Computers and Electronics in Agriculture 178 (2020) 105701

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture

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

A detection and severity estimation system for generic diseases of tomato T

greenhouse plants
Patrick Wspanialy , Medhat Moussa
⁎

University of Guelph, Canada

ARTICLE INFO ABSTRACT

Keywords: The management of plant disease is a significant economic and environmental factor in the production of
Greenhouse tomato greenhouse tomato plants. Human expertise for assessing the presence and extent of disease is important in
Severity estimation creating and implementing management plans, but it is difficult and expensive to acquire. In this paper, we
Disease detection present a new computer vision system to automatically recognize several diseases, detect previously unseen
Machine learning
disease and to estimate per-leaf severity. Training and testing of models used several modified versions of the
nine types of tomato disease of the PlantVillage tomato dataset and showed how different leaf properties impact
disease detection.

1. Introduction which cataloged thousands of plant leaf disease images from twelve
crop species, including nine tomato diseases. Previous deep learning
Tomato disease management is a challenging process, requiring models trained on the PlantVillage dataset achieved high degrees of
continual attention throughout the crop cycle and accounts for a sig­ accuracy when recognizing previously seen symptoms of a particular
nificant fraction of total production costs (Peet and Welles, 2005). disease (Mohanty et al., 2016; Brahimi et al., 2018). Studies using a
Earlier detection can help reduce the cost of treatment, lower the en­ reduced number of PlantVillage images and other smaller datasets also
vironmental impact of chemical inputs, and mitigate risks of yield loss. showed similarly high accuracy rates (Barbedo, 2018; Fuentes et al.,
Current disease detection techniques are limited by the time required 2017; Alfarisy et al., 2018). However, for datasets with classes of si­
for expert laborers to manually locate and assess disease, which is milar visual appearance, such as those of different tomato leaf disease,
complicated by the volume of plants found in commercial greenhouses smaller datasets are expected to result in lower-performing models.
and the small size of disease symptoms at their earliest stages. The Since the performance of models trained on progressively smaller da­
expense and time required typically limits disease scouting to an in­ tasets remains high, this may suggest that their images contained
frequent schedule or sparse sampling, which can miss early localized spuriously correlated features, extraneous to disease. In this paper, we
symptoms and have a significant impact on the severity of an outbreak. propose a new method for detecting diseases in a tomato greenhouse.
Investigations into automated detection methods have included The proposed method is focused on recognizing generic features that
molecular analysis, spectroscopy, and analysis of volatile organic are associated with diseases regardless of disease type. This enables
compounds but are expensive and impractical to apply at commercial detection of new instances that might not have been seen before.
operating scales (Martinelli et al., 2015). Studies using visible features Quantifying the level of disease is another important challenge in
imaged with conventional RGB cameras have shown the ability for integrated disease management. Since it is often impossible to com­
machine learning systems to recognize the presence of known plant pletely eradicate disease once an epidemic begins, treatment to de­
disease using deep convolutional neural network models (Mohanty crease its spread can mitigate its impact on yield. Without a quantifi­
et al., 2016; Alfarisy et al., 2018). Deep learning models typically re­ able measure, it is difficult to gauge the effectiveness of management
quire thousands of data points to accurately generalize predictions, practices or to predict and prepare for future epidemics. Measuring
while only small plant disease datasets were publicly available until the disease incidence or severity are two common methods for quantifying
introduction of the PlantVillage dataset (Hughes and Salathe, 2015) its degree. Incidence, the simpler of the two methods, is a measurement

⁎
Corresponding author.
E-mail addresses: [email protected] (P. Wspanialy), [email protected] (M. Moussa).

https://doi.org/10.1016/j.compag.2020.105701
Received 31 January 2020; Received in revised form 21 April 2020; Accepted 5 August 2020
0168-1699/ © 2020 Published by Elsevier B.V.
P. Wspanialy and M. Moussa Computers and Electronics in Agriculture 178 (2020) 105701

The two methods presented can help reduce associated labor costs
while increasing the accuracy and precision of disease scouting.
Additionally, it can enable larger automated scouting operations, which
can help reduce sample size bias.

