A Deep Learning–Based Method for Paddy Leaf Disease Detection and Growth Stage–Specific Treatment Recommendation

Summary

This study presents a deep learning–based method for predicting paddy leaf growth stages and classifying diseases, delivering stage-specific treatment recommendations through an automated, user-friendly decision support system for effective crop disease management.

Abstract

Paddy leaf diseases significantly affect rice yield and quality, making early detection and proper treatment essential for precision agriculture. This study proposes a deep learning–based decision support system for paddy leaf disease detection, growth-stage prediction, and stage-specific treatment recommendations. The dataset used in this study comprises paddy leaf images collected from multiple sources and categorized by growth stage and disease class. The dataset was divided into training (80%), validation (10%), and test (10%) sets to ensure proper model evaluation. For growth stage prediction, a lightweight Convolutional Neural Network (CNN) model was developed, while disease classification was performed using transfer learning models, including VGG16, ResNet50, InceptionV3, and MobileNetV2. An ensemble method based on average probability voting was used to improve classification performance. The models were evaluated using accuracy, precision, recall, and F1-score on an independent test set. The experimental results show that the ensemble model achieved higher accuracy compared to individual models, demonstrating improved robustness and generalization. The proposed system was implemented as a Streamlit web application that provides disease detection, growth-stage prediction, and treatment recommendations. The proposed integrated framework can support farmers and agricultural experts in making timely and accurate disease management decisions.

Introduction

India, being an agrarian economy, depends heavily on agriculture, with rice (paddy) serving as a staple crop for a large portion of the population. However, paddy cultivation is significantly affected by leaf diseases such as blast, bacterial leaf blight, brown spot, and sheath blight, which reduce yield and cause economic losses. Early detection of these diseases is essential, as traditional manual inspection methods are time-consuming and often inaccurate. To address this issue, this study proposes a deep learning–based decision support system for paddy leaf growth stage prediction, disease classification, and treatment recommendation. The system uses a neural network model to predict the growth stage of the paddy leaf and multiple deep learning models, including CNN, VGG16, ResNet50, InceptionV3, and MobileNetV2, for disease classification. These models are combined using an ensemble approach to improve accuracy and robustness, achieving a classification accuracy of 98.99%. Based on the predicted disease and growth stage, the system provides stage-specific treatment recommendations. The proposed system is implemented as a Streamlit-based web application that allows users to upload leaf images and receive real-time predictions and recommendations. This integrated approach provides a practical and scalable solution for precision agriculture, helping farmers reduce crop losses and improve productivity.

Several studies have applied machine learning and deep learning techniques for crop recommendation, disease detection, and agricultural decision support. Ensemble deep learning methods have been used to improve crop and fertilizer recommendation accuracy by integrating disease prediction, showing better performance than single-model approaches¹. Machine learning–based recommendation systems using decision trees and support vector machines have also been developed to suggest crops, fertilizers, and disease treatments while incorporating farmer feedback for improved adaptability².

For disease detection, hybrid CNN models combined with feature extraction techniques such as CLAHE and GLCM have improved paddy disease classification accuracy³. Other approaches have used clustering and Random Forest algorithms with texture-based feature extraction to enhance disease detection performance⁴. Deep learning models, particularly CNNs and transfer learning architectures, have achieved high accuracy in plant disease detection and enabled real-time and smartphone-based diagnosis systems^5,6,7. Traditional image processing and classification tree methods using color, texture, and shape features have also shown effective disease detection with lower computational complexity^8,9.

Ensemble learning has been widely used in crop recommendation systems by combining classifiers such as Random Forest, Naïve Bayes, and Support Vector Machines to improve prediction accuracy and robustness¹⁰. CNN-based models have also been used for paddy disease detection and integrated with IoT systems for real-time agricultural recommendations¹¹. Machine learning models such as K-Nearest Neighbors and regression techniques have been applied in predictive decision support systems¹². Additionally, crop recommendation systems based on soil, seasonal, and productivity factors have been developed to support data-driven farming decisions ¹³.

