PlantDoc: A Dataset for Visual Plant Disease Detection

Davinder Singh*, Naman Jain*, Pranjali Jain*, Pratik Kayal*

Sudhakar Kumawat, Nipun Batra
Indian Institute of Technology Gandhinagar, Gujarat, India 382 355
{davinder.singh,naman.j,pranjali.jain,pratik.kayal,sudhakar.kumawat,nipun.batra}@iitgn.ac.in

ABSTRACT food security and have broad economic, social, and environmental
India loses 35% of the annual crop yield due to plant diseases. Early impacts [5].
detection of plant diseases remains difficult due to the lack of lab Timely disease detection in plants remains a challenging task
infrastructure and expertise. In this paper, we explore the possibility for farmers. They do not have many options other than consulting
of computer vision approaches for scalable and early plant disease fellow farmers or the Kisan helpline [17]. Expertise in plant diseases
detection. The lack of availability of sufficiently large-scale non-lab is necessary for an individual to be able to identify the diseased
data set remains a major challenge for enabling vision based plant leaves. Furthermore, in most cases it is necessary to have a lab
disease detection. Against this background, we present PlantDoc: infrastructure to identify a diseased leaf.
a dataset for visual plant disease detection. Our dataset contains In this work, we explore the possibility of using computer vision
2,598 data points in total across 13 plant species and up to 17 classes for scalable and cost-effective plant disease detection. Computer
of diseases, involving approximately 300 human hours of effort in vision has made tremendous advances in the past few years through
annotating internet scraped images. To show the efficacy of our various advances in deep convolutional neural networks. While
dataset, we learn 3 models for the task of plant disease classification. training large neural networks can be very time consuming, the
Our results show that modelling using our dataset can increase the trained models can classify images very quickly, which makes them
classification accuracy by up to 31%. We believe that our dataset also suitable for consumer applications on smartphones. Image
can help reduce the entry barrier of computer vision techniques in processing for detecting plant diseases opens up new avenues to
plant disease detection. combine the knowledge of deep learning approaches with real-
world problems in agriculture, and hence, facilitates advancements
KEYWORDS in agricultural knowledge, the yield of crops, and disease control.
Majority of existing vision-based solutions require high-
Deep Learning, Object Detection, Image Classification
resolution images with a plain background. In contrast, as the
ACM Reference Format: majority of Indian farmers use low-end mobile devices with natural
Davinder Singh*, Naman Jain*, Pranjali Jain*, Pratik Kayal* and Sudhakar background and lighting conditions, we focus on images in natural
Kumawat, Nipun Batra. 2020. PlantDoc: A Dataset for Visual Plant Disease environmental conditions with non-trivial background noise and
Detection. In 7th ACM IKDD CoDS and 25th COMAD (CoDS COMAD 2020),
provide the best possible query resolution for crops and plants.
January 5–7, 2020, Hyderabad, India. ACM, New York, NY, USA, 5 pages.
https://doi.org/10.1145/3371158.3371196
Against this background, we highlight our two main contributions:
i) development of PlantDoc: a dataset of 2,598 images across 13
plant species and 27 classes(17-10, disease-healthy) ii) benchmark-
1 INTRODUCTION
ing the curated data set and showing its utility in disease detection
Annually the Earth’s population increases by about 1.6%, and so in non-controlled environments. To the best of our knowledge,
does the demand for plant products of every kind [16]. The pro- this is the first such dataset containing data from non-controlled
tection of crops against plant diseases has a vital role to play in settings.
meeting the growing demand for food quality and quantity [22]. We evaluated our dataset using various classification and object
In terms of economic value, plant diseases alone cost the global detection architectures mentioned in Section 4 to establish the
economy around US$220 billion annually [1]. According to the requirement of a dataset in non-controlled settings. The results
Indian Council of Agricultural Research, more than 35% of crop suggested that lab-controlled dataset cannot be used to classify or
production is lost every year due to Pests and Disease [15]. Food detect images in real-scenario. We found that fine-tuning the models
security is threatened by an alarming increase in the number of on PlantDoc reduces the classification error by up to 31%. Thus,
outbreaks of pests and plant diseases. These diseases jeopardize our dataset can potentially be used to build an application which
*Equal Contribution.
detects and classifies 27 plant disease/healthy classes efficiently.

