Concept Explainability for Plant Diseases Classiﬁcation

Jihen Amara

1 a

, Birgitta K

onig-Ries

1,2 b

and Sheeba Samuel

1,2 c

Heinz Nixdorf Chair for Distributed Information Systems, Department of Mathematics and Computer Science,

Friedrich Schiller University Jena, Jena, Germany

Michael-Stifel-Center for Data-Driven and Simulation Science, Jena, Germany

Keywords:

Plant Disease Classiﬁcation, Explainable Artiﬁcial Intelligence, Convolutional Neural Networks, Testing with

Concept Activation Vectors (TCAV).

Abstract:

Plant diseases remain a considerable threat to food security and agricultural sustainability. Rapid and early

identiﬁcation of these diseases has become a signiﬁcant concern motivating several studies to rely on the in-

creasing global digitalization and the recent advances in computer vision based on deep learning. In fact, plant

disease classiﬁcation based on deep convolutional neural networks has shown impressive performance. How-

ever, these methods have yet to be adopted globally due to concerns regarding their robustness, transparency,

and the lack of explainability compared with their human experts counterparts. Methods such as saliency-

based approaches associating the network output to perturbations of the input pixels have been proposed to

give insights into these algorithms. Still, they are not easily comprehensible and not intuitive for human users

and are threatened by bias. In this work, we deploy a method called Testing with Concept Activation Vectors

(TCAV) that shifts the focus from pixels to user-deﬁned concepts. To the best of our knowledge, our paper is

the ﬁrst to employ this method in the ﬁeld of plant disease classiﬁcation. Important concepts such as color,

texture and disease related concepts were analyzed. The results suggest that concept-based explanation meth-

ods can signiﬁcantly beneﬁt automated plant disease identiﬁcation.

1 INTRODUCTION

Plant diseases are important factors as they result in

serious reduction in quality and quantity of agricul-

tural products. Therefore, early detection and diag-

nosis of these diseases are important. In our prior

work, we built a deep learning model based on con-

volutional neural networks (CNN) to identify diseases

from images of plant leaves (Amara et al., 2017) au-

tomatically. While successful, this type of model is a

black-box predictor preventing the acquisition of any

explanation for the predictions. We believe that the

availability of an explainable model that can rapidly

and accurately identify and quantify plant diseases

would have a signiﬁcant impact on scientiﬁc research

and smart crop production. Humans need to know

and understand about the detection, symptoms and

diagnosis process in addition to the high accuracy

of the plant disease classiﬁcation models. Hence, it

is widely believed that coupling black-box models

with interpretability techniques would increase their

https://orcid.org/0000-0003-3675-259X

https://orcid.org/0000-0002-2382-9722

https://orcid.org/0000-0002-7981-8504

adoption in the industry, agriculture, healthcare, and

other high-stakes ﬁelds (Molnar, 2020). In this paper,

we deﬁne interpretability as the ability to explain or

present in understandable terms to a human, as sug-

gested by (Doshi-Velez and Kim, 2017). We will use

explainability and interpretability interchangeably in

the paper. This urgent need for model interpretabil-

ity led to a proliferation of proposed methods. These

methods, that we will review in Section 2, follow a

common strategy which is simply highlighting pixels

that were relevant for a certain class classiﬁcation by

a neural network. However, they suffer from various

drawbacks. It has been shown that they are not as reli-

able as expected and are susceptible to human conﬁr-

mation biases (Ghorbani et al., 2019). Consequently,

a new line of research has focused on producing ex-

planations in the form of high-level ”concepts” (Kim

et al., 2018). Hence, our goal in this work is to inves-

tigate the usefulness of these methods that focus on

producing semantic and human-understandable ex-

planations for our use case plant diseases classiﬁca-

tion. We believe that semantics can enhance inter-

pretability in many areas. Instead of just displaying

numbers or saliency maps on image regions, these

246

Amara, J., König-Ries, B. and Samuel, S.

Concept Explainability for Plant Diseases Classiﬁcation.

