Agro AI: Intelligent Nutrient and Disease Management System for

Sustainable Farming

P. Veera Prakash, Gowthami M, Lokesh B, Anila R

and Harsha Vardhan P

Department of CSE, Srinivasa Ramanujan Institute of Technology, Rotarypuram Village, BKS Mandal,

Anantapur, Andhra Pradesh, India

Keywords: NPK Prediction, Disease Detection, Deep Learning, ResNet-34, BiGRU, CNN, GAN, Image Processing,

Time Series Data, Precision Agriculture.

Abstract: Agriculture plays a key role in food security, and managing nutrients and detecting diseases are crucial for

food production. Farmers traditionally struggle to identify nutrient deficiencies and diseases accurately

despite the possibility of using technology, due to the cost of a solution that can provide accurate and real-

time data. As an answer to this gap AGRO-AI proposes an intelligent system that predicts NPK (Nitrogen,

Phosphorus, Potassium) values in a specific field and disease detection on paddy leaves with the help of

image processing and deep learning. The dataset consisted of repeated sampling (i.e. three samples per week)

over 20 weeks to allow for analysis of variations in nutrient data and disease symptoms. A hybrid model is

used in the system for NPK prediction that consists of ResNet-34 and BiGRU whereas CNN and GAN

models are used for disease prediction. The results are analyzed in more detail using confusion matrices,

accuracy trends, and precision-recall curves. According to the predictions, AGRO-AI provides practical

recommendations for natural and chemical fertilizers and also suggests disease management policies. Trained

on data through October 2023, this research ultimately serves to connect technology with an application that

can empower farmers with needed data in real-time for optimized crop management to lead to increased yield

productivity.

1 INTRODUCTION

No data is sufficient to substitute for agriculture,

especially in the areas where rice is the major crop.

Paddy crop needs an optimum dose of macro and

micro nutrients such as N, P and K for better growth.

Also, disease detection and management with control

measures are fundamental to reduce yield losses. The

traditional approach involving visual inspection,

chemical analysis, etc. is tedious, prone to human

errors and leads to delayed corrective action.

AI and other image processing methods enable

solutions to provide much faster and reliable insights

into crop health. By applying AI on paddy leaf

images through paddy leaf nutrient deficiency

detector, farmers can receive substantial information

on prediction of nutrient deficiency or diseases that

require timely action. CNN and GAN have proven to

be the most efficient methods for disease detection

and classification as evidenced by studies conducted.

Likewise, hybrid models have reported

improvements in identifying nutrient deficiencies by

leveraging the use of CNNs with transfer learning.

Unfortunately, many of the existing solutions use pre-

collected datasets rather than real time data, resulting

in limited ability to adapt to constantly changing

conditions on the field. Moreover, models prioritize

only nutrient deficiency prediction or disease

detection, which leads to a partial solution. To

overcome these issues, in this work we present

AGRO-AI, a twofold architecture able to accomplish

both: NPK prediction, using ResNet-34 and BiGRU

models, and disease detection, using a CNN-GAN

architecture. AGRO-AI differs from previous

methods since it is based on a real time, time series

dating over 20 weeks, with three samples per week,

in order to make precise and dynamic predictions.

The primary objectives of this research are:

• Developing a deep learning-based system for

accurate NPK prediction using paddy leaf

images.

Prakash, P. V., Gowthami, M., Lokesh, B., Anila, R. and P., H. V.

Agro AI: Intelligent Nutrient and Disease Management System for Sustainable Farming.

DOI: 10.5220/0013932500004919

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

531-537

ISBN: 978-989-758-777-1

531

• Implementing a robust disease detection

model for early identification and

classification of paddy diseases.

• Providing actionable recommendations for

nutrient management and disease treatment

using both natural and chemical fertilizers.

• Delivering a scalable, real-time agricultural

decision support system for farmers.

By introducing a comprehensive AI-driven

approach, AGRO-AI addresses the limitations of

traditional agricultural diagnostics. It empowers

farmers with timely, data-driven insights, facilitating

proactive decision-making to enhance crop

productivity and sustainability. This research

contributes to the advancement of precision

agriculture and supports global efforts toward

ensuring food security.