2. Dataset

The primary source of data used in this paper were images from the
unmodified color version of tomato leaves in the PlantVillage dataset
(Hughes and Salathe, 2015). The dataset consists of images of single
leaves removed from their plants with naturally occurring or inoculated
Fig. 1. Illustrated difference between disease incidence and severity. The top disease. Images were taken with natural lighting against a grey back­
example shows three plants, two of which are exhibiting low levels of disease. ground near the sampling site. The subset of tomato images contains
The bottom example shows three plants, again with two exhibiting disease but
16415 diseased leaf images exhibiting nine different conditions and
at higher degrees. The severity measurement reflects this change, while in­
1590 healthy leaves, with their conditions classified by experienced
cidence does not.
observers. The original PlantVillage dataset was captured using a
twenty-megapixel camera from four to seven different orientations to
of disease frequency, usually at the level of individual plants. Severity, compensate for directional lighting variation. Each image was later
in contrast, is a more precise value, measuring the ratio of diseased area rotated to have the leaf tips pointed upwards. The version of the dataset
based on leaf coverage. Fig. 1 illustrates how the two measurements 1
used in this study had been scaled down to 256 × 256px and rotations
differ. Both techniques can be affected by errors in sampling, but se­ of the same leaf were removed. Example images from the ten classes of
verity introduces additional errors when estimating leaf coverage (Bock tomato leaves are shown in Fig. 3.
et al., 2010). Severity is often compared against standard area diagrams The original study of this dataset (Mohanty et al., 2016) in­
showing prototypical examples of symptomatic leaves (Barratt and vestigated the impact of color and background information on the
Horsfall, 1945; Sconyers et al., 2006). performance of classification models, showing that color was important
The surveying aids traditionally used by scouts to assess severity, for in improving accuracy, but that backgrounds did not have an effect.
example Fig. 2, do not produce results precise enough to discriminate When removing backgrounds, the authors simultaneously applied color
between significant differences in diseased area (Parker et al., 1995). correction to the leaf area which may have introduced bias if adjust­
Improvements to comparative diagrams were developed as a digital ments were not uniform across every class. Fig. 4 shows an example of
assistant but required users to manually select the color ranges asso­ the differences between the original color image and the segmented
ciated with disease in images for the symptomatic areas to be calculated version used in Mohanty et al. (2016).
(Pethybridge and Nelson, 2015). While interactive labeling is suitable
for smaller samples, it still requires significant manual input and is
3. Disease detection
impractical when applied to large samples of commercially sized
greenhouses. Another approach to estimating disease progress was in­
In order to explore the contribution image properties to disease
vestigated using a machine learning model to automatically classify
detection and to investigate detection methods that are agnostic to the
leaves into discreet early, middle, and end-stage classification of disease
type of disease, we tested several variations of the PlantVillage dataset.
(Wang et al., 2017). This technique provides more information than just
The original PlantVillage dataset was used to generate variations to
binary detection, but the resolution is not fine-grained enough for
investigate the contribution of leaf surface features, leaf shape, and
tracking small changes in disease progression. Current methods are
background content to the accuracy of disease classification. In order to
reliant on user input or provide a course account of severity, which
investigate the ability for a system to recognize previously unseen
cannot be used for monitoring large areas. In this paper, we propose a
disease, the dataset was split into nine different disease-healthy binary
new method of estimating severity by detecting and measuring the
subsets, each one leaving out one of the nine diseases. The corre­
proportion of symptomatic leaf area automatically.
sponding test-sets for each leave-one-out dataset contained only sam­
ples from the left-out disease. Each dataset variant was nominally split
1.1. Paper contribution
into 60% training, 5% validation, and 25% test subsets. The test subsets
were split evenly between classes, with excess samples moved to the
There are two primary contributions in this paper.
training dataset. A description of the dataset variations is shown in
Table 1, and example images are shown in Fig. 5.
1. We propose a disease detection model that learns to detect instances
In order to avoid overfitting, which can impact performance due to
of disease based on generic features. The algorithm utilizes the Res-
poor generalization beyond training samples, we will use several
Net deep learning architecture. Performance shows that it can detect
methods for data augmentation. Data augmentation is a technique used
new instances of diseases not seen before.
to artificially increase variations in images to reduce model overfitting.
2. We propose a new model for estimating disease severity. The model
Datasets were augmented during the training stage of each model to
utilizes a modified U-Net deep learning architecture to estimate the
reduce the effect of rotation, translation, size, and illumination. The
severity of disease based on a hybrid deep learning model.
following modifications were randomly applied to images: vertical and
horizontal flipping, scaling ± 20%, translation ± 20%, rotation ± 45
degrees, shear ± 20 degrees, modification of hue and saturation
by ± 20%, and a gaussian blur with a standard deviation between 0 and
3. Since augmentations were applied dynamically during training, the
number of dataset samples increased proportionally to training length.
This results in an effectively infinitely sized dataset, but with progres­
sively lower variation between samples as it grows.

