A hybrid deep learning model with adaptive feature fusion for automated rice leaf disease detection and classification

Abstract

Many countries greatly rely on agriculture as a means of livelihood and economic growth. Even the most industrialized countries need food, medicine, clothing, and shelter produced by crops. Rice is one of the most significant and widely grown crops worldwide. Nonetheless, the severely impacted crops in rice production are those of bacteria, fungi, and viruses, which decrease yield and quality. Manual disease detection is hectic, challenging, and, in most cases, inaccurate. Recent advances in deep learning and computer vision have demonstrated significant potential to improve the detection and classification of diseases. This study proposes a deep learning hybrid model for the automated detection and classification of rice leaf diseases. This method consists of five key stages: image preprocessing, segmentation, augmentation, multi-feature extraction via adaptive fusion, and classification. There are five rice leaf diseases to discuss and recognize: Blight, brown spot, sheath blight, tungro, and leaf blast. The first step is global contrast enhancement, which improves image quality. After that, the segmentation is performed using Otsu’s Thresholding to extract the leaf area. Then, the modified VGG16 and modified ResNet50 networks are used in parallel to extract features using a transfer-learning approach. The adaptive fusion technique combines these features to obtain a dominant, proper feature representation. Lastly, the classification is done using an adaptive fusion score technique. Experimental results show excellent performance, with class-wise Precision in the range of 95.5–100%, class-wise recall in the range of 97.4–100%, and overall test accuracy of 98.5%.

Introduction

Agriculture is often referred to as the oldest profession. Long before the advent of science, humans began cultivating crops, laying the foundation for society. As scientific knowledge advanced, it became clear that crops are living organisms that can breathe, reproduce, and are susceptible to diseases¹. The economic development of any nation heavily relies on its agricultural industry, with most people depending on farm products for their raw materials. Rice is one of the most widely cultivated crops globally, grown in nearly 100 countries. About 700 million tons of rice are produced on 158 million hectares of rice farms, with Asia the leading continent in rice production². However, the agriculture sector is facing significant challenges due to environmental changes. These changes have led to an increase in plant diseases, which in turn affect their quality and yield. Various infections affect rice crops, including brown spot, leaf blast, rice curl disease, Hispa, Blight, and tungro, which are among the most prevalent and deadly³. Symptoms of these diseases can be seen on the crop’s leaves, necks, stems, fruits and roots. The presence of crop diseases can adversely affect both productivity and the environment. It’s estimated that plant diseases contribute to 10–15% of all agricultural production losses⁴. In severe cases, farms can experience crop yield losses of up to 50%, posing a significant risk to farmers and the country’s economy⁵. With high production and consumption levels, rice crops are particularly vulnerable to these diseases, which can severely impact both their quality and quantity.

The pathogens that cause rice diseases are diverse, including bacteria, fungi, viruses, and nematodes, all of which pose severe threats to global rice production. One of the most common of these is Bacterial Leaf Blight (BLB), which is caused by Xanthomonas oryzae pv oryzae. It is characterized by water-filled lesions that spread through the leaves and eventually cause leaf drying, resulting in significant yield shrinkage^6–8. Magnaporthe oryzae, a fungus, is another significant pathogen that causes Rice Blast disease, known to be one of the most destructive because it causes lesions on leaves, stems, and panicles and may lead to a loss of up to 75% of yields if left uncontrolled. Sheath Blight is a fungus caused by Rhizoctonia solani that mainly attacks the sheaths and leaves of plants, causing long lesions, weakening plant structure, and reducing productivity. Additionally, Rice Tungro Disease is a viral disease transmitted by leafhoppers that reduces chlorophyll content, resulting in stunted growth, reduced tillering, and yield losses. Collectively, these diseases pose significant challenges to sustainable rice production and underscore the need for efficient detection and control methods to maintain worldwide food security^9–11.

Rice diseases exhibit a diverse range of symptoms depending on the pathogen and infection stage. Typical symptoms include leaf discolouration, lesions, wilting, and deformities. To provide some examples, bacterial leaf blight produces water-soaked spots on leaves, rice blast produces small circles with greyish cores, sheath blight produces long ones along the leaf sheaths, and rice tungro makes plants yellow and dwarf^12,13. Historically, rice disease diagnosis has primarily depended on visual examination by plant pathologists. Even today, in villages and remote areas, rice disease is detected by the naked eye. This manual observation and recognition of diseases is a time-consuming, biased, and inaccurate process^14–16, as it depends on the individual’s skills and experience and is thus highly prone to human error. Five kinds of common and deadly rice leaf infections: Blight, brown spot, sheath blight, tungro and leaf blast are addressed in this work. The description of these diseases is shown in Fig. 1.

