Cell Phone Image Based Plant Disease Classification

See discussions, stats, and author profiles for this publication at: https://www.researchgate.
net/publication/306234913
Cell phone image-based plant disease classification
Chapter · October 2015

DOI: 10.4018/978-1-4666-9435-4.ch014
CITATIONS READS
0 165
5 authors, including:
Marion Neumann Christian Bauckhage

Washington University in St. Louis University of Bonn
16 PUBLICATIONS 98 CITATIONS 303 PUBLICATIONS 4,964 CITATIONS
SEE PROFILE SEE PROFILE
Some of the authors of this publication are also working on these related projects:
lectures on game AI View project
lectures on pattern recognition View project
All content following this page was uploaded by Marion Neumann on 13 September 2016.
The user has requested enhancement of the downloaded file.

Computer Vision and
Pattern Recognition in
Environmental Informatics
Jun Zhou
School of Information and Communication Technology, Griffith University,
Australia
Xiao Bai
School of Computer Science and Engineering, Beihang University, China
Terry Caelli
Department of Electrical and Electronic Engineering, The University of
Melbourne, Australia
A volume in the Advances in Environmental

Engineering and Green Technologies (AEEGT)
Book Series
Published in the United States of America by
Information Science Reference (an imprint of IGI Global)
701 E. Chocolate Avenue
Hershey PA, USA 17033
Tel: 717-533-8845
Fax: 717-533-8661
E-mail: cust@igi-global.com
Web site: http://www.igi-global.com
Copyright © 2016 by IGI Global. All rights reserved. No part of this publication may be reproduced, stored or distributed in
any form or by any means, electronic or mechanical, including photocopying, without written permission from the publisher.
Product or company names used in this set are for identification purposes only. Inclusion of the names of the products or
companies does not indicate a claim of ownership by IGI Global of the trademark or registered trademark.
Library of Congress Cataloging-in-Publication Data
Computer vision and pattern recognition in environmental informatics / Jun Zhou, Xiao Bai, and Terry Caelli, editors.
pages cm
Includes bibliographical references and index.
ISBN 978-1-4666-9435-4 (hardcover) -- ISBN 978-1-4666-9436-1 (ebook) 1. Wildlife monitoring--Technological innova-
tions. 2. Computer vision. 3. Pattern recognition systems. 4. Environmental sciences--Information technology. I. Zhou,
Jun, 1974- editor. II. Xiao, Bai, editor. III. Caelli, Terry, editor.
QL83.17.C66 2016
333.950285--dc23
2015028789
This book is published in the IGI Global book series Advances in Environmental Engineering and Green Technologies
(AEEGT) (ISSN: 2326-9162; eISSN: 2326-9170)
British Cataloguing in Publication Data

A Cataloguing in Publication record for this book is available from the British Library.
All work contributed to this book is new, previously-unpublished material. The views expressed in this book are those of the
authors, but not necessarily of the publisher.
For electronic access to this publication, please contact: eresources@igi-global.com.

295
Chapter 14
Cell Phone Image-Based
Plant Disease Classification
Marion Neumann Benjamin Klatt
Universtiy of Bonn, Germany Central Institute for Decision Support Systems in
Crop Protection, Germany
Lisa Hallau
University of Bonn, Germany Kristian Kersting
TU Dortmund University, Germany
Christian Bauckhage
Fraunhofer IAIS, Germany
ABSTRACT
Modern communication and sensor technology coupled with powerful pattern recognition algorithms for
information extraction and classification allow the development and use of integrated systems to tackle
environmental problems. This integration is particularly promising for applications in crop farming,
where such systems can help to control growth and improve yields while harmful environmental impacts
are minimized. Thus, the vision of sustainable agriculture for anybody, anytime, and anywhere in the
world can be put into reach. This chapter reviews and presents approaches to plant disease classification
based on cell phone images, a novel way to supply farmers with personalized information and processing
recommendations in real time. Several statistical image features and a novel scheme of measuring local
textures of leaf spots are introduced. The classification of disease symptoms caused by various fungi or
bacteria are evaluated for two important agricultural crop varieties, wheat and sugar beet.
INTRODUCTION
In the presence of increasing environmental challenges such as water scarcity, climate change, concerns
about food supply, safety and the reduction of adverse environmental impacts, sustainable agriculture
is an extremely important resort to cope with a rapidly growing world population. The scope of farm-
ing that is sustainable outreaches short-term yield maximization and efficiency steered exploitation of
resources. Instead the focus lies on farming methods that secure long-term yields of products satisfy-
ing the food and fiber needs of animals and humans while enhancing the quality of the environment.
One important way to achieve this goal is to control the outbreak and spread of plant diseases causing
DOI: 10.4018/978-1-4666-9435-4.ch014
Copyright © 2016, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.

Cell Phone Image-Based Plant Disease Classification
significant reduction in the quantity and quality of farming products. Accordingly, it is crucial to apply
targeted rather than broad and overdosed preventive control to minimize soil and water damage. Targeted
disease control, however, relies on two major requirements. First, plant diseases need to be identified
reliably, even for early stage outbreaks. Second, the information has to reach crop producers and farm-
ers promptly, especially in remote areas that are difficult to access by experts. The approach presented
in this chapter aims to achieve both requirements, accurate plant disease classification and real-time
forecast based on the particular state in the field.
In addition to providing farmers with treatment recommendations, accurate plant disease classifica-
tion is also pivotal in order to monitor the outbreak and spread of diseases. This information can then
again be used to predict yield losses and to forecast the temporal and spatial disease spread facilitating
the coordination of countermeasures. While there exist many methods to classify plant diseases as for
instance molecular techniques, non-invasive approaches such as visual assessment, are more favorable.
Modern techniques of visual disease assessment are based on digital photography combined with image
analysis (Bock, Poole, Parker, & Gottwald, 2010; Camargo & Smith, 2009a) or hyperspectral imaging
(Rumpf et al., 2010; Liu, Wu, & Huang, 2010). Visual image analysis also contributes to the under-
standing of biological processes such as the molecular mechanisms of photosynthesis (Rascher, Nichol,
Small, & Hendricks, 2007) or the growth rate and spread of symptoms on the disease carrier (Duncan
& Howard, 2000; Rossi et al., 2000).
Nowadays advances in cell phone technology, data storage and transmission, as well as network
coverage allow smart phone users to communicate almost any kind of information in real time from
anywhere, such as images taken with the built-in camera. Further, efficient and accurate pattern recog-
nition approaches are able to extract and process information encoded in image data in order to retrieve
information about its content in real time. It is therefore not difficult to imagine an app that automatically
applies face detection and annotates a picture of a group of friends taken with a smart phone (Ng, Sav-
vides, & Khosla, 2005; Hadid, Heikkilä, Silvén, & Pietikäinen, 2007). The image could then instantly be
posted on a social network platform or sent to the persons it contains. Unfortunately, equivalent services
are not available for agricultural application scenarios.
Whereas “passive” information systems providing farmers with instructions on how to identify ma-
jor plant diseases are available as smart phone applications, it is not possible for a farmer to get instant
personalized feedback on the actual state of the plants in his field. Our goal is to develop a system to
solve the specific problem of disease classification based on plant cell phone images. Mobile devices
such as smart phones equipped with cameras provide the technical opportunity to achieve this goal.
However, new challenges arise for image processing and analysis. These challenges are due to the fact
that implementations of image processing algorithms on a cell phone have to comply with particular
characteristics such as constrained battery life, restricted computational power, or limited bandwidth.
The work reported here results from a project on using cell phone images in an agricultural scenario,
where farmers take pictures of plants they suspect to be infected by a disease such as shown in Figures
1 and 2. Information extracted from the images is then sent to a central server and analysis results are
supposed to be reported back to the farmer while still in the field. In this setting, efficient and reliable
image analysis is pivotal. Given the weak connection strengths out in the fields or the increased fees for
high volume data transfer, it is hardly possible to transmit several pictures of sufficient resolution. If,
on the other hand, the extraction of regions of interest or even the feature computation were performed
by an app running on the cell phone, transmission times and costs can be reduced considerably. In this
296

Figure 1. Cell phone camera images of sugar beet leaves showing leaf spots caused by Cercospora
beticola (cerc), Ramularia beticola (ram), Pseudomonas syringae (pseu), Uromyces betae (rust), Phoma
betae (phom), and combined infestation of Cercospora and Phoma (from upper left to lower right image)
Figure 2. Cell phone camera images of wheat leaves showing symptoms caused by Puccinia triticina
(leaf_rust), Puccinia striiformis (stripe_rust), Pyrenophora tritici-repentis (hlb), Blumeria graminis
(mildew), and Septoria tritici (septoria) (from left to right image)
297