DOI: 10.5220/0011667900003417

In Proceedings of the 18th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2023) - Volume 4: VISAPP, pages

246-253

ISBN: 978-989-758-634-7; ISSN: 2184-4321

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

methods output explanations that are understandable

by humans and based on interpretable concepts. For

a diseased plant leaf, instead of outputting a single

probability (e.g., 90% probability of having the late

blight disease), this type of algorithm would, for ex-

ample, output ”high texture irregularity, high amount

of black, yellow and brown areas on top of the leaf”

etc. One of the methods that belong to this family

is Testing with Concept Activation Vectors (TCAV)

(Kim et al., 2018). This method tests the sensitivity

of a trained deep neural model to a deﬁned concept of

interest. It also provides a global explanation for the

model. The central idea of TCAV is to evaluate how

responsive a CNN is to input patterns representing a

concept (e.g., color or texture) linked to the prediction

output of the CNN (e.g., the class ”late blight” dis-

ease). Therefore, in our work, interpretability refers

to a quantitative explanation of which plant disease

concepts are most important for accurate plant dis-

ease classiﬁcation by CNNs. Our research contribu-

tions are summarized as follows. First, to the best

of our knowledge, this study is the ﬁrst attempt at a

comprehensive understanding of what semantic con-

cepts the CNN learns during the plant disease diag-

noses. This is a critical issue for the vast prolifera-

tion of deep learning techniques in plant phenotyping

tasks. It can give insights about CNN models for plant

image analysis and help increase trust in such mod-

els. Second, we have presented a concept dataset that

can be used in the future to test concept methods with

plant disease image classiﬁcation. The remainder of

this paper is organized as follows. Section 2 presents

related work on interpretability and its application to

plant disease classiﬁcation, while Section 3 describes

the trained networks, the TCAV theory, and the uti-

lized datasets. Section 4 presents our experimental

results regarding prediction accuracy and model in-

terpretation. Finally, Section 5 depicts our conclusive

remarks and possible future work.

2 RELATED WORK

In recent years, there has been an increasing inter-

est in explainability and interpretability approaches to

deep learning. Two main sets of methods have been

proposed, which are saliency-based methods (Zeiler

and Fergus, 2014) and concept-based methods (Kim

et al., 2018; Zhou et al., 2018). The ﬁrst is based on

simply highlighting relevant pixels for a certain class

classiﬁcation by a neural network. Saliency-based

methods are also called feature attribution methods.

In the case of image classiﬁcation, features are in-

put pixels, and such methods aim to give each pixel

a value that can be understood as the pixel’s rele-

vance to the image’s classiﬁcation. Few works in the

literature focus on interpreting deep learning mod-

els for plant disease classiﬁcation using saliency-

based methods. For example, different papers have

tried to apply these visualization methods and present

a comparison study when applied to plant diseases

(Brahimi et al., 2018; Toda and Okura, 2019; Kinger

and Kulkarni, 2021). Other works have focused on

using visualization methods to extract the descrip-

tion of plant diseases from trained CNN (Sladojevic

et al., 2016; Ballester et al., 2017). In addition, some

papers tried to present novel visualization methods

for plant disease classiﬁcation (Ghosal et al., 2018;

Brahimi et al., 2019). These methods are beneﬁcial

because they give visual explanations, making it easy

to see the critically highlighted pixels. However, these

methods are considered fragile and sensitive to ad-

versarial perturbation (Ghorbani et al., 2019). Other

work has shown how these methods could be highly

unreliable (Kindermans et al., 2019). Since these

methods create importance maps based on individ-

ual input samples, they provide only local interpre-

tations and cannot explain the network’s decisions on

a global scale (Lucieri et al., 2020). In addition, these

methods’ lack of expressiveness to users is an essen-

tial drawback. For instance, the importance of a single

pixel in the classiﬁcation does not bring a meaningful

explanation, and it is also contrived by the number of

features (Molnar, 2020). Hence, concept-based ap-

proaches were proposed to address these limitations

(Zhou et al., 2018; Kim et al., 2018). Concepts can

be colors, objects, or abstract ideas. Users can deﬁne

these concepts without the need to train the network

on them. Hence, they are understandable by humans,

and they are not limited to the neural network feature

space. One of these methods is the TCAV approach,

which was proposed by (Kim et al., 2018). We will

describe this method in detail in the following section.