Fig. 1.

Rice leaf diseases and their symptoms.

While the integration of image processing and machine learning has already brought about a tremendous advancement in detection and analysis of plant diseases¹⁷, recent advancements in deep learning, particularly Deep Convolutional Networks a specialized branch of machine learning, have led to significant improvements in image detection and classification accuracy^18,19. In the study²⁰, the authors conducted advanced research in plant disease detection using CNN architectures and achieved promising results, ranging from 60 to 100%.

CNNs are designed to automatically extract complex features from images and videos through a hierarchical learning process, inspired by the human visual cortex. Their impact spans technological advancements, from medical imaging to autonomous vehicles. Crop disease detection and classification are also not untouched by the application of CNNs. The results of plant disease detection using deep learning can be further improved by focusing only on the leaf region of the input image before applying deep learning²¹. The segmentation process helps locate the leaf portion in drone-captured images from real fields. Single CNN models may not capture all important features, such as fine details (texture, shape, and pattern) and important global contextual features (spatial hierarchy). In this study, we have presented a hybrid CNN architecture that employs an adaptive feature fusion technique to select the most relevant features from the input dataset.

The fundamental contributions of the study are discussed below:

• Suggest the Adaptive Multi-Feature Fusion-based Hybrid Network (AMFF-Hybrid Net), which integrates modified VGG16 and modified ResNet50 to extract fine-grained texture and deep contextual features simultaneously in an efficient manner.

• Adopt multi-feature fusion, which integrates the joint feature space into a single representation, thereby improving discriminative capacity and classification accuracy.

• Apply Otsu background removal to remove the backgrounds, reduce noise, and focus the learning on the areas of disease value.

• Provide improved image analysis capabilities in the hybrid model, in comparison to the single-network methods, when dealing with complex image analysis situations.

The remaining structure of papers is presented as follows: sections. The literature review in “Literature review” section provides an overview of each study and presents a research gap. “Materials and methodology’ section explains the methodology for the proposed work. The experimental setup and outcomes are discussed in “Experimental procedure and findings” section. “Conclusion and future scope” section discusses the conclusion regarding plant leaf disease detection and future scope.

Literature review

This section presents a thorough review of recent plant disease detection techniques and the state-of-the-art models used in related studies, aiming to understand their underlying mechanisms and to highlight potential limitations and weaknesses.

Due to various infections affecting the paddy crops, farmers have experienced significant losses. The pain of farmers motivated²² to propose a method for recognizing paddy disease problems using Convolutional Neural Networks. The authors used a dataset of 500 rice disease images from HALRS in China. In their study, four models were compared, with the following accuracy results: Back Propagation (BP) achieved 92%, Support Vector Machine (SVM) attained 91%, Particle Swarm Optimization (PSO) reached 88%, while the CNN model achieved the highest accuracy of 95%. Some researchers experimented with image preprocessing techniques alongside ML/DL methods to detect rice diseases. Authors in²³ proposed a deep learning method for recognizing rice sheath blight (RSF) infection. Their research involved collecting 693 leaf pictures from real crop fields. The proposed technique focused on designing a new CNN model by merging an SVM classifier with the Histogram of Oriented Gradients. This new structure was then compared with established models such as AlexNet and VGGNet-11 to assess its advantages. However, the SVM-HOG approach demonstrated an accuracy rate of 89.5%; the proposed CNN-based method aimed to improve upon that.

To improve robustness, model authors begin collecting images from real crop fields. Authors in²⁴ developed an automated plant disease diagnosis system using a CNN model on aerial imagery, highlighting its effectiveness for large-scale disease detection. The model was evaluated on real crop-field-captured maze leaf images. An accuracy of 95.1% was achieved. However, its reliance on aerial images limits its ability to detect finer disease symptoms that may be more apparent in close-up views. Authors in²⁵ proposed a DenseNet-based corn disease diagnosis system, achieving 98.06% accuracy while using fewer parameters and requiring less computation time than existing CNN frameworks. The proposed model achieved an accuracy of 98.06%. However, the model faces computational limitations, posing memory and processing constraints for installation on low-resource edge devices. Authors in²⁶ utilized Faster R-CNN and Mask R-CNN to build a paddy disease classification model. They collected 1500 images of paddy leaves with 3 types of disease from various paddy crop fields. Analysis of results revealed that Mask R-CNN performed better, achieving accuracies of 95%, 96%, and 94.5% for brown spot, blast, and sheath blight, respectively.

An ensemble of pretrained deep CNN models was proposed by²⁷ for the detection of illnesses affecting apple leaves. Authors have used an EfficientNet NoisyStudent, an EfficientNetB7, and a DenseNet121 pretrained model to perform ensemble learning. The proposed model was evaluated on a dataset comprising 3,642 apple leaf samples across 4 classes. Their method improved apple disease classification accuracy (96.25%), but when several illnesses were present, its identification performance deteriorated to 90%. In the same direction, authors in²⁸ addressed the issue of early detection of plant diseases to facilitate food security and sustainable agricultural objectives. They suggested two deep transfer learning models, PDDNet-AE and PDDNet-LVE, that use an ensemble of nine pre-trained CNNs for more robust feature extraction. The experiments were conducted on the PlantVillage data set (15 classes and 54,305 images), achieving high accuracies of 96.74 (AE) and 97.79 (LVE). The proposed model achieved better generalization, lower overfitting, higher accuracy, and earlier disease detection without expert involvement. The weakness of the approach is the requirement for large, labelled datasets for training and the limited validation using images in real-world, uncontrolled environments.

Many researchers used transfer learning to enable fast learning without sacrificing accuracy. Authors in²⁹ proposed a CNN-based automated rice disease detection technique. The dataset consisted of 984 images sourced from various online platforms, including UCI and Kaggle. The proposed system used transfer learning for disease detection and classification. They evaluated four pre-trained models: Inception-ResNet-V2, which achieved 92.68% accuracy; ResNet-101, which achieved 91.52% accuracy; VGG-19, which achieved 81.43% classification accuracy; and Xception, which achieved 89.42% classification accuracy. Further, the authors in³⁰ presented a transfer-learning-based VGGNet architecture to recognize paddy diseases. They created their own dataset of 1509 paddy leaf images, of which 647 were used for testing, achieving 92.46% accuracy. Transfer Learning with predefined parameters for fine-tuning the model’s efficiency has been enhanced by the VGG Net, which, on the other hand, cannot achieve high accuracy with a limited dataset. With an enhanced dataset, authors in³¹ also implemented a transfer learning strategy using InceptionResNetV2 to detect rice leaf diseases. 5200 images were sourced from the Kaggle platform to train and validate the suggested model. The proposed model’s classification features have been improved, and it now achieves 95.67% accuracy.

In another study, researchers used generative data augmentation to address data scarcity and introduce artificial noise, improving the model’s robustness. Authors in³² presented a DC-GAN-MDFC-ResNet architecture that enabled detecting crop diseases at the initial stage by applying generative data augmentation and compensating for residual features. It focused on a finer classification of bacterial leaf blight, leaf streak, and panicle blight using the PlantVillage data. The model achieved 95.99% accuracy and produced higher-quality synthetics than current deep learning models. Pros of this approach include improved feature acquisition and detection as the number of features increases, and reduced data requirements; however, it requires greater real-time deployment overhead. Similarly, authors in³³ presented a Rice DenseNet model that outperformed baseline classification architectures when tested on data generated by a GAN, achieving 98.74% accuracy. They proposed three types of datasets and compared the baseline transfer learning models: Mobile Net, ResNet-50, InceptionV3, ResNet101 and Xception, with the proposed Rice DenseNet model.

Researchers proposed other CNN-based models to boost accuracy. Author³⁴ carried out real-time plant disease diagnosis work by applying CNN-based frameworks to overcome the limitations of traditional methods, providing quick, precise diagnosis with 98.32% accuracy, 97.99% F1-score, and 42.6 ms/image inference time. The system has low tolerance to illumination irregularities and partial blockages; hence, it is not very robust in the field environment. Authors in³⁵ suggested a deep learning automated system to be used to detect paddy leaf disease to improve agricultural output. Images are preprocessed and segmented, and features are extracted using a pre-trained MobileNetV3 model, then optimized using a hybrid GKSO Simulated Annealing algorithm. The CatBoost classifier is used to perform disease classification. Experimental evidence indicates an accuracy of 98.52%, which is better than the current state-of-the-art strategies.

Researchers propose a multi-scale fusion-based approach to provide a Twin CNN framework³⁶. This method enhances the rice leaf disease by multi-scale fusion of VGG16, ResNet50, and InceptionV3 features, using PCA dimensionality reduction, and attains 96.8% accuracy. Although the Twin CNN framework demonstrates good results, its most significant shortcomings are the relatively small dataset and high computational costs, which prevent its application online or in real time on resource-limited edge devices.

Even though current models of plant disease detection using deep learning have shown promising results, a significant number of them have just one backbone network, which restricts them in the simultaneous acquisition of fine-grained texture features and deep contextual semantics. Although some related works^27,28,36 proposed a multi-feature fusion approach, they fail to address model efficiency. The existing methods tend to be less adaptive in multi-feature fusion; hence, they are less optimal in learning representations in complex background leaf images. In addition, a lack of focus on efficient background removal leads to the models learning irrelevant aspects and reducing their robustness in the real field of use. The classical techniques of image preprocessing, such as Otsu-based segmentation, when fused with deep hybrid architectures, have yet to be fully explored to improve disease-specific feature learning and disease diagnosis efficiency with high accuracy as shown in Table 1.

Table 1.

Related work for plant disease diagnosis.

Ref	Objective	Approach	Key points	Plant Species	Accuracy	Limitation
24	Automated maize disease diagnosis	CNN architecture	Highly effective at detecting diseases at scale using aerial imagery	Maize	95.1%	Restricted to aerial images, which may overlook subtle infection symptoms
25	Corn Disease diagnosis	Optimized- DenseNet	Using fewer parameters and lower computation time compared to existing CNN architectures	Corn plants	98.06%	-Still Computational Constraints - memory and processing challenges for low-resource edge devices
27	Detection of infections affecting apple leaves	Ensemble learning of a pretrained CNN architecture	Enhanced performance with an easy-to-use Web application	Apple plants	96.25%	The model’s performance may be sensitive to dataset quality and diversity
30	Paddy disease diagnosis	Transfer learning with the VGG Net pretrained model	The authors created their own dataset consisting of 1509 paddy leaves	Rice	92.46%	It cannot produce high accuracy in a limited set of data
31	Rice disease diagnosis	Transfer learning with pretrained models	Promising results with an enhanced dataset by InceptionV3	Rice	95%	It cannot produce high accuracy even with an enhanced dataset
32	Robust model for Rice disease diagnosis	A hybrid residual architecture with GAN	Robust with noise due to generative data augmentation	Rice	95.99%	More overhead to deploy in real-time
33	Robust model for rice disease diagnosis	Rice dense net with transfer learning on GAN-generated data	Robust with noise due to generative data augmentation	Rice	98.74%	Real field robustness is not validated
34	Real-time disease diagnosis	CNN-based frameworks	Analyzed the efficiency (42.6 ms/image inference)	Multi crops	98.32%	Poor tolerance to changes in illumination and partial occlusions
35	Automated paddy disease detection	Features extraction with MobileNetV3 classification using CatBoost	Hybrid optimization improves feature quality	Paddy	98.52	The authors did not not analyzed efficiency
36	Rice disease detection	Twin CNN with multi-feature fusion	Multi-scale feature fusion	Rice	96.8%	High computational costs

Materials and methodology

In this section, we discuss the materials and methods used for this study. We describe the data set, data preprocessing, data augmentation, data segmentation, baseline models, and the proposed Hybrid model. Figure 2 shows a basic diagram of the proposed approach. We have arranged the details of all concepts and techniques in the following subsections.

Fig. 2.

Proposed rice disease detection system.

Roadmap

The basic steps required for model design are discussed as follows:

1. The rice plant leaf disease images are sourced from the Kaggle platform.

2. After the collection of datasets, the next step is image resizing and contrast stretching (image preprocessing) to get a quality image.

3. After preprocessing, we have applied image augmentation techniques like translation, scaