Convolutional Neural Network Based Crop Monitoring

Sasikala C.

, Sainath Reddy R.

, Sree Ram Vijaya Vikram R.

Pavan Kumar Ram Prasad D.

and Satish G.

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

Anantapur District‑515701, Andhra Pradesh, India

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

Anantapur District, 515701, Andhra Pradesh, India

Keywords: Machine Learning, Crop Monitoring, Ripeness Classification, Disease Classification, Histogram of Oriented

Gradients, Convolutional Neural Networks, Sustainable Agriculture.

Abstract: To ensure sustainable crop production, farmers need to focus on efficient farming practices such as crop

health, soil health, pest control and yield analysis. This process relies on reliable monitoring of disease and

ripeness classification. This paper provides a machine learning system that classifies the crop images based

on ripeness and detect diseases. Here for Feature extraction, we use Histogram of oriented gradients, for

ripeness classification we use logistic regression and for disease classification we use Convolutional Neural

Networks. To implement this system, we are using flask-based web interface where it ensures seamless

deployment and we have visual tools like bar chart, pie chart to improve readability. More-over this system

provides insights to nutrient management to optimize yields and reduce crop losses. A voice enabled feature

enables that farmers can retrieve the information about yield analysis, nutrient management, remedial

measures, and disease classification. This system improves efficient crop monitoring where it can minimize

the errors from manual inspection, to maintain sustainable agricultural productivity and it supports decision

making based on data to enhance crop health and yield impact.

1 INTRODUCTION

Agricultural productivity is the core element of food

security and economic stability. To maintain the

proper balance between yield and sustainability, we

need to be aware of crop diseases and quality

monitoring inefficiencies. Ensuring timely

identification of the diseases and proper assessment

of ripeness is important to meet the demands of global

food production. To implement this process, we have

technological innovations, especially in machine

learning and deep learning models offer solutions to

these problems, which enable farmers to make

informed decisions and maintain good crop

production.

This proposed work involves machine learning

models to classify crop diseases and ripeness,

addressing crucial aspects regarding agricultural

monitoring. Features are extracted using Histogram

of Oriented Gradients (HOG); we classify crop

ripeness using logistic regression, and we classify

crop diseases using convolutional neural networks.

The provided system performs very well against all

metrics, where it achieves a high accuracy

demonstrated by an F1 score of 92.3%. It ensures

reliability and effectiveness in real-world

applications.

A seamlessly integrated web-based solution

developed by using Stream-lit and Flask. It ensures

that accessibility and ease of use are available to

users. The system provides various visualizations to

enhance decision-making, including disease yield

impact analysis, disease distribution pie charts,

disease progression timelines, and nutrient

requirement analysis. These insights help farmers and

researchers to improve decision-making ability so

that in the future they will get better results, and also

they will evaluate classification accuracy, disease

trends, and crop health in real time.

Additionally, disease classification, the solution

offers a recommendation system where it includes

nutrient management recommendations to optimize

yield and minimize crop issues. It offers a voice-

enabled feature that helps farmers to analyze disease

856

C., S., R., S. R., R., S. R. V. V., D., P. K. R. P. and G., S.

Convolutional Neural Network Based Crop Monitoring.

DOI: 10.5220/0013956600004919

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 5, pages

856-861

ISBN: 978-989-758-777-1

effects and predicted yields by delivering analyses

and suggesting corrective actions. This solution

automates the process of monitoring crops, which

reduces manual errors, enhances classification

accuracy, and tackles agricultural issues related to

diseases. This leads to greater productivity and

improved crop quality and plays a part in global food

security.

2 RELATED WORKS

In precision agriculture, classifying plant diseases

with the help of convolutional neural networks is

essential. Sharma, R., & Jain, A. (2020) To identify

crop diseases from visual attributes such as color,

shape, and texture, we employ image processing

methods like convolutional neural networks. In the

past, many existing machine learning models, like

support vector machine (SVM), k-nearest neighbor

(K-NN), and random forests, relied on manual

inspection for feature extraction, with time being a

critical factor as well. Shah, M. et al, (2019).

Convolutional neural network (CNN) is considered as

the best technique in this field, and it automatically

extracts spatial features and improves highly accurate

disease classification. Earlier, Mohanty et al (2016)

achieved over 99% accuracy in identifying 38 crop

diseases using CNNs. Barbedo, J. G. A. (2019) We

have open datasets like Plant Village; it includes all

types of diseases where we identify common tomato

diseases.

Simonyan, K., & Zisserman, A. (2014). To

improve harvesting techniques and maintain quality

control, ripeness classification is essential. Existing

manual methods lead to errors, are time-consuming,

have limited accuracy, and depend on human

judgment, while new technology innovations like

machine learning models and computer vision

techniques are used to classify ripeness based on

features like shape, color, and texture. Qin, Z.