case, however, elaborate image processing techniques being both robust and limited to the restricted
computational resources need to be applied.
Addressing these issues, we present a cascade of efficient image preprocessing, region detection,
feature extraction, and classification steps tailored for the recognition of pathogens that infect crop plants
(Agrios, 1997). To deal with changes in illumination, scale and perspective changes we take a statistical
machine learning approach using field data that significantly varies between environments and contexts.
The data considered for learning and evaluation are cell phone images from unconstrained settings –
different camera types, different resolutions, with no constraints on how the image was acquired. Image
preprocessing and region detection is performed to extract regions of interest being likely to show disease
symptoms. In this step a color-filter based approach is introduced which is suitable for being performed
on the cell phone in order to reduce the data volume to be sent to the central sever for further processing.
Feature computation is the core step in our pipeline. A wide range of possible techniques can be
considered to extract meaningful descriptors. On the one hand, the input regions contain complex infor-
mation, such as intensity and color values, edge contours, or intensity changes, which can be revealed by
applying various transformations to the input region images. Specifically, we will consider RGB color
channels, image gradients and local binary patterns (LBPs) (Ojala, Pietikäinen, & Mäenpää, 2002). On
the other hand, a variety of features can be extracted from the original image and its transformations.
We will introduce first- and second-order statistical texture features and a simple technique to leverage
the local structure specific to certain infection symptoms on plant leaves. Local feature computation is
beneficial, as the phenotype of some disease symptoms can show circular bands of similar appearance.
Thus, by applying multiple erosion steps to the extracted regions features can be computed locally on
several parts of the original input regions. This yields what we call erosion band features. The erosion
band features are generated through iterative erosions of regions of interest. Each iteration forms what
is called an erosion band. Consequently, each erosion band provides a mask to extract any kind of image
feature from the region masked by the band. Once all the features are computed any classifier such as
nearest neighbor models or support vector machines (Cristianini & Shawe-Taylor, 2000) can be trained
and used for predictions. Whereas the discussion of different classifiers goes beyond the scope of this
chapter, we will briefly introduce support vector machine classification as an example method.
In the second part of this chapter, our general pattern recognition pipeline will be exemplified and
evaluated in two application scenarios: The cell phone image-based classification of disease symptoms
on wheat leaves and on sugar beet leaves. Wheat is one of the most produced cereals in the world and
is a widespread staple food. Sugar beet is a widely cultivated commercial crop used to produce table
sugar. Unfortunately, fungal and bacterial attacks frequently reduce both types of crop yields. An early
recognition of disease onsets assisted by an automated and “personalized” system may limit the amount
of fungicides needed for pest control and can thus reduce costs and environmental burden. An extensive
evaluation in both applications shows great performances for disease classification with accuracies of
80% for a dataset of 302 cell phone images of wheat leaves and disease detection rates of 95% and disease
classification rates of 84% for the considered sugar beet data comprising 2957 regions extracted from
495 images. The data was recorded with six different cell phone camera types at multiple times of the
growing season showing different stages of the disease infections.
The present chapter is a significant extension of previously presented work (Neumann, Hallau, Klatt,
Kersting, & Bauckhage, 2014). It provides a more detailed review of related approaches and, more im-
portantly, a first application of the pattern recognition pipeline to cell phone image-based wheat disease
classification. More precisely, after reviewing related work, the next section details the pattern recognition
298

pipeline for general cell phone image-based plant disease classification. Therein, efficient yet informa-
tive features including the recently developed erosion band feature (Neumann et al., 2014) for leaf spots
with circular disease phenotypes are introduced. In the subsequent sections, two application scenarios
are introduced and an extensive feature study comparing previously described texture features computed
on the whole regions as well as on local erosion bands is presented. Further, we show how to perform
feature selection leading to a reduced – and hence computationally more efficient – set of features with
good performance on both tasks.
BACKGROUND
The classification of plant diseases based on cell phone images is related to several lines of research.
In general, it is a method of computer aided plant phenotyping (Vankadavath et al., 2009). In the fol-
lowing, we summarize related research focusing on image-based analysis of infected plants, as well as
developments in image processing on cell phones.
The assessment of plant diseases based on visual information can roughly be divided into three lines
of approaches, manual assessment by experts, automated assessment based on digital images, such as
images recorded in RGB or other color spaces, and automated assessment based on hyperspectral imag-
ing. Manual assessment has some obvious drawbacks, it is error-prone and it requires expert knowledge.
While hyperspectral images (Fu, Robles-Kelly, Tan, & Caelli, 2006; Rascher et al., 2007; Rumpf et al.,
2010; Liu et al., 2010; Römer et al., 2012) certainly provide a lot of insights towards the understanding
of biological processes as for instance photosynthesis or drought stress, image acquisition requires ex-
tremely controlled settings and expensive equipment. As our aim is to provide a system for plant disease
classification for smart phone users directly applicable on the field, we restrict this review to approaches
based on the analysis of digital images.
Image-based approaches to plant disease classification consist of two main image processing steps,
region detection and feature computation. Region detection is mostly based on color information, as
healthy plant parts are often green whereas diseased regions tend to be more of yellowish, brownish,
reddish, grayish, or whitish color. Consequently, threshold-based methods on the color channels of the
input images provide good segmentations of regions of interest (Camargo & Smith, 2009b). Other ap-
plied methods are k-means clustering (Al-Hiary, Bani-Ahmad, Reyalat, Braik, & Alrahamneh, 2011;
Omrani et al., 2014), or binarization coupled with erosion and dilation (Siddiqi, Ahmad, & Sulaiman,
2009). Note that if the extracted regions are used in a classification scenario, we can also cope with
wrongly extracted regions actually showing healthy plant parts as we can add an additional class to our
classification problem subsuming non-infected regions (Al-Hiary et al., 2011; Neumann et al., 2014).
This is not possible when considering the task of disease quantification, where region detection is one
of the main processing steps. Also referred to as disease severity estimation or disease rating this task is
closely related to disease classification. However, the goal is different to ours as instead of distinguishing
diseases based on a feature representation of the detected symptoms, the stage of a disease is quantified
(Weizheng, Yachun, Zhanliang, & Hongda,, 2008; Bock et al., 2010; Afridi, Liu, & McGrath, 2014).
The construction of visual features lies at the core of image-based disease classification. Visual fea-
tures in classical image processing are based on shapes, colors, textures, or relative content locations,
among others. However, for the classification of regions showing disease symptoms, studies have shown
that shape- and size-based descriptors yield poor performance (Camargo & Smith, 2009a; Bauckhage &
299

Kersting, 2013). This is not surprising given that disease symptoms drastically change over time. Usu-
ally disease infected leaf spots first appear as few small spots, growing from larger more pronounced
regions to even coadunate areas until they cover the whole plant or plant part. This means that the same
disease appears in various shapes and sizes, depending on the stage of disease. Thus, feature computa-
tion mainly relies on texture descriptors (Haralick, Shanmugam, & Dinstein, 1973), which have proven
very successful in several disease classification applications (Pydipati, Burks, & Lee, 2005; Camargo
& Smith, 2009a; Al-Hiary et al., 2011).
Given the importance of the problem there are numerous works reporting approaches to plant dis-
ease classification from images. Most of these approaches are tailored to a specific problem, that is to
a specific plant and specific diseases as for instance diseases on rice (Yao et al., 2009), citrus plants
(Pydipati et al., 2005), cotton crops (Camargo & Smith, 2009a), or apple leaves (Omrani et al., 2014).
Other research focuses on the analysis of images showing entire plants or entire fields (Doudkin, Inyutin,
Petrovsky, Vatkin, & Sobkowiak, 2007; Siddiqi et al., 2009; Afridi et al., 2014). Most of these approaches,
however, rely on images taken in a controlled setting with sufficient resolution, fixed distances, light-
ing and backgrounds to ease segmentation. In contrast to these existing approaches, image acquisition
in our scenario is performed by non-expert users resulting in cell phone images in unconstrained set-
tings – different camera types, different resolutions, no constraints on how to take the image. Further,
data processing is performed on a cellular device requiring resource constraint computations and the
consulting system should provide advice in an online fashion while the farmer is still in the field. This
application scenario is novel to image-based plant disease classification and requires more elaborate
techniques for efficient region detection, feature computation, and reliable disease prediction. Thus, in
this chapter we introduce a pattern recognition pipeline aiming at minimal computational costs for region
detection and feature computation with the goal to perform its steps directly on a smart phone. Further,
we present a technique to compute local features reflecting the spatial structure of disease symptoms
first introduced in (Neumann et al., 2014). The use of more elaborate features is specifically necessary
when classifying early-stage symptoms where the leaf spots caused by the considered diseases have a
very similar appearance making them even indistinguishable by human experts.
Beyond real-time face recognition (Ng et al., 2005), which is nowadays implemented in every camera
application for smart phones, there are surprisingly little existing approaches to image processing on
cell phones. Existing work focuses on three application fields: traffic, tourism, and biology. Roters et al.
(2011) present a method to analyze live videos on mobile phones in order to localize traffic lights and
classify their state using resource-efficient features and time-based verification to handle occlusions. A
relational bag-of-visual-words approach incorporating context information such as location and direc-
tion captured by mobile devices was introduced to tackle the problem of mobile landmark recognition
(Chen, Yap, & Zhang, 2014). In relation to tourist information services, a system based on detecting and
identifying urban objects from cell phone images (Fritz, Seifert, & Paletta, 2006) has bene developed
with some success. Whereas these applications share the issues of constraint computational resources
and unconstrained settings, such as difficult illumination or varying distances, with the problem of cell
phone image-based plant disease classification, their application fields are inherently different. In the
field of biology, there are only few existing approaches addressing for example mobile plant recogni-
tion (Kumar et al., 2012; Liu, Coquin, Valet, & Cerutti, 2014). Both Kumar et al. (2012) and Liu et al.
(2014) use shape-based descriptors, which are fast and easy to compute. Unfortunately, it is difficult
if not impossible to use shape features to classify symptoms of plant diseases, which is the main focus
of the approach presented in this chapter. Another interesting application in the context of biology is
300

cellphone microscopy (Arpa, Wetzstein, Lanman, & Raskar, 2012). Here, the goal is similar to ours,
namely providing users in remote areas with decision support. However, Arpa et al. (2012) focus on the
development of mobile microscopes, whereas we are interested in deriving predictive features from cell-
phone images of plant leaves in an agricultural scenario. In the following, we give a detailed description
of the pattern recognition steps necessary to achieve this goal.
PATTERN RECOGNITION PIPELINE FOR PLANT DISEASE CLASSIFICATION
Cell phone image-based plant disease classification essentially consists of three steps: image prepro-
cessing and region detection, feature extraction, and class prediction. Whereas preprocessing and region
extraction is applied to the whole input image, feature computation and classification are performed on
each extracted region of interest. As we assume that the regions of interest comprise only small segments
of the input image, we do not have to send the entire image to a central server for classification. Figure
3 illustrates our proposed pipeline, where image preprocessing and region detection are performed on
the cell phone directly. Due to memory- and time-intense computations, the whole training procedure
including feature computation, selection, and classifier training is performed on the server. In the on-
line phase when we want to predict the disease(s) for a new input image, feature computation can be
performed on the cell phone directly or on a central server. This choice depends on the computational
cost of the feature computation. Prediction is most conveniently performed on the server as this allows
us to update the trained classifier (SVM model) without updating the app on the cell phone. Moreover,
the server connection allows us to monitor the outbreak of diseases locally and spatially. In the follow-
ing, we describe each step in this general pattern recognition pipeline in detail. Specific settings and
parameters for the disease classification on wheat and sugar beet leaves will be provided in the section
describing the respective application.
Figure 3. Pattern recognition pipeline for cell phone image-based plant disease classification
301