Studies have also examined rice production constraints and disease management strategies, highlighting the importance of sustainable disease control and improved agricultural practices¹⁴. However, most existing studies focus only on disease detection or crop recommendation individually, and very few integrate disease classification, growth stage prediction, and treatment recommendation into a single decision support system. This gap highlights the need for an integrated and practical agricultural decision support framework¹⁵.

The novelty of this research lies not only in disease classification but in the development of a complete decision support system for paddy disease management¹⁶. The proposed system integrates paddy leaf growth stage prediction, disease classification using an ensemble of deep learning models, and growth stage–specific treatment recommendation. Unlike existing studies that focus only on disease detection, this work links disease prediction with actionable treatment advice based on the plant growth stage¹⁷. This integrated approach makes the system more practical for real-world agricultural use and supports precision agriculture by providing timely and stage-appropriate intervention¹⁸.

The main goal of the given research is to create the next stage deep learning system to detect the paddy leaf disease earlier and to precisely predict the growth stages of the leaf and recommend the treatment in time, as well as selectively. The system attains a high accuracy in disease classification of 98.99 by fine engineering and assembling state of art models such as CNN, VGG16, ResNet50, InceptionV3, MobileNetV2 amongst others. It incorporates a neural network that predicts the growth stage and gives specific herbicide or pesticide suggestions regarding the analysis. The technology is available to farmers through a user-friendly Streamlit web application that enables real-time predictions as well as treatment advice¹⁹. The study should facilitate decisions related to disease management through timely predictions and treatment recommendations with possible applicability to agrarian economies such as India.

Even though multiple studies are performed on the detection of plant diseases in terms of machine learning and deep learning, the majority of existing research attempts at disease classification without providing the consideration of the development stage of the plant and suggestions of treatment. It is not the case, however, that in actual practice in agriculture the treatment instructions are based not only on the nature of the disease but also on the stage in which the crop is at the time of treatment, since the degree and type of pesticides differ according to the stage of the crop. Very little research incorporates disease detection, growth stage prediction as well as treatment recommendation all in one automated system. To bridge this gap, the current study suggests a combined deep learning-based architecture that will incorporate leaf stage prediction of paddy, disease classification in a set of deep learning models, and recommendation of specific treatment based on a growth stage through a Streamlit-executed decision support system²⁰. Such integration allows the proposed system to become more practical and useful in the real-world precision agriculture. This study proposes a stage-aware disease diagnosis framework in which growth stage prediction and disease classification are combined with a treatment recommendation module to form a complete decision support pipeline²¹.

Although several deep learning and ensemble learning approaches have been proposed for plant disease detection, most existing studies focus only on disease classification and do not consider the growth stage of the plant. However, in practical agriculture, the treatment method, pesticide type, and dosage often depend not only on the disease type but also on the plant growth stage. Existing disease detection systems do not provide stage-specific treatment recommendations, which limits their practical applicability in precision agriculture²². Furthermore, many existing studies focus only on improving classification accuracy using deep learning models but do not integrate the prediction system into a practical decision support framework that can be used by farmers or agricultural experts. To address these limitations, this study proposes an integrated decision support system that combines paddy leaf growth stage prediction, disease classification using ensemble learning, and stage-specific treatment recommendation. The proposed system is implemented as a web-based application to support real-time agricultural decision-making²³.

The main contributions of this study include the development of a deep learning–based model for accurately predicting the growth stages of paddy leaves using image data, enabling better crop monitoring and management. It also involves the implementation of transfer learning techniques to effectively classify paddy leaf diseases by leveraging pre-trained models, improving accuracy while reducing training complexity²⁴. To further enhance the reliability of disease detection, an ensemble learning approach is employed, combining multiple models to achieve superior classification performance. Additionally, the study introduces a stage-specific treatment recommendation system that provides targeted solutions based on both the identified disease and the growth stage of the crop, ensuring more precise and effective agricultural interventions. Finally, all these components are integrated into a Streamlit-based web application, offering a user-friendly platform for real-time decision support, allowing farmers and stakeholders to easily access predictions and recommendations²⁵.