2 RELATED WORK

Artificial intelligence and machine learning

techniques have gained prominence in precision

agriculture because of their utility in predicting NPK

and disease detection. Before feeding the model with

the data set, you need data sets that can help in

recognizing the nutrient deficiency in plants and

diseases in crops, Various models like CNNs, GAN,

and hybrid model have shown succor in nutrient

deficiency and crop disease identification. Transfer

Learning: Transfer learning has also been a

significant area of research, with researchers looking

at methods to leverage pre-trained models to enhance

accuracy and reduce training time.

Though these methods show great promise, a key

limitation is that they rely on fixed datasets. Most

models are based on pre-collected images which do

not have variations in real-time, limiting their

application on dynamic agricultural environment.

Moreover, the existing literature tends to develop

models that solely address either nutrient prediction

or disease detection, overlooking the potential for a

combined model that could simultaneously deliver a

holistic perspective.

AGRO-AI overcomes such limitations by

utilizing ResNet-34 and BiGRU when predicting the

required NPK through real-time data and time series

analysis. The final design is an effective hybrid of

CNN and GAN, used for disease detection. Making

sure that the system will sense Nutrients Deficiency

& Diseases, providing farmers with actionable

insights for timely intervention. The time-series data

allows us to take advantage of multiple samples taken

over multiple weeks (3 each week and this over 20

weeks) to increase the accuracy and reliability of

predictions.

Table 1: Comparison of soil and crop prediction models, highlighting algorithms used and key features (Source: author).

No. Paper Title Algorithm Use

Key Features

Soil NPK Levels Characterization

Using Near Infrared and ANN

ANN

Uses NIR Spectroscopy for

soil analysis

Crop Prediction using NPK

Sensors and Machine Learning

Decision Tree

Predicts crops based on NPK

levels using sensors

3 Soil NPK Levels Characterization

ANN with Back

Propagation

Achieved high correlation (R

= 0.998)

Crop Prediction using NPK

Sensors

Decision Tree

Uses real-time sensor data for

crop prediction

5 Soil Characterization Approach ANN

Relates soil absorbance data

to NPK levels

6 Soil Testing Methodology ANN

Verified with traditional

chemical analysis

7 Real-time Crop Prediction Decision Tree

Uses ESP-32 for real-time

data transmission

8 Soil Spectroscopy for Agriculture ANN

Non-invasive soil testing

metho

9 Machine Learning for Agriculture Decision Tree

Predicts best crop using

environmental data

10 NPK Level Determination ANN

Demonstrates ANN efficiency

redictin

soil nutrients

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

532

Table 1 summarizes a comparative analysis of

related studies showing the algorithms employed for

research focus as well as key features. This is what

allows to better show why AGRO-AI is needed and

how it is a more complete and integrated solution to

a faster advancing field such as that of precision

agriculture.

3 DATA COLLECTION AND

PREPROCESSING

The data used in this study was acquired in real-time

for over 20 weeks, during which three paddy leaf

samples from every week were collected. These were

selected so they would represent different growth

stages and nutrient availabilities. The environmental

parameters such as temperature, humidity, and soil

moisture were also logged in conjunction with the

leaf images to completely represent the data. Actual

NPK (Nitrogen, Phosphorus, and Potassium)

concentrations were obtained by way of laboratory

analysis to provide true ground truth figures for

model training.

The data was cleaned and processed using

various techniques to remove outliers or irrelevant

information and to convert the data into a format more

suitable for training machine learning models. Images

were resized to normalize resolution disparity, and

noise reduction techniques were applied to the

images to mitigate non-essential visual noise. Also,

transformations such as rotation, flipping, and

brightness elements to the training images were

added to the model to increase the diversity of the data

and increase the robustness of the model.

The possible vegetation indices were calculated in

the preprocessing step (Visible Atmospherically

Resistant Index (VARI), Green Leaf Index (GLI), and

Excess Green Index (ExG)) from the data. However,

above indices turned out to be important features for

model predictions that closely provided essential

information on the overall health status and nutrient

content of paddy leaves.

And then, this dataset was structured as a time

series for the temporal changes of leaf characters.

The model uses a time series approach that allowed it

to learn patterns and trends over time, improving the

accuracy in predicting NPK values and accuracy in

disease detection too. To show how the data

compared some visualizations of the sample data

shows how the differences were society across the

various stages of plant growth.