TCAV was successfully used with a different applica-

tion in the medical ﬁeld (Lucieri et al., 2020). To our

knowledge, concept-based explanation methods have

not previously been explored for plant disease classi-

ﬁcation networks. Hence, in this work, we adopt the

TCAV method to the problem of plant disease classi-

ﬁcation. We also present a set of concepts that could

be used in this case.

Concept Explainability for Plant Diseases Classiﬁcation

247

3 MATERIALS AND METHODS

3.1 Dataset and Networks for Plant

Diseases Classiﬁcation

3.1.1 Datasets

The Plant Village dataset is a public repository that

contains 54,323 images of 14 crops and 38 different

types of plant diseases (Hughes et al., 2015). It has

been extensively used by the community of plant dis-

ease image classiﬁcation. We used only images of

tomato leaves from this dataset to train the CNNs.

The total number of images is 18,160, divided into ten

classes (nine diseases and a healthy class). Finally,

the data was separated into three sets, containing 80%

of the data in the training set, and the remaining 20%

is divided equally between the testing and validation

sets. Figure 1 presents one example of each disease

class, and Table 1 summarizes our dataset.

Figure 1: Sample images from the Plant Village Dataset.

(a) Bacterial Spot, (b) Early Blight, (c) Healthy, (d) Late

Blight, (e) Leaf Mold, (f) Septoria Leaf Spot, (g) Spider

Mites, (h) Target Spot, (i) Mosaic Virus, and (j) Yellow Leaf

Curl Virus.

Table 1: The image numbers of each tomato disease class

in the Plant Village dataset.

Classes Number of Images

Bacterial Spot 2127

Early Blight 1000

Healthy 1591

Late Blight 1909

Leaf Mold 952

Septoria Leaf Spot 1771

Spider Mites 1676

Target Spot 1404

Mosaic Virus 373

Yellow Leaf Curl Virus 5357

Total 18,160

3.1.2 Models

In this work, we will focus on two widely used CNN

architectures: InceptionV3 (Szegedy et al., 2016) and

Vgg16 (Szegedy et al., 2015). We selected these

architectures due to their high use in the plant dis-

ease classiﬁcation literature (Lee et al., 2020). To

train these networks, we took advantage of ﬁne-tuned

transfer learning (Weiss et al., 2016), which is based

on transferring the knowledge gained from training

the models on a more signiﬁcant dataset to a smaller

one. The models were created and loaded with pre-

trained weights on the ImageNet dataset (Deng et al.,

2009). In addition, excluding top layers was per-

formed by deﬁning new layers on the top of the mod-

els. The altered top architecture consists of three

dense layers with corresponding dropout layers. For

training and optimizing the weights on the tomato dis-

ease dataset, we froze the ﬁrst 51 convolutional layers

and made the rest trainable for InceptionV3, and we

unfroze the ﬁnal convolutional layers of Vgg16 and

made them trainable. Training optimization was car-

ried out via stochastic gradient descent optimizer with

a learning rate of 0.0001 and momentum of 0.9. We

used a batch size of 20 and 20 epochs for training.

We use data augmentation techniques to increase the

dataset size in training while including different vari-

ations. These variations consist of transformations

such as random rotations, zooms, translations, shears,

and ﬂips to the training data as we train. Both models

were implemented using Keras (Chollet, 2021), and

were saved for subsequent performance testing and

interpretability analysis. We experimented on a server

with a GPU that consists of two NVIDIA Tesla V100

with 128 GB of RAM.

3.2 Network Explanation Through

Concepts

This section will introduce a deep neural network ex-

planation method called Testing with CAV (TCAV).

TCAV requires three types of data samples: samples

representing particular plant disease concepts (con-

cept samples), samples known to be speciﬁc plant dis-

ease classes (disease-class samples), and samples se-

lected randomly to supervise the training quality (to

ensure the stability of the results). In contrast to the

original work (Kim et al., 2018), where random ex-

amples may by chance include the concept, we en-

sure that the selected set is strictly negative. Below is

a detailed description of TCAV theory and our study’s

used concepts.