Permission to make digital or hard copies of all or part of this work for personal or 2 RELATED WORK
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation Our related work can be broadly categorized into: i) techniques for
on the first page. Copyrights for components of this work owned by others than ACM plant disease detection; and ii) datasets advancing research in plant
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, disease detection.
to post on servers or to redistribute to lists, requires prior specific permission and/or a
fee. Request permissions from [email protected].
CoDS COMAD 2020, January 5–7, 2020, Hyderabad, India 2.1 Techniques for plant disease detection
© 2020 Association for Computing Machinery.
ACM ISBN 978-1-4503-7738-6/20/01. . . $15.00 Prior work by Sankaran et al. [19] proposed using reliable sensors
https://doi.org/10.1145/3371158.3371196 for monitoring health and diseases in plants under field conditions.

249
CoDS COMAD 2020, January 5–7, 2020, Hyderabad, India D. Singh*, N. Jain*, P. Jain*, P. Kayal*, S. Kumawat, N. Batra

Apple Bell Pepper Blueberry Cherry Corn Grape Potato

Black Rot Bacterial Healthy Powdery Mildew Gray Spots Black Rot Early Blight

PVD

PlantDoc

Figure 1: Samples from various classes in the PlantDoc Dataset show the gap between lab-controlled and real-life images

However, plant disease detection using sensors has the potential to 200
benefit only a few farmers because of the substantial hardware cost
175
and lack of expertise to operate such sensors. In contrast, prior work
150

Frequency of Images
by Patil et al. [18] extracted shape features for disease detection in
sugarcane leaves obtaining a final average accuracy of 98.60%. In a 125
similar work, Patil et al. [3] used texture features, namely inertia,
homogeneity, and correlation obtained by calculating the gray level 100
co-occurrence matrix on the image and color extraction for disease 75
detection on maize leaves. Recent work [8] has looked into neural
50
networks for the identification of three different legume species
based on the morphological patterns of leaves veins. Likewise, fea- 25
ture extraction and Neural Network Ensemble (NNE) have been 0
used for recognizing tea leaf diseases with a final testing accuracy

Corn GraCorn eBaliltghhyt

Grape CBolarn Rpuostt

RaostapbtoerLrate Bliligghhtt
App althy
Bell PeppApplele SRcuasbt
Healthy

Late lthy
Grape Hcek Rot
Peach Healthy

Tomato TMomatoBMligohldt
Early Balthy

maatorlyHeBalight

S toriaosaic Virus
us
BluebeerrryLeaf Spoyt

owdery HMealthhyy
Bacteriael althy
Spot
awberry H ildew

YellLoewafVSirpot
of 91% [25]. A host of other recent works have looked at convolu- Bell Pepp er Health

ealt
y Leaf S
Apple He

tional neural network variants for disease detection using plant leaf Cherry H

SquashSoPyabeany H
images [7, 21]. These works are limited to a particular crop, which

E
is a significant limitation. Also, the datasets used in the works have

TomTaoto
Potato

Tomato

Tomeapto
not been made public, thereby, impacting reproducibility.