4 METHODOLOGY

Architecture and Implementation of AGRO-AI for

NPK Prediction and Disease Detection of Paddy

Leaves Models such as ResNet-34 and BiGRU are

used for nutrient prediction, and CNN and GAN

models are applied for classification of disease.

4.1 Model Architecture Overview

4.1.1 NPK Prediction Model

The NPK prediction model uses a hybrid architecture

of feature extraction through ResNet-34 and sequence

extraction through BiGRU. Model inputs are the

time series of vegetation indices (VARI, GLI, ExG)

from weeks 1 to 20 The ResNet-34 highlighted the

visual features from the images, while BiGRU

described the temporal dependencies for predicting

nitrogen, phosphorus, and potassium accurately.

Figure 1: Confusion Matrix of NPK prediction model.

Figure 2: Model Accuracy over Epochs for NPK prediction.

Agro AI: Intelligent Nutrient and Disease Management System for Sustainable Farming

533

Figure 3: Model Loss over Epochs for NPK prediction.

Figures 1, 2, and 3 demonstrate the performance

metrics and training progress of the NPK prediction

model. The confusion matrix shows the classification

accuracy, while the accuracy and loss graphs

illustrate the model’s learning curve.

4.1.2 Disease Detection Model

For disease classification, a CNN model is used to

extract spatial features from leaf images.

Additionally, a GAN model enhances disease

classification by generating high-quality synthetic

samples, increasing the dataset’s diversity. Five

classes of diseases are identified: bacterial blight,

brown spot, rice blast, sheath blight, and tungro virus.

Figure 4: Confusion Matrix for Disease Detection.

These visualizations provide insights into the

model’s classification performance, showcasing both

the accuracy and error patterns. The ROC and

precision-recall curves further demonstrate the

classifier's effectiveness across different disease

categories.

Figure 5: Model Accuracy over Epochs for Disease

Detection.

Figure 6: Model Loss over Epochs for Disease Detection.

Figure 7: ROC Curve for Disease Detection.

Figure 8: Precision-Recall Curve for Disease Detection.

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

534

4.2 Model Training and Validation

Both models were trained using a combination of

supervised learning techniques. For NPK prediction,

the dataset was divided into 80% for training and 20%

for testing. A similar split was applied for disease

detection models. The training process utilized

different loss functions based on the task — Mean

Squared Error (MSE) for NPK prediction, ensuring

accurate nutrient value predictions, and Categorical

Cross-Entropy for disease classification to optimize

multi-class predictions. To achieve efficient learning,

the Adam optimizer was used in both cases,

leveraging adaptive learning rates for faster

convergence. Key hyperparameters included a batch

size of 32 and an initial learning rate of 0.001.

Additionally, early stopping was implemented to

prevent overfitting, ensuring the models generalize

well on unseen data.

4.3 Evaluation Metrics

Different evaluation metrics were used to evaluate the

models performance. To measure the overall

correctness of the predictions, we calculated

accuracy, defined as the ratio of correct predictions

to total predictions. Overall performance of the model

was assessed using Precision and Recall mechanisms

applied to the positive cases that were correctly

identified, where precision is the ratio of relevant

instances amongst the retrieved instances, and recall

is the ratio of relevant instances that were

successfully retrieved. To account for both metrics,

we used the F1-Score, which provides a harmonic

mean of precision and recall.

Also, the ROC Curve was used to graphically

represent the trade-off between true positive versus

false positive rates, giving insight to the model’s

performance in classification. And as the Precision-

Recall Curve tends to be more informative than the

ROC curve in the exposition of class imbalance, it

was performed to reassert the accuracy of the model.

The Figures 1 to 8 are the illustrations for these

metrics, showing the results of the NPK prediction

and disease classification by using the proposed

AGRO-AI models.

5 RESULTS AND DISCUSSION

The results from the AGRO-AI models for both NPK

prediction and disease detection are presented and

analyzed in this section. Visual representations in

Figures 1 to 8 illustrate the performance of the models

using various evaluation metrics.

5.1 NPK Prediction Results

The predicted nitrogen, phosphorus and potassium

(NPK) levels of paddy leaves showed a strong

performance in their ability to utilize deep learning

models such as ResNet-34 and BiGRU. The number

of true positive (TP), true negative (TN), false

positive (FP), and false negative (FN) rates can be

seen in the figure below (Figure 1) With our

accuracy over epochs (Figure 2) steadily increasing

during training and the loss over epochs (Figure 3)

largely decreasing indicating that we are learning, it

is time to move on to testing our new model.