Fig. 2. Example disease surveying aid Sconyers et al. (2006). 1

https://github.com/spMohanty/PlantVillage-Dataset.

2
P. Wspanialy and M. Moussa Computers and Electronics in Agriculture 178 (2020) 105701

Fig. 3. Examples of images from the ten classes found in the PlantVillage tomato dataset.

A residual network with 50 layers (ResNet-50), Fig. 6, was chosen as

it has shown to be a good generalized model, outperforming many other
architectures on the ImageNet challenge (He et al., 2015). Various
network architectures were tested, with results comparable to that of
ResNet-50, confirming the findings of Brahimi et al. (2018). The net­
work was pre-trained on the ImageNet dataset, consisting of millions of
images (Deng et al., 2009). This process trains the model to identify
Fig. 4. Arithmetic difference between color and segmented images used in
visual features common to all objects, including corners, lines, and
Mohanty et al. (2016). If the image was segmented without modification of the other patterns. The weights of this pre-trained model are then used, in a
leaf, the difference. would be empty. process called transfer learning, as the starting point for disease de­

Table 1
PlantVillage dataset variations used in this study.
Dataset Variation Description

Color The original 256 × 256 color images used in Mohanty et al. (2016). Nine classes of disease and one healthy class.
Color-masked The color version masked using the segmentations from Mohanty et al. (2016).
No color correction performed.
Silhouette The segmentation masks from Mohanty et al. (2016).
Background-only Segmentation masks from Mohanty et al. (2016), heavily dilated and blurred, thresholded, inverted, and applied to color images.
Color-masked-binary The color-masked dataset with only two classes, diseased and healthy.
Leave-one-out (9 variations) The color-masked-binary dataset but with one of the nine diseases left out. Nine variations, each with a different disease left out.

tection. The model’s output layer was resized from the 1000 classes
found in ImageNet to match the number of classes in each PlantVillage
dataset variation, and its weights were randomly assigned. For ex­
ample, when training the model to differentiate between each of the
nine different diseases and healthy leaves, the output size was 10
(9 + 1). When trained to distinguish between the binary option of
Fig. 5. Example images from dataset variations. healthy or diseased, the output size was 2. Models were trained using a

Fig. 6. Data-flow of ResNet model used to classify between one of the color-masked binary datasets. The structure is similar for models with mode than two classes,
with only an expansion of the final output layer.

3
P. Wspanialy and M. Moussa Computers and Electronics in Agriculture 178 (2020) 105701

Fig. 7. Confusion matrix for the color dataset. Fig. 9. Confusion matrix for the background-only dataset.