3.2.1 TCAV Theory

To better understand the concepts used by the trained

CNN to classify images as either healthy or diseased

leaves, we use the concept activation vectors (CAVs)

method (Kim et al., 2018) deﬁned as follows: Im-

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

248

portance of a ‘concept’ C (e.g., brown) to an image

class k (e.g., late blight) is found by taking the direc-

tional derivative of class predictions (for class k) at

each layer l of a CNN in the direction of (with respect

to) a CAV. Hence, a CAV, is a vector v

in the em-

bedding space of a CNN denoting the direction that

encodes the given concept C in the activation space of

a neural network layer l. To ﬁnd CAV of a concept

C, we need to prepare two datasets: a concept dataset,

which represents C, and a random dataset that does

not contain the concept. A binary concept classiﬁer

is trained to separate the activations generated by the

concept set from those generated by the random set

at a speciﬁc hidden layer l. The CAV v

is then de-

ﬁned as the normal to the hyperplane separating the

two classes at a particular layer l. Finally, given an

image input x, we can measure its conceptual sensi-

tivity by computing the directional derivative S of the

prediction in the direction of the CAV v

for concept

c,k,l

(x) = lim

ε→0

l,k

( f

(x) +εv

) − h

l,k

( f

(x))

= ∇h

l,k

( f

(x)).v

(1)

where f

maps the input x to the activation vector

of the layer l and h

l,k

maps the activation vector to

the logit output of class k. Then, to measure the in-

ﬂuence of a CAV on a class of input images, a metric

called TCAV score is computed. It employs the direc-

tional derivatives S

c,k,l

(x) to compute the contextual

sensitivity of a concept towards the whole inputs X

for class k. The TCAV score is given by:

TCAV

c,k,l



x ∈ X

c,k,l

(x) > 0



(2)

Hence, the TCAV score denotes the ratio of class k’s

inputs that are positively inﬂuenced by concept C.

Additionally, the authors perform a statistical signif-

icance test of TCAV scores to make sure that only

meaningful CAV s are taken into account. They com-

pute multiple CAVs between concept images and a

batch of random images. In addition, they train ran-

dom CAVs where both concept set and random set are

random images. Then they perform a two-sided t-test

of the TCAV scores based on these multiple samples.

The resulting concept is considered signiﬁcant for the

class prediction if they can reject the null hypothesis

of a TCAV score of 0.5. This helps to make sure that

concept CAVs and random CAVs are signiﬁcantly dif-

ferent from each other. TCAV authors further perform

Bonferroni correction (p < α/m,m = 2) for multiple

comparisons between all concept-random pairs to re-

duce potential for false positives (incorrect rejection

of the null hypothesis, or a Type I error) to prevent

mistaking as signiﬁcant a truly insigniﬁcant concept

(Kim et al., 2018). In our work, TCAV is imple-

mented using Keras rather than the original authors’

Tensorﬂow code (Kim et al., 2018).

3.2.2 Concepts Used for Analysis

Identiﬁcation of plant diseases is commonly accom-

plished through visual inspection of the disease’s ef-

fect on the plant. This effect is considered a symptom

and could be detectable as a change of color or texture

of the leaf caused by the pathogen. In our work, we

would like to see if the CNN learned representation

of various concepts and those often used by plant dis-

eases experts are complementary. Hence, we choose

to test the concepts of color, texture, and late blight

disease pattern described below. Also, to show the

feasibility of this methodology and to have a general

disease pattern, we focus on one disease class: late

blight disease, which affects both potatoes and toma-

toes. The concepts used in this work to interpret the

deep classiﬁers are deﬁned below in accordance with

the description in plant stress phenotyping literature

(Isleib, 2012).

Color Concepts. To train TCAV, we used colors

as concepts such as red, brown, blue, yellow, and

green. We wanted to test our trained models’ sensitiv-

ity to these colors especially green, brown, and yellow

which can present the difference between healthy and

diseased states. We will hence gain more comprehen-

sion of how our internal network behaves. Random

images used are grayscale images of diseased plant

leaves other than tomato. We produced 100 pictures

per color synthetically by generating the color chan-

nel randomly.

Describable Texture Dataset (DTD) Concepts

(Cimpoi et al., 2014). Texture is an essential part of

plant disease identiﬁcation as it depicts more knowl-

edge of the infected leaf region. Infected leaves can

turn dry and present signs like crackedness, wrin-

kles, bumpiness, ﬁbrousness, etc. Hence, this work

presents a dataset of texture concepts that could be

used with plant diseases and are also understandable

to humans. The dataset is extracted from the DTD

database (Cimpoi et al., 2014). DTD is a texture

database collected “in the wild” with semantic at-

tributes. It contains a list of 47 categories inspired

by human perception. The concepts concerning plant

diseases that we choose from DTD are: blotchy,

bumpy, cracked, ﬁbrous, pitted and wrinkled. For

the TCAV experiment, we chose 100 images for each

concept category. For the random images, we selected

healthy images of leaves other than Tomato. Figure 2

presents some examples from the DTD categories.

Concept Explainability for Plant Diseases Classiﬁcation

249

Figure 2: Concepts related to plant disease symptoms

extracted from DTD texture database (blotchy, bumpy,

cracked, pitted, ﬁbrous, wrinkled) (Cimpoi et al., 2014).

Late Blight Disease Concepts. Late blight is a devas-

tating disease that can affect potato and tomato plants.

It is caused by the water mold Phytophthora infestans

(Britannica, 2020). The ﬁrst symptoms of the disease

can start with light to dark green spots in the leaves,

which rapidly develop into large dark brown or black

lesions. Leaf lesions are also frequently surrounded

by a yellow chlorotic halo. Leaves can become dry

and shriveled, and crops can be severely damaged. In

our experiment, we use the images of potato leaves in-

fected by the late blight disease as a visual representa-

tion of the disease concept. These images will depict

the disease pattern that appears on the infected leaves.

It will help us test the network’s ability to learn the

representation of the disease independently from the

leaf shape or type. We used healthy leaves of plants

other than tomatoes and potatoes for the random im-

ages.

4 RESULTS AND ANALYZES

4.1 Network Training Results

This section will present the results of the models

evaluation on our dataset. Table 2 shows the found

results. The performance of the trained models is

evaluated using recall, precision, and accuracy met-

rics (Davis and Goadrich, 2006).

Table 2: Performance measures for pre-trained models

Vgg16 and InceptionV3.

InceptionV3 Vgg16

Accuracy 0.92 0.98

Precision 0.93 0.98

Recall 0.92 0.98

F1-score 0.92 0.98

As shown in Table 2, Vgg16 and InceptionV3

trained on the Tomato disease dataset have achieved

good results with a predominance from Vgg16

achieving an accuracy of 98% compared to Incep-

tionV3 with an accuracy of 92%.

4.2 TCAV Experiments Result

We will concentrate on the quantitative evaluation of

the TCAV score. The TCAV score quantiﬁes a given

concept’s positive or negative inﬂuence on a speciﬁc

target class. For InceptionV3, in each experiment,

TCAV scores were computed within 9 layers (mixed3,

mixed5 to mixed9), which are the concatenation lay-

ers at the end of each inception module. The num-

bers denotes the location depth of a layer in the net-

work. This can help us see how different network

depths affected the ﬁnal classiﬁcation. For Vgg16,

TCAV scores were computed within three layers (ﬂat-

ten, dense, and dense 1). We tested the ﬁnal layers

added on top of the model, which go from the shal-

lowest (ﬂatten) to the deepest (dense 1). We train

CAVs on the activations extracted from these chosen

layers.

4.2.1 Importance of Colors Concepts

Figure 3: TCAV scores for color concepts red, brown,

blue, yellow and green in InceptionV3 with layers mixed

3, mixed 5, mixed 6, mixed 7, mixed 8 and mixed 9.

Figure 4: TCAV scores for color concepts red, brown, blue,

yellow and green in Vgg16 with layers dense 1, dense and

ﬂatten.

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

250

Results for InceptionV3 are shown in Figure 3, where

we tested layers mixed 3, mixed 5, mixed 6, mixed

7, mixed 8, and mixed 9. The layers that should be

chosen for testing are those who passed the statistical

testing (t-test) mentioned above successfully for most

of the concepts. In consistency with plant diseases ex-

perts description (Isleib, 2012) we found that brown

and yellow were both signiﬁcant for identifying late

blight disease, with TCAV scores as high as 0.83 for

brown in mixed 9 and 0.64 for yellow in mixed 6.