Protocol

The following protocol illustrates the process of paddy leaf disease detection and treatment recommendation system building and deployment step by step. The steps taken need to be used consecutively to replicate the findings presented in this paper.

Dataset preparation

Paddy leaves stage dataset: This data was based on PaddyNet4. The scene comprises 560 images as classified under four different development stages, namely: Stage 2, Stage 3, Stage 4 and Stage 5. The classification depends on the color and maturity of the paddy leaves, which depend on the concentration of nitrogen, among other factors, and the development of the plant. The presented dataset is important in knowing the right stage of the paddy plant (Figure 1). PaddyNet is accessible publicly via the website of data.mendeley.com/datasets/. Paddy leaves disease dataset: This dataset was available as several separate publicly accessible datasets on Kaggle and combined to create one unified dataset with a total of 6,920 images, each of which belonged to one of six disease types shown on (Figure 2). Treatment dataset: The dataset was critically developed by retrieving and integrating information of various publicly available datasets and contains data on various management practices of paddy diseases at four growth phases²¹. The main objective of this dataset is to come up with recommendations to prevent and control the paddy disease. The data is one of the resources that are vital towards fighting the diseases in the paddy plants by selecting the right kind of pesticides^{19,22,23,24,25}

Data augmentation

The data augmentation was used to enhance the training dataset and reduce overfitting. The augmentation methods that have been employed in the current study are rotation (up to 30), width and height moving (up to 20%), zooming (up to 20%), shearing, and horizontal flipping. The use of these augmentation methods assists the model in acquiring the essentials of the model and enhances the generalization of the model when it is tested on unseen data. Data augmentation was done on the training dataset, whereas the validation and test datasets were left as they were in order to review the fair performance of the system. The size of the original dataset had to be augmented with data because the size of the original dataset is rather small, and the approach was needed to enhance the generalization and minimize overfitting by introducing the model to the differences in orientation, scale, and position of leaf images.

Image preprocessing

The stage and disease data collected were passed through a number of preprocessing steps to make the data consistent and enhance the performance of the models. All the pictures from both datasets were scaled to 128 × 128 × 3 to ensure a workable fit in any deep learning architecture. The pixel values were scaled to the interval [0,1] so as to hasten convergence when training. Moreover, labels of diseases and their stages were transformed into numbers to evaluate and train the model more powerfully. These preprocessing methods helped to extract features better, a better classification rate, and the robustness of the model to different environmental conditions. The size of the image used, 128 × 128, was compromised between computation and feature loss. Concrete Smaller image size minimizes training time and computational cost, and does not eliminate information (visual features) to classify the disease and stage when utilizing deep learning models.

Dataset splitting

The processed data were divided into three subsets (Table 1): Training Set (80%), Validation Set (10%), Test Set (10%). To avoid an unbalanced distribution of classes in all three subsets, stratified random sampling was used to ensure that both the stage dataset (560 images) and disease dataset (6,920 images) had an equal distribution of classes. There was no data leakage among the subsets.

Model training

In the classification of the stage of the growth of paddy leaves and diseases, two steps model training method was used.

Stage classification model

The classification method of leaf growth stage has four classes and a relatively low data size, thus the architecture of the Convolutional Neural Network (NN) was not very expanded but considered lightweight. A simpler architecture can be used to provide reduced over-fitting and computational power, at the expense of having high accuracy in the case of classification using color and texture features. The new CNN structure will have three convolutional layers using the ReLU activation functions. The convolutional layers are alternated with the max-pooling layers (2×2) to downscale the spatial dimensions, as well as to identify the prominent features. The feature maps undergo a convolutional layer, after which they are flattened, and the fully connected dense layer with dropout regularization is run to discourage overfitting. Lastly, multi-class classification of the four growth stages is done with the help of a softmax output layer. The Adam optimizer, which has a learning rate of 0.001 and categorical cross-entropy as a loss function, was used to train the model. The number of the batch was set to 10, and the model was trained in 10 epochs. Validation accuracy and validation loss were also used to monitor the performance of the model since they are used to make sure that the model is generalized.