(2016).To identify ripeness, colour is a key indicator,

with colour space transformations (eg. RGB to HSV)

and histogram analysis to access ripeness stages.

Hinton, G.E., et al, (2012). New advancements

involve CNNs to train on labelled datasets to

achieving high accuracy in classifying ripe, unripe,

old and damaged crops. Ivanovici, M. et al, (2024)

Lighting variations are addressed through techniques

and solved using data processing techniques.

Recommendation systems provide actionable

insights to farmers to mitigate diseases and to analyze

yield impact. These systems suggest soil nutrients,

yield impact, nutrient requirements, remedial

measures, and a voice-enabled feature where it can

help farmers to interact and know more about the

harvesting problems. Earlier, research by Singh et al.

(2018) implements a hybrid recommendation system

that combines rule-based systems and filtering to

recommend best practices to farmers regarding plant

diseases. Moreover, nutrient management has been

highlighted as a key indicator in improving harvesting

and recovery.

Bochtis, D. et al, (2018). Image-based

classification and recommendation systems have

faced so many challenges, such as environmental

factors (lighting, background clutter), image quality

like low-resolution images, blurry images, and not

being suitable for large datasets [10]. In the future, the

advanced technologies used for agricultural disease

management involve integrating IoT sensors, cloud

computing, and AI platforms for data collection and

analytics. This method improves model performance

and gets better results for disease classification and

recommendations while improving scalability.

Moreover, the incorporation of multilingual voice

outputs enhances the accessibility of this work for

farmers aiming at sustainable crop production.

3 DESIGNED SYSTEM

The designed system is used to show the detailed

description on how the images are analyzed and it

also focuses on two main objectives: ripeness

detection and disease classification. Figure 1

illustrates how the system works on the images

dataset and the images are divided into two types:

Ripe Images and Disease images, we need to select

which types of images we are choosing and when the

ripeness classification option is chosen the system

performs ripeness classification over the image and

determines how ripen is the fruit or vegetable and if

the disease detection option is selected the system

performs disease classification and identifies what

kind of disease it is and displays the severity of the

disease over a graph scale.

After the disease classification and ripeness

classification, the next step is visualizations where the

learned data is represented in an interpretable format

so that farmers can easily understand it by observing

it. Users can then have three options to make action

selections, such as prediction, recommendation, or

both. Prediction can give brief description about the

disease name and yield analysis, while

recommendation provides actionable information,

including nutrient requirements and management

suggestions and while selecting both, it provides both

Convolutional Neural Network Based Crop Monitoring

857

prediction and recommendation analysis in a

comprehensive way offering in depth evaluation. The

final step involves a voice response feature,

delivering the results in a user-friendly manner and an

understandable way, providing actionable insights to

the user.

Figure 1: Proposed system for crop monitoring.

Figure 1 - The proposed system involves

technologies such as machine learning and image

processing to improve agricultural productivity and

decision-making. It effectively leverages advanced

methods to get actionable insights, supporting

farmers to have efficient crop yield and sustainable

farming. This will reduce manual errors and increase

crop production, which results in better productivity,

and this will benefit farmers to get better yield

production.

4 IMPLEMENTATION

4.1 Preprocessing Techniques

Picture Scaling and Normalization: Picture scaling is

used to resize the image depending on their size as the

data can be in different sizes typically, they are of

224x224 pixels, to ensure the linearity in the dataset,

normalization is used in resizing the images pixel

values to the range of (0,1) this improves model’s

intersection.

Data Augmentation: Data Augmentation is used

to improve the diversity of training data and also

reduces overfitting, it also involves some techniques

such as rotation, flipping, zooming and adjustments

on brightness are applied. This step not only increases

the size of dataset but also allows the model to

normalize in a much simpler way by learning from

different examples.

Label Encoding: Each and every image in the

dataset is labelled and also helps to indicate or

identify the type of class such as healthy, blight or

ripe. These labels are again converted into numerical

values in order to serve as inputs during the model’s

training.

Noise Removal and Image Filtering: Lack of

better resolution in image and irrelevance in the

content or excessive noise are eliminated in order to

maintain the dataset’s integrity. This step promises

that the model is properly trained with data which is

more precise and accurate.

4.2 Methodologies

An experimental design is chosen in this research.

The reason behind this is to improve and validate a

hybrid system that detects multiple diseases in a

tomato plant and also performs ripeness

classification. The ultimate goal was to integrate

feature extraction using Histogram of Oriented

Gradients (HOG) and to perform image classification

using Convolutional Neural Networks (CNN) in

order to achieve the prediction that are more precise

and trustworthy. This research methodology also

involves preparation and preprocessing of datasets,

feature extraction, model training, deployment and

evaluation. A web application that was interactive

towards the users was developed to make the system

more user friendly.