Also, green was signiﬁcant in layer mixed 6 with a

TCAV score of 0.69. In contrast, the rest of the col-

ors were less important. The bars marked by a star

‘*’ indicate TCAV scores that did not pass the sta-

tistical testing. For Vgg16 (Figure 4), the important

color concepts captured by the model were green and

brown. To our surprise, it seems red was also crucial

for the model in classifying the disease. Green had a

TCAV score of 1 for the three tested layers dense 1,

dense, and ﬂatten. While brown had TCAV scores of

0.9 and 0.87 for the dense and ﬂatten layers. More-

over, red had TCAV scores of 1 and 0.9 for layers

dense 1 and dense. Hence, even though Vgg16 had

better classiﬁcation performance than InceptionV3,

the latter was better at capturing color concepts simi-

lar to those used by experts to classify plant diseases.

4.2.2 Importance of Describable Texture

Dataset (DTD) Concepts

To show our trained models’ sensitivity to textures re-

lated to plant disease symptoms, we experiment on a

texture recognition dataset (DTD). As shown in Fig-

ure 5 and Figure 6, we got a good insight into what

texture concepts exactly were important for Incep-

tionV3 and Vgg16 when looking to identify a plant

disease. TCAV scores show that concepts such as

blotchy, cracked, wrinkled, bumpy, ﬁbrous, and pit-

ted are highly relevant to the late blight disease class.

In the case of InceptionV3 (Figure 5), we found that

Figure 5: Conceptual importance (TCAV scores) of DTD

concepts blotchy, cracked, wrinkled, bumpy, ﬁbrous, pitted

in InceptionV3.

Figure 6: Conceptual importance (TCAV scores) of DTD

concepts blotchy, cracked, wrinkled, bumpy, ﬁbrous, pitted

in Vgg16.

the layers that could capture these important texture

concepts were mixed 3 and 9, while for the rest of the

layers, TCAV scores were insigniﬁcant. This ﬁnding

can raise the possibility that some layers can be omit-

ted and consequently reduce the parameters needed

for the network training. Especially in the case of

plant disease classiﬁcation, where mobile phones can

be used for such tasks, reducing the number of net-

work parameters is exceptionally beneﬁcial for mem-

ory and calculation efﬁciency. In the other case, re-

sults show that Vgg16 (Figure 6) is sensitive to all

the DTD concepts compared to InceptionV3. TCAV

scores were as high as 1 and 0.9 for the tested layers

dense, dense 1, and ﬂatten, which indicate the sen-

sitivity of these layers to such concepts when classi-

fying late blight disease. In the case of our training

set, images are about a single leaf in a uniform back-

ground, which makes recognizing the change of tex-

tures on the leaf sufﬁcient to solve the task of plant

disease classiﬁcation. Hence, the results show that

different texture detectors emerge, which explains

why the texture concepts dominate in the tested lay-

ers especially in the case of Vgg16. This discov-

ery is consistent with the former qualitative ﬁnding

from (Toda and Okura, 2019) where feature visual-

ization methods showed that the model focused on

learning the visual cues (textutre) of the disease le-

sions rather than objects or shapes. Hence, TCAV

permit for quantitative conﬁrmation of this previous

qualitative discovery. Results look very much as ex-

pected and are aligned with the description in plant

stress phenotyping literature (Isleib, 2012).

4.2.3 Importance of Late Blight Disease

Concepts

To assess the importance of the late blight concept

for both InceptionV3 and Vgg16 when classifying the

disease, we experimented with images of potato with

late blight disease and healthy images of tomato. In-

Concept Explainability for Plant Diseases Classiﬁcation

251

Figure 7: Conceptual importance (TCAV scores) of late

blight disease concept in InceptionV3.

Figure 8: Conceptual importance (TCAV scores) of late

blight disease concept in Vgg16.

terestingly, from Figure 7 we can see that, for the In-

ceptionV3, the late blight concepts contribute posi-

tively to the tomato late blight disease class. While

the healthy concepts don’t have any signiﬁcant con-

tribution. To our surprise, this was the opposite for

Vgg16 (Figure 8) where the late blight concepts did

not play a role. A possible explanation for these re-

sults may be the fact that the network did not focus on

disease regions but crop-speciﬁc characteristics such

as leaf shape and also for the fact that InceptionV3 is

a deeper network than Vgg16 which helped in captur-

ing better the concepts. These ﬁndings further support

the idea of training models with common diseases re-

gardless of crop type for more generalizability (Lee

et al., 2020). Also, it shows that regardless of the

accuracy that indicates a good learning performance,

DL models could inherently learn or fail to learn rep-

resentations from the data which an expert might con-

sider important (Isleib, 2012). This enforces the im-

portance of including explainability in the deep learn-

ing workﬂow.