Disease classification model based on ensemble learning

For disease classification, a series of deep learning architectures was adopted to improve prediction accuracy. Initial development of a custom CNN model consisted of four layers, namely, three convolutional layers, three max-pooling layers, a dense fully connected layer, and dropout regularization. MobileNetV2 employs depth-wise separable convolutions, inverted residual blocks, and inverted bottleneck with linear residue. The model has a small size of only 3.4 million parameters and can be used in a resource-constrained environment. The models started using ImageNet activations, and the convolutional layers were frozen so that the learned feature representations are not lost. The softmax classification layer was attached to a fully connected layer that had dropout regularization. All the models were trained during 20 epochs with a batch size of 32, and the optimizer and loss were identical to those of the CNN model. In an attempt to improve on classification, the average voting ensemble learning method was used. The results were the average probabilities predicted by all five models (CNN, VGG16, ResNet50, InceptionV3, and MobileNetV2) to obtain the final class predictions. This combination technique was useful in lessening bias in models and enhancing generalization. The last combination model was more accurate in classifications. The reason why the ensemble approach was selected is that learning of the various feature representations is learned using the same data set by the different deep learning models. The ensemble also decreases model bias and variance and increases the overall prediction issues and overall generalization performance as compared to independent models.

Prediction strategy

In the process of predicting paddy leaf disease and its associated stage of growth, the ensemble-based classification method was utilized. The five deep learning models (CNN, VGG16, ResNet50, InceptionV3, and MobileNetV2) were independently trained to classify the diseases and pooled together by utilizing an average voting ensemble method. This procedure entailed the forecasting of probabilities of each disease category of all models and the average probability distribution, which adds precision to predictability and minimizes bias in the model. The final prediction was taken as the disease class with the most avg probability. Further, a different CNN-based stage classification model was trained and used to forecast the growth-stage of the paddy leaf. Both classification models would take in input images of size 128x 128 then normalize the pixel values and then use the models to make an inference. The final predictions of both leaf stage and disease were then mapped onto an existing set of treatment recommendations to offer valid treatment in order to manage the disease.

Example Output: "Leaf stage: 3, Disease: Tungro"

Treatment recommendation

To make the paddy leaf disease classification more practical, a treatment recommendation system was integrated into the application based on Streamlit. The system takes the forecasted disease and leaf stage to offer specific treatment recommendations to make sure that the disease is managed successfully. The model of ensemble disease classification (including CNN, VGG16, ResNet50, InceptionV3, and MobileNetV2) classifies the disease, whereas the model of classifying the leaf stage is another model. The modeled stage of the leaf and disease are compared with the data of treatment that has recommended doses (Preventive measures and Application of Pesticide).

Methodological framework and contribution

The methodology of the presented system is a pipeline of multi-stages, which combines image classification and decision support. The framework consists of five large steps, namely, promotion of images and augmentation, predicting leaf growth stage, using deep learning models trained on the images to classify the disease, combining the prediction results to obtain an ensemble decision, and recommending the treatment based on the type and growth stage of an illness. The methodological contribution of this work is geared at the implementation of many prediction models and a treatment recommendation module into one decision support pipeline. The system is based not just on the classification of diseases but also on incorporating the information of the stage of plant growth so as to provide treatment recommendations at a particular stage. The system is more relevant to the process of agricultural decision-making in the real world because of this stage-conscious recommendation strategy. The input image will first undergo preprocessing before being fed through the stage classification model and disease classification models. The disease model results are averaged using an ensemble technique, and the disease prediction and growth stage are carried to the end to obtain correct treatment recommendations based on the identification of relevant treatment records in the treatment database.

