Hybrid deep learning for smart paddy disease diagnosis using self supervised hierarchical … – Nature

As a staple grain, paddy is essential to the world’s food security, particularly in nations like India, where it is a primary dietary component and a key economic driver. According to the International Rice Research Institute (IRRI), rice is the staple food for more than half of the world’s population and supports the livelihood of approximately 4 billion people globally^1,2. Given the current per capita availability, paddy production needs to increase by 40–70%, from 116 to 145 and 190 million tons of milled paddy by 2030 and 2050 to meet growing demands Pradhan et al³. However, crop diseases pose a significant challenge to paddy production, affecting yield and quality. Crop diseases pose a significant challenge to paddy production, affecting both yield and quality. For example, a severe outbreak of paddy blast disease in Tamil Nadu in 2019 resulted in a yield loss of up to 30%, significantly disrupting local rice supply and causing economic hardship for farmers⁴. Early detection of such diseases is essential to reduce yield loss and ensure food security. Furthermore, incorporating both temporal and frequency domain analyses can provide deeper insights into disease progression. Major causes of crop loss such as plant diseases and pest infestations are known to be heavily influenced by environmental and agronomic factors^5,6.

Poor water management creates conditions conducive to disease and pest proliferation⁷. Similarly, inadequate soil nutrition weakens plant health, making crops more vulnerable to infections⁸. Furthermore, climate variability leads to plant stress, increasing their susceptibility to biotic diseases and consequently reducing productivity⁹. The increased cultivation of high-yielding, fertilizer-responsive paddy varieties has further contributed to a rise in disease prevalence, especially during monsoon seasons¹⁰. Recent developments in deep learning (DL) and artificial intelligence (AI) have transformed the identification of plant diseases by providing automated, precise, and real-time disease diagnosis systems. Among these, deep learning models such as^11,12 Convolutional Neural Networks (CNNs), Multi-Layer Perceptrons (MLPs), and Long Short-Term Memory (LSTM) networks have shown remarkable effectiveness in identifying diseases.

Literature survey

A study on stress-tolerant breeding was conducted by Pradhan et al. in³. Strong performance has been shown by deep learning models like CNN and LSTM¹³ used a CNN-LSTM hybrid to identify apple illness with 92.3% accuracy, while Patil & Kumar¹⁴ introduced Paddy-Fusion, increasing accuracy to 94.7%. CNN-LSTM was used to improve illness severity categorization by Kukreja et al.¹⁵ and Kaur et al.¹⁶, who achieved 91.5% and 93.2% accuracy, respectively. Jiang et al.¹⁷ used image-based detection techniques that incorporated deep learning and SVM, achieving 90.1% accuracy. Kukreja. January¹⁸: presents a new model that blends Long Short-Term Memory (LSTM) networks with Convolutional Neural Networks (CNN). This model uses temporal and spatial data to accurately (94.06%) determine the severity of the condition. Dubey & Choubey^19,20 used MBi-LSTM with adaptive feature selection, increasing classification accuracy to 95.4%. While Badgujar et al.²¹ enable AI-based categorisation, real-world datasets such as “PaddyDoctor,” created by Petchiammal et al. ²², The accuracy of the CNN-LSTM model is 94.5%, which is higher than that of the conventional CNN-based models, which are between 88 and 90% accurate. AI is beneficial to precision agriculture, as discussed in Agriculture 5.0 by Malekloo et al.²³ and Mesías-Ruiz et al.²⁴. The LSTM classifier and MLP-CNNs produced the greatest accuracy score of 99.2%. Choubey and Dubey²⁵. Time-series forecasting for disease prediction, like CNN-LSTM-MLP models by Ghimire et al.²⁶, achieved 92.8% accuracy, and Kontopoulou et al.³³ examined ARIMA vs. ML models and found that ML outperformed conventional techniques. In their exploration of transfer learning using ResNet, VGG, and Inception, Shafiq & Gu²⁵ and Pillai et al.²¹ achieved 96.1% accuracy in the categorization of plant diseases. According to Lu et al.³², hybrid models such as CNN-BiGRU enhanced the identification of paddy diseases with an accuracy of 94.2%, while Thapliyal et al.⁴⁰ employed CNN-LSTM to detect maize diseases with an accuracy of 92.5%. To diagnose paddy nutrient levels during the panicle initiation stage, Liao et al.²⁴ created a CNN-LSTM model, which outperformed conventional image-based techniques (88–91%) with an accuracy of 94.2%. Similar to this, Daniya & Vigneshwari³¹ demonstrated that hybrid feature extraction enhances illness identification by proposing a deep neural network that integrates texture-based (GLCM) and deep features, with 91.8% accuracy. In the meantime, Farooqui et al.³² improved illness classification accuracy to 93.5%³³ by using a concatenated deep feature method with a modified LSTM, outperforming conventional CNN-LSTM models (89–91%).