The results highlight the model's effectiveness in

learning complex patterns through the integration of

image processing and sequential modeling techniques

within the time series context. Vegetation indices

(VARI, GLI, and ExG) used in this study were

important in achieving a higher accuracy for the

model.

5.2 Disease Detection Results

The CNN and GAN models were well suited for

disease detection as they accurately classified paddy

leaf diseases with a high percentage of accuracy. The

confusion matrix (Figure 4) in the below depicts that

the model can distinguish between healthy and

diseased leaves with good accuracy. As depicted in

the accuracy over epochs chart (Figure 5), there is an

apparent upward trend of accuracy line with epochs

while loss over epochs chart (Figure 6) depicts the

downward journey of loss, indicating better

convergence of the model.

To provide a more thorough assessment of

classification performance, the ROC curve (shown in

Figure 7) illustrates the rate of true positive

predictions against false positive predictions for the

model, which is maximally enclosed due to a

favorable area under the curve (AUC). Finally, the

precision-recall curve in Figure 8 confirms that the

model is effective, especially in the case of a class

imbalance.

5.3 Comparative Analysis

These suggest that the AGRO-AI models provide

higher accuracy prediction and robustness in disease

detection compared to existing methods. This

substantial gain in prediction accuracy can be

attributed to the application of deep learning

Agro AI: Intelligent Nutrient and Disease Management System for Sustainable Farming

535

architectures, alongside vegetation indices and

temporal data.

The developed NPK prediction model had good

agreement with lab test results (Showed close

correlation with lab test results) thus indicating its

real-world applicability. The disease detection model

showed good recognition of the patterns relating to

the diseases which is essential for early diagnosis and

intervention.

5.4 Insights and Implications

AGRO indicates the system specializes in

agriculture, and AI suggests that the model leverages

the power of artificial intelligence. The farmers can

predict how much NPK to use and thus optimally use

the fertilizer, thus saving the fertilizer wastage and

ensure a better growth of agricultural crops. This

minimizes problems of over-fertilization, thus

increasing yield and also reducing stress on the

environment due to the excessive use of fertilizers.

Diseases in a particular stage are accurately

classified, therefore directing the system towards the

disease early detection and treatment which helps to

prevent it from spreading and also helps to reduce

crop loss. These AGRO-AIs can be preferably scaled

up to facilitate integration of the smart agricultural

systems that need real-time monitoring and decision

support. Additionally, the size of the dataset can be

further increased in terms of number of crops and

environmental diversity, thereby increasing the

robustness and accuracy of the models. These

innovations will inform the AGRO-AI approach and

lead to radical improvements in nutrient management

and disease detection through the application of

machine learning as a pathway to more sustainable

and efficient agronomic systems.

The findings demonstrate the potential value of

AGRO-AI as an integrated tool for nutrient

management and the identification of paddy disease,

aiding in the advancement of sustainable and

precision agriculture by embracing technology to

enhance yield and global food security.

6 FUTURE SCOPE

AGRO-AI: A smart farming technology experts

from India and USA Listen to audio version of this

article 0:00 / 1:05 1X Throughout the year, online

surveys are conducted for farmers to capture their

needs. More can be done to improve the functions.

This can be improved by expanding the images the

dataset contains to include images of various crop

varieties and variations in climatic conditions. To

improve prediction accuracy, more vegetation indices

could be used, and multispectral or hyperspectral

images could also be utilized.

Moreover, coupling AGRO-AI with Internet of

Things (IoT) devices for monitoring and enable real-

time decision making would provide timely insights

to farmers. To make this practically usable, creating

some mobile/web-based apps that the farmers can

easily reach predictions and recommendations would

be an asset. Optimizing the system's performance

using transfer learning and fine-tuning on larger

datasets might involve experiments with parameter

tuning and advanced architectures.

Also, to evaluate how precision nutrient

management and early disease detection through

AGRO-AI systems can improve the economic and

environmental sustainability of farming and to

reflect on future research directions. AGRO-AI can

also be further developed and applied to a wider

range of crops and areas of agriculture, making it an

adaptable and revolutionary solution in contemporary

agriculture.