Fig. 8. Confusion matrix for the color-masked dataset. Fig. 10. Confusion matrix for the silhouette dataset.

stochastic gradient descent optimizer with a weighted loss2 to com­ Table 2

pensate for uneven distributions of samples in the training dataset, and PlantVillage dataset variation classification accuracies.
used a rectified linear unit (ReLU) for its activation function. Images Dataset Variation Classification accuracy
were grouped into batches of 32 for each pass through the network to
help generate weight updates representative of the entire dataset in­ Color 0.970
Color-masked 0.940
stead of individual samples. The training phase was divided into Background-only 0.635
epochs, each approximating a complete pass through the entire dataset. Silhouette 0.523
The first five epochs were trained with updates only to weights in the
final output layer. Initiating training without modifying the pre-trained
portion was intended to prevent large irrelevant updates to the

2 min (class weights )

class loss weight = class weight
.

4
P. Wspanialy and M. Moussa Computers and Electronics in Agriculture 178 (2020) 105701

Table 3
Leave-one-out dataset classification accuracies.
Left-out-disease dataset variation Classification accuracy

Baseline (color-masked-binary) 0.987

Bacterial Spot 0.994
Early Blight 0.997
Late Blight 0.961
Leaf Mold 0.997
Mosaic Virus 0.997
Septoria Leaf Spot 0.991
Spider Mites 0.788
Target Spot 0.774
Yellow Leaf Curl Virus 0.998

Fig. 12. Confusion matrices for the color-masked-binary and leave-one-out

datasets (continued).

Fig. 13. Confusion matrices for disease recognition on the Google “tomato leaf”
dataset 13.

To explore the effect of the background, the background-only da­

taset removed nearly all of the image content, leaving only small pat­
ches of background. The model trained on the background-only dataset
Fig. 11. Confusion matrices for the color-masked-binary and leave-one-out
still achieved a classification accuracy of 0.635, while the model trained
datasets.
on the binary mask of the leaves in the silhouette dataset scored 0.523.
Confusion matrices for the background-only and silhouette datasets are
transferred weights caused by the random assignment of the final layer. shown in Figs. 9 and 10. A summary of the results for each of the four
This was followed by fine-tuning the complete network for an addi­ datasets is shown in Table 2.
tional 25 epochs. The initial learning rate was set to 1 × 10 3 and was A further ten models were trained to explore the ability for a model
reduced by a factor of ten when loss reached a plateau. Plateaus were to generalize to symptoms beyond any conditions previously seen. Since
automatically detected when the loss stopped improving for three previously unseen symptoms cannot be placed into individual groups,
epochs.2 the color-masked dataset was split into only two classes - healthy and
disease. The color-masked version of the dataset was chosen to reduce
3.1. Results the bias introduced from background content. A baseline model was
trained, which included all the available disease classes from the ori­
A model was trained and tested for each of the dataset variations. ginal dataset into the disease meta-class, scoring an accuracy of 0.987.
The results for the original color dataset served as a baseline to compare Each of the nine other models were trained on a dataset which left-
with the variation datasets, and to explore how properties of the images out one of the diseases from the disease meta-class during training. The
contribute to the overall accuracy. When trained on the original color models were tested on images only from either previously unseen
dataset with ten classes, the classification accuracy on an evenly-split healthy samples or the unseen disease class. The classification ac­
test set was 0.970. With the background masked off, the classification curacies of the models were largely unchanged for seven of the nine
accuracy dropped to 0.940. Confusion matrices for the color and color- diseases, and decreased for two.
masked datasets are shown in Figs. 7 and 8. A summary of the results for each of the leave-one-out datasets is

5
P. Wspanialy and M. Moussa Computers and Electronics in Agriculture 178 (2020) 105701

Fig. 14. Annotation example of a leaf afflicted with late blight disease.

