease with 0.497 confi-
of S-CNN model (Fig. 8(b)) with higher number of images hav- dence as compared to 0.437 confidence for target spot. How-
ing (CTrue CSecondBest) greater than 0.93 as compared to F-CNN ever, when the image was segmented into two sub-images
model (Fig. 8(a)) is repeated. An increase of 0.145 in mean of containing distinctly different symptoms, shown in Fig. 10,
confidence difference, 0.141 in median of confidence differ- the model correctly detected target spot (Fig. 10(a)) and leaf
ence and decrease of 0.052 in standard deviation of confi- mold (Fig. 10(b)) with 0.883 and 0.899 confidence, respectively.
dence difference between S-CNN and F-CNN shown in This is expected since the S-CNN model is only fed regions
Table 1 once again highlight improvements in model with symptoms of only one disease and is thus less likely to
performance. mis-classify.
Since standard deviation is more sensitive to values further Fig. 11(a) shows a histogram of difference in confidence for
away from the mean, we use the more robust indicator quar- S-CNN and F-CNN models for images with tomato target spot
tile coefficient (QCoD) to measure the dispersion in results of disease type. The difference was calculated between the cor-
F-CNN and S-CNN. QCoD is calculated as QCoD = (Q3 Q1)/ responding model output for each image meaning positive
(Q3 + Q1). Here Q3 is the 3rd quartile (3/4th of total observations numbers signify improvement whereas negative numbers
have values lower than this) and Q1 is the 1st quartile (1/4th of signify reduction in confidence going from F-CNN to S-CNN.
total observations have values lower than this). Lower values For 82% of the images, S-CNN resulted in a higher confidence
of QCoD mean smaller dispersion and vice versa. than F-CNN and for all the images, the resulting confidence
As seen in Table 1, QCoD of confidence values shows that was on average 0.236 higher for S-CNN than F-CNN. Around
dispersion for F-CNN is 3 times greater than for S-CNN. Sim- 70% of images where the model performed badly had multiple
ilarly, QCoD of confidence difference for F-CNN is more than 5 symptoms in the segmented area. An example of multiple
times greater than for S-CNN. These much smaller values for symptoms can be seen in Fig. 11(b), where the confidence
S-CNN show a remarkable improvement in model perfor- for S-CNN was only 0.263. This was resolved when a smaller
mance as compared to F-CNN.

Fig. 8 – Histogram of difference in confidence for the true class (CTrue) and second best class (CSecondBest) for correctly classified
images with tomato target spot disease type. (a) F-CNN model. (b) S-CNN model.
572 Information Processing in Agriculture 7 (2 0 2 0) 5 6 6–57 4

Fig. 9 – Example of a tomato leaf highlighting presence of multiple disease symptoms. (a) Leaf image. (b) Confidence for each
disease class using F-CNN.

Fig. 10 – Segmented regions of the leaf image in Fig. 9 and the corresponding classification results using S-CNN. (a) Region
with target spot disease type. (b) Region with leaf mold disease type. The images are rotated clockwise by 90° as compared to
Fig. 9.

cropped area with only symptoms of target spot was chosen, individual spots/lesions instead of full leaf images. Even
resulting in confidence of 0.939 for S-CNN. though the S-CNN model gives much better results than the
When tested on the independent set of data taken from F-CNN model, the authors acknowledge that the datasets
the internet and from fields previously unknown to the mod- used to train the models might not contain all the disease fea-
els, S-CNN achieved an accuracy of 98.6% as compared to tures observed in nature. For example, if the spores in case of
42.3% as reported earlier for F-CNN. This remarkable powdery mildew are somehow absent due to friction damage/
improvement in accuracy of the S-CNN model is attributed winds, it becomes harder for the model to detect the disease
to the better ability of the model to learn features because correctly since the visual symptoms change remarkably.
the image dataset has been prepared to get rid of other factors Vastly different extents of damage might also contribute in
(for example, background). Our results are in line with the lowering the prediction accuracy. Better datasets with enough
findings from Barbedo [26], where the author observed better images representing all the different manifestations of dis-
performance for models trained with and tested against ease symptoms are thus required to be developed/curated.
 Information Processing in Agriculture 7 ( 2 0 2 0 ) 5 6 6 –5 7 4 573

Fig. 11 – Comparison of S-CNN and F-CNN outputs. (a) Histogram of difference in confidence for S-CNN and F-CNN models for
images with tomato target spot disease type. (b) Example cropped image where S-CNN model performed badly.