5 CONCLUSIONS

The inﬂuence of artiﬁcial intelligence in modern agri-

culture is increasing with the incorporation of new

technologies such as machine learning and robotics to

boost crop abundance and quality. Plant pathologists

are starting to use deep learning techniques to help

in disease classiﬁcation, surveillance, and manage-

ment. However, they have yet to be adopted globally

due to concerns regarding robustness, transparency,

and lack of explainability compared with their hu-

man experts counterparts. Hence, in our work, we

adopted a concept-based explanation method called

TCAV for plant disease classiﬁcation to gain more in-

sights into the model and which concepts affect its

decisions. Based on transfer learning techniques, we

have trained two famous networks, Vgg16 and Incep-

tionV3. Important concepts such as color, texture,

and concept-related disease were analyzed. Our re-

sults show that deep learning based models can learn

and use similar disease-related concepts for predic-

tion as plant pathologists use. Nevertheless, the study

could be extended to grant a more comprehensive in-

terpretation of the TCAV scores for this particular use

case. It would be great to exploit more granular la-

beling of disease concepts and explore more disease

classes to draw further insight into the model’s classi-

ﬁcation mechanism and to give better validation of its

decisions. Future work could incorporate plant dis-

ease datasets from the ﬁeld with real background, and

then, we can use TCAV to ensure the model is un-

biased towards the background. Also, ﬁnding a way

to deﬁne concepts automatically could simplify the

process by eliminating the necessity for manual an-

notations and could allow revealing new knowledge

for plant disease experts or unexpected biases from

the network. We would also like to investigate ways

of adding semantics through context and Knowledge

Graphs into TCAV.

ACKNOWLEDGEMENTS

We would like to thank Chiheb Karray for helpful dis-

cussions. We thank the German Academic Exchange

Service (DAAD) for supporting the work of Jihen

Amara and the Carl Zeiss Foundation for the ﬁnancial

support of the project “A Virtual Werkstatt for Digiti-

zation in the Sciences (K3)” within the scope of the

program line “Breakthroughs: Exploring Intelligent

Systems for Digitization”- explore the basics, use ap-

plications”. The computational experiments were per-

formed on resources of Friedrich Schiller University

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

252

Jena supported in part by DFG grants INST 275/334-

1 FUGG and INST 275/363-1 FUGG.

REFERENCES

Amara, J., Bouaziz, B., Algergawy, A., et al. (2017). A deep

learning-based approach for banana leaf diseases clas-

siﬁcation. In BTW (workshops), volume 266, pages

79–88.

Ballester, P., Correa, U. B., Birck, M., and Araujo, R.

(2017). Assessing the performance of convolutional

neural networks on classifying disorders in apple tree

leaves. In Latin American Workshop on Computa-

tional Neuroscience, pages 31–38. Springer.

Brahimi, M., Arsenovic, M., Laraba, S., Sladojevic, S.,

Boukhalfa, K., and Moussaoui, A. (2018). Deep learn-

ing for plant diseases: detection and saliency map vi-

sualisation. In Human and machine learning, pages

93–117. Springer.

Brahimi, M., Mahmoudi, S., Boukhalfa, K., and Mous-

saoui, A. (2019). Deep interpretable architecture for

plant diseases classiﬁcation. In 2019 Signal Process-

ing: Algorithms, Architectures, Arrangements, and

Applications (SPA), pages 111–116. IEEE.

Britannica, T. (2020). Editors of encyclopaedia. Argon.

Encyclopedia Britannica.

Chollet, F. (2021). Deep learning with Python. Simon and

Schuster.

Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., and

Vedaldi, A. (2014). Describing textures in the wild.

In Proceedings of the IEEE conference on computer

vision and pattern recognition, pages 3606–3613.

