Smart Agri Assist: Enhancing Leaf Disease Recognition Using Deep

Learning Techniques

Sirisha Pagadala

, Harshitha Reddy Peddakotla

, Kusuma Mara

, Greeshma Teja Gaddam

and Mythri Thammineni

Department of CSE, Srinivasa Ramanujan Institute of Technology, Rotarypuram Village, B K Samudram Mandal,

Anantapuramu - 515701, Andhra Pradesh, India

Department of CSE (AI & ML), Srinivasa Ramanujan Institute of Technology, Rotarypuram Village, B K Samudram

Mandal, Anantapuramu - 515701, Andhra Pradesh, India

Keywords: Plant Disease Diagnosis, Deep Learning, MobileNet, Image Classification, Treatment Recommendations.

Abstract: Implementing disease prediction systems for potato leaves would improve agricultural productivity and crop

yield through early identification. This project is focused on detecting blight diseases using Artificial

Intelligence (AI) and Deep Learning techniques. The proposed system can analyse leaf images, as shown in

Figure 1, using a MobileNet-based architecture to extract the critical features making sure enough attention

is paid towards colour, texture as well as shape in order to do leaf classification. Along with a dense layer,

this also requires a SoftMax layer to ensure that the model provides an output with a confidence score

corresponding to each diagnosis in order to establish the reliability of the output. It further offers personalized

therapeutic recommendations, including advice on organic and chemical pathologies and mechanical damage

management, aiding farmers in tracking up on the specified pathology and damage. This indicates how deep

learning is substantially changing agriculture and also holds the potential to help farmers make better decisions

to minimize crop loss and increase sustainability in agricultural practices.

1 INTRODUCTION

Potatoes are an important staple crop grown

worldwide crucial to food security, but potato

production is frequently threatened by diseases such

as Early Blight and Late Blight, leading to significant

yield losses. Early and accurate disease detection is

vital for the possible treatment of the crops and

reducing economic losses. Traditional visual checks

require trained scientists, are labor intensive and

reliant on subjective training and experience, thus not

widely adopted by farmers. Get involved in changing

computational intelligence and neural network

methods in the fight against the disease and the

automation of disease recognition.

In this project we will build an AI model that, by

looking at pictures of the leaves, we can identify

potato diseases. It is based on a MobileNet

architecture, an efficient one for visual analysis.

Basically, the system is trained on potato leaf dataset

containing images with different variety of potato

leaves, and it classifies the diseases into three classes

Figure 1: Potato Leaves.

Initial fungal infection (Early Blight) 2. Advance

fungal infections (Late Blight) and 3. the healthy

leaves (Healthy). Data augmentation was used to

Pagadala, S., Peddakotla, H. R., Mara, K., Gaddam, G. T. and Thammineni, M.

Smart Agri Assist: Enhancing Leaf Disease Recognition Using Deep Learning Techniques.

DOI: 10.5220/0013874000004919

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 1, pages

819-826

ISBN: 978-989-758-777-1

819

make the model more robust to variation in image

quality. The system not only provides classification

but also confidence scores to help communicate

prediction reliability, and precision treatment

recommendations, in terms of both organic and

chemical options for specific diseases. Crop

management is enhanced and accidents minimized

It.) The project brings AI into agriculture to improve

agricultural performance, support sustainable

solutions, and strengthen global food resilience.

2 RELATED WORKS

Liu, J., Cheng, Q., Gong, W., et al. (2022) A deep

learning-based proposed tomato and potato disease

recognition based on transfer learning which provides

a good classification accuracy. The model showed

robust performance across a range of lighting and

weather conditions, suggesting its potential for

agricultural use."

Arshaghi, A., Ashourian, M. & Ghabeli, L. (2023)

Deep learning methods (e.g., ResNet, VGG) potato

disease detection and classification were examined.

Both ResNet and Inception v4 produced the best

results compared to traditional machine learning

methods on a curated dataset of potato leaf images.

Kumar, A., Patel, V.K. (2023) In potato leaves

disease identification, developed a hierarchy deep

learning-based CNN. The proposed model not only

improved classification performance but also

addressed the scalability challenge of computational

complexity, making it suitable for real-world

deployment.

Sharma, A., Zhang, L., & Tanwar, S. (2021) A new

deep learning framework used for the early detection

of late blight in potato crops. The system, which

relied on hyperspectral imaging and CNN

architectures, predicted the onset of plant disease

before the manifestation of visible symptoms,

supporting preventive farming practices.

Yuan, D., Wu, C., & Li, J. (2020) A hybrid CNN

model was suggested in that combined traditional

image processing techniques with deep learning

methods for detecting potato leaf disease. In that case,

the work increased the efficiency of the feature

extraction, especially for late blight and early blight

diseases.

Kong, G., Wang, H., Wang, L., et al. (2022)

incorporated deep learning and edge computing for

accurate real-time potato late blight identification in

the field. This system effectively reduced latency and

the usage of bandwidth, and hence it became feasible

for IoT-based agriculture monitoring.

Shrestha, R., Gaire, A., & Moh, S. (2021) developed a

lightweight efficient CNN model for potato disease

recognition, which is suitable for low-power device

deployment. The model showed a balance between

accuracy and computational resource demands.

Parihar, N., Rani, A., Gupta, M., et al. (2020) Deep

CNNs were used in for potato disease classification

which, on public datasets, yield state-of-the-art

results. The approach highlighted the application of

data augmentation methods in scenarios where

training data is not abundantly available.

Zhang, L., Zhang, Y., & Zhu, Z. (2022) investigated

the detection of potato plant diseases with deep neural

networks with emphasis on its applicability at an

industrial scale. The model produced consistent

results in diverse environmental conditions, proving

its adaptability.

Yang, J., Xie, Y., & Wei, J. (2021) developed a new

CNN framework for detecting potato plant diseases

using attention-based modules to increase feature

localization. The method yields greater accuracy than

baselines.

Du, J., Ma, X., & Li, B. (2022) introduced attention

mechanism-based CNN for potato disease

classification that highlights the area where diseases

are present which gave better interpretability with

precision in disease localization. The effect of dataset

size on model performance has also been explored in

the study.

Zhang, Y., Chen, S., & Sun, Q. (2020) CNN-based

approach to identify potato late blight presented and

maintained external, real-world defense against it.

The result authenticated the accuracy of the model

for the early diagnosis of the disease which can save

the crop damage.

Sharma, S., Sharma, A., & Gupta, A. (2021) Recent

advances in deep learning for potato disease detection

were surveyed, which

revealed gaps in generalizability and real-time

processing. The review highlighted the requirement

of lightweight models designed for edge devices.

Wang, L., Liu, L., & Li, Y. (2020) developed an

efficient deep learning model for early potato disease

detection, emphasizing computational optimization

for farm-level use. The system achieved high

accuracy with minimal hardware requirements.

Ma, R., Hu, J., & Li, Y. (2022) proposed a CNN-

based late blight identification method, showcasing

its effectiveness in controlled and field environments.

The study highlighted the role of preprocessing in

improving model robustness.

Shen, L., Zhang, J., & Huang, X. (2021) validated a

deep learning approach for potato disease recognition

under varying lighting and occlusion conditions. The

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

820

model maintained consistent performance, proving its

practicality for real-world agriculture.

Li, M., Yang, Z., & Wang, Y. (2020) automated

potato disease detection using CNNs, focusing on

integration with precision agriculture systems. The

work outlined challenges in deploying AI models for

non-technical end-users.

3 PROPOSED WORK

3.1 Dataset

To develop a reliable website to detect and classify

potato leaf infections, we curated a diverse dataset

comprising images of healthy and those showing

signs of diseased. The dataset was sourced from

platforms like Kaggle as shown in Table 1, which

hosts a wide range of datasets for machine learning

applications. We utilized the Plant Village dataset,

which includes JPG color images of 256x256 pixels

representing healthy and diseased leaves of 14 plant

species. For this study, we selected a subset of 152

images of disease-free leaves combined with 2000

images showing symptoms of Blight-infected leaves.

3.1.1 Early Blight

A fungal infection caused by Alternaria solani. It

begins as small black spots that gradually enlarge into

larger dark brown or black patches. These spots are

usually round or oval and often appear along the

edges of leaf veins. In some cases, a black fungal

growth can also be seen on the undersides of the

leaves. This disease primarily affects potatoes,

causing them to rot, especially in warm weather

(above 26°C) or when the plants are stressed due to

poor nutrition or excessive heat. Figure 2 shows the

Early Blight Disease of Potato Leaves.

Figure 2: Early Blight Diseased Potato Leaves.

3.1.2 Late Blight

A highly damaging disease caused by the pathogen

Phytophthora infestans. It severely affects potato

plants, particularly in cool and moist environments.