Recent advancements in deep learning have significantly improved the accuracy and reliability of plant disease detection systems. Among these, proposed a multi-attention convolutional neural network (CNN) integrated with transfer learning techniques for the early classification of rice diseases. Their model leverages attention mechanisms at multiple stages of the CNN pipeline to focus on disease-relevant features in leaf images, improving both interpretability and precision. Furthermore, the use of transfer learning enabled the model to benefit from large-scale pre-trained weights, reducing the need for massive domain-specific data. Their results showed a substantial improvement in early detection accuracy, particularly under varied lighting and background conditions, making the model more applicable to real-world agricultural environments.

In another recent study, Zhou et al. ³⁴ introduced a hybrid interpretable deep ensemble model designed specifically for rice disease identification under real field conditions. Their framework combines multiple deep learning models, including CNNs and attention-based networks, into a cohesive ensemble that enhances robustness and performance. A key contribution of this work is its emphasis on model interpretability, which is crucial in agriculture for gaining trust from end-users such as farmers and agronomists. The authors also validated their model on a complex dataset collected in natural field environments, which included challenges such as occlusion, noise, and varied lighting. The ensemble method outperformed individual models in terms of accuracy, stability, and resilience to noisy inputs, highlighting its practicality for deployment in smart farming systems.

While we acknowledge that many deep learning models have demonstrated high accuracy in plant disease classification, our choice of the Self-Supervised Deep Hierarchical Reconstruction (SSDHR) model integrated with LSTM and Symmetric Fusion Attention (SFA) was made deliberately based on the specific limitations of existing models and the unique requirements of early paddy disease detection.

Specifically, our decision is justified by the following reasons:

1. 1.

   Limitation of existing models in temporal learning:

   Most existing feed-forward deep learning models primarily focus on extracting spatial features. Although they can identify disease-affected regions, they are less effective in detecting diseases at early stages because they lack temporal awareness. Our framework addresses this limitation by incorporating LSTM, which captures temporal dependencies and enables early disease prediction through continuous monitoring.

2. 2.

   Self-supervised learning advantage

   We employ a Self-Supervised Deep Hierarchical Reconstruction (SSDHR) approach to learn robust representations from unlabelled data, which is particularly useful in agricultural settings where annotated datasets are scarce or expensive to obtain. This distinguishes our model from traditional supervised-only approaches.

3. 3.

   Enhanced feature discrimination through SFA:

   The proposed Symmetric Fusion Attention (SFA) module fuses spatial and temporal attention to emphasize subtle but crucial disease features. This is a key differentiator from other models, which often rely solely on spatial cues and overlook inter-frame correlations.

4. 4.

   High accuracy and stability with XGBoost:

   To further enhance the classification stability, we integrate XGBoost (XGB) as the final decision layer, which complements the learned deep features and provides improved generalization performance. Our framework achieves 99.25% accuracy across 13 distinct paddy classes, including multiple diseases and healthy samples, indicating strong performance.

In summary, while many models offer good accuracy, our model was chosen for its ability to bridge the spatial and temporal gap, perform early detection, and operate effectively with minimal supervision. These advantages make it better suited for real-world agricultural applications where early intervention is critical.

Materials

Paddy Doctor dataset data collected from IEEE Data port, following picture cleaning, our dataset consists of 16,225 images. Figure 2 depicts the data set, data collecting, and annotation procedure. Detailed information is shown in Table 1. In Fig. 1. violin plot displays the age distribution for each label. Labels are shown on the x-axis, while age is shown on the y-axis.

Paddy classes are distributed on days (age).

Full size image

The dataset used in our study was obtained from the publicly available repository titled “Paddy Disease and Pest Image Dataset” on IEEE Data Port. The dataset comprises 16,225 high-quality images across 13 classes, including 12 paddy disease and pest categories along with healthy samples. Below are the detailed aspects of the data collection methodology:

• Source & location: The images were collected from actual paddy fields near Tirunelveli district, Tamil Nadu, India, during the cropping season from February to April 2021.

• Crop stage: Paddy crops ranged in age from 40 to 80 days, covering the most disease-prone growth phases.