Validation strategy and overfitting prevention

In order to provide stringent validation of the developed models, the data set was split into three disjunctive subsets: training set (80%), validation set (10%), and test set (10%). The models were trained using the training set, tuned using the validation set, and evaluated their final performance using the test set only. The test set remained absolutely independent to have an impartial analysis of the model. The following methods were used to guarantee that the model does not overfit and enhance its ability to generalize. To increase the level of diversity in the dataset, first, data augmentation methods (rotation, zooming, shifting, shearing, and horizontal flipping) were used. Second, dropout regularization was used in the fully connected layers of the CNN models to reduce overfitting. Third, pretrained models (VGG16, ResNet50, InceptionV3, and MobileNetV2) were transferred using the frozen convolutional layers technique to preserve the learned features and minimize the chances of overfitting caused by the small size of the dataset. Lastly, validation loss and accuracy were used to monitor model performance during training to be able to ensure that the model is not overfitting to the training data. These are the validation and regularization schemes that ensure that the model proposed generalizes very well to unseen data and offers good performance.

Proposed system

The suggested one is a deep learning solution to the paddy leaf disease management, which incorporates the leaf stage forecasting and classification of the disease. It starts with the preprocessing of paddy leaf images in terms of resizing them to a common height and the pixel values. The leaf stage of growth is predicted using a neural network, and an average voting technique is used to use an ensemble of deep learning models (CNN, VGG16, ResNet50, InceptionV3, MobileNetV2) to classify diseases. The system uses the forecasted stage and disease to prescribe the right doses of herbicides or pesticides to be used in the treatment. The workflow of the proposed system is shown in (Figure 3). The previous subsections (A through I) explain each of the individual elements of this pipeline, that is why more description is not done here. The Streamlit-based application developed in this study is a prototype decision support tool intended to demonstrate the practical applicability of the proposed model. The system allows users to upload paddy leaf images and receive disease predictions, growth stage information, and treatment recommendations. Large-scale usability testing and field deployment with farmers will be conducted as part of future work to evaluate the system’s real-world effectiveness and usability.

Results

Stage classification model

The stage prediction model achieved a training accuracy of 95.88% (loss: 0.1172) and a validation accuracy of 95.24% (loss: 0.1751), indicating that the model is learning consistently and making fair predictions when applied to unseen validation samples. A validation accuracy of 95.24% was achieved, with a validation loss of 0.1751. The model achieved a test accuracy of 94.05 percent and a test loss of 0.1759 on the test data, which supports the robu...

Discussion

The current paper suggests a novel decision support system which involves the prediction of growth stage of paddy leaves, classification of the diseases with the help of an ensemble of deep learning models (VGG16, MobileNetV2, and ResNet50), and stage-based recommendation of treatment based on a web-based application. The ensemble model got an accuracy of 98.99% in the classification of the model which is higher than that of individual models, as previously reported that models in individual strategy have sensitivity to ...

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
Dataset	Paddy Leaf Stage Dataset (PaddyNet4)	560 images across 4 growth stages (Stage 2–5)
Dataset	Paddy Leaf Disease Dataset	6920 images across 6 disease classes from Kaggle
Dataset	Treatment Dataset	Compiled dataset with disease management practices by growth stage
Preprocessing	Image Resizing	Images resized to 128x128x3
Preprocessing	Normalization	Pixel values scaled to [0,1]
Preprocessing	Label Encoding	Converted categorical labels to numeric
Augmentation	Rotation	Up to 30 degrees
Augmentation	Width/Height Shift	Up to 20%
Augmentation	Zoom	Up to 20%
Augmentation	Shear & Flip	Shearing and horizontal flipping
Model	CNN (Stage Classification)	Lightweight CNN with 3 conv layers
Model	CNN (Disease Classification)	Custom CNN with conv + pooling layers
Model	VGG16	Transfer learning model
Model	ResNet50	Transfer learning model
Model	InceptionV3	Transfer learning model
Model	MobileNetV2	Lightweight transfer learning model
Technique	Ensemble Learning	Average probability voting of 5 models
Technique	Transfer Learning	Using pretrained ImageNet weights
Technique	Dropout Regularization	To prevent overfitting
Training	Optimizer	Adam (learning rate 0.001)
Training	Loss Function	Categorical Cross-Entropy
Training	Batch Size	10 (stage model), 32 (disease models)
Training	Epochs	10 (stage), 20 (disease)
Evaluation	Metrics	Accuracy, Precision, Recall, F1-score
Deployment	Streamlit	Web application for real-time prediction