This fast-spreading disease harms both leaves and

tubers, leading to substantial reductions in potato

yields. Figure 3 shows Late Blight Diseased Potato

Leaves.

Table 1: Summary of Plant Village Dataset.

Label Category Total Images

Training

Samples

Validation

Samples

Test

Samples

1 Late Blight 1000 800 100 100

2 Early Blight 1000 800 100 100

3 Healthy 152 122 15 15

Total

2152 1722 215 215

Smart Agri Assist: Enhancing Leaf Disease Recognition Using Deep Learning Techniques

821

Figure 3: Late Blight Diseased Potato Leaves.

3.2 Data Preprocessing

Hence, the subsequent preprocessing steps were

carried out to maintain the integrity of the dataset:

3.2.1 Image Resizing

All images were resized to a constant 224x224 to

ensure consistency.

3.2.2 Normalisation

This was performed by calibrating the pixel

intensities with the number of possible colours per

pixel, which in this case was 255, constraining the

pixel values in the set of [0,1]

3.2.3 Training Data Enrichment

To maximize model adaptivity, techniques such as

image rotation, resizing, and horizontal flipping were

utilized during the training of the model in order to

maximize the variability of the dataset and minimize

risk of overfitting.

3.3 Model Development

3.3.1 Model Selection

MobileNet was selected because of its efficient

design and great performance for image classifier

tasks.

3.3.2 Training

80:20 split, with larger partition used for training and

smaller partition used for testing. The model was

trained using backpropagation with validation data to

track progress and avoid overfitting.

3.3.3 Evaluation

The performance metrics used to determine the

effectiveness of the model were overall detection

accuracy, positive prediction rate, negative

prediction rate, and the unified score.

3.4 User Interface Development

3.4.1 Web Application Framework

The system was built using Django, a powerful

framework for developing web applications.

3.4.2 Image Upload Functionality

A user-friendly interface was designed, enabling

users to upload images of potato leaves for disease

classification.

3.4.5 Result Display

The application presents disease predictions along

with confidence scores and recommends suitable

treatments based on the classification.

4 METHODOLOGY

The MobileNet model applies a systematic approach

to classify input images into several classes based on

their features. Fastai is a powerful and flexible deep

learning library designed to facilitate high-

performance image classification. Here’s a detailed

breakdown of how it operates step by step in figure

4.1 Input Image Processing

The user uploads an image of a potato leaf, and the

process begins. So, to make it consistent, the input

image is resized into a fixed size (128×128 or

224×224 in some versions). This breaks down the

images to reduce the variations in their sizes. Takes

pixel normalization next. This step helps increase the

efficiency of the training, as pixel intensities are now

more homogeneous in range, resulting in lower

complexity of computation, as well as increased

stability of the learning.

4.2 Feature Extraction using

Convolutional Layers

MobileNet uses Depthwise Separable Convolutions

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

822

instead of regular convolutional layers, which greatly

decrease the computational cost without

compromising accuracy. This is achieved through

two fundamental operations:

Figure 4: Overview of MobileNet Architure.

4.2.1 Depth wise Convolution

Convolutional layer: A convolutional layer performs

a similar operation for a filter, except all input

channels are transformed with the same operation.

However, in a depth wise convolution, a single filter

is applied over each channel (Red, Green, Blue). This

means that the model learns (example: right and left

edges for pixels of the red color) independently per

color (channel), a computation is now much easier. It

eliminates the complexities of processing the image

entirely in RGB in exchange for maintaining core

features like edge, texture, and pattern details while

reducing computational overhead by processing the

RGB in parallel with three-edge channels.

4.2.2 Pointwise Convolution

Figure 5: Depthwise Separable Convolution Layers.

This is followed up by a pointwise convolution using

1×1 filter. Small filters learn inter-channel

relationships from extracted channel-wise features.

And this process is critical for establishing

informative feature representations. Figure 5

Depthwise convolution is about spatial feature while

pointwise convolution is the operation to merge these

facts to an effective yet informative features.

4.3 Activation Function (ReLU)

To introduce non-linearity and improve the feature

learning the Rectified Linear Unit (ReLU) function

is applied after each convolution layer. It zeros any

negative pixel values and helps retain only the metrics

we need to learn. This activation function was

helpful as it allowed the model to accurately process

details of potato leaf images, contributing to the

improvement of its results.

4.4 Feature Pooling (Max Pooling)

Max Pooling is used to reduce the final feature map

dimensionality, preserving only the most important

information. Due to this process the model pipeline

becomes efficient by down sampling the image

representation but keeping main features. Max

pooling enhances feature conservation and avoids

overfitting by maintaining only the highest value

across each section of the feature map.