4.2.1 Feature Extraction Using HOG

The Histogram of Oriented Gradients (HOG) is used

to extract features that are more robust in nature while

performing the tasks. Figure 2 illustrates the working

of feature extraction using HOG.

The primary step for feature extraction begins

with an original image that contains raw data, after

that aligning the image according to the standards.

The image resized has a fixed dimension of 128x128

pixels this process can be applied to multiple images

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

858

by applying the consistent feature. Later the input

image is divided into non-overlaying cells with a

pixel size of 16x16, this step is important as it

captures all the local features, structure and texture of

the image. The fifth step Gradient computation is an

important step in HOG it measures the pixel intensity

of an image its main purpose is to detect sharp edges,

shapes, texture of an image. After the Gradient

computation step the fifth step is Feature Vector

Generation which is used to divide the image into

small cell of 8x8 size where each cell has HOG in

different direction, the final result is a feature vector

representing edges, contours, patterns, and texture of

an image.

Figure 2: Feature extraction.

4.2.2 CNN Architecture

CNN plays an important role to process and classify

the image based on the disease. Figure 3 explains the

architecture of convolutional neural network, the

input for the model leaf image dimensions is

128x128x3 and passes through several layers, Input

layer is the first layer where images are represented in

RGB and all the images are resized into 128x128x3

pixels for uniformity, after this convolution layer 1

model is used, this uses 32 filters with a dimension of

3x3 along with RELU function. It takes of extra

features like edges, shapes, patterns, etc.., the next

step is Max Pooling Layer 1 where the pooling size is

2x2 filter is used to reduce the dimensions for a

feature map to 64x64x32 while retaining essential

features this layer also reduces the noisy input data.

Now the Convolution Layer 2 is used where the

filters are increased from 32 to 64 to detect more

complexes, kernel size of 3x3 is used along with the

RELU function, this layer is used to extract specific

disease patterns. The fifth step is Max Pooling Layer

2 this layer performs with pooling size of 2x2 to

reduce further dimensions for feature map 32x32x64

and retains all essential features, now the Convolution

Layer 3 is used where the filters are increased from

64 to 128 to identify more complex patterns, it also

uses 3x3 kernel along with RELU function this layer

is used to detect irregular textures and patterns in the

image. Flatten layer is the next step where multi-

dimensional maps convert into 1D array, this process

is done for the next layer’s input process which is

dense layer. Dense layer performs to learn high level

relationships between extracted features. It requires

512 neurons along with the RELU activation

function. Dropout regularization is also performed to

prevent overfitting. The final step Output Layer

consisting of 10 neurons, corresponding to disease

classes where a SoftMax activation function is used

to assign probabilities for each disease class and it

enables multi-class classification.

Figure 3: Convolutional neural network.

Convolutional Neural Network Based Crop Monitoring

859

5 RESULTS AND ANALYSIS

This section is very crucial for any model. Results and

Analysis section is to analyze model performance. It

involves analyzing the experiments outcomes and

comparing the proposed solution with existing model.

So that we can get an idea about the proposed model

performance.

5.1 Evaluation Metrics

Accuracy: Accuracy is a crucial evaluation metric to

assess model performance. This model achieves an

accuracy score of 93% indicates the correctness in

classifying plant diseases and ripeness.

Precision: Precision is an essential evaluation

metric to assess model performance. This model

achieves 92% indicates that a plant has a disease or a

certain ripeness level. It means this model avoids

making incorrect positive predictions.

Recall: Recall is an evaluation metric to assess

model performance. This model achieves a recall

score of 90% which indicates the model is good at

identifying the correctly disease plant.

F1-Score: F1-score is an evaluation metric to

assess model performance. This model achieves a F1-

score of 92.3% which indicates that this model is well

balanced between precision and accuracy. It also

ensures that the model is both accurate and consistent.

Specificity: For this evaluation metric, it achieves

94% score which explains the model's performance in

identifying healthy plants.

5.2 Graphical Representation

5.2.1 Training and Validation Accuracy

Analysis

Figure 4 shows the Training accuracy and Validation

accuracy changes of the model performance over four

epochs during the training process. In this figure, X-

axis represents the number of epochs and Y-axis

represents the accuracy values. As this figure

describes how the model is classifying plant diseases

over training and validation processes.

Blue line describes the training accuracy which

increases gradually with each epoch highlighting the

model's performance to adapt and improve on the

training data. Orange line represents the validation

accuracy which steadily increases, highlights how the

model performs against unseen data.