Table 4 annotated, 200 from each of the nine class. Each disease class dataset
Mean and standard deviation ( ) of disease severity for each tomato disease was spit into subsets for training, validation, and testing, at proportions
dataset. of 70%, 10%, and 20%, respectively. Image augmentation was ex­
Disease Mean Severity ( ) panded to create background invariance by randomly replacing back­
grounds with images from the Describable Textures Dataset
Early Blight (Alternaria solani) 0.118 (0.079) (Describable Textures Dataset). Annotations were made using a tool 3
Late Blight (Phytophthora infestans) 0.395 (0.229)
developed for annotating plant parts, producing results compatible with
Leaf Mold (Passalora fulva) 0.157 (0.121)
Septoria Leaf Spot (Septoria lycopersici) 0.119 (0.072)
the common objects in context (COCO) (Lin et al., 2014) data format.
Target Spot (Corynespora cassiicola) 0.038 (0.037) Image pixels were annotated into one of three classes; health, disease,
Bacterial Spot (Xanthomonas campestris pv. vesicatoria) 0.167 (0.141) and background. Table 4 shows the mean severity of each of the nine
Mosaic Virus 0.233 (0.109) conditions. A separate set of datasets were created for each disease class
Yellow Leaf Curl Virus 0.258 (0.111)
to evaluate the effects of their symptom expression on model perfor­
Spider Mites (Tetranychus urticae) 0.315 (0.153)
mance independently.

Fig. 15. Modified U-Net using VGG-16 as an image

feature backbone. Convolution with 3×3 kernels
and padding takes place between horizontally ad­
jacent blocks with ReLu activation. Max-pooling
with a 2×2 pooling window is done between sets of
vertical blocks in the down-path. The last step of
each lateral layer is concatenated to its corre­
sponding up-sampled block in the up-path. The
numbers to the left of the figure represent the
number of filters used.

shown in Table 3. Their corresponding confusion matrices are shown in A U-Net architecture, Fig. 15, originally used to segment a small
Figs. 11 and 12. dataset of microscopic images of cells was used to segment diseased leaf
In order to determine how well the models generalized, an external area. U-Net’s characteristic “U” shape architecture is built with a tra­
dataset was created by collecting the first 100 images found using a ditional convolutional neural network in its downsampling path and is
Google image search for “tomato leaf”. Leaves were manually seg­ followed by a mirrored upsampling step. At the end of upsampling, the
mented (see Fig. 14) and evaluated for health status using the model original resolution of the input is restored, assigning class labels it each
trained on the color-masked dataset. The Google dataset had an 80% pixel of the image. Ronneberger et al. (2015) This fine-grained class
bias towards diseased leaves, which mirrored the resulting confusion prediction is known as semantic segmentation. The model’s original
matrix show in Fig. 13, indicating poor generalization. downsampling step was replaced with the VGG16 (Simonyan and
Zisserman, 2014) architecture, functioning as an image feature ex­
4. Severity estimation tractor. Weights in the VGG16 portion were pre-trained on the Im­
ageNet dataset and weights in the upsampling path were randomly set
The goal of a disease severity estimation is to precisely and accu­ with a Xavier initializer (Glorot and Bengio, 2014). Models for each
rately measure the proportion of leaf area showing disease symptoms, disease dataset were trained for a total of 30 epochs. The first ten
or epochs only adjusted the weights in the upsampling path, since gradient
updates from the random weights could disrupt the pre-trained Im­
diseased area
severity = ageNet weights of the downsampled path. The final 20 epochs were
total leaf area (1) trained using the complete set of weights. Training loss was weighted in
The PlantVillage dataset contains several thousand samples classi­
fied by disease but without annotation of diseased area. The process of
segmenting images is labor-intensive, and so only 1800 images were 3
https://github.com/uoguelph-ri/coco-annotator.

6
P. Wspanialy and M. Moussa Computers and Electronics in Agriculture 178 (2020) 105701

Fig. 16. Example outputs of the disease severity estimation model for early and late blight, leaf mold, and Septoria leaf spot. The input images are in the first column,
followed by ground truth annotations, and model predictions. Numbers in parentheses are the ground truth and predicted severity estimations.