Tomato
tr
P

Tomato
2.2 Datasets for plant disease detection
The PlantVillage dataset(PVD) [14] is the only public dataset for
plant disease detection to the best of our knowledge. The data set Leaf Classes
curators created an automated system using GoogleNet [23] and
AlexNet [12] for disease detection, achieving an accuracy of 99.35%. Figure 2: Statistics of PlantDoc Dataset
However, the images in PlantVillage dataset are taken in laboratory
setups and not in the real conditions of cultivation fields, due to
which their efficacy in real world is likely to be poor. In contrast,
accurate plant disease detection in the farm setting. We downloaded
we curate real-life images of healthy and diseased plants to create
images from the internet since collecting large-scale plant disease
a publicly available dataset.
data through fieldwork requires enormous effort. We collected
about 20,900 images by using scientific and common names of 38
3 THE PLANTDOC DATASET classes mentioned in the dataset by Mohanty et al. [14].
The PlantVillage dataset contains images taken under controlled Four users filtered the images by selecting images based on their
settings. This dataset limits the effectiveness of detecting diseases metadata on the website and guidelines mentioned on APSNet [2].
because, in reality, plant images may contain multiple leaves with APS compiled a list of peer-reviewed literature corresponding to
different types of background conditions with varying lighting each plant disease. We referred APS’ prior literature and accord-
conditions (shown in Figure 1). Against this background, we now ingly classified images. Some of the most important factors for
describe our curated dataset and discuss the techniques used for classification were the color, area and density of the diseased part
curation. and shape of the species. We removed inappropriate (such as non-
leaf plant, lab controlled and out-of-scope images) and duplicate
3.1 Data Collection images across classes downloaded due to web search. Every image
To account for the intricacies of the real world, we require models was checked by two individuals according to the guidelines to re-
trained on real-life images. This fact motivated us to create a dataset duce labeling errors. Finally, to have sufficient training samples, we
by downloading images from Google Images and Ecosia [6] for removed the classes with less than 50 images. Figure 2 shows the

250
PlantDoc: A Dataset for Visual Plant Disease Detection CoDS COMAD 2020, January 5–7, 2020, Hyderabad, India

Model PreTrained Weights Training Set Test Set Accuracy F1-Score

(Set %) (Set %)
VGG16 ImageNet C-PD (80) C-PD(20) 44.52 0.44
VGG16 ImageNet PVD C-PD (100) 19.73 0.18
VGG16 ImageNet+PVD C-PD (80) C-PD (20) 60.41 0.60
InceptionV3 ImageNet C-PD (80) C-PD (20) 46.67 0.46
InceptionV3 ImageNet PVD C-PD (100) 30.78 0.28
InceptionV3 ImageNet+PVD C-PD (80) C-PD (20) 62.06 0.61
InceptionResNet V2 ImageNet C-PD (80) C-PD (20) 49.04 0.49
InceptionResNet V2 ImageNet PVD C-PD (100) 39.87 0.38
InceptionResNet V2 ImageNet+PVD C-PD (80) C-PD (20) 70.53 0.70

Table 2: Training on controlled dataset (PlantVillage - PVD)

gives poor performance on real world images. Performance
Figure 3: An Image with bounding boxes and its cropped
on real world images can be improved by training on real
leaves
world images from our dataset