Davis, J. and Goadrich, M. (2006). The relationship be-

tween precision-recall and roc curves. In Proceed-

ings of the 23rd international conference on Machine

learning, pages 233–240.

Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-

Fei, L. (2009). Imagenet: A large-scale hierarchical

image database. In 2009 IEEE conference on com-

puter vision and pattern recognition, pages 248–255.

Ieee.

Doshi-Velez, F. and Kim, B. (2017). Towards a rigorous sci-

ence of interpretable machine learning. arXiv preprint

arXiv:1702.08608.

Ghorbani, A., Abid, A., and Zou, J. (2019). Interpretation

of neural networks is fragile. In Proceedings of the

AAAI conference on artiﬁcial intelligence, volume 33,

pages 3681–3688.

Ghosal, S., Blystone, D., Singh, A. K., Ganapathysubrama-

nian, B., Singh, A., and Sarkar, S. (2018). An explain-

able deep machine vision framework for plant stress

phenotyping. Proceedings of the National Academy

of Sciences, 115(18):4613–4618.

Hughes, D., Salath

e, M., et al. (2015). An open access

repository of images on plant health to enable the

development of mobile disease diagnostics. arXiv

preprint arXiv:1511.08060.

Isleib, J. (2012). Signs and symptoms of plant disease: Is

it fungal, viral or bacterial. Michigan State University

Extension.

Kim, B., Wattenberg, M., Gilmer, J., Cai, C., Wexler, J.,

Viegas, F., et al. (2018). Interpretability beyond fea-

ture attribution: Quantitative testing with concept ac-

tivation vectors (tcav). In International conference on

machine learning, pages 2668–2677. PMLR.

Kindermans, P.-J., Hooker, S., Adebayo, J., Alber, M.,

Sch

utt, K. T., D

ahne, S., Erhan, D., and Kim, B.

(2019). The (un) reliability of saliency methods. In

Explainable AI: Interpreting, Explaining and Visual-

izing Deep Learning, pages 267–280. Springer.

Kinger, S. and Kulkarni, V. (2021). Explainable ai for deep

learning based disease detection. In 2021 Thirteenth

International Conference on Contemporary Comput-

ing (IC3-2021), pages 209–216.

Lee, S. H., Go

eau, H., Bonnet, P., and Joly, A. (2020). New

perspectives on plant disease characterization based

on deep learning. Computers and Electronics in Agri-

culture, 170:105220.

Lucieri, A., Bajwa, M. N., Braun, S. A., Malik, M. I., Den-

gel, A., and Ahmed, S. (2020). On interpretability of

deep learning based skin lesion classiﬁers using con-

cept activation vectors. In 2020 international joint

conference on neural networks (IJCNN), pages 1–10.

IEEE.

Molnar, C. (2020). Interpretable machine learning. Lulu.

com.

Sladojevic, S., Arsenovic, M., Anderla, A., Culibrk, D., and

Stefanovic, D. (2016). Deep neural networks based

recognition of plant diseases by leaf image classiﬁ-

cation. Computational intelligence and neuroscience,

2016.

Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S.,

Anguelov, D., Erhan, D., Vanhoucke, V., and Rabi-

novich, A. (2015). Going deeper with convolutions.

In Proceedings of the IEEE conference on computer

vision and pattern recognition, pages 1–9.

Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., and Wo-

jna, Z. (2016). Rethinking the inception architecture

for computer vision. In Proceedings of the IEEE con-

ference on computer vision and pattern recognition,

pages 2818–2826.

Toda, Y. and Okura, F. (2019). How convolutional neural

networks diagnose plant disease. Plant Phenomics,

2019.

Weiss, K., Khoshgoftaar, T. M., and Wang, D. (2016).

A survey of transfer learning. Journal of Big data,

3(1):1–40.

Zeiler, M. D. and Fergus, R. (2014). Visualizing and under-

standing convolutional networks. In European confer-

ence on computer vision, pages 818–833. Springer.

Zhou, B., Sun, Y., Bau, D., and Torralba, A. (2018). Inter-

pretable basis decomposition for visual explanation.

In Proceedings of the European Conference on Com-

puter Vision (ECCV), pages 119–134.

Concept Explainability for Plant Diseases Classiﬁcation

253