7 CONCLUSIONS

So, AGRO-AI system could accurately recommend

nutrients and detect disease in paddy crops. While

NPK prediction was done using deep learning

models such as ResNet-34 and BiGRU, CNN and

GAN were used as deep-learning models for disease

classification. Using real-time data collected over the

course of 20 weeks has also increased the model’s

predictive capability by including vegetation indices.

The findings show how AGRO-AI could help

enable growers through actionable recommendations

that result in more precise fertilizer application and

timely approaches to disease management.

Additionally, the system can be used whenever

needed in smart-abseture-agricultural systems for

continuous monitoring and decision-making.

AGRO-AI contributes significantly to sustainable

agriculture by reducing yield losses and resource

consumption.

In short, AGRO-AI is a real solution for the

modern agricultural landscape at the crossroads of

efficient production and sustainability. As thousands

of years of cultivated datasets are built upon as per

advancement of this technology, it can change the

entire face of agricultural management and food

sustainability.

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

536

REFERENCES

P. V. Prakash and M. Srivenkatesh, “Hybrid Deep Learning

Algorithms for Predicting Nutrient Deficiencies in

Paddy Crops using CNN and Super Resolution

Generative Adversarial Neural Networks”, Int J Intell

Syst Appl Eng, vol. 12, no. 16s, pp. 520–526, Feb. 2024.

J. S. R. and M. S. G. Premi, “Nutrient Deficiency of Paddy

Leaf Classification using Hybrid Convolutional Neural

Network,” *Int. J. Electr. Electron. Res.*, vol. 12, no.

1, pp. 286–291, 2024.

J. Zhou and H. Zhang, “Customised Convolutional Neural

Network with Transfer Learning for Multi-Nutrient

Deficiency Identification in Paddy Leaves,” *J. Plant

Health*, vol. 12, no. 3, pp. 45–56, 2023.

X. Li and Y. Chen, “Using Deep Convolutional Neural

Networks for Image-Based Diagnosis of Nutrient

Deficiencies in Rice,” *Front. Plant Sci.*, vol. 12, pp.

748040, 2021.

S. Wang and Q. Liu, “Improved Tunicate Swarm

Optimization Based Hybrid Convolutional Neural

Network for Rice Leaf Disease and Nutrient Deficiency

Classification,” *Agronomy*, vol. 14, no. 8, p. 1851,

2023.

D. Patel and P. Mehta, “Classification Of Nutrition

Deficiency In Rice Plant Using CNN,” in *IEEE Conf.

Comput. Intell.*, 2022, paper 9873082.

T. Zhang and Y. Wang, “PND-Net: Plant Nutrition

Deficiency and Dis ease Classification using Graph

Convolutional Network,” *Sci. Rep.*, vol. 14, p. 66543,

2024.

L. Huang and J. Zhao, “Deep Learning-Based Detection of

Nutrient Deficiencies in Rice Leaves,” *Comput.

Electron. Agric.*, vol. 198, p. 107006, 2023.

A. Singh and P. Kaur, “Automated Identification of

Nutrient Deficiencies in Paddy Crops Using Machine

Learning,” *J. Agric. Inform. *, vol. 13, no. 1, pp. 12–

25, 2022.

H. Chen and F. Gao, “Image-Based Recognition of Rice

Leaf Nutrient Deficiencies Using Deep Learning,”

*Plant Methods*, vol. 17, p. 88, 2021.

S. Yadav and R. Verma, “Detection of Nutrient Deficiency

in Rice Plants Using Transfer Learning Techniques,”

*Int. J. Adv. Comput. Sci. Appl.*, vol. 14, no. 5, pp.

234–240, 2023.

T. T. Nguyen and Q. Pham, “Application of Deep Learning

in Diagnosing Nutrient Deficiencies in Rice Leaves,”

*J. Inf. Telecommun. *, vol. 6, no. 2, pp. 123–135,

2022.

S. Ghosh and A. Das, “Identification of Nutrient Deficiency

in Paddy Leaves Using Image Processing and Machine

Learning Techniques,” *J. Agric. Eng.*, vol. 50, no. 2,

pp. 67–75, 2023.

J. Park and S. Lee, “Real-Time Detection of Nutrient

Deficiencies in Rice Plants Using Convolutional Neural

Networks,” *Sensors*, vol. 22, no. 15, p. 5634, 2022.

Agro AI: Intelligent Nutrient and Disease Management System for Sustainable Farming

537