4.5 Fully Connected (Dense) Layers

After the extraction of high-level features through

convolution and pooling, classification is the next

step. Features are first flattened to create a 1D vector

and passed through fully connected (dense) layers.

Layers are learned and adjusted as a classification

process.

4.6 Softmax-Driven Classification

Approach

The Softmax function is applied to generate

Input Image

Image

Preprocessing

Depthwise

Convolution

Pointwise

Convolution

ReLU

Activation

Max Pooling

Fully

Connected

Layer

Softmax

Activation

Output

Smart Agri Assist: Enhancing Leaf Disease Recognition Using Deep Learning Techniques

823

probability scores for each class and the class with

the greatest probability is selected as the final

prediction.

4.7 Output & User Interface

After the classification process is finished, the system

then shows the expected disease category as well as

the confidence level. The application further

recommends suitable treatments according to the

classification outcome. The predictions are displayed

on a web interface running on Django, giving users

an easy-to-apply platform for disease identification.

5 RESULTS AND EVALUATION

5.1 Training and Validation accuracy

Figure 6. graph show us the learning process of our

deep learning system, for potato disease

classification, for each epoch The trend at early stage

is very steep, which means good feature extraction

and gradients optimization. Likewise, the accuracy of

the model on the validation set increases with time,

indicating that model generalizes well on new data.

After 10th epoch, the accuracy levels off; therefore,

we can tell that the model have learned the necessary

features for classification and are minimizing the

errors. The validation curve remains stable with only

minor fluctuations, showing that the data is not too

complex and not getting overfit, as the validation

accuracy remains high and close to the accuracy of

the training dataset. When the upper right threshold is

plotted against the lower right threshold, variable, the

close distance between the two curvatures indicates

that the model is well- regularized and the

hyperparameters have been optimally calibrated.

Figure 6: Accuracy Trends During Training and Validation.

5.2 Training and Validation Loss

Figure 7 All epochs' performance of our potato deep

learning classifier on the left, we see a steep decline

in both loss metrics, indicating our model is learning

and moving toward a minimum by successfully

updating its parameters during the backpropagation

and gradient descent updates.

Figure 7: Loss Trends Across Epochs.

The training loss has reached a steady low value

after around 10 epochs, indicating that convergence

has been reached. The validation loss seems to

fluctuate a little, I think because of the nature of the

validation dataset. The losses converge tightly,

indicating the model's robustness and reliability. The

characteristics of these loss patterns suggest that the

model is successfully implementing a well-

regularized learning process for agricultural disease

classification.

5.3 Confusion Matrix

Confusion matrix which shows performance of the

deep learning framework in classifying potato

diseases by using predicted labels and actual labels.

It measures the classification accuracy and helps find

entries misclassified across categories. The high true

positive is confirmed from 110 healthy, 125 late

blight, and 17 early blight samples predicated

correctly from the model. We see only a couple of

misclassifications with very little confusion. The fact

that there are no false positives in the Healthy

category shows that the model is doing good to tune

in to non-infected leaves. The proposed model shows

promising results in Figure 8, it can be relied upon for

automated potato disease detection in the field of

agriculture.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

824

Figure 8: Prediction Outcome Grid.

5.4 User Interface

Figure 9: Image Upload Page.

Figure 9 shows the user uploading image page and Figure

10 shows the disease prediction and treament

recommendation.

Figure 10: Disease Prediction and Treatment

Recommendation.

6 DISCUSSION

The proposed potato disease detection system, based

on the MobileNet architecture, demonstrated high

accuracy (over 99%) in classifying healthy and

diseased potato leaves. This performance can be

attributed to the use of a well-curated dataset and data

augmentation techniques, which improved the

model's ability to generalize. The lightweight nature

of MobileNet proved beneficial for real-time

applications, as it reduced computational overhead

while maintaining high precision, making it more

efficient than heavier models like VGG16 or ResNet.

The web-based implementation using Django

provided an accessible and user-friendly interface for

farmers, allowing them to upload images and receive

instant disease predictions along with confidence

scores. The inclusion of treatment recommendations

(both organic and chemical) further enhanced the

system's practicality, bridging the gap between

disease diagnosis and actionable solutions. However,

the model's effectiveness depends on the diversity and

quality of the training data. Future iterations could

benefit from incorporating more disease variations,

environmental conditions, and regional-specific

datasets to improve robustness.