• Camera & resolution: Images were captured using the CAT S62 Pro smartphone, which supports both RGB and infrared imaging. The original resolution was 1080 × 1440 pixels, and the images were later resized to 480 × 640 pixels to optimize computational efficiency during training.

• Lighting conditions: All images were taken under natural lighting conditions, ensuring realistic field variability.

• Annotation process: The dataset was manually annotated by agricultural experts, including agronomy specialists and plant pathologists. Each leaf was categorized into one of the disease classes or as healthy.

• Initial cleaning: From an initial collection of over 30,000 images, noisy, blurred, and redundant samples were removed through manual inspection, resulting in the final curated set of 16,225 images.

• Metadata collection: Additional metadata such as paddy variety and crop age were also documented for each sample to enable more contextual analysis.

Dataset Link: https://ieee-dataport.org/documents/paddy-doctor-visual-image-dataset-automated-paddy-disease-classification-and-benchmarking.

The dataset comprises a total of 16,225 high-quality images, covering 13 categories (12 disease/pest classes and 1 healthy class). Each class contains 1248 images, ensuring minimal class imbalance. This distribution was carefully maintained during dataset creation to support fair model training and evaluation is shown in Table 2.

A violin plot representation of the distribution of plant ages across 13 different paddy crop classes, including both healthy (normal) and diseased categories such as blast, hispa, tungro, bacterial leaf blight, brown spot, among others is shown in Fig. 1. The x-axis denotes the class labels, while the y-axis shows the crop age (in days), ranging from approximately 40 to 85 days. The shape and width of each violin illustrate the density of samples at specific age intervals, with wider sections indicating higher sample concentrations. Median values are shown as white dots, and interquartile ranges (IQR) are marked by thick black bars, providing insights into the central tendency and variability of each class. Notably, classes such as blast, bacterial panicle blight, and black stem borer exhibit relatively narrow age distributions, indicating that these diseases predominantly affect crops at specific stages. In contrast, the normal and leaf roller classes demonstrate broader age variability, suggesting their occurrence across a wider growth range. Moreover, the multimodal distributions observed in classes like tungro and yellow stem borer suggest disease incidence across multiple crop stages. This figure underscores the age dependency of various paddy diseases and highlights the need for incorporating temporal information in the model.

Methods

The proposed method integrates a Self-Supervised Deep Hierarchical Reconstruction (SSDHR) network, a Symmetric Fusion Attention (SFA) module, and a Long Short-Term Memory (LSTM) network to generate robust features before classification using XGBoost. Accurate detection of paddy diseases requires extracting spatial features by SSDHR and temporal features by LSTM from images and time-series data and the SFA module refines the features. The details of the network are shown in Fig. 2.

Working flow of the proposed model.

Full size image

To improve quality, data pre-processing (resizing, normalisation, augmentation) comes first is used. A symmetric fusion attention approach is used to decrease noise and retain pertinent information after a multi-scale CNN extracts spatial features. Temporal illness development is captured by concatenating and reshaping features for LSTM processing. Accuracy is increased by utilising XGBoost to classify the refined features. The model categorises paddy illnesses, such as Healthy, Tungro Virus, Brown Spot, Blast, and Bacterial Blight. This comprehensive strategy guarantees precise, instantaneous illness identification.

Self-supervised classification deep hierarchical reconstruction network with symmetric fusion attention

The planned SSDHR classifier network structure. To provide a larger feature pool for plant disease classification, the authors of³⁵ simultaneously used classification features y and the reconstructed latent representation z for closed-set classification as well as identification. Latent representations are extracted from every step of the middle-level layers in the classification network in the article, as a result of the loss of data details in the higher-level layers of supervised deep networks.

In Fig. 3., two symmetric branches have convolution kernels of different sizes (3 × 3 and 7 × 7). The features “z1, z2, z3, …, z6” represent the pooled results of multiple reconstructed latent representations. “Attention” represents the proposed SFA module. The feature “y” represents classification features.

The architecture of SSDHR network.

Full size image

The model’s resilience and perception are improved by adding two parallel branches that use convolutional kernels of varying sizes for multi-scale feature extraction. The channel dimension is used to concatenate characteristics from both branches and recreate intermediate layers for richer representations. These are categorized by processing them through completely connected layers. A Symmetric Fusion Attention (SFA) module enhances performance by improving focus on important aspects. Mixed Centre Loss also improves accuracy by refining features for more biased and condensed representations.

