Volume 9, Issue 11, November– 2024 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24NOV1036

Deep Learning Based Detection of Corn Stalk

Diseases with XAI
Ravishka Ranasinghe1 Guhanathan Poravi2
School of Computing School of Computing
Asia Pacific Institute of Information Technology Asia Pacific Institute of Information Technology
Colombo 02, Sri Lanka Colombo 02, Sri Lanka

Abstract:- This research aims at developing a deep disease detection system that leverages deep learning for
learning model for corn stalk disease detection (for accurate predictions and Grad-CAM for explainability.
anthracnose disease) using explainable AI approaches,
Grad-CAM. Based on the CNN deep learning models, the II. LITERATURE REVIEW
proposed system is developed. For inputs including
images of corn stalks, the system prepares them,  Corn Stalk Disease Detection
generates visual descriptions of the layer, and correctly Corn stalk diseases have unique patterns, therefore, most
categorizes the image as being related to a healthy corn of the existing disease detection systems find it difficult to
plant or a diseased one. Overall, it plays a part in identify them with higher accuracy. Therefore, disease
enhancing explainability of model predictions to the end detection in corn stalks relies on manual inspection, which is
user, especially the uninitiated in the aspect of some level time consuming, subjective and has more chance for errors.
of understanding. However, the system significantly saves Also, another key issue is that these diseases are identified
training time and computational expense by using when the symptoms are very visible, which means late
transfer learning without a decline in accuracy. detection [4]. By that time, it’ll be too late for diagnosis. Also,
another challenge in making a system to identify corn stalk
Keywords:- Deep Learning, Grad-CAM, Convolutional diseases is lack of properly labelled, high quality, publicly
Neural Networks, Image Classification, Explainable AI. available dataset. This is one of the main reasons for not
having much research on corn stalk diseases.
I. INTRODUCTION
 Image Processing and Deep Learning in Agricultural
Corn is among the most productive and dominant crops Disease Detection
grown around the world today. However, corn production Edge detection, contrast enhancement, noise reduction
faces significant challenges due to various diseases, including processors are some pre-processing techniques used in image
those affecting the stalk [1]. Stalk is the main part of the corn processing. [5].These methods also focus on the subject
plant, and it should be very strong in terms of handle the selection, what type of features of corn stalks indicate
weight of the corn ears and creating paths for the nutrients to disease, like lesions or discoloration. Disease symptoms are
run in them [2]. Anthracnose disease is a common fungal generally identified using deep learning models, mainly CNN
disease that softens the stalks, reduces yield and, in severe after some level of processing. CNNs can be trained directly
cases, results in the failure of the crop [3].One of the on raw image data to provide raw hierarchical features that
important factors in the disease control and reduction of their are ideal for disease detection of complex patterns in plant
effects is its initial and proper diagnosis. disease.[6]

 Anthracnose Disease in Corn Stalks Deep learning models such as AlexNet, VGGNet,
Currently there is limited research specifically targeting RESNet, have been used for plant disease detection, obtaining
corn stalks diseases, especially anthracnose disease. These high accuracy for plant types [4]. Given big datasets of
diseases have unique patterns, existing systems fail to labeled images from which these models can be trained, it can
identify. Therefore, disease detection in corn stalks relies on be taught to distinguish between healthy plants and diseased
manual inspection. But it is time consuming, subjective and plants based on feature, if trained on.
can be prone to errors more often.
Transfer learning refers to the practice of utilizing a pre-
 Motivation existing model that has been trained on a different but similar
Most existing research in the domain of corn disease problem. The corn stalk disease dataset is employed to further
detection has primarily focused on corn leaves rather than the train algorithms the foundational architecture of which was
stalk. Out of the very few research, there is no evidence for created on vast image datasets. This strategy builds from
anthracnose detection. This disease is causing significant yield existing information to not require a large amount of data and
losses, impacting food security and economic well-being, resources to train effective models from scratch.[7]
particularly for smallholder farmers. To address this
challenge, this research proposes an AI-driven corn stalk

IJISRT24NOV1036 www.ijisrt.com 1554

Volume 9, Issue 11, November– 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24NOV1036

Image augmentation is a method used to enhance the III. EXISTING WORK

variety of the training dataset without the need to gather more
data [5]. To increase reliability of machine learning models Several works have already shown the capability of deep
and prevent model overfitting on maize stalks, image learning in detecting plant diseases. [10] for instance used a
preprocessing is applied as rotation, scaling or flipping of the deep convolutional neural network (CNN) to categorize 38
image. Normalization, contrast adjustment and noise different plant diseases at an accuracy over 99%. Similar to
reduction are among the data pre-processing methods predicting diseases on corn stalks where signs that are evident
performed on the images in order to improve the efficiency of include lesions, discolorations, or fungal growths on the stalks
machines learning algorithms. On one hand, preprocessing and leaves, the study centred its research on the ability of
ensures that the input photos are in a uniform format, and image pattern recognition of the CNN to predict the patterns
eases recovery of better features by the models.[6] in the images of the leaves.