Additionally, while the current system focuses on

image-based detection, integrating real-time

monitoring through IoT devices or mobile

applications could enhance its usability in field

conditions. Edge computing optimizations could also

be explored to enable offline functionality,

particularly in regions with limited internet

connectivity.

7 CONCLUSIONS

This study successfully developed an efficient deep

learning-based system for detecting potato diseases

using the MobileNet algorithm. The model achieved

high accuracy, demonstrating its potential for real-

world agricultural applications. The lightweight

architecture ensured fast processing, making it

suitable for deployment in resource-constrained

environments. The accompanying web application

provided an intuitive platform for farmers to diagnose

diseases and access treatment recommendations,

thereby supporting better crop management

decisions.

Future work should focus on expanding the

dataset to include more disease types and

environmental variations, improving model

Smart Agri Assist: Enhancing Leaf Disease Recognition Using Deep Learning Techniques

825

generalizability. Further optimizations, such as edge

deployment and real-time monitoring capabilities,

could enhance the system's scalability. By addressing

these aspects, the proposed solution can evolve into a

more comprehensive tool for precision agriculture,

ultimately contributing to reduced crop losses and

improved farm productivity.

REFERENCES

Liu, J., Cheng, Q., Gong, W., et al. (2022). Deep learning-

based tomato and potato diseases recognition. Journal

of Food Engineering, 357, 109771.

Arshaghi, A., Ashourian, M. & Ghabeli, L. (2023). Potato

diseases detection and classification using deep

learning methods. Multimed Tools Appl, 82, 5725–

5742.

Kumar, A., Patel, V.K. (2023). Classification and

identification of disease in potato leaf using hierarchical

based deep learning convolutional neural network.

Multimed Tools Appl, 82, 31101–31127.

Sharma, A., Zhang, L., & Tanwar, S. (2021). A novel deep

learning approach for early detection of late blight

disease in potato crops. Computers and Electronics in

Agriculture, 182, 106075.

Yuan, D., Wu, C., & Li, J. (2020). Potato leaf disease

detection using a hybrid convolutional neural network.

IEEE Access, 8, 72671-72677.

Kong, G., Wang, H., Wang, L., et al. (2022). Identification

of potato late blight disease based on deep learning and

edge computing. IEEE Internet of Things Journal, 9(6),

5046-5054.

Shrestha, R., Gaire, A., & Moh, S. (2021). An efficient

potato disease classification model using convolutional

neural networks. In Proceedings of the IEEE

International Conference on Computer Vision (ICCV)

Workshops, 2021.

Parihar, N., Rani, A., Gupta, M., et al. (2020). Deep

convolutional neural network for the classification of

potato diseases. In Proceedings of the International

Conference on Computer Vision and Pattern

Recognition (CVPR), 2020.

Zhang, L., Zhang, Y., & Zhu, Z. (2022). Potato disease

detection using deep neural networks. Computers in

Industry, 134, 103590.

Yang, J., Xie, Y., & Wei, J. (2021). Potato leaf disease

detection using a novel CNN architecture. Computers

and Electronics in Agriculture, 184, 106095.

Du, J., Ma, X., & Li, B. (2022). Potato leaf disease

classification using a deep convolutional neural

network with attention mechanism. IEEE Transactions

on Image Processing, 31, 173-183.

Zhang, Y., Chen, S., & Sun, Q. (2020). Detection of potato

late blight disease using a convolutional neural

network-based method. Plant Disease, 104(10), 2671-

2678.

Sharma, S., Sharma, A., & Gupta, A. (2021). A survey of

recent advances in potato disease detection using deep

learning techniques. Journal of Agricultural Science

and Technology, 23(4), 815-834.

Wang, L., Liu, L., & Li, Y. (2020). An efficient deep

learning model for early detection of potato diseases

using CNN. Journal of Agricultural Information

Science, 42(2), 59-69.

Ma, R., Hu, J., & Li, Y. (2022). A novel method for the

identification of potato late blight disease based on

convolutional neural networks. Computers in Industry,

152, 103488.

Shen, L., Zhang, J., & Huang, X. (2021). Potato disease

recognition using a deep learning-based approach.

Plant Pathology, 70(4), 943-950.

Li, M., Yang, Z., & Wang, Y. (2020). Automated detection

of potato diseases using convolutional neural networks.

International Journal of Agricultural and Biological

Engineering, 13(6), 103-11.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

826