Symmetric fusion attention

Symmetric Fusion Attention (SFA) is a custom-designed attention mechanism developed to enhance the model’s ability to capture discriminative features from paddy leaf images by integrating both low-level and high-level spatial information. The “symmetric” nature refers to the balanced fusion of features extracted from multiple layers of the network (e.g., early and late convolutional layers), ensuring that fine-grained texture and contextual patterns are preserved. This attention module selectively emphasizes disease-relevant regions in the image, improving classification performance, especially in visually complex or overlapping disease cases.

The general layout of the attention module, which is suitable for parallel structured networks is shown in Fig. 4. To preserve information throughout each channel and lower computational overhead, restructure a portion of the channels for each input feature X (in {mathbb{R}}) ^CxHxW into batch dimensions^36,37. Instead of adding more branching structures, use the network’s initial parallel branches to capture multi-scale information. Additionally, each branch attention to both spatial and temporal aspects.

SFA model architecture.

Full size image

Figure 4 shows how these two attentions are computed. As a result, temporal attention,

“G” stands for the split groupings. One-dimensional vertical global pooling and one-dimensional horizontal global pooling are denoted by “X Avg Pool” and “Y Avg Pool,” respectively. Temporal attention is represented by “T attention” and frequency attention by “F attention.” The terms “3 × 3 Input” and “7 × 7 Input” denote inputs with various scales frequency attention, and coordinate attention are the three sub-branches that each branch of the symmetric structure simultaneously includes. Coordinate attention is a variation on CA attention^38,39, in which inputs are processed via one-dimensional vertical and horizontal global pooling rather than global average pooling. This allows for the preservation of exact positional information along one spatial guidance while collecting long-range dependencies along another. At height h, the c-th channel’s output can be shown as:

The output of the c-th channel at width w can be represented as:

where yc indicates the input features at the c-th channel.

We perform 1 × 1 convolutional processes and a sigmoid activation function after pooling the reshaped input features in two directions. This produces two sets of coordinate attention weights (P_{a}^{h} in {mathbb{R}}^{{frac{C}{G} times H times 1}}) and (P_{a}^{w} in {mathbb{R}}^{{frac{C}{G} times w times 1}}) in distinct directions. Coordinate attention can be thought of as independently integrating time and frequency information into the channel dimension when the model’s input is spectrograms. Furthermore, the following are the formulas for the previously introduced frequency attention Fa (in {mathbb{R}}^{1times Htimes 1}) and temporal attention is Ta (in {mathbb{R}}^{1times wtimes 1}) :

where AvgPool (1,1) represents adaptive average pooling, which reduces each channel’s feature map to 1 × 1, (σ) represents the sigmoid activation function, (.)T1,3 and (.)T1,2 represents transpose operations on dimensions 1 and 3, and dimensions 1 and 2, respectively. XG (in {mathbb{R}}^{{frac{C}{G} times w times H}}) represents the reshaped input feature map.

By applying temporal attention, frequency attention, and coordinate attention to XG and then weighting them, the output X’G (in {mathbb{R}}^{{frac{C}{G} times w times H}}) is obtained:

where (alpha ,beta ,gamma) are learnable parameters, and softmax function is applied to normalize them. Following that, information aggregation across several spatial dimension orientations is done cross-spatially. To achieve outputs, the two symmetrical branches employ temporal, frequency, and coordination attention simultaneously as shown in Fig. 5. The outputs of the two parallel branches are then subjected to two-dimensional global average pooling, and the pooled outputs are then reshaped to form ({mathbb{R}}^{{1xfrac{C}{G}}}). The outputs of both branches are simultaneously reshaped to dimensions ({mathbb{R}}^{{frac{C}{G}_{xHW} }}) ²¹. To obtain two spatial attention maps on each branch, respectively, global average pooling is applied to the outputs, then softmax is applied and cross-branch matrix dot product operations are performed. The final shared spatial attention weights for the two branches are calculated by adding the two spatial attention maps and using the sigmoid function. To create outputs that are the same size as the inputs for both branches, these weights are multiplied by the inputs of the two branches.

Temporal attention and frequency/spatial attention.

Full size image

LSTM

LSTM networks allow for long-term dependency learning by processing sequential data by keeping a memory cell under the control of input, forget, and output gates. They work well for language translation, speech recognition, and time series forecasting. They can also be used in conjunction with CNNs to analyze images and videos.

In Fig. 6 shows Three gates provide control over the memory cell: the input gate, forget gate, and output gate. These gates determine which data should be input into, taken out of, and output from the memory cell. Information added to the memory cell is regulated by the input gate. The data that is deleted from the memory cell is managed by the forget gate. Additionally, what data is output from the memory cell is controlled by the output gate. In order to learn long-term dependencies, LSTM networks are able to selectively preserve or reject information as it moves through the network.