One limitation of the S-CNN model is that it can fail in R E F E R E N C E S

cases of regions containing symptoms from multiple dis-
eases. Masking or cropping to isolate single disease symp-
toms might alleviate this problem. Another related [1] Pydipati R, Burks TF, Lee WS. Statistical and neural network
limitation of the S-CNN model is that it can be very sensitive classifiers for citrus disease detection using machine vision.
to the quality of segmentation, which depends on the human Trans ASAE 2005;48(5):2007–14.
[2] Sanyal P, Patel SC. Pattern recognition method to detect two
involved in segmentation. For example in Fig. 11(b), the deci-
diseases in rice plants. Imag Sci J 2008;56(6):319–25.
sion of the rater was to segment the image as shown, but [3] Kurniawati NN, Abdullah SNHS, Abdullah S, Abdullah S.
cropping the image to isolate the symptoms resulted in accu- Investigation on image processing techniques for diagnosing
rate results. paddy diseases. In: 2009 International Conference of Soft
The most essential building block for a CNN model is the Computing and Pattern Recognition. p. 272–7.
underlying data used to train it and thus synchronous efforts [4] Camargo A, Smith J. An image-processing based algorithm to
automatically identify plant disease visual symptoms.
for collecting data are required. In this field, crowdsourcing
Biosyst Eng 2009;102(1):9–21.
efforts such as Plantix [17], Plant Village [20] and Digipathos
[5] Story D, Kacira M, Kubota C, Akoglu A, An L. Lettuce calcium
[29] need special mention. Furthermore, data archival tech- deficiency detection with machine vision computed plant
niques which provide unique Digital Object Identifiers (DOIs) features in controlled environments. Comput Electron Agric
for linking such as the Materials Data facility [30] need to be 2010;74(2):238–43.
developed. [6] Pugoy RADL, Mariano VY. Automated rice leaf disease
detection using color image analysis. Proceedings SPIE, Third
International Conference on Digital Image Processing, 2011.
4. Conclusions [7] Phadikar S, Sil J, Das AK. Rice diseases classification using
feature selection and rule generation techniques. Comp
Most deep learning models for automatic disease detection Electron Agric 2013;90:76–85.
suffer from reduced performance when applied to real world [8] Kruse OMO, Prats-Montalban JM, Indahl UG, Kvaal K, Ferrer
images previously unseen. In this paper, we show the feasibil- A, Futsaether CM. Pixel classification methods for identifying
ity of training a convolutional neural network (CNN) model and quantifying leaf surface injury from digital images.
Comp Electron Agric 2014;108:155–65.
using segmented and annotated images instead of full
[9] Clement A, Verfaille T, Lormel C, Jaloux B. A new colour vision
images. When the same CNN model is trained using the seg-
system to quantify automatically foliar discolouration caused
mented images (S-CNN) as compared to training using full by insect pests feeding on leaf cells. Biosyst Eng
images (F-CNN), model performance on independent data 2015;133:128–40.
increases from 42.3% to 98.6%. Furthermore, quantitative [10] Barbedo JGA, Koenigkan LV, Santos TT. Identifying multiple
analysis of self-classification confidence showed a significant plant diseases using digital image processing. Biosyst Eng
improvement with 82% of test dataset showing increase in 2016;147:104–16.
[11] Mohanty SP, Hughes DP, Salathe M. Using deep learning for
confidence. As better datasets become available in the future,
image-based plant disease detection. Front Plant Sci
pre-processing of images before model training in CNN can 2016;7:1419.
prove invaluable to achieve high real-world performance. [12] Fuentes A, Yoon S, Kim SC, Park DS. A robust deep-learning-
based detector for real-time tomato plant diseases and pests
Declaration of Competing Interest recognition. Sensors. 2017;17:2022.
[13] Wang G, Sun Y, Wang J. Automatic image-based plant disease
severity estimation using deep learning. Comput Intellig
Author declares that there is no conflicts of interest. Neurosci 2017;2017:2917536.
574 Information Processing in Agriculture 7 (2 0 2 0) 5 6 6–57 4