As this progress indicates that there is a parallel

improvement in both training and validation accuracy

which explains the model balanced learning process.

Gap between two accuracies is relatively small which

shows there is no chance for overfitting and

underfitting. This reflects that the model adapts well

to the validation dataset.

Figure 4: Training accuracy vs validation accuracy.

5.2.2 Training and Validation Loss Analysis

Figure 5 shows the Training loss and Validation loss

changes of the model performance over four epochs

during the training process. In this figure, the X-axis

represents the number of epochs and the Y-axis

represents the loss of values. As this figure assesses if

your model is actually learning patterns or just

identifying patterns and also prevents from

overfitting.

Blue line explains the training loss which

decreases gradually with each epoch illustrating the

model is learning from the training data. Orange line

represents the validation loss which decreases rapidly

over epochs represents improved performance on

unseen validation data.

As both training loss and validation loss decrease

over epochs suggests that the model is learning

effectively and prevents overfitting.

Figure 5: Training loss vs validation loss.

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

860

5.2.3 Disease Impact Severity Analysis

Figure 6 indicating heatmap showing the impact

severity of various diseases on tomato yield. The X-

axis represents the yield impact severity as percentage

and Y-axis represents lists of diseases.

Color scale is used to represents how the disease is

impacted. The color ranges from red (low impact) to

green (high impact), along with intermediate shades

indicating severity levels. The healthy tomato is

marked with the highest yield impact (100%), the

diseases show the least yield impact as 40% and 45%

respectively.

This visualization helps in identifying which

diseases has most significant yield impact, providing

a brief overview of their impact severity.

Figure 6: Disease impact severity heatmap.

6 CONCLUSION AND

PROSPECTS

This research provides an in-depth explanation for

disease classification and ripeness classification.

Additionally, it includes visualizations to understand

in detail and recommendations to improve crop yield

and to reduce manual error. A voice-enabled feature

is featured to provide detailed explanations about

disease classification, ripeness classification, and

recommendations to farmers.

For disease classification, we use a convolutional

neural network to analyze images, and to extract

features for ripeness classification, we use a

histogram of oriented gradients.

This system provides an interface that integrates

visual insights through graphical representation.

These visualizations help in identifying diseases and

crop yield and maximizing nutrients to improve yield.

This results in reducing manual errors, improving

decision making, promoting sustainable farming

practices and enhancing productivity and crop health.

The proposed system has solved so many issues

that were addressed previously, but apart from the

existing solution, there are several areas where the

proposed work needs to be expanded so that it

maximizes productivity and promotes sustainable

farming.

The key areas for expansion include Extension to

other crops, Real-Time Monitoring and IoT

Integration, Enhanced Dataset Diversity, Multi-

Language Voice Assistance and Integration with

Drone-Based Crop Monitoring.

REFERENCES

Barbedo, J. G. A. (2019). Plant disease identification from

individual lesions and spots using deep learning.

Biosystems Engineering, 180, 96-107.

El Sakka, M., Mothe, J., & Ivanovici, M. (2024). Images

and CNN applications in smart agriculture. European

Journal of Remote Sensing, 57, 2352386.

Hossain, M.A., & Sajib, M.S.A. (2019).Classification of

image using convolutional neural network (CNN).

Global Journal of Computer Science and Technology,

19, 13–14.

Jha, K., Doshi, A., Patel, P., & Shah, M. (2019). A

comprehensive review on automation in agriculture

using artificial intelligence. Artificial Intelligence in

Agriculture, 2, 1-12.

Krizhevsky, A., Sutskever, I., & Hinton, G.E. (2012).

Imagenet classification with deep convolutional neural

networks. Advances in Neural Information Processing

Systems, 25, 1097-1105.

Liakos, K.G., Busato, P., Moshou, D., Pearson, S., &

Bochtis, D. (2018). Machine learning in agriculture: A

review. Sensors, 18, 2674.

Mohanty, S. P., Hughes, D. P., & Salathé, M. (2016). Using

deep learning for image-based plant disease detection.

Frontiers in Plant Science, 7, 1419.

Qin, Z. (2016). Precision Agriculture Technology for Crop

Farming. CRC Press.

Sharma, R., & Jain, A. (2020). Fruit Ripeness Detection

Using Image Processing and Machine Learning

Techniques: A Review. International Journal of

Applied Sciences and Technology, 2(5), 1–10.

Simonyan, K., & Zisserman, A. (2014). Very Deep

Convolutional Networks for Large-Scale Image

Recognition. arXiv preprint arXiv:1409.1556.

Convolutional Neural Network Based Crop Monitoring

861