 Explainable AI in Agriculture Similarly [4] applied deep learning models such as

Deep learning models reach high accuracy, but their AlexNet and VGGNet to plant disease detection, obtaining
‘black box’ nature prevents user from understanding how the high accuracy for plant types. Large datasets of labeled
decision is being made. In agriculture especially farmers and images from which these models can be trained, can teach it
agricultural experts want to trust the predictions made by to differentiate healthy plants from diseased plants based on
these models and this becomes a problem. Explanation of AI feature, if trained on.
(XAI) explains this problem by giving insights into the
model’s decision-making process. To produce visual Sometimes deep learning models achieve very high
explanations for predictions of the models XAI techniques accuracy, but this approach puzzles the user because the
such as Grad-CAM, LIME, SHAP can be used as it detailed rationale of how the decision was made cannot be
highlights the part of an image that most affected the model’s explained. In agriculture especially farmers and agricultural
prediction. [8] experts want to trust the predictions made by these models
and this becomes a problem.. [9] research has shown that
Gradient based methods like Grad-CAM produce Grad-CAM can provide meaningful explanations for deep
heatmaps on input images and highlight the areas that learning models in different domains such as plant disease
influenced model’s predictions in identifying plant diseases detection.
[9]. LIME explains individual predictions by approximating
intricate models with comprehensible ones, offering insights IV. PROPOSED APPROACH
into the features (e.g., leaf colour, texture) that influenced the
disease detection decision. SHAP measures the impact of The author proposed to develop a deep learning-based
each variable on the model's output, facilitating the system for detecting corn stalk diseases, with a focus on
identification of specific characteristics (e.g., size, spots, or anthracnose, by utilizing ResNet50 for disease prediction and
colour of lesions) influenced illness classification. incorporating Grad-CAM for explainability, ensuring high
accuracy and interpretability for real-world agricultural
Explainable Artificial Intelligence (XAI) methodologies applications.
are progressively utilised in agricultural disease identification
to enhance the transparency, interpretability, and reliability of  Dataset
AI-driven judgements for farmers and researchers. Through The author made a dataset by combining two publicly
incorporating the XAI into the disease detection system, the available datasets in Kaggle and labelled it with the help of
model is not a black box as it assists disease identification domain experts. The made dataset consisted of 600 images as
while explaining its decision-making process. In particular, 300 images of stalks with anthracnose disease and 300
the field where the deployment of such systems can have images of healthy stalks.
critical implications, accurate predictions are crucial.

Fig 1 Samples of the Dataset

IJISRT24NOV1036 www.ijisrt.com 1555

Volume 9, Issue 11, November– 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24NOV1036

 Data Pre-Processing  Training the Model

The author has split the dataset into three sets as 70% of The author trained the model for 30 epochs on a dataset
the total dataset used for training, 15% used for validation and evaluated the performance using training accuracy/loss
and the remaining 15% used for testing. Additionally, the and validation accuracy/loss. The model achieved high
author has mentioned that data augmentation techniques were accuracy on both training and validation data which indicates
used during the training process to increase diversity of the that it learned to make accurate decisions. At the same time
training set and improve model’s ability to generalize to new, the loss values also decreased gradually which shows the
unseen data. confidence in model’s predictions. The also tried different
hyperparameters to reduce overfitting.

Fig 2 Training and Validation Loss

 Hyper Parameter Tuning V. EVALUATION

The author has used EarlyStopping which monitors the
validation loss and stops if there is no improvement for 5  Classification Report
consecutive epochs. Also, the author For the evaluation of the model’s performance, the
usedReduceLROnPlateau and the minimum learning rate is classification report for test data is used. It shows metrics like
set to 0.00001. These hyperparameter tunings helped to precision, recall and F1-Score. The model achieved an
prevent overfitting and ensure model is generalized well to accuracy of 94% and a recall of 0.91. The precision of the
new data. model is 0.98. The F1-Score which is a harmonic mean of
precision and recall is 0.95, indicating that the model has a
good balance between precision and recall.

Fig 3 Classification Report

IJISRT24NOV1036 www.ijisrt.com 1556

Volume 9, Issue 11, November– 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24NOV1036

 Confusion Matrix and 45 diseased). Out of the 45 anthracnose diseased stalks,

Confusion Matrix is used for the evaluation of the 41 were predicted correctly and only 4 were misclassified.
model. The low number of misclassifications reflected in Out of the 45 healthy stalks only 1 was misclassified.
high accuracy. 90 images were used for testing (45 healthy

Fig 4 Confusion Matrix

 Grad-CAM Heatmaps important features which was used by the model for the
The author has used Grad-CAM (Gradient weighted predictions. The generated Grad-CAM heatmaps have shown
Class Activation Mapping) to explain the predictions made that the model is accurately identifying the symptoms for
by the model. Grad-CAM is an XAI (Explainable Artificial anthracnose disease in corn stalks. By comparing with the
Intelligence) technique which helps for understanding of original image most of the regions that contains symptoms
model’s decision-making process by highlighting the are highlighted in the heatmaps.

IJISRT24NOV1036 www.ijisrt.com 1557

Volume 9, Issue 11, November– 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24NOV1036

Fig 5 Heatmaps vs Original Image

VI. DISCUSSION ACKNOWLEDGMENT

The research done by the author highlights the Mr. Guhanathan Poravi thanks to his critiques,
challenging task of identifying corn stalk diseases, especially suggestions, and research viewpoints that the author was able
anthracnose disease. This is a significant issue in the domain to successfully conduct this research.
of agricultural disease identification, where most of the
systems either focuses on leaves or the whole plant. The REFERENCES
research is an important contribution to this field as it
addresses a critical challenge and proposes a solution for it. [1] N. Ansori, A. Rachmad, E. Mala, S. Rochman, and Y.
Panca Asmara, ‘Corn stalk disease classification using
In addition, for the proposing of the corn stalk disease random forest combination of extraction features’,
identification system, the author also has incorporated XAI Commun. Math. Biol. Neurosci., 2024, doi:
techniques like Grad-CAM to improve interpretability. This 10.28919/cmbn/8404.
is particularly important as explainability is a key factor in [2] L. S. Farmer Musings of a Pig, ‘The Importance of
ensuring the reliability and trustworthiness of the system. Corn’, Latham Hi-Tech Seeds. Accessed: Apr. 04,
2024. [Online]. Available:
VII. FUTURE ENHANCEMENTS https://www.lathamseeds.com/2012/06/the-
importance-of-corn/
As for the future enhancements, the author will develop [3] T. A. Jackson-Ziems, J. M. Rees, and R. M. Harveson,
a mobile application to allow farmers in remote areas to ‘Common Stalk Rot Diseases of Corn’, 2014.
quickly upload images of corn stalks and receive predictions [4] K. P. Ferentinos, ‘Deep learning models for plant
in real time. Also, the author will explore more XAI disease detection and diagnosis’, Comput. Electron.
techniques like a hybrid approach to improve the Agric., vol. 145, pp. 311–318, Feb. 2018, doi:
interpretability more. 10.1016/j.compag.2018.01.009.
[5] N. Ahmad, H. M. S. Asif, G. Saleem, M. U. Younus,
VIII. CONCLUSION S. Anwar, and M. R. Anjum, ‘Leaf Image-Based Plant
Disease Identification Using Color and Texture
In conclusion, this research successfully developed a Features’, Wirel. Pers. Commun., vol. 121, no. 2, pp.
deep learning-based corn stalk disease detection system using 1139–1168, Nov. 2021, doi: 10.1007/s11277-021-
ResNet50 and Grad-CAM for interpretability, achieving high 09054-2.
accuracy on unseen data and providing valuable visual [6] T. Islam, ‘Plant Disease Detection using CNN Model
explanations for model decisions. Through rigorous and Image Processing’, Int. J. Eng. Res., vol. 9, no.
evaluation and testing, the system has demonstrated its ability 10, Oct. 2020.
to address a real-world agricultural challenge, contributing to
both research and practical use.

IJISRT24NOV1036 www.ijisrt.com 1558

Volume 9, Issue 11, November– 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24NOV1036

[7] M. Fraiwan, E. Faouri, and N. Khasawneh,

‘Classification of Corn Diseases from Leaf Images
Using Deep Transfer Learning’, Plants, vol. 11, no.
20, p. 2668, Oct. 2022, doi: 10.3390/plants11202668.
[8] M. H. K. Mehedi et al., Plant Leaf Disease Detection
using Transfer Learning and Explainable AI. 2022.
[9] R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam,
D. Parikh, and D. Batra, ‘Grad-CAM: Visual
Explanations from Deep Networks via Gradient-Based
Localization’, in 2017 IEEE International Conference
on Computer Vision (ICCV), Oct. 2017, pp. 618–626.
doi: 10.1109/ICCV.2017.74.
[10] S. P. Mohanty, D. P. Hughes, and M. Salathé, ‘Using
Deep Learning for Image-Based Plant Disease
Detection’, Front. Plant Sci., vol. 7, Sep. 2016, doi:
10.3389/fpls.2016.01419.

IJISRT24NOV1036 www.ijisrt.com 1559