[14] Johannes A, Picon A, Alvarez-Gila A, Echazarra J, Rodriguez- [23] Too EC, Yujian L, Njuki S, Yingchun L. A comparative study of
Vaamonde S, Navajas AD, et al. Automatic plant disease fine-tuning deep learning models for plant disease
diagnosis using mobile capture devices, applied on a wheat identification. Comput Electron Agric 2019;161:272–9.
use case. Comput Electron Agric 2017;138:200–9. [24] Suryawati E, Sustika R, Yuwana RS, Subekti A, Pardede HF.
[15] Barbedo JGA. Impact of dataset size and variety on the Deep structured convolutional neural network for tomato
effectiveness of deep learning and transfer learning for plant diseases detection. In: 2018 International Conference on
disease classification. Comput Electron Agric 2018;153:46–53. Advanced Computer Science and Information Systems
[16] Kamilaris A, Prenafeta-Boldu FX. Deep learning in (ICACSIS). IEEE; 2018. p. 385–90.
agriculture: A survey. Comput Electron Agric 2018;147:70–90. [25] Zhang S, Huang W, Zhang C. Three-channel convolutional
[17] Plantix. https://plantixnet/en 2019; Accessed: 2019-08-20. neural networks for vegetable leaf disease recognition.
[18] Sharma P, Berwal YPS, Ghai W. Krishimitr (farmers friend): Cognit Syst Res 2019;53:31–41.
Using machine learning to identify diseases in plants. In: [26] Barbedo JGA. Plant disease identification from individual
2018 IEEE International Conference on Internet of Things and lesions and spots using deep learning. Biosyst Eng
Intelligence System (IOTAIS), Bali. p. 29–34. 2019;180:96–107.
[19] Ferentinos KP. Deep learning models for plant disease [27] Abadi M, Agarwal A, Barham P, Brevdo E, Chen Z, Citro C,
detection and diagnosis. Comput Electron Agric et al. TensorFlow: Large-scale machine learning on
2018;145:311–8. heterogeneous systems. 2015. URL: http://tensorflow.org/;
[20] Hughes DP, Salathe M. An open access repository of images Accessed: 2019-08-20.
on plant health to enable the development of mobile disease [28] Liu B, Zhang Y, He D, Li Y. Identification of apple leaf diseases
diagnostics through machine learning and crowdsourcing. based on deep convolutional networks. Symmetry
2015. ArXiv:1511.08060. 2018;10:11.
[21] Durmusß H, Günesß EO, Kırcı M. Disease detection on the leaves [29] Barbedo JGA, Koenigkan LV, Halfeld-Vieira BA, Costa RV,
of the tomato plants by using deep learning. In: 2017 6th Nechet KL, Godoy CV, et al. Annotated plant pathology
International Conference on Agro-Geoinformatics. IEEE; databases for image-based detection and recognition of
2017. p. 1–5. diseases. IEEE Lat Am Trans 2018;16(6):1749–57.
[22] Zhang X, Qiao Y, Meng F, Fan C, Zhang M. Identification of [30] Blaiszik B, Chard K, Pruyne J, Ananthakrishnan R, Tuecke S,
maize leaf diseases using improved deep convolutional Foster I. The materials data facility: data services to advance
neural networks. IEEE Access 2018;6:30370–7. materials science research. JOM 2016;68(8):2045–52.