the same manner as that for disease detection to compensate for the 4.1. Results
unequal amounts of pixels belonging to the diseased, healthy, and
background classes. Severity semantic segmentation models were trained on each of the
The performance of disease segmentation was calculated using the individual disease’s training datasets and tested on their corresponding
Jaccard index, also known as the intersection over union (IoU) mea­ test sets. Figs. 16 and 17 show example predictions for each of the nine
sure. The index is defined as the ratio between the intersection of the diseases. Disease severity estimation performance was measured using
model prediction and the truth, and their union and used to measure the absolute difference between ground truth and predicted values of
their similarity, Eq. (2). The predicted area is generated by the computer severity, i.e. severity error. This measure only considers how accurate
model after being trained, while the truth area is labeled manually by a the severity estimation was and does not consider if the specific areas
human. were marked correctly, unlike the Jaccard index. Mean performances of
each model measured using the Jaccard index, and severity error are
predicted truth shown Table 5.
J (predicted, ground truth) =
predicted truth (2)

5. Discussion

The original PlantVillage dataset was collected over several years

7
P. Wspanialy and M. Moussa Computers and Electronics in Agriculture 178 (2020) 105701

and in multiple locations which likely contributed bias into disease

classes, particularly in the background areas. Further bias may have
been introduced by using consistent intra-class lighting and camera
settings, creating correlated features extraneous of disease symptoms.
The results showed that disease detection fell when backgrounds were
removed and that models trained on only background portions still
scored significantly higher than the expected value of 0.1 for an evenly-
split ten-class dataset of random data, indicating that the background
area was strongly biased towards the disease class of the leaf.
The silhouette dataset showed that leaf shape is also correlated with
disease, with detection results again well above random chance. This
suggests that the shape of diseased leaves can provide enough in­
formation about their condition to make it possible to detect disease
with symptoms on the underside of leaves, or with poor lighting of leaf
surface.
The trained binary disease-health model performed with higher
accuracy than the original ten-class version, as would be expected with
fewer output classes. In the leave-one-out disease datasets, the two
types of disease performing the poorest, spider mites, and target spot,
are closest in appearance to those of healthy leaves. In these two cases,
the models showed no ability to discriminate between the diseased and
healthy leaves, nearly always predicting the healthy class. With the
exceptions of very subtle disease symptoms, these results demonstrate
the feasibility of detecting previously unseen diseases of tomato leaves
with a model trained on a general disease-healthy dataset.
When evaluated images external to the PlantVillage dataset, model
performance indicated that there was not enough data variation to
generalize to diseased leaf examples outside of the PlantVillage dataset.
The original study using PlantVillage showed that using only twenty
percent of the data still produced well-performing systems, even
without transfer learning from larger datasets (Mohanty et al., 2016).
This indicates that there was likely a lack of variation in the diseased
leaf samples. Other studies using smaller datasets also showed high
accuracy when tested on data similar to that which was used during
training (Barbedo, 2018; Fuentes et al., 2017; Alfarisy et al., 2018). A
larger, more diverse dataset would help in the training of more gen­
eralizable models by reducing correlations within classes of features
unrelated to symptoms.
Models trained for semantic segmentation for disease severity pre­
dictions generally produced estimates higher than the ground truth,
with a severity error of 11.8% which is comparable to human evalua­
tion using standard area diagrams (Nutter and Esker, 2006). Severity
estimation using proportional area worked best when applied to dis­
Fig. 17. Example outputs of the disease severity estimation model for mosaic crete, localized symptoms such as those caused by fungi. Proportional
and yellow leaf curl virus, bacterial spot, target spot, and spider mites. The area was less suitable for conditions with systemic symptoms, such as
input images are in the first column, followed by ground truth annotations, and those caused by viruses. Virus symptoms include progressive chlorosis,
model predictions. Numbers in parentheses are the ground truth and predicted stunting, and leaf curl affecting large portions simultaneously and
severity estimations. uniformly. Bacterial spot, with symptoms similar to those of fungi, also
worked well. Spider mite symptoms are more evenly distributed on
Table 5 leaves, like those of viruses, which made area estimation less effective.
Mean Jaccard Index and severity error for each tomato disease. Standard de­ Ordinal or interval class scores should be used instead of proportional
viation ( ) is shown in parentheses. area when symptoms are systemic.

Disease Mean Jaccard ( ) Mean Severity Error ( )

6. Conclusions
Early Blight 0.506 (0.140) 0.068 (0.065)
Late Blight 0.606 (0.115) 0.127 (0.090) This study showed three important results. The shape of leaves
Leaf Mold 0.437 (0.153) 0.113 (0.145) alone have predictive power for disease recognition, a general disease
Septoria Leaf Spot 0.487 (0.089) 0.060 (0.062)
model can be trained to recognize previously unseen tomato leaf dis­
Target Spot 0.243 (0.103) 0.072 (0.052)
Bacterial Spot 0.408 (0.178) 0.092 (0.090) eases, and disease severity estimation can achieve results comparable to
Mosaic Virus 0.434 (0.228) 0.150 (0.154) those of human evaluation. With the understanding of the limitations of
Yellow Leaf Curl Virus 0.395 (0.130) 0.216 (0.145) the PlantVillage dataset and its image capturing techniques, future data
Spider Mites 0.217 (0.211) 0.169 (0.066)
collection initiatives can improve the variation of their datasets by re­
ducing accidental bias, particularly in their backgrounds. The general
disease-healthy binary model can, in effect, be interpreted as a healthy-
leaf recognizer. Such a model can help reduce the effort required in
building a comprehensive dataset to detect general disease, and has the

8
P. Wspanialy and M. Moussa Computers and Electronics in Agriculture 178 (2020) 105701

ability to recognize diseases not trained on before. imaging. Crit. Rev. Plant Sci. 29 (2), 59–107. https://doi.org/10.1080/
Additionally, this paper introduced a new annotated proportional 07352681003617285.
Brahimi, M., Arsenovic, M., Laraba, S., Sladojevic, S., Boukhalfa, K., Moussaoui, A., 2018.
disease severity dataset and associated estimation model. Fungal and Deep learning for plant diseases: detection and saliency map visualisation. In: Zhou,
bacterial diseases are most suitable for disease severity estimation using J., Chen, F. (Eds.), Human and Machine Learning: Visible, Explainable, Trustworthy
proportional area measures, while ordinal categories are more suitable and Transparent, Human-Computer Interaction Series, Springer International
Publishing, Cham, 2018, pp. 93–117. doi:10.1007/978-3-319-90403-0_6.
for systemic diseases like those caused by viruses and insects. The re­ Deng, J., Dong, W., Socher, R., Li, L., Li, K., Fei-Fei, L., 2009. ImageNet: A large-scale
sults of this study can be applied practically by integrating the models hierarchical image database. In: 2009 IEEE Conference on Computer Vision and
into an automated surveying system, reducing costs and measurement Pattern Recognition, pp. 248–255. https://doi.org/10.1109/CVPR.2009.5206848.
Describable Textures Dataset, https://www.robots.ox.ac.uk/ vgg/data/dtd/.
bias, and increasing precision and greenhouse coverage. Fuentes, A., Yoon, S., Kim, S.C., Park, D.S., 2017. A robust deep-learning-based detector
for real-time tomato plant diseases and pests recognition. Sensors (Basel,
CRediT authorship contribution statement Switzerland) 17(9). doi:10.3390/s17092022.
Glorot, X., Bengio, Y., 2014. Understanding the difficulty of training deep feedforward
neural networks 8.
Patrick Wspanialy: Conceptualization, Methodology, Software, He, K., Zhang, X., Ren, S., Sun, J., 2015. Deep Residual Learning for Image Recognition,
Validation, Formal analysis, Investigation, Resources, Data curation, arXiv:1512.03385 [cs]arXiv:1512.03385.
Writing - original draft, Writing - review & editing, Visualization. Hughes, D.P., Salathe, M., 2015. An open access repository of images on plant health to
enable the development of mobile disease diagnostics, arXiv:1511.08060 [cs]
Medhat Moussa: Conceptualization, Resources, Writing - review & arXiv:1511.08060.
editing, Supervision, Project administration, Funding acquisition. Lin, T.-Y., Maire, M., Belongie, S., Bourdev, L., Girshick, R., Hays, J., Perona, P.,
Ramanan, D., Zitnick, C.L., Dollár, P., 2014. Microsoft COCO: Common Objects in
Context, arXiv:1405.0312 [cs]arXiv:1405.0312.
Declaration of Competing Interest Martinelli, F., Scalenghe, R., Davino, S., Panno, S., Scuderi, G., Ruisi, P., Villa, P.,
Stroppiana, D., Boschetti, M., Goulart, L.R., Davis, C.E., Dandekar, A.M., 2015.
Advanced methods of plant disease detection. A review. Agronomy Sustain. Develop.
The authors declare that they have no known competing financial 35 (1), 1–25. https://doi.org/10.1007/s13593-014-0246-1.
interests or personal relationships that could have appeared to influ­ Mohanty, S.P., Hughes, D.P., Salathé, M., 2016. Using deep learning for image-based
ence the work reported in this paper. plant disease detection. Front. Plant Sci. 7. doi:10.3389/fpls.2016.01419.
Nutter, F.W., Esker, P.D., 2006. The role of psychophysics in phytopathology: the weber-
fechner law revisited. Eur. J. Plant Pathol. 114 (2), 199–213. https://doi.org/10.
Appendix A. Supplementary material 1007/s10658-005-4732-9.
Parker, S.R., Shaw, M.W., Royle, D.J., 1995. The reliability of visual estimates of disease
severity on cereal leaves. Plant. Pathol. 44 (5), 856–864. https://doi.org/10.1111/j.
Supplementary data associated with this article can be found, in the
1365-3059.1995.tb02745.x.
online version, at https://doi.org/10.1016/j.compag.2020.105701. Peet, M.M., Welles, G., 2005. Greenhouse Tomato Production 48.
Pethybridge, S.J., Nelson, S.C., 2015. Leaf Doctor: a new portable application for quan­
References tifying plant disease severity. Plant Dis. 99 (10), 1310–1316. https://doi.org/10.
1094/PDIS-03-15-0319-RE.
Ronneberger, O., Fischer, P., Brox, T., 2015. U-Net: Convolutional Networks for
Alfarisy, A.A., Chen, Q., Guo, M., 2018. Deep Learning based classification for paddy pests Biomedical Image Segmentation, arXiv:1505.04597 [cs]arXiv:1505.04597.
& diseases recognition. In: Proceedings of 2018 International Conference on Sconyers, L.E., Kemerait, R.C., Brock, J., Phillips, D.V., Jost, P.H., Sikora, E.J., Gutierrez-
Mathematics and Artificial Intelligence, ICMAI ’18, ACM, New York, NY, USA, 2018, Estrada, A., Mueller, J.D., Marois, J.J., Wright, D.L., et al., 2006. Asian soybean rust
pp. 21–25. doi:10.1145/3208788.3208795. development in 2005: A perspective from the Southeastern United States. APSnet
Barbedo, J.G.A., 2018. Impact of dataset size and variety on the effectiveness of deep Features.
learning and transfer learning for plant disease classification. Comput. Electron. Simonyan, K., Zisserman, A., 2014. Very Deep Convolutional Networks for Large-Scale
Agric. 153 (Complete), 46–53. https://doi.org/10.1016/j.compag.2018.08.013. Image Recognition, arXiv:1409.1556 [cs]arXiv:1409.1556.
Barratt, R., Horsfall, J., 1945. An improved grading system for measuring plant disease. Wang, G., Sun, Y., Wang, J., 2017. Automatic Image-Based Plant Disease Severity
Phytopathology 35, 655. Estimation Using Deep Learning, https://www.hindawi.com/journals/cin/2017/
Bock, C.H., Poole, G.H., Parker, P.E., Gottwald, T.R., 2010. Plant disease severity esti­ 2917536/abs/ (2017). doi:10.1155/2017/2917536.
mated visually, by digital photography and image analysis, and by hyperspectral