Architecture of LSTM.

Full size image

XGBoost classifier

After feature extraction, the final feature vector F is used as input for XGBoost (eXtreme Gradient Boosting), an optimized gradient boosting algorithm that constructs an ensemble of decision trees. XGBoost predicts the disease class yi’ for a given input xi by aggregating multiple decision trees:

where, K is the number of trees, ({f}_{k}left({x}_{i}right)) represents the k-th decision tree, and (y_{{hat{i}}}) is the predicted disease label. Each tree is trained iteratively to minimize the objective function:

where, (lleft( {y_{i} ,y_{{hat{i}}} } right)) is the loss function and (Omega left({f}_{k}right)) is the regularization term controlling tree complexity. The gradient boosting update rule is:

By integrating SSDHR + SFA + LSTM for feature extraction with XGBoost for classification, the proposed method achieves state-of-the-art accuracy in detecting 13 paddy types. This hybrid approach enhances model generalizability, making it suitable for real-time crop health monitoring and early disease detection.

Loss function

In this section, we present the total loss function for the proposed LSTM-integrated SSDHR model for disease detection. The total loss function obtained by combining center loss, reconstruction loss, and mixed center loss of LSTM with the classification loss (AdaCos), of SSDHR model.

One-class Centre loss

The Centre Loss aims to increase the discriminative power of features by increasing the distance between feature points of different classes and decreasing the distance between feature points of the same class in the feature space. This is done by minimizing the separation between each sample and its corresponding class center. For disease recognition in paddy crops, this loss helps the model focus on learning robust disease features, improving classification accuracy.

where, fj ∈ Rd is the jth feature vector belonging to the yjth class. μyj ∈ Rd is the center of deep features for the yjth class. d is the feature dimension. This loss ensures that the deep features of the samples are clustered around the class centers, improving class separation.

One-class Centre loss (for normal samples)

The One-Class Centre Loss computes the distance between each feature and a global feature center for all normal samples (non-disease cases).

where, μ is the global feature center for all normal samples.

Mixed Centre loss

The Mixed Centre Loss combines both the center loss for the disease class and the one-class center loss for normal samples, giving us the final mixed loss.

where, λ₁,λ₂ are scalar weights to balance the influence of both losses.

Reconstruction loss and classification loss

The network integrates reconstruction loss (based on Mean Squared Error) and classification loss (using AdaCos loss). This combination ensures that the network can both reconstruct the input features and classify them correctly.

Reconstruction loss (MSE)

The network has two parallel branches for reconstructing the latent representations of the input data. The Mean Squared Error (MSE) for each branch is computed as follows:

For the first branch:

For the second branch:

where, A_j is the input feature (such as a spectrogram of disease features), Aj′ and ‘‘Aj′′ are the reconstructed features from the first and second branches, respectively.

AdaCos loss (classification loss)

AdaCos Loss ensures that the model maintains a margin between different classes. This helps in reducing intra-class variance and maximizing inter-class variance. The probability Pi,j that sample Aj belongs to class j is computed using the following formula:

where, cosθi,j is the cosine similarity between input feature Xi and the class center wj, S ~ is a scaling factor.

The dynamically adaptive scaling parameter at step t is computed as:

Total loss function

The final total loss function Ltotal for the SSDHR model, combining all components, is given by:

This comprehensive loss function ensures that the model performs both feature learning and disease classification effectively while accounting for the temporal nature of disease progression using the LSTM architecture.

Implemented proposed SSDHR integrated with LSTM architecture was implemented and validated with the Paddy data set. Including temporal and spatial attention mechanisms enhances the feature selection and improve the metrics such as accuracy, precision, recall, and F1-score are analyzed, and the results are compared with baseline deep learning models to demonstrate the effectiveness of our approach.

Figure 7 is a confusion matrix, a widely used tool for evaluating the performance of classification models. The classification performance parameters, such as precision, recall, and F1-score, are shown in the table for both healthy samples and different crop disease groups. The precision values, which range from 0.9841 to 0.9993, demonstrate how well the model can identify each condition without producing an overwhelming number of false positives.

Performance comparison matrix between predicted and actual paddy disease classes.

Full size image

These values were obtained over five independent runs using fivefold cross-validation, which ensures a robust assessment of the model’s performance variability across different data splits is shown in Table 3.

The model has a great ability to capture true positives, as seen by the continuously high recall values (which assess its ability to properly recognize all instances of an illness), with a minimum of 0.9881. Additionally, the F1 scores, which strike a compromise between recall and precision, are above 0.9870, indicating strong classification performance in every class. The model’s total accuracy of 0.9953 shows remarkable classification reliability.

There are three phases to the disease’s progression (40–80 days). Stage 1 (Early): Minor symptoms, such as tiny patches or yellowing. Stage 2 (Moderate): spreading infection, leaf bending, and larger lesions. Stage 3 (Severe): Significant wilting, necrosis, and damage that reduces yield. Control depends on early detection. The disease courses of brown spot, white stem borer, bacterial leaf blight, and bacterial panicle blight are depicted in Fig. 8. Bacterial leaf blight (a) and white stem borer (b, c) go from minor leaf discoloration to complete leaf destruction.

Progression of paddy disease in three stages.

Full size image

From minor lesions to significant leaf loss and plant stress, Brown Spot (Row 2) develops. Grain sterility and total panicle drying are the severe symptoms of Bacterial Panicle Blight (Row 3), which begins with minor discoloration and progresses to a major illness. Beginning with modest white spikes, White Stem Borer (Row 4) causes damage to the larvae and, in more severe cases, slowed growth, yield loss, and large white spikes.

Table 4 examines the parameters, complexity, strengths, and shortcomings of six deep learning models: Inception-ResNet v2, VGG19, ResNet101, Inception v3, Xception, and SSDHR Classifier. Despite its speed, Inception-ResNet v2 (164 layers) is prone to overfitting. Although VGG19 (19 layers) has a significant computational cost, it obtains good accuracy. Although ResNet101 (101 layers) is resource-intensive (163.2 MB), it is generalisable. Though it runs the danger of overfitting, Inception v3 (48 layers) effectively retrieves characteristics. With depth-wise separable convolutions, Xception (312 layers) improves efficiency but maintains complexity. Although SSDHR (18 layers) is precise and lightweight, its feature extraction depth is constrained. The table illustrates the trade-offs between generalisability, depth, accuracy, and efficiency.

Plant disease detection with different machine learning models is the main topic of Table 5. According to Turkoglu et al., LSTM performed better than SVM in detecting apple illness (98.20%). With an accuracy of 95.31%, Patil & Kumar’s CNN-LSTM hybrid (Paddy-Fusion) outperformed unimodal models. Severity was successfully categorized by CNN-LSTM models developed by Lamba et al. (92%) and Kukreja et al. (94.06%). CNN-LSTM’s effectiveness in treating paddy illnesses was validated by Kaur et al. (94.84%). The accuracy of Choubey & Dubey’s CNN + ABi-LSTM with feature selection was 98.86%. Jiang et al. showed that CNN was more than 95% successful in identifying paddy illnesses. Outperforming all models, the suggested SSDHR classifier had the greatest accuracy (99.25%).

The comparative analysis presented in the Table 6 highlights the performance of various state-of-the-art models for paddy disease classification. Among the evaluated models, the proposed model achieved the highest accuracy of 99.25%, along with precision, recall, and F1-score all at 99.2%, demonstrating its robustness and reliability in early disease detection. Models such as CNN + ABi-LSTM + SVM-RFE + ARO and Inception-ResNet v2 also exhibited strong performance, with accuracies of 98.86% and 98.2% respectively. Traditional CNN and LSTM hybrid models like those proposed by Patil & Kumar and Kukreja et al. showed relatively lower accuracies of 95.31% and 94.06%, respectively. Other deep learning architectures such as Xception and ResNet101 performed well, with accuracies of 97.2% and 97.7%, but still lagged behind the proposed framework. The consistent outperforming of the proposed model across all metrics indicates its superior ability to extract and utilize both spatial and temporal features for more accurate and early disease identification, making it highly suitable for real-world agricultural applications.

1. Candelaria, N. P. The International Rice Research Institute (IRRI) as an International Agricultural Research Center (IARC): From the Cold War to One CGIAR.

2. Arvas, Y. E., Golden Rice Project and its impact on global nutritional security. In Crop Biofortification: Biotechnological Approaches for Achieving Nutritional Security Under Changing Climate 13–32 (2025).

3. Pradhan, S. K., Das, S. R. & Patra, B. C. Advances in paddy breeding: Stress tolerance (2021).

4. The Hindu (2019). Blast disease affects paddy crop in Tamil Nadu. Available at: https://www.thehindu.com/news/national/tamil-nadu/blast-disease-in-paddy/article29734353.ece (Accessed: 28 July 2025).

5. Savary, S. et al. The global burden of pathogens and pests on major food crops. Nat. Ecol. Evolut. 3(3), 430–439 (2019).

   Article  Google Scholar 

6. Emerich, S. A. Food security through disaster planning in Tanzania: how agriculture-based disaster preparedness can improve food security in Tanzania (2022).

7. Kulshrestha, M. & Mittal, A. K. Diseases associated with poor water and sanitation: Hazards, prevention, and solutions. Rev. Environ. Health 18(1), 33–50 (2003).

   Article  PubMed  Google Scholar 

8. Ojha, R. K. & Jha, S. K. Role of mineral nutrition in the management of plant diseases. In Farmers’ Prosperity through Improved Agricultural Technologies (eds Singh, H. K. et al.) 241–261 (Jaya Publishing House, 2021).

   Google Scholar 

9. Martinez-Feria, R. A. & Basso, B. Unstable crop yields reveal opportunities for site-specific adaptations to climate variability. Sci. Rep. 10(1), 2885 (2020).

   Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

10. Shamim, M. et al. Economical and environmental impact of paddy fungal diseases on global food security. In Fungal Diseases of Paddy and Their Management, 1–30 (Apple Academic Press, 2024).

11. El-Shahat, D., Tolba, A. & Kumar, A. K. An improved deep learning model for detecting paddy diseases. Optim. Agric. 1, 1–10 (2024).

    Article  Google Scholar 

12. Jesie, R. S., Premi, M. G. & Jarin, T. Comparative analysis of paddyleaf diseases sensing with a hybrid convolutional neural network model. Measur. Sens. 31, 100966 (2024).

    Article  Google Scholar 

13. Turkoglu, M., Hanbay, D. & Sengur, A. Multi-model LSTM-based convolutional neural networks for detection of apple diseases and pests. J. Ambient. Intell. Humaniz. Comput. 13(7), 3335–3345 (2022).

    Article  Google Scholar 

14. Patil, R. R. & Kumar, S. Paddy-fusion: A multimodality data fusion framework for paddy disease diagnosis. IEEE Access 10, 5207–5222 (2022).

    Article  Google Scholar 

15. Lamba, S., Baliyan, A., Kukreja, V. & Tripathy, R. An ensemble (CNN-LSTM) model for severity detection of bacterial blight paddy disease. In Mobile Radio Communications and 5G Networks: Proceedings of Third MRCN 2022, Singapore: Springer Nature Singapore 159–171 (2023).

16. Kaur, A. et al. Paddy sheath rot disease detection and severity classification: A novel framework leveraging CNN-LSTM models for multi-classification. In 2024 IEEE International Conference on Interdisciplinary Approaches in Technology and Management for Social Innovation (IATMSI), 1–5 Vol. 2 (2024).

17. Jiang, F., Lu, Y., Chen, Y., Cai, D. & Li, G. Image recognition of four paddy leaf diseases based on deep learning and support vector machine. Comput. Electron. Agric. 179, 105824 (2020).

    Article  Google Scholar 

18. Kukreja, V., Sharma, V. & Sharma, R. Automated paddy panicle blast disease severity classification using hybrid CNN-LSTM model. In 2024 Fourth International Conference on Advances in Electrical, Computing, Communication and Sustainable Technologies (ICAECT) 1–6 (2024).

19. Dubey, R. K. & Choubey, D. K. An efficient adaptive feature selection with deep learning model-based paddyplant leaf disease classification. Multimed. Tools Appl. 83(8), 22639–22661 (2024).

    Article  Google Scholar 

20. Lamba, S. et al. A novel fine-tuned deep-learning-based multi-class classifier for severity of paddyleaf diseases. Front. Plant Sci. 14, 1234067 (2023).

    Article  PubMed  PubMed Central  Google Scholar 

21. Litjens, G. et al. ‘A survey on deep learning in medical image analysis’. Med. Image Anal. 42, 60–88 (2017).

    Article  PubMed  Google Scholar 

22. https://kisanvedika.bighaat.com/major-diseases-of-paddy/.

23. Malekloo, A., Ozer, E., AlHamaydeh, M. & Girolami, M. Machine learning and structural health monitoring overview with emerging technology and high-dimensional data source highlights. Struct. Health Monit. 21(4), 1906–1955 (2022).

    Article  Google Scholar 

24. Liao, F. et al. A hybrid CNN-LSTM model for diagnosing paddy nutrient levels at the paddy panicle initiation stage. J. Integrat. Agric. 23(2), 711–723 (2024).

    Article  CAS  Google Scholar 

25. Shafiq, M. & Gu, Z. Deep residual learning for image recognition: A survey. Appl. Sci. 12(18), 8972 (2022).

    Article  CAS  Google Scholar 

26. Ghimire, S. et al. Deep learning CNN-LSTM-MLP hybrid fusion model for feature optimizations and daily solar radiation prediction. Measurement 202, 111759 (2022).

    Article  Google Scholar 

27. Mesías-Ruiz, G. A., Pérez-Ortiz, M., Dorado, J., De Castro, A. I. & Peña, J. M. Boosting precision crop protection towards agriculture 5.0 via machine learning and emerging technologies: A contextual review. Front. Plant Sci. 14, 114 3326 (2023).

    Article  Google Scholar 

28. Petchiammal, Kiruba, B., Murugan, & Arjunan, P. Paddydoctor: A visual image dataset for automated paddydisease classification and benchmarking. In Proceedings of the 6th Joint International Conference on Data Science & Management of Data (10th ACM IKDD CODS and 28th COMAD) 203–207 (2023).

29. Dhawan, N., Kukreja, V., Sharma, R., Vats, S., & Verma, A. Deep learning based sugarcane downy mildew disease detection paddy. Using CNN-LSTM ensemble model for severity level classification. In 2023 14th International Conference on Computing Communication and Networking Technologies (ICCCNT) 1–5 (IEEE 2023).

30. Daniya, T. & Vigneshwari, S. J. T. C. J. Deep neural network for disease detection in paddy plant using the texture and deep features. Comput. J. 65(7), 1812–1825 (2022).

    Article  Google Scholar 

31. Farooqui, N. A., Mishra, A. K. & Mehra, R. Concatenated deep features with modified LSTM for enhanced crop disease classification. Int. J. Intell. Robot. Appl. 7(3), 510–534 (2023).

    Article  Google Scholar 

32. Lu, Y., Wu, X., Liu, P., Li, H. & Liu, W. Paddy disease identification method based on improved CNN-BiGRU. Artif. Intell. Agric. 9, 100–109 (2023).

    Google Scholar 

33. Kontopoulou, V. I., Panagopoulos, A. D., Kakkos, I. & Matsopoulos, G. K. A review of ARIMA vs. machine learning approaches for time series forecasting in data-driven networks. Fut. Internet 15(8), 255 (2023).

    Article  Google Scholar 

34. Zhou, et al. A hybrid interpretable deep ensemble model for rice disease identification under real field conditions. Inf. Sci. https://doi.org/10.1016/j.ins.2025.122506 (2025).

    Article  Google Scholar 

35. LeCun, Y., Bengio, Y. & Hinton, G. ‘Deep learning’. Nature 521(7553), 436 (2015).

    Article  ADS  CAS  PubMed  Google Scholar 

36. Vankdothu, R., Hameed, M. A. & Fatima, H. A brain tumor identification and classification using deep learning based on CNN-LSTM method. Comput. Electric. Eng. 101, 107960 (2022).

    Article  Google Scholar 

37. Jabbar, M. A., Kantipudi, M. P., Peng, S. L., Reaz, M. B. I. & Madureira, A. M. Machine Learning Methods for Signal (River Publishers, 2022).

    Google Scholar 

38. Pillai, R. et al. Brain tumor classification using VGG 16, ResNet50, and Inception v3 transfer learning models. In 2023 2nd International Conference for Innovation in Technology (INOCON). (IEEE, 2023).

39. Alzakari, S. A., Alhussan, A. A., Qenawy, A. S. T., & Elshewey, A. M. Early detection of Potato Disease using an enhanced convolutional neural network-long short-term memory Deep Learning Model. Potato Research 1–19 (2024).

40. Thapliyal, N., Aeri, M., Kumar, A., Kukreja, V., & Sharma, R. Combining spatial and temporal analysis: A CNN-LSTM hybrid model for maize disease classification. In: 2024 IEEE International Conference on Computing, Power and Communication Technologies (IC2PCT) Vol. 5 1529–1533. (IEEE, 2024).

Download references

Authors and Affiliations

1. Vignan’s Foundation for Science, Technology and Research, Guntur, 522213, India

   Nagamani Thanuboddi & Usha Rani Nelakuditi

Contributions

Nagamani Thanuboddi: methodology, and writing the original draft. Usha Rani. Nelakuditi: conceptualization, review, and editing.

Corresponding author

Correspondence to Nagamani Thanuboddi.

Competing interests

No, I declare that the authors have no competing interests as defined by Nature Research, or other interests that might be perceived to influence the results and/or discussion reported in this paper.

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and permissions