Figure 4. Preprocessing and region detection steps for a cell phone camera image of a beet leaf show-
ing symptoms of Phoma betae: original cell phone image, maximal RGB values, color-filtered binary
image, binary region image and extracted regions (from left to right image)
Step 1: Image Preprocessing and Region Detection
The images to be analyzed in our scenario are recorded under uncontrollable conditions. Whenever
farmers in the field take snapshots of plant leaves, scene illumination, camera angle, focal length, and
distance to object are essentially random variables (cf. Figures 1 and 2). In order to be able to circumvent
undesirable effects due to some of these variables, the default setting of our system is to consider rather
high-resolution images. Given an RGB image I of 5 resp. 8 megapixels in jpg format, the following
preprocessing steps, illustrated in Figure 4, yield useful intermediate representations for later analysis:
1. Preprocessing: Scale the input image I to an image D of fixed size; this facilitates further computa-
tion; cf. first image in Figure 4(a). Note that the scaling factor depends on the sizes of the images
in the database as well as on the distance of the camera to the plant leaves.
2. Color Filter: Compute a binary image B from D such that the foreground pixels are the pixels
(most likely) showing diseased regions. The choice of color-filter depends on the appearance of
the disease symptoms. The second image in Figure 4 illustrates the maximal color values in an
example image D showing a sugar beet leaf and the third image in Figure 4 shows the color-filtered
binary image when selecting all pixels with maximal red value.
3. Binary Region Image: Compute a binary region image B’ from B; this step is inherently region
appearance dependent and can subsume median filtering, connected component analysis, hole fill-
ing, filtering of regions adjacent to the image borders, as well as filtering of regions with undesired
shapes. For all these computations only the binary image B is required. An example binary region
image B’ is shown in the fourth image in Figure 4.
4. Region Image: Suppress all pixels in D that are background pixels in B’; the information in the
new image R forms the basis for the subsequent analysis; an image R is shown in the last image in
Figure 4.
Note that up to this point none of the steps in our processing cascade makes use of floating point
operations. As each of these intermediate steps can be computed using only integer arithmetic, they can
be implemented even on older generation phones without processors with floating point units.
302

Step 2: Feature Extraction
Feature computation and further classification is now performed on the region image R. Depending on
the appearance of the disease symptoms, features can be computed either on each connected component
separately or on one (possibly disconnected) region of interest. Both approaches have advantages and
disadvantages. Whereas the first approach allows one to classify infection symptoms from multiple
diseases appearing on the same plant or plant leaf as illustrated in the last image in Figure 1, it fails to
produce expressive features for diseases with less compact symptoms scattered on the plant as shown in
the first image in Figure 2. Hence, a region ri should be defined as either a connected component in B’
or as all foreground pixels in B’ depending on the application. In the first case, we get several regions per
input image; in the latter case one region per image is to be considered. To achieve scale-independence
each region ri is now scaled to a fixed size keeping its aspect ratio.
As we are not assuming a particular stage of infection, the regions of interest, which are now ei-
ther leaf spots or scattered disease symptoms, are typically indistinguishable by their forms and sizes.
Hence, established shape descriptors and shape recognition methods, such as form factors (Gonzales
& Woods, 2008), bounding box signatures (Bauckhage, 2006), or entropies of curves (Dupain, Kamae,
& Mendes-France, 1986; Ebrahim, Ahmed, Chau, & Abdelsalam, 2007) lack of discriminative power
for our classification task. We therefore resort to texture descriptors computed from various statistics
on the color channels of the region pixels and various transformations over color pixel values. In the
following, we introduce several image transformations producing input values for statistical analysis,
as well as various first- and second-order statistics in order to measure simple intensity levels or more
elaborate texture appearance.
Image Information and Transformations
As the main characteristics of leaf spots caused by plant diseases, such as illustrated in Figures 1 and
2, are based on color values and intensity changes, we analyze color-, intensity-, and gradient-based
features. From the original RGB values of the region image R the following values can be computed for
each region of interest ri by simple image transformations:
• Red (R), green (G), blue (B) channels,

• Intensities (INT),
• LBPs of intensities (INT LBP),
• Gradient magnitudes (Gmag),
• Gradient directions (Gdir), and
• LBPs of gradient magnitudes (Gmag LBP),
where LBP is the local binary pattern transformation (Ojala et al., 2002). Intensity values can be
computed from the RGB image by taking a possibly weighted average of the color channels. The most
basic LBP transformation is defined on a neighborhood of eight pixels and the LBP value for a pixel p
is given by
( )
LBP (p) = ∑ s val (pn ) − val (p ) 2n
n =0
303

Figure 5. Images of red, green, and blue channel, intensity image, intensity LBP image, gradient direc-
tion image, gradient magnitude image and gradient magnitude LBP image of an example Cercospora
region (from upper left to lower right image)
where pn is the n-th neighbour of pixel p, val(x) is the respective value of pixel x, as for instance its
intensity value, and s(x) = 1 for x ≥ 1 and s(x) = 0 otherwise. LBP descriptors provide an efficient
way to characterize local texture elements such as horizontally, vertically or diagonally running edges
or corners. There are numerous accounts in the recent literature where LBPs were reported to achieve
highly accurate classification results (Ojala et al., 2002; Huang, Li, & Wang, 2004; Maturana, Mery, &
Soto, 2009). Gradient magnitudes and directions are computed from the directional gradients [Gx,Gy]
T
with respect to both image axes, where Gx is defined along the pixel columns and Gy is defined along
the pixel rows of an image. The gradient magnitude and is given by
T
Gmag = Gx ,Gy  = Gx2 + Gy2
 
G 

and the gradient direction is Gdir = tan−1  y  (Jähne, 2005; Gonzales and Woods, 2008). Note that all
Gx 
transformations result again in images, that is, the values can be stored in a pixel grid of the same size
as the input image. The resulting images of all considered transformations for a RGB region of an ex-
ample leaf spot are shown in Figure 5.
Statistical Measures
To analyze color, intensity, and gradient information of the extracted regions, we can now compute simple
and co-occurrence based statistics for the newly derived input images described above.
First-Order Statistics computed from the respective normalized histogram hi of n bins lead to simple
characteristics of image texture. The following statistics are commonly used:
n
• Mean: ∑ kh
k =1
ik
= µi
n
∑ h (k − µ )
2
• Var: ik i
k =1
304

n
• Entropy: −∑ hik log (hik )
k =1
where k indicates the respective bins and the total number of bins is set to an integer n ≤ 256.
Second-order statistics computed from the spatial co-occurrence of pixel values analyze relationships
among pairs of pixels such as homogeneity, contrast, and correlation. In contrast to the simple first-order
statistics, these measures constitute more powerful texture descriptors (Haralick et al., 1973). Given the
number of occurring values or the number of thresholded values m of an input image, the m × m co-
occurrence matrix Cθ,d is measuring the distribution of co-occurring values at a given offset defined by
an angle θ ∈ {0◦, 45◦, 90◦, 135◦} and the distance of the pixel relation d
1, if val (i, j ) = x and val (i + ∆i, j + ∆j ) = y 

 
C θ,d (x , y ) = ∑  
i, j 
 0, else 
 
where ∆i and ∆j are to be computed from θ and d, and ∑ i, j is the double sum over all image rows
and columns. Typically d is set to a small value such as d ∈ {1,...,5}. Now, let the mean of Cθ,d be given
1 1
by µ = 2 ∑ i, j C θ.d (i, j ) , the mean of the row and column sums be given by µi = ∑ i i ∑ j C θ,d (i, j )
m m
1
and µj = ∑ j j ∑ i C θ,d (i, j ) , and the standard deviation of the row and column sums be given by
m
1 1
σi = ∑ i (i − µi ) ∑ j C θ,d (i, j ) and σ j = ∑ j ( j − µj ) ∑ i C θ,d (i, j ) .
2 2
m m
Then, the second-order statistics are defined as:
∑ C (i, j )
2
• Angular 2ND Moment, also Energy Feature: θ,d
i, j
m −1
• Contrast: ∑k ∑ 2
C θ,d (i, j )
k =0 i , j :i − j=k
(i − µ )( j − µ )C (i, j )
∑
i j θ,d
• Correlation:
i, j σi σ j
• Entropy: − ∑ C θ,d (i, j ) log C θ,d (i, j )
i, j
1
• Inverse Difference Moment: ∑ C θ,d (i, j )
1 + (i − j )
2
i, j
∑C (i, j )(i − µ)
2
• Variance, also Sum of Squares: θ,d
i, j
2m −1
• Sum Average: ∑ k p (k )
k =1
+
2m −1
• Sum Entropy: − ∑ p+ (k ) log p+ (k ) = se
k =1
305

m
• Difference Entropy: − ∑ p− (k ) log p− (k )
k =1
2m −1
∑ (k − se ) p+ (k )
2
• Sum Variance:
k =1
2m −1
∑ (k − se ) p− (k )
2
• Difference Variance:
k =1
where p+ (k ) = ∑ i, j C θ,d (i, j ) with k = i + j and p− (k ) = ∑ i, j C θ,d (i, j ) with k = |i − j|. To achieve
rotation invariance, each feature value is derived as the average of the statistics computed from Cθ,d over
all angles θ. Given an input image, as for instance the intensity image, it is common to use all or a sub-
set of these second-order statistics as texture feature. Using N of the above defined statistics results in a
N-dimensional feature vector depending on the distance d. We will call such texture features TEXTUREd.
Note that the computational cost of each introduced feature varies. Therefore, given a specific ap-
plication, we will have to consider the trade-off between efficiency and expressivity of a feature set.
Whereas simple first-order features can be computed on a cell phone for instance, statistics based on the
co-occurrence matrix need more computational resources. Similar concerns hold for classification as
prediction can be achieved easily based on a single measurement, cf. Specific Example II in (Bauckhage
& Kersting, 2013). However, it gets considerably harder for multi-dimensional feature vectors.
Feature Selection
Considering all possible combinations of statistical measures and input images leads to a high-dimensional,
complex feature space being both costly to compute and prone to overfitting. Similar to an existing
approach speeding up the training process by instance selection in a multiple-instance learning setting
(Fu, Robles-Kelly, & Zhou, 2011), we want to select a subset of statistical measures from input images
allowing for robust classification. This selection process, however, is non-trivial as we need to consider
all possible sets of value/statistics pairs to derive a meaningful feature set. For example if we consider
all 8 image transformations as input values as introduced above and all 3 simple statistics, as well as 3
texture features (e.g. for d = 1, d = 3, and d = 5) we get 48 single features. This considerably small set
of candidate features leads to 248 possible feature sets being extremely costly to evaluate. Therefore, we
propose to use a combination of forward selection and backward elimination to retrieve an expressive
yet efficiently computable feature set (Guyon & Elisseeff, 2003). Forward selection starts with an empty
feature set and greedily selects the best performing feature to add. Backward elimination starts with a
given set of seed features, which could be all possible features or a subset thereof, and removes the ones
that are not improving performance step by step. Usually both approaches are used after one another.
Note, however, that the order in which they are applied matters as this can lead to different feature sets.
In the following, we briefly introduce erosion band features, a technique to compute statistical measures
locally reflecting the structural appearance of plant leaf spots.
Erosion Band Features
Following experts’ descriptions of disease phenotypes caused by various plant pathogens affecting crops
such as sugar beets, Neumann et al. (2014) introduced erosion band features for plant disease classifica-
306

Figure 6. Erosion of the example Cercospora region shown in Figure 5; upper row shows the erosion
bands and lower row the masked region image of the example spot.
tion. Similar to erosion band signatures (Vazquez, Yang, & Slabaugh, 2013), we want to extract more
fine-grained texture features following the circular structure of the leaf spots. Given the erosion width
( 0)
w, binary erosion is applied iteratively to the extracted regions. Given the extracted region ri = ri and
(t )
Ew a spherical structuring element, the erosion band at the t-th erosion step bi is calculated by
(t ) (t −1) (t ) (t ) (t −1)
bi = ri − ri with ri = ri ⊖ Ew
where ⊖ is the morphological erosion operator and t = (1, . . ., tmax). The erosion width can be deduced
from a given number of iterations and the region dimensions. Figure 6 shows the erosion masks and
eroded RGB regions of the leaf spot depicted in Figure 5 using four erosion steps. For each erosion step
the desired texture descriptors can now be computed locally on the respective values of the masked
regions. By concatenating all features we can derive a more expressive texture descriptor, however, by
facing higher computational costs as the dimension of an erosion band feature is tmax times larger than
the dimension of the according feature computed on the whole image region. This computational effort,
however, leads to texture features reflecting the circular structure of the phenotypes of the considered
(t )
diseases. Note, that by limiting tmax and filling the erosion band bi max in the last iteration, we ensure a
fixed feature dimension across all regions. Erosion band feature extraction is a general approach that can
be applied to any input image information and any statistical measure. Before we discuss the performance
of the introduced features in two application scenarios, we cover the last step in our pattern recognition
pipeline, namely disease classification.
Step 3: Classification
Based on the feature representations of extracted regions caused by known diseases we can now learn a
classifier which can then be used to predict plant diseases for unseen target cases. Common classification
algorithms used in pattern recognition approaches are nearest-neighbor models, naive Bayes classifiers,
neural networks, or kernel-based learning approaches such as support vector machines (Bishop, 2006,
307

Hastie, Tibshirani, & Friedman, 2009). In our application scenarios we will train a one-vs-one multi-
class support vector machine (SVM) using the radial basis function (RBF) kernel
 x − x ′ 
k (x , x ′) = exp − ( )
 = exp γ x − x ′ 2
 2σ 2 
1
where σ is the kernel width and γ = − . The SVM is a sparse maximum margin classifier essen-
2σ 2
tially selecting the support vectors, which are training points most informative for the distinction of
classes. The decision boundary and support vectors can be learned by solving a convex optimization
problem based on a kernel representing the similarity among data points. For a detailed introduction of
support vector machines we refer to (Cristianini & Shawe-Taylor, 2000).
Now, we have introduced the general pipeline for our pattern recognition system predicting plant dis-
eases based on their visual symptoms recorded by (cell phone) images. In the following, we will exemplify
its steps for two specific tasks: the classification of disease symptoms on wheat and sugar beet leaves.
CLASSIFICATION OF WHEAT DISEASE SYMPTOMS
In this section, we apply our pattern recognition approach to the classification of disease symptoms on
wheat leaves. Specifically, we will present region extraction steps and texture features suitable to dis-
tinguish cell phone images showing wheat leaves with five different symptoms: leaf rust, wheat stripe
rust, yellow leaf spots, powdery mildew, and septoria blotch as illustrated in Figure 2. We begin with
introducing the problem and data basis.
Problem Introduction and Data
Being used to produce for example bread, biscuits, cookies, cakes, breakfast cereals, noodles, and cous-
cous, wheat is one of the world’s most important staple foods. Unfortunately, various diseases caused by
different fungi create significant yield losses of the worldwide wheat production. For example, in 2007,
leaf rust caused a 14% loss in winter wheat yields in Kansas, the leading wheat-producing state in the
US (Bolton, Kolmer, & Garvin, 2008). Whereas fungicides are able to control the infections, their use
causes environmental damage and thus should be kept to a minimum. As different diseases require dif-
ferent treatments, accurately predicting the disease symptoms on wheat leaves is an important problem.
In particular, we believe that a system, which is able to predict wheat diseases accurately and instantly
from personalized example images – taken with a smart phone on the field – can considerably help
farmers and crop producers to decide about specific treatments.
In our experimental evaluation we consider a dataset of 203 cell phone images of wheat leaves taken
with multiple smart phones cameras such as the Samsung GT-I9300, the Sony Ericsson LT18i, and the
Motorola XT910. The classification task is to distinguish symptoms of the following five diseases af-
fecting wheat plants.
308

Figure 7. Preprocessed images of wheat leaves showing disease symptoms of all five considered classes:
leaf_rust, stripe_rust, hlb, mildew, and snb (left to right image)
• Puccinia triticina is a fungal pathogen and causal agent of leaf rust. Leaf rust is characterized by
orange-to-brown, round-to-ovoid pustules with diameters of about 1.5 mm. Pustules are scattered
on both the upper and the lower leaf surfaces and can merge to coadunate regions in later stages.
• Puccinia striiformis f.sp. tritici, causal agent of stripe rust or yellow rust, is a fungus showing yel-
low- colored stripes running parallel along the leaf veins.
• Helminthosporium leaf blights (HLB), also yellow leaf spot or tan spot, is caused by both
Drechslera tritici-repentis and Pyrenophora tritici-repentis. Small, dark brown spots of 1 to 2
mm length characterize early lesions. In later stages, these lesions extend very quickly to oval-to-
elongated blotches of several centimeters with light to dark brownish appearance.
• Blumeria graminis is a fungus and causal agent of wheat mildew or powdery mildew. The fungus
forms white to grayish patches of powdery appearance.
• Septoria tritici is a fungus showing oval, longish spots of yellowish/greenish later brownish color
also called septoria nodorum blotch (SNB).
So, we tackle a 5-class classification problem with the classes leaf_rust (44), stripe_rust (75), hlb
(32), mildew (14), snb (38); the numbers in parentheses reflect the class frequencies in our database.
Figure 7 shows example images of all considered disease classes.
Pattern Recognition (Steps 1 and 2)
We performed the following preprocessing and region detection to the input images:
1. Preprocessing: The input image I is manually cropped to remove unnecessary background such
as ground or other plants, cf. Figure 7. Then the cropped images D are scaled to have 75 pixels in
y direction (pixel rows).
2. Color Filter: For the classification of symptoms caused by plant pathogens on wheat leaves, pixels
with a red value higher than both the green and the blue values as well as pixels with high values in
all three color channels are considered to be foreground pixels. The latter filter allows us to retrieve
309

regions with whitish appearance, which is necessary to detect mildew. This step produces a binary
image B.
3. Binary Region Image: For this specific application B’ = B as the images D were already cropped
to include the leaf only. As the symptoms can be either compact, scattered or coadunate, further
processing is not useful for the classification of wheat disease symptoms. Further, we will consider
one possibly disconnected region per image, so ri = B’.
4. Region Image: To get R we suppress all pixels in D that are background pixels in B’.
Now, we consider the following input images and statistics: INT, INT LBP, R, G, B, RG, RB, GB,
RGB, Gmag, Gdir, and Gmag LBP, and MEAN, VAR, MEAN & VAR, and ENTROPY as first-order statistics
and a 7-dimensional feature, TEXTUREd, including the following second-order statistics: CORRELA-
TION, ENTROPY, INVERSE DIFFERENCE MOMENT, VARIANCE, SUM ENTROPY, DIFFERENCE
ENTROPY, and SUM VARIANCE. The number of histogram bins to compute the simple statistics is set
to n = 256 and d is varied in {1,3,5}. Thus, the initial set of possible features consists of 12 × 7 value/
statistics pairs leading to 84 possible single features.
Experimental Protocol (Step 3)
Feature set performance is evaluated by running C-SVM classification using libSVM (http://www.csie.
ntu.edu.tw/~cjlin/libsvm/) on 20 randomly generated but fixed data splits. Support vector machine (SVM)
classification is described above in Step 3 of our pattern recognition pipeline. For all experiments we
learn the SVM-cost parameter c and the kernel parameter γ via a 10-fold cross validation on the respec-
tive training splits. The considered parameter ranges are c ∈ {2−3, 2−2, . . ., 26} and γ ∈ {0.1, 0.5, 1.0,
5.0, 10.0}. Performance is measured by average accuracy. Feature selection is performed as follows: the
best performing single features out of 6 groups of features are selected as seed set. To force diversity
in the seed set we divided the features in two groups with respect to the statistical measure: first- and
second order statistics and further separated intensity based values from color and gradient information
leading to three groups of input data. On the seed set comprising 6 features we then performed backward
elimination based on the achieved average accuracies leading to the 41-dimensional feature set selec-
tion consisting of four of the initial feature sets: INT + TEXTURE5, RGB + TEXTURE5, Gmag LBP +
TEXTURE1, and RGB + MEAN & VAR. A subsequent application of forward selection did not add
back any of the previously removed features.
Classification Results
The classification results are reported in Table 1. The best performing single feature set was TEXTURE5
computed on all three color channels RGB giving an overall average accuracy of 76.6%. This leads to
the conclusion, that all three color channels encode classification-relevant texture information. Using all
possible combinations of values and statistics, results in a 192-dimensional feature with a performance
of 79.0% accuracy. However, we can do even better with respect to both feature computation time and
accuracy when considering only a subset of these features. The subset SELECTION retrieved by the
feature selection approach described above consists only of a 41-dimensional feature and achieves an
average accuracy of 80.1%. Whereas the predictive performance is not significantly better (under a paired
t-test with p-value 0.05) than using all features the computational effort is significantly reduced as the
310

Table 1. Average accuracies ± standard errors in % and feature dimensionality of the best performing
single feature (RGB + TEXTURE5), a feature set consisting of all features (ALL), the selected features
(SELECTION), most frequent class (MFC) and random (RAND) for disease classification on wheat.
The best result is indicated in bold.
RGB+Texture5 All Selection MFC Rand

acc ± stderr 76.6 ± 3.3 79.0 ± 2.7 80.1 ± 3.2 37.0 ± 3.2 20.0 ± 0.0
dim 21 192 41 - -
Feature dimensionality is reduced by 78.6% from 192 to 41. This is an important insight when con-
sidering feature extraction and subsequent classification to be performed on a cellular device directly.
The confusion matrix
32 2 9 3 3 
 
 2 25 2 0 1 
 
 
M =

7 2 63 0 2 

 2 2 0 10 0 
 
 1 1 1 1 32
shows convincing per class performance (classes top to bottom: leaf_rust, hlb, stripe_rust, mildew, snb);
each class can be classified correctly with a probability of at least 71%.
To summarize, these results show that cell phone image-based plant disease classification for wheat
crops is feasible. In particular, we showed that it is possible to achieve good classification performance
with a small set of texture features computed on color, intensity and gradient magnitude LBPs even on
a small set of training examples. Note that extending the training data can most likely increase classifi-
cation performance and robustness with respect to illumination, size, and resolution changes. Now, we
will apply the same pattern recognition pipeline to classify diseases affecting sugar beet crops.
CLASSIFICATION OF SUGAR BEET DISEASE SYMPTOMS
In the following, we consider another challenging image analysis task in agriculture where cell phone
camera images of beet leaves are analyzed as to the presence of plant diseases. Aiming at minimal com-
putational costs on the cellular device and highly accurate prediction results, we exemplify efficient image
preprocessing and detection of potential disease regions and show how to achieve robust classification
based on texture features. In particular, we find that leveraging the circular structure of disease spots
on beet leaves is highly beneficial. A combination of descriptors derived on multiple erosion bands of
the RGB color channels, as well as, the local binary patterns of gradient magnitudes of the extracted
regions accurately distinguishes between symptoms of five diseases, including infections caused by the
fungi Cercospora beticola, Ramularia beticola, Uromyces betae, and Phoma betae, and the bacterium
Pseudomonas syringae pv. aptata.
311

In order to determine a small but meaningful feature ensemble for robust leaf spot classification, we
performed an extensive feature study on a subset of 296 regions of the full database. First, we evaluate
each combination of statistical measure and input value on the whole region and further on the erosion
bands as described above. Then, a feature ensemble of the best performing but most diverse feature/
value combinations is created from the first evaluation step. This feature ensemble, called erosion feature
ensemble, is then evaluated on a comprehensive dataset of 2957 regions extracted from 495 cell phone
images.
Problem Introduction
Suagr Beet (Beta vulgaris subsp. vulgaris) is a widely cultivated commercial crop used to produce,
for example, table sugar. Unfortunately, fungal and bacterial attacks frequently reduce yields. An early
recognition of disease onsets assisted by our pattern recognition approach can trigger targeted control
reducing costs and environmental burden. Further, a cell phone based classification system can provide
instant help for farmers and crop producers in remote areas, which may be too time and cost consuming
to be reached by experts. In particular, we attempt to automatically recognize symptoms of five common
kinds of diseases of sugar beets; Figure 1 shows example images of all considered disease symptoms.
• Cercospora beticola is a fungal plant pathogen. Infected beet plants show leaf spots that are round
blemishes with a definite edge between infected and healthy leaf tissue; while the border of a spot
is typically darker and of brownish-reddish hue, spot interiors appear bright with dark conidio-
phores and hyaline conidia (spores).
• Ramularia beticola, also a fungal pathogen infecting beet plants, showing irregular leaf spots with
light interior surrounded by a light-brown to brown border. The interior shows white spores. Leaf
spots at a pronounced stage are likely to chap or form coadunate regions.
• Pseudomonas syringae pv. aptata, the only considered bacterial pathogen commonly infecting
beets, shows light to dark brown spots potentially with a light interior. Spots may have a brown to
black edge.
• Uromyces betae is a fungal pathogen causing sugar beet rust. Typical symptoms are small reddish-
brownish spots surrounded by yellow areas.
• Phoma betae is a soil-borne fungus and plant pathogen showing rather large leaf spots with con-
centric rings of irregular shape and less pronounced edges; borders of a spot are yellowish or light
brown, spot interiors are characterized by darker brownish hues; when several spots are present,
they can grow together and form larger blotches.
Region Detection (Step 1)
We performed the following preprocessing and region detection to a dataset of 495 input images recorded
with six different cell phone camera types; the steps are illustrated in Figure 4.
1. Preprocessing: First, each input image I is scaled by 25%; the resulting image D is of size 484×648×3
resp. 612×816×3 depending on the used smart phone camera.
312

2. Color Filter: Then a binary image B is computed from D such that the foreground pixels are the
pixels with a red value higher than both the green and the blue value; this pronounces reddish/
brownish image regions which may hint at leaf spots caused by fungus or bacteria.
3. Binary Region Image: To compute a binary region image B’ from B we performed connected
component analysis, hole filling, and filtering of regions adjacent to the image borders, as well
as filtering of regions where the width/height ratio of their bounding boxes is not in the interval
(0.7, 2); the latter is motivated from biological expertise: since leaf spots tend to be compact, too
elongated regions can be discarded from further analysis.
4. Region Image: To get R all pixels in D that are background pixels in B’ are suppressed.
Feature computation and further classification is performed on each detected region separately. A
region ri is defined as a connected component in B’. Each ri is scaled such that its larger image dimen-
sion measures 75 pixels to achieve scale-invariance.
Data, Features, and Experimental Protocol (Steps 2 and 3)
From all available regions we now derive two datasets, one for feature selection and one for evaluation.
The first one, called STUDY, is a subset of all available regions extracted from 495 cell phone camera
images applying the region detector described above. It was derived by manually selecting a balanced
number of regions for each class and considering regions consisting of one single leaf spot only as il-
lustrated in Figure 5. This dataset is used in our feature study and hence, can be seen as training dataset
for the feature selection process. The second dataset, full, is composed of all 2957 extracted regions and
it also contains regions consisting of several coadunate leaf spots as illustrated in Figure 8. This dataset
also shows high class imbalance, as we did not meet many beet plants infected by Rust and Phoma dur-
ing data acquisition. The class frequencies of both datasets are listed in Table 2.
Figure 8. Example regions with coadunate spots of three disease symptoms: Pseudomonas syringae pv.
aptata, Cercospora beticola and Ramularia beticola (from left to right image)
Table 2. Number of regions per class for the two considered datasets STUDY and FULL
Classes
Dataset cerc ram pseu rust phom n-inf Total
STUDY 57 57 44 47 36 55 296
FULL 1003 255 494 72 55 1108 2957
313

We use the following second-order statistics, CORRELATION, ENTROPY, INVERSE DIFFER-

ENCE MOMENT, VARIANCE, SUM ENTROPY, DIFFERENCE ENTROPY, and SUM VARIACNE
resulting in a 7-dimensional texture descriptor. We also tested ANGULAR SECOND MOMENT and
DIFFERENCE VARIANCE, however, adding these features to the descriptor decreased classification
performance in an initial set of experiments on the training data. Further, the same input values and
simple statistics are chosen as for the classification of wheat disease symptoms in the previous section.
Moreover, the same experimental protocol as described in the previous section is considered. However,
the kernel parameter γ is optimized in {0.01, 0.05, . . . 5.0, 10.0}. For erosion band features we set tmax =
4 in all experimental evaluations. Using less erosion steps produces less discriminant features, whereas
setting tmax > 4 unnecessarily increases the feature dimensionality.
Further, we train the SVM classifier on 6 classes, where 5 classes are leaf spots caused by the previ-
ously introduced pathogens Cercospora beticola (cerc), Ramularia beticola (ram), Pseudomonas syringae
pv. aptata (pseu), Uromyces betae (rust), and Phoma betae (phom). To handle regions extracted by the
region detector not belonging to one of these classes we consider a sixth class of non-infected regions
(n-inf). Examples of this class could be healthy leaf parts, such as reflections or leaf veins, dirt on the
leave, holes, or earth on the ground. These regions occur as we apply a simple and efficient region de-
tector being feasible on the smart phone. To analyse disease detection rates, we therefore also evaluate
a classification task, where we consider the classes inf and n-inf, where inf subsumes all regions labeled
as being caused by one of the considered diseases. For this binary task we also follow the experimental
protocol described above and use a SVM with the same kernel settings as described for the multi-class
problem.
Feature Study
Before we describe the feature ensemble construction, we discuss the results of our feature study on the
dataset STUDY. Feature performance is measured by average accuracy. The results for all single value/
statistics combinations are shown in Table 3.
First, we observe that the performance is universally improved by computing the texture features
on the erosion bands instead of on the whole regions. This result proves the usefulness of erosion band
features for leaf spot classification. Further, we see that LBP images of intensities (INT LBP) and gra-
dient magnitudes (Gmag LBP) can improve the performance compared to just using INT and Gmag when
considering simple statistics, such as mean and entropy, confirming results presented in (Ojala et al.,
2002). In general, second-order statistics, denoted by TEXTUREd, lead to better performance compared
to simple first-order statistics. Considering distances d > 1, however, did decrease accuracy, leading to
the conclusion that for leaf spot classification in our scenario the co-occurrence of neighboring pixel
pairs is most informative about the disease. Note, that second-order texture descriptors as well as erosion
band features are of higher dimensionality. The dimensionality of each feature/value pair is given by the
product of feature and value dimensionality indicated in the second row and third column of Table 3.
Note that simply using all feature/value combinations results in a 960-dimensional input space leading
to a very costly to compute classification cascade.
314

Table 3. Average accuracies (%) obtained by 20-fold cross validation on the dataset STUDY; Member-
ship in the candidate feature set for the erosion feature ensemble is indicated in bold.
INT INT R G B RG RB GB RGB Gmag Gdir Gmag

LBP LBP
dim (1) (1) (1) (1) (1) (2) (2) (2) (3) (1) (1) (1)
Whole Region MEAN (1) 35.1 42.4 32.7 41.5 42.9 42.2 42.9 43.1 47.0 21.6 32.4 35.5
VAR (1) 23.3 20.0 21.4 30.3 23.6 36.8 35.7 34.4 43.5 30.5 25.7 19.6
MEAN & VAR (2) 37.5 42.0 30.6 37.1 47.9 44.1 49.6 46.9 54.3 41.2 33.4 36.1
ENTROPY (1) 35.6 43.3 33.3 34.7 30.3 35.3 36.1 45.3 44.5 34.8 49.0 44.2
TEXTURE1 (7) 57.6 52.8 59.3 56.7 58.8 65.9 65.5 64.4 68.2 51.9 48.7 48.7
TEXTURE3 (7) 50.0 48.8 51.2 50.3 48.5 62.8 57.0 54.5 62.4 48.2 49.3 45.8
TEXTURE5 (7) 49.5 50.8 48.9 43.5 46.9 57.7 54.6 52.6 57.0 49.3 46.2 41.2
Erosion MEAN (4) 54.5 50.5 49.2 54.6 52.6 54.3 53.2 56.0 54.6 42.8 30.7 43.9
Bands
VAR (4) 45.1 29.0 42.5 45.1 45.5 52.6 50.7 44.5 52.3 36.7 38.2 32.8
MEAN & VAR (8) 58.7 51.0 56.6 60.1 55.4 60.8 59.0 60.4 61.7 48.3 37.1 42.8
ENTROPY (4) 43.2 54.0 44.5 42.8 47.2 47.3 50.0 50.3 55.3 47.2 43.6 52.9
TEXTURE1 (28) 67.7 55.4 69.0 65.8 64.8 71.8 68.8 71.3 71.9 48.9 51.4 53.7
TEXTURE3 (28) 62.4 54.4 67.0 67.1 60.7 69.1 66.8 63.8 67.7 47.2 51.4 44.2
TEXTURE5 (28) 63.8 57.8 66.2 65.5 62.5 69.1 68.8 64.2 70.9 47.2 48.0 46.6
Erosion Feature Ensemble
To build a small, yet powerful and diverse feature ensemble from the results of our feature study, we
consider the following grouping of statistics/value pairs: first, we contrast computation on the whole
region versus on erosion bands; second, we divide the values in intensity-based, color-based and gradient-
based; and then we use the natural distinction of simple, i.e., first-order versus second-order statistics.
This leads to 12 statistics/value groups including erosion and non- erosion based descriptors. Now, we
consider the best performing single feature in each group to be a candidate for our feature ensemble. The
performance of these candidates is marked bold in Table 3; blocks indicate the groups. The final feature
ensemble is derived from forward selection on the candidate set. This feature ensemble is then pruned by
backward elimination. With this approach we traced the combination of erosion band texture descriptors
with d = 1 on the three color channels (RGB) and on the gradient magnitude LBPs (Gmag LBP). This
best performing feature ensemble, named erosion feature ensemble (EFE), has 112 dimensions and its
classification accuracy is 75.2% on the dataset STUDY. In the following section we evaluate this feature
ensemble on the dataset FULL covering all extracted regions from the available cell phone images.
Results
We performed a 20-fold cross validation to evaluate EFE on the dataset FULL. We study two tasks, dis-
ease detection to predict whether a leaf spot is caused by an infection or not and disease classification,
where we predict one out of six classes (cerc, ram, pseu, rust, phom, and n-inf) per region.
315

For disease detection we consider two classes, inf and n-inf, where the former subsumes all infected
regions. Using EFE we achieve 93.3% accuracy on FULL (93.9% on STUDY), cf., . When detecting
diseases we are especially interested in a high recall. We detect 95.4% (FULL), respectively 99.2%
(STUDY), of all infected leaf spots as being caused by a disease. Without feature selection (ALL) detec-
tion accuracy is worse for STUDY and only slightly higher for FULL while requiring more computation
time. These results clearly show that erosion band features are highly appropriate for the identification
of leaf spots caused by fungal and bacterial diseases.
For disease classification EFE leads to an average accuracy of 83.8% whereas the use of all descriptor/
value combinations (ALL) achieves 87.6% on the dataset FULL. This is a highly convincing result as
the classification problem has six classes. Random class assignment results in an accuracy of 16.7% and
predicting the most frequent class yields 37.5%. ALL does outperform EFE, however, feature computa-
tion and prediction times increase by an order of magnitude as the dimension of the considered feature
space increases from 112 to 960. These results, summarized in Table 4, confirm the power of erosion
based texture descriptors for leaf spot classification and justify our proposed feature selection method.
For full the confusion matrix to analyze the per class performance of EFE is
1009 26 2 58 15 6 
 
 39 891 0 94 28 14
 

 2 0 65 1 0 0 
M =  
 44 64 5 335 17 3 
 
 9 24 0 6 163 16
 
 2 1 0 0 2 16
where the classes are n-inf, cerc, rust, pseu, ram, and phom (top to bottom resp. left to right), and the
off-diagonals of the rows (columns) show the number of false positives (false negatives). All classes
except phom have classification accuracies above 67%. The misclassified leaf spots caused by Phoma
were frequently labeled as Cercospora or Ramularia. This is due to their similar appearance and the
lack of sufficient training data in this class; less than 2% of the considered regions have the label phom.
These results suggest that we can achieve even higher performance rates for this class and the allover
evaluation once having access to more training data for Phoma.
Table 4. Average accuracies and standard errors in % (average prediction time in seconds) of the erosion
feature ensemble (EFE), all features (ALL), most frequent class (MFC) and random (RAND) for disease
detection (DETECT) and classification (CLASSIFY)
EFE ALL MFC RAND

DETECT STUDY 93.9 ± 1.9 (0.00″) 92.5 ± 1.6 (0.00″) 81.3 50.0
FULL 93.3 ± 0.4 (0.02″) 95.4 ± 0.5 (0.15″) 62.6 50.0
CLASSIFY STUDY 75.2 ± 1.9 (0.00″) 77.0 ± 2.6 (0.01″) 19.3 16.7
FULL 83.8 ± 0.7 (0.04″) 87.6 ± 0.6 (0.38″) 37.4 16.7
316

SOLUTIONS AND RECOMMENDATIONS
Personalized and instantaneous classification of crop diseases is an important problem. In this chapter,
we described and showcased a pattern recognition pipeline using smart phone images to achieve accurate
plant disease classification in real time. While our approach yields promising results already, we believe
that the integration of spatio-temporal information from other sources may further improve predictions.
In general, more and more sensors are being embedded into mobile devices such as smartphones, and this
is enabling a new generation of environmental sensing applications. To turn this technological makeover
of agriculture into a success, more pattern recognition approaches should be ported to smartphones.
Ultimately, this should yield novel, energy-efficient approaches and theoretical insights.
FUTURE RESEARCH DIRECTIONS
Geo-information processing plays a particular role in computational agriculture and precision farming.
Research in this area considers mobile access to geographically aggregated crop information (Kumar,
Dave, Bhadauriya, & Chaudhary, 2013), region specific yield prediction (Ruß & Brenning, 2010), or
environmental impact analysis (Gocht & Roder, 2011). Coupling disease classification as well as tasks
such as yield prediction with these approaches in order to improve prediction results is an exciting goal
for future research.
The success of erosion-based texture features suggests applying similar techniques to disease clas-
sification of other plants, such as cotton crops in the future. Feature-wise the exploration of other color
spaces such as the HSV representation seems promising (Camargo & Smith, 2009a). Further, we plan
to integrate the introduced pattern recognition pipeline in an automatic consultation system provided as
a smart phone app helping farmers in the field to correctly and instantly classify plant diseases. The use
of such a system will reduce costs and environmental burden due to a more targeted treatment.
The development of increasingly powerful mobile devices will certainly provide the opportunity
to implement more sophisticated pattern recognition approaches including more expressive features
directly on the cellular device. The resource-efficient implementation of features based on wavelets or
the Fourier transformation (Jähne, 2005) deserves certainly more attention.
CONCLUSION
In this chapter, we introduced a general pattern recognition pipeline to cell phone image based plant
disease classification. The general approach was then exemplified on two application scenarios: disease
prediction for wheat and sugar beets, two important commercial crops.
Region detection is based on a simple color filter, which is both flexible – to either retrieve compact
or scattered infection regions – and efficient to be executed on a cellular device. The considered features
are statistical texture descriptors derived from intensities, LBPs, color channels, gradient magnitudes,
and gradient magnitude LBPs. In both applications a combination of texture descriptors on the RGB
color channels as well as gradient magnitude LBPs lead to detection rates of over 90% and classification
accuracies of over 80%. For the classification of compact leaf spots, as present on sugar beet leaves, we
presented a feature derivation scheme based on statistical texture features computed on multiple erosion
317

bands of the regions of interest. In addition, we conducted an extensive feature study and systematically
derived the erosion feature ensemble, a discriminative, diverse, and robust feature ensemble for leaf spot
classification. The introduced approach has several benefits: The feature design allows for easy and ef-
ficient computation while achieving high predictive performance. It is a general framework applicable
to a wide range of plants and diseases and it goes beyond the use of descriptors simply computed on the
whole extracted regions as previously used for plant disease classification.
ACKNOWLEDGMENT
This work is supported by the Federal Ministry of Food, Agriculture and Consumer Protection (BMELV)
based on a decision of the German Federal Office for Agriculture and Food (BLE) under the innovation
support program; grant number “2815411310”.
REFERENCES
Afridi, M. J., Liu, X., & McGrath, J. M. (2014). An automated system for plant-level disease rating in
real fields. Proceedings of the 22nd International Conference on Pattern Recognition (pp. 148-153).
Los Alamitios, CA: IEEE Computer Society. doi:10.1109/ICPR.2014.35
Agrios, G. (1997). Plant Pathology (4th ed.). New York, NY: Academic Press.
Al-Hiary, H., Bani-Ahmad, S., Reyalat, M., Braik, M., & Alrahamneh, Z. (2011). Fast and accurate
detection and classification of plant diseases. International Journal of Computers and Applications,
17(1), 31–38. doi:10.5120/2183-2754
Arpa, A., Wetzstein, G., Lanman, D., & Raskar, R. (2012). Single lens off-chip cellphone microscopy.
In 2012 IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops
(pp. 23-28). Los Alamitios, CA: IEEE Computer Society. doi:10.1109/CVPRW.2012.6239195
Bauckhage, C. (2006). Tree-based signatures for shape classification. Proceedings of the 2006 IEEE
International Conference on Image Processing (pp. 2105-2108). Piscataway, NJ: IEEE Signal Process-
ing Society. doi:10.1109/ICIP.2006.312823
Bauckhage, C. & Kersting, K. (2013). Data mining and pattern recognition in agriculture. KI, 27(4),
313-324.
Bishop, C. M. (2006). Pattern Recognition and Machine Learning (Information Science and Statistics).
New York, NY: Springer.
Bock, C. H., Poole, G. H., Parker, P. E., & Gottwald, T. R. (2010). Plant disease severity estimated visu-
ally, by digital photography and image analysis, and by hyperspectral imaging. Critical Reviews in Plant
Sciences, 29(2), 59–107. doi:10.1080/07352681003617285
Bolton, M. D., Kolmer, J. A., & Garvin, D. F. (2008). Wheat leaf rust caused by puccinia triticina.
Molecular Plant Pathology, 9(5), 563–575. doi:10.1111/j.1364-3703.2008.00487.x PMID:19018988
318

Camargo, A., & Smith, J. (2009a). Image pattern classification for the identification of disease caus-
ing agents in plants. Computers and Electronics in Agriculture, 66(2), 121–125. doi:10.1016/j.com-
pag.2009.01.003
Camargo, A., & Smith, J. (2009b). An image-processing based algorithm to automatically identify plant
disease visual symptoms. Biosystems Engineering, 102(1), 9–21. doi:10.1016/j.biosystemseng.2008.09.030
Chen, T., Yap, K., & Zhang, D. (2014). Discriminative Soft Bag-of-Visual Phrase for Mobile Landmark
Recognition. IEEE Transactions on Multimedia, 16(3), 612–622. doi:10.1109/TMM.2014.2301978
Cristianini, N., & Shawe-Taylor, J. (2000). An Introduction to Support Vector Machines: And Other Kernel-
based Learning Methods. New York, NY: Cambridge University Press. doi:10.1017/CBO9780511801389
Doudkin, A. A., Inyutin, A. V., Petrovsky, A. I., Vatkin, M. E., & Sobkowiak, B. (2007). Three-level
neural network for data clusterization on images of infected crop field. Journal of Research and Ap-
plications in Agricultural Engineering, 52(1), 5–7.
Duncan, K. E., & Howard, R. J. (2000). Cytological analysis of wheat infection by the leaf blotch
pathogen mycosphaerella graminicola. Mycological Research, 104(9), 1074–1082. doi:10.1017/
S0953756299002294
Dupain, Y., Kamae, T., & Mendes-France, M. (1986). Can One Measure the Temperature of a Curve.
Archive for Rational Mechanics and Analysis, 94(2), 155–163. doi:10.1007/BF00280431
Ebrahim, Y., Ahmed, M., Chau, S. C., & Abdelsalam, W. (2007). An efficient shape representation and
description technique. Proceedings of the 2007 IEEE International Conference on Image Processing
(Vol. 6, pp. 441-444). Piscataway, NJ: IEEE Signal Processing Society. doi:10.1109/ICIP.2007.4379616
Fritz, G., Seifert, C., & Paletta, L. (2006). A Mobile Vision System for Urban Detection with Informa-
tive Local Descriptors. Proceedings of the 2006 IEEE International Conference on Computer Vision
Systems (p. 30). Piscataway, NJ: IEEE Computer Society. doi:10.1109/ICVS.2006.5
Fu, Z., Robles-Kelly, A., Tan, R. T., & Caelli, T. (2006). Invariant Object Material Identification via
Discriminant Learning on Absorption Features. In Conference on Computer Vision and Pattern Recog-
nition Workshop. (pp. 140-140). Piscataway, NJ: IEEE Computer Society.
Fu, Z., Robles-Kelly, A., & Zhou, J. (2011). MILIS: Multiple Instance Learning with Instance Selec-
tion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(5), 958–977. doi:10.1109/
TPAMI.2010.155 PMID:20733226
Gocht, A., & Roder, N. (2011). Salvage the treasure of geographic information in farm census data. Paper
presented at the 2011 International Congress of the European Association of Agricultural Economists,
Zurich, Switzerland.
Gonzales, R., & Woods, R. (2008). Digital Image Processing (3rd ed.). Upper Saddle River, NJ: Pearson
Prentice Hall.
Guyon, I., & Elisseeff, A. (2003). An introduction to variable and feature selection. Journal of Machine
Learning Research, 3, 1157–1182.
319

Hadid, A., Heikkilä, J. Y., Silvén, O., & Pietikäinen, M. (2007). Face and eye detection for person au-
thentication in mobile phones. Proceedings of the 2007 First ACM/IEEE International Conference on
Distributed Smart Cameras (pp. 101–108). Piscataway, NJ: IEEE. doi:10.1109/ICDSC.2007.4357512
Haralick, R. M., Shanmugam, K., & Dinstein (1973). Textural Features for Image Classification. Systems,
Man and Cybernetics, IEEE Transactions on, 3(6), 610-621.
Hastie, T., Tibshirani, R., & Friedman, J. (2009). The elements of statistical learning: data mining, infer-
ence and prediction (2nd ed.). New York, NY: Springer. doi:10.1007/978-0-387-84858-7
Huang, X., Li, S. Z., & Wang, Y. (2004). Shape localization based on statistical method using extended
local binary pattern. Proceedings of the 2004 IEEE First Symposium on Multi-Agent Security and Sur-
vivability (pp. 184-187). Piscataway, NJ: IEEE.
Jähne, B. (2005). Digital Image Processing (6th ed.). Berlin, Germany: Springer.
Kumar, N., Belhumeur, P. N., Biswas, A., Jacobs, D. W., Kress, W. J., Lopez, I. C., & Soares, J. V. (2012).
Leafsnap: A computer vision system for automatic plant species identification. Proceedings of the 12th
European Conference on Computer Vision (pp. 502-516). Berlin, Germany: Springer. doi:10.1007/978-
3-642-33709-3_36
Kumar, V., Dave, V., Bhadauriya, R., & Chaudhary, S. (2013). Krishimantra: Agricultural recommen-
dation system. Proceedings of the 3rd ACM Symposium on Computing for Development (p. 45). New
York, NY: ACM.
Liu, H., Coquin, D., Valet, L., & Cerutti, G. (2014). Leaf species classification based on a botanical
shape sub-classifier strategy. Proceedings of the 22nd International Conference on Pattern Recognition
(pp. 1496-1501). Los Alamitios, CA: IEEE Computer Society. doi:10.1109/ICPR.2014.266
Liu, Z.-Y., Wu, H.-F., & Huang, J.-F. (2010). Application of neural networks to discriminate fungal infec-
tion levels in rice panicles using hyperspectral reflectance and principal components analysis. Computers
and Electronics in Agriculture, 72(2), 99–106. doi:10.1016/j.compag.2010.03.003
Maturana, D., Mery, D., & Soto, A. (2009). Face recognition with local binary patterns, spatial pyramid
histograms and naive Bayes nearest neighbor classification. Proceedings of the 2009 International Con-
ference of the Chilean Computer Science Society (pp. 125-132). Los Alamitios, CA: IEEE Computer
Society. doi:10.1109/SCCC.2009.21
Neumann, M., Hallau, L., Klatt, B., Kersting, K., & Bauckhage, C. (2014). Erosion band features for
cell phone image based plant disease classification. Proceedings of the 22nd International Conference
on Pattern Recognition (pp. 3315-3320). Los Alamitios, CA: IEEE Computer Society. doi:10.1109/
ICPR.2014.571
Ng, C. K., Savvides, M., & Khosla, P. K. (2005). Real-time face verification system on a cell-phone us-
ing advanced correlation filters. Proceedings of the Fourth IEEE Workshop on Automatic Identification
Advanced Technologies (pp. 57-62). Los Alamitios, CA: IEEE Computer Society.
320

Ojala, T., Pietikäinen, M., & Mäenpää, T. (2002). Multiresolution Gray-scale and Rotation Invariant
Texture Classification with Local Binary Patterns. IEEE Transactions on Pattern Analysis and Machine
Intelligence, 24(7), 971–987. doi:10.1109/TPAMI.2002.1017623
Omrani, E., Khoshnevisan, B., Shamshirband, S., Saboohi, H., Anuar, N.B., & Nasir, M.H.N.M. (2014).
Potential of radial basis function-based support vector regression for apple disease detection. Measure-
ment, 55, 512–519. doi:10.1016/j.measurement.2014.05.033
Pydipati, R., Burks, T. F., & Lee, W. S. (2005). Statistical and neural network classifiers for citrus dis-
ease detection using machine vision. Transactions of the ASABE, 48(5), 2007. doi:10.13031/2013.19994
Rascher, U., Nichol, C. J., Small, C., & Hendricks, L. (2007). Monitoring spatio-temporal dynamics of
photosynthesis with a portable hyperspectral imaging system. Photogrammetric Engineering and Remote
Sensing, 73(1), 45–56. doi:10.14358/PERS.73.1.45
Römer, C., Wahabzada, M., Ballvora, A., Pinto, F., Rossini, M., & Panigada, C. et al. (2012). Early
drought stress detection in cereals: Simplex volume maximisation for hyperspectral image analysis.
Functional Plant Biology, 39(11), 878–890. doi:10.1071/FP12060
Rossi, V., Battilani, P., Chiusa, G., Giosuè, S., Languasco, L., & Racca, P. (2000). Components of rate-
reducing resistance to Cercospora leaf spot in sugar beet: Conidiation length, spore yield. Journal of
Plant Pathology, 82(2), 125–131.
Roters, J., Jiang, X., & Rothaus, K. (2011). Recognition of traffic lights in live video streams on mo-
bile devices. IEEE Transactions on Circuits and Systems for Video Technology, 21(10), 1497–1511.
doi:10.1109/TCSVT.2011.2163452
Rumpf, T., Mahlein, A. K., Steiner, U., Oerke, E. C., Dehne, H. W., & Plümer, L. (2010). Early detection
and classification of plant diseases with Support Vector Machines based on hyperspectral reflectance.
Computers and Electronics in Agriculture, 74(1), 91–99. doi:10.1016/j.compag.2010.06.009
Ruß, G., & Brenning, A. (2010). Data Mining in Precision Agriculture: Management of Spatial In-
formation. In Computational Intelligence for Knowledge-Based System Design (pp. 350–359). Berlin,
Germany: Springer.
Siddiqi, M. H., Ahmad, I., & Sulaiman, S. B. (2009, April). Weed recognition based on erosion and dila-
tion segmentation algorithm. Proceedings of the 2009 International Conference on Education Technology
and Computer (pp. 224-228). Los Alamitios, CA: IEEE Computer Society. doi:10.1109/ICETC.2009.62
Vankadavath, R. N., Hussain, A. J., Bodanapu, R., Kharshiing, E., Basha, P. O., & Gupta, S. et al. (2009).
Computer aided data acquisition tool for high-throughput phenotyping of plant populations. Plant Meth-
ods, 5(1), 18. doi:10.1186/1746-4811-5-18 PMID:20003250
Vazquez, E., Yang, X., & Slabaugh, G. (2013). Erosion band signatures for spatial extraction of features.
Machine Vision and Applications, 24(4), 695–705. doi:10.1007/s00138-012-0422-8
321

Weizheng, S., Yachun, W., Zhanliang, C., & Hongda, W. (2008). Grading method of leaf spot disease
based on image processing. Proceedings of the International Conference on Computer Science and
Software Engineering (Vol. 6, pp. 491-494). Los Alamitios, CA: IEEE Computer Society. doi:10.1109/
CSSE.2008.1649
Yao, Q., Guan, Z., Zhou, Y., Tang, J., Hu, Y., & Yang, B. (2009). Application of support vector machine
for detecting rice diseases using shape and color texture features. Proceedings of the 2009 International
Conference on Engineering Computation (pp. 79-83). Los Alamitios, CA: IEEE Computer Society.
doi:10.1109/ICEC.2009.73
KEY TERMS AND DEFINITIONS
Co-Occurrence Matrix: The co-occurrence matrix measures the distributions of co-occurring pixel
values in an image at a given angle and distance.
Descriptor (also Visual Descriptor): A descriptor is a numerical vector containing visual charac-
teristics, so-called features, of images.
Erosion Band Feature: Texture descriptor concatenating features computed on local image regions
derived from subsequent erosion steps.
Feature Selection: The process of selecting discriminative features given a specific classification
model and a set of labeled training data.
Gradient Image: An image with values indicating the gradient information of an original image.
Gradient information can be the gradient magnitude (computed in either dimension of the image or the
norm of the respective gradient vector) or the gradient direction.
Intensity Image: Single channel gray level image with values from 0 to 255. Can be computed from
RGB images by averaging the three color channels.
Local Binary Pattern (LBP): Computer vision feature characterizing local texture elements such
as horizontally, vertically or diagonally running edges or corners.
Pathogen: An infectious agent which can cause diseases to other organisms.
Statistical Texture Feature: Vector valued feature representing image texture as a quantitative
measure of the arrangement of intensities in an image or image region.
322
View publication stats

Cell Phone Image Based Plant Disease Classification

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Cell Phone Image Based Plant Disease Classification

Uploaded by

Copyright:

Available Formats

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

Cell phone image-based plant disease classiﬁcation

Chapter · October 2015

Marion Neumann Christian Bauckhage

SEE PROFILE SEE PROFILE

lectures on game AI View project

lectures on pattern recognition View project

The user has requested enhancement of the downloaded ﬁle.

A volume in the Advances in Environmental

British Cataloguing in Publication Data

For electronic access to this publication, please contact: eresources@igi-global.com.

PATTERN RECOGNITION PIPELINE FOR PLANT DISEASE CLASSIFICATION

Step 1: Image Preprocessing and Region Detection

Step 2: Feature Extraction

Image Information and Transformations

• Red (R), green (G), blue (B) channels,

1, if val (i, j ) = x and val (i + ∆i, j + ∆j ) = y 

Erosion Band Features

CLASSIFICATION OF WHEAT DISEASE SYMPTOMS

Problem Introduction and Data

Pattern Recognition (Steps 1 and 2)

Experimental Protocol (Step 3)

RGB+Texture5 All Selection MFC Rand

CLASSIFICATION OF SUGAR BEET DISEASE SYMPTOMS

Region Detection (Step 1)

Data, Features, and Experimental Protocol (Steps 2 and 3)

We use the following second-order statistics, CORRELATION, ENTROPY, INVERSE DIFFER-

INT INT R G B RG RB GB RGB Gmag Gdir Gmag

Erosion Feature Ensemble

EFE ALL MFC RAND

SOLUTIONS AND RECOMMENDATIONS

FUTURE RESEARCH DIRECTIONS

KEY TERMS AND DEFINITIONS

View publication stats

You might also like

For electronic access to this publication, please contact: eresources@igi-global.com.