statistics of the final dataset having a total of 27 classes spanning

over 13 species with 2,598 images. 4.2 Plant image classification
To build an application for the object detection task, we need Our main goal was to construct a model which can detect a leaf in
exact bounding regions containing the leaf in the entire image. an image and then classify it into the particular classes shown in
Hence, we used the LabelImg tool [24] to make the bounding boxes Figure 2. We performed two main experiments, which we discuss
around the leaves (Figure 3) in all the images. In real scenarios, the after describing our experimental settings.
image may have multiple leaves or a combination of diseased and
healthy leaves. We labeled all the leaves in the image explicitly with 4.2.1 Experimental settings. For training the networks, we used
their particular classes. While labeling the boxes, we made sure stochastic gradient descent with momentum 0.9, categorical cross-
that the entire leaf should be present inside the box and the area of entropy loss, and a learning rate of 0.001. All weights were ini-
the bounding box should not be smaller than 1/8th (approximately) tialized with the orthogonal initializer. We applied common data
of the image size. After labeling, the information about all the augmentation techniques such as rotation, scaling, flipping etc. on
coordinates of boxes in an image and their respective class label the input images. All images were resized to 100 × 100, before feed-
were stored separately in an XML file corresponding to each image. ing into the networks. For pre-trained models, we used the weights
Cropped-PlantDoc Dataset: To show the differences between provided in Keras trained on ImageNet.
our dataset and PlantVillage, we built another dataset called the 4.2.2 Plant image classification using raw images (uncropped). Our
Cropped-PlantDoc (C-PD) by cropping the images using bounding first experiments aims to understand classification accuracy on
box information. Similar to PlantVillage, cropped images contains the uncropped PlantDoc dataset. We evaluated the performance of
only the leaf but these images are of low-quality, have small-size VGG16 [20] using different training sets on PlantDoc as shown in
and varying backgrounds. The total number of leaf images after Table 1.
cropping 2,598 images turns out to be 9,216 i.e. 9,216 bounding
boxes. 4.2.3 Plant image classification using cropped images. Further, we
evaluate the performance of several popular CNN architectures on
4 BENCHMARKING PLANTDOC DATASET the Cropped-PlantDoc dataset that have recently achieved state-
We now discuss two benchmark set of experiments on our dataset: of-the-art accuracies on image classification tasks on the popular
i) plant image classification; and ii) detecting leaf within an image. datasets, such as ImageNet [4], CIFAR-10 [11], etc. Table 2 gives the
complete list of the architectures that we used for benchmarking
our Cropped-PlantDoc dataset. This experiment was conducted to
PreTrained Weights Training Set Test Set Accuracy F1-Score
(Set %) (Set %) verify the performance of PlantVillage in real-setting.
ImageNet PlantDoc (80) PlantDoc (20) 13.74 0.12
ImageNet PVD PlantDoc (100) 15.08 0.15 4.3 Leaf Detection
ImageNet+PVD PlantDoc (80) PlantDoc (20) 29.73 0.28
The aim of our next experiments is to evaluate the performance
Table 1: Transfer Learning doubled the accuracy after fine- of Faster R-CNN with InceptionResnetV2 model and MobileNet
tuning on Uncropped PlantDoc dataset model on our PlantDoc Dataset as shown in Table 3. We use mean
average precision (mAP: higher is better) to evaluate the models
and compare it with scores on COCO dataset since no evaluation
4.1 System configuration exists in the domain of plant disease.
All our experiments used NVidia V100 GPU with a system of 32 GB 4.3.1 Experimental Setting. Object Detection models require train-
RAM and 8 CPU cores. We used Keras with Tensorflow backend ing for a much longer duration. For training Faster R-CNN with
as the deep learning framework. We plan to make the fully repro- Inception Resnet v2 network, we used Momentum optimizer keep-
ducible Github repository1, 2 for code and dataset public at the time ing a degrading learning rate with an initial value of 0.0006. For
of camera ready. training the MobileNet network, with RMSprop as optimizer – we
1 https://github.com/pratikkayal/PlantDoc-Object-Detection-Dataset took an initial learning rate of 0.0005 with decay steps as 25000 and
2 https://github.com/pratikkayal/PlantDoc-Dataset decay factor as 0.95. While training, data augmentation like random

251
CoDS COMAD 2020, January 5–7, 2020, Hyderabad, India D. Singh*, N. Jain*, P. Jain*, P. Kayal*, S. Kumawat, N. Batra

Model PreTrained Weights mAP (at 50% iou)

MobileNet COCO 32.8
MobileNet COCO+PVD 22.4
Faster-rcnn-inception-resnet iNaturalist 36.1
Faster-rcnn-inception-resnet COCO 38.9
Table 3: Leaf detection mAP
∗ iou refers to intersection over union

Figure 5: Leaf detection results in our mobile application

(a) (b)

Figure 6: Tomato Bacterial leaf spot(a)and Septoria(b) looks

similar and are hard to label using visual features alone
Figure 4: Saliency and Activation Maps shows the affected
parts of disease in a leaf.
6 APPLICATION BUILDING
We were able to adapt the above solution to a mobile environment
(Figure 5) by using models that very significant reduce complexity,
horizontal flip and random SSD crop was applied on input images. without sacrificing the effective accuracy. This allowed us to achieve
We split our dataset into 2,360-238 based on training-testing. We the best possible performance, given that the application should
took the pre-trained weights and fine-tuned on training set of Plant- predict the bounding boxes and classes in real time in a mobile CPU.
Doc. As aforementioned, we provide train-test splits of the dataset We have build application that utilizes MobileNets Object Detection
for consistent evaluation and fair comparison over the dataset in Network [10] due to its efficiency and competitive accuracy. The
future. network builds on top of the SSD framework [13].

5 RESULTS AND DISCUSSION 7 LIMITATIONS

Figure 4 shows saliency map and gradient activation map of the The dataset has been curated with care, but due to lack of extensive
Corn Leaf Blight and Tomato Bacterial Spots respectively. As ex- domain expertise, there are some images in the dataset which can
pected, the neural network is learning to focus on the set of visual potentially be wrongly classified (shown in Figure 6). Further, to
features which are correlated with disease such as the blemishes train highly accurate models for disease detection, we may require
in the leaf (lines in Corn Leaf Blight and spots in Tomato Bacterial a dataset with more number of images in each class. But, due to
Spots). The network even learns the shape of the leaf (shown in non-availability of public dataset and lack of real-life scenario for
second row of Figure 4) to help it distinguish between species. field work, our approach gives a feasible direction to tackle the
As predicted, the results in Table 1 clearly shows that real case on-going problem of disease detection.
scenarios have low accuracy when processed initially with Ima-
geNet or PlantVillage. Also, Table 1 and Table 2 clearly shows low 8 CONCLUSIONS AND FUTURE WORK
accuracy achieved by training on PlantVillage and testing on Plant- In this paper, we addressed the problem of detection of dis-
Doc. Model fails to produce accurate results due to background eased/healthy leaves in images using state of the art object detection
noise, images with leaf from multiple classes in a dataset and low- models. One of the main contributions of our work is to propose
resolution leaf images. an entirely new dataset for plant disease detection called PlantDoc.
Table 3 shows that Faster R-CNN with InceptionResnetV2 per- Our benchmark experiments show the lack of efficacy of models
forms the best with an mAP of 38.9. It is interesting to see learnt on controlled datasets, thereby, showing the significance of
that MobileNet performance is decreased when pre-trained on real-world datasets such as ours. Applying image segmentation
COCO+PlantVillage compared to the model where pre-training techniques to extract leaf out of the images can potentially enhance
was done only on COCO. This attributes to the fact that PlantVil- the utility of the dataset. We believe that this dataset is an important
lage is not contributing towards better results. MobileNet gives an first step towards computer vision enabled scalable plant disease
mAP of 22 when evaluated on COCO dataset which has significantly detection.
more classes [9].

252
PlantDoc: A Dataset for Visual Plant Disease Detection CoDS COMAD 2020, January 5–7, 2020, Hyderabad, India

REFERENCES [13] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed,
[1] GN Agrios. 2005. Plant pathology 5th Edition: Elsevier Academic Press. Burling- Cheng-Yang Fu, and Alexander C Berg. 2016. Ssd: Single shot multibox detector.
ton, Ma. USA (2005), 79–103. In European conference on computer vision. Springer, 21–37.
[2] APSNet. 2019. Resources for Plant Diseases. https://www.apsnet.org/edcenter/ [14] Sharada P Mohanty, David P Hughes, and Marcel Salathé. 2016. Using deep
resources/commonnames/Pages/default.aspx learning for image-based plant disease detection. Frontiers in plant science 7
[3] Sanjay B Patil, K Shrikant, and Bodhe . 2011. Betel Leaf Area Measurement Using (2016), 1419.
Image Processing. International Journal on Computer Science and Engineering [15] T Mohapatra. 2018. ICAR News July-September 2018. Pub-
(IJCSE) 3 (01 2011). lished in monthly newsletter, https://www.icar.org.in/sites/default/files/
[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. 2009. Imagenet: ICARNewsJulySeptember2018.pdf.
A large-scale hierarchical image database. In 2009 IEEE conference on computer [16] E-C Oerke, H-W Dehne, Fritz Schönbeck, and Adolf Weber. 2012. Crop production
vision and pattern recognition. Ieee, 248–255. and crop protection: estimated losses in major food and cash crops. Elsevier.
[5] Messe Düsseldorf. [n. d.]. SAVE FOOD. ([n. d.]). https://www.messe-duesseldorf. [17] Government of India. 2019. Kisan Knowledge Management System. https:
com/cgi-bin/md_home/lib/pub/tt.cgi/SAVE_FOOD.html?oid=121&lang=2& //dackkms.gov.in/account/login.aspx
ticket=g_u_e_s_t [18] Sanjay B Patil and Shrikant K Bodhe. 2011. Leaf disease severity measurement
[6] Ecosia. 2019. Search Engine. https://www.ecosia.org/?c=en using image processing. International Journal of Engineering and Technology 3, 5
[7] Alvaro Fuentes, Sook Yoon, Sang Kim, and Dong Park. 2017. A robust deep- (2011), 297–301.
learning-based detector for real-time tomato plant diseases and pests recognition. [19] Sindhuja Sankaran, Ashish Mishra, Reza Ehsani, and Cristina Davis. 2010. A
Sensors 17, 9 (2017), 2022. review of advanced techniques for detecting plant diseases. Computers and
[8] Guillermo L Grinblat, Lucas C Uzal, Mónica G Larese, and Pablo M Granitto. Electronics in Agriculture 72, 1 (2010), 1–13.
2016. Deep learning for plant identification using vein morphological patterns. [20] Karen Simonyan and Andrew Zisserman. 2014. Very deep convolutional networks
Computers and Electronics in Agriculture 127 (2016), 418–424. for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014).
[9] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun [21] Srdjan Sladojevic, Marko Arsenovic, Andras Anderla, Dubravko Culibrk, and
Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. 2017. Mobilenets: Darko Stefanovic. 2016. Deep neural networks based recognition of plant diseases
Efficient convolutional neural networks for mobile vision applications. arXiv by leaf image classification. Computational intelligence and neuroscience 2016
preprint arXiv:1704.04861 (2017). (2016).
[10] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun [22] Richard N Strange and Peter R Scott. 2005. Plant disease: a threat to global food
Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. 2017. MobileNets: security. Annu. Rev. Phytopathol. 43 (2005), 83–116.
Efficient Convolutional Neural Networks for Mobile Vision Applications. CoRR [23] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir
abs/1704.04861 (2017). arXiv:1704.04861 http://arxiv.org/abs/1704.04861 Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. 2015.
[11] Alex Krizhevsky et al. 2009. Learning multiple layers of features from tiny images. Going deeper with convolutions. In Proceedings of the IEEE conference on computer
Technical Report. Citeseer. vision and pattern recognition. 1–9.
[12] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. 2012. Imagenet classifica- [24] Tzutalin. 2015. LabelImg. Free Software: MIT License. https://github.com/
tion with deep convolutional neural networks. In Advances in neural information tzutalin/labelImg
processing systems. 1097–1105. [25] Zhi-Hua Zhou and SF Chen. 2002. Neural network ensemble. CHINESE JOURNAL
OF COMPUTERS-CHINESE EDITION- 25, 1 (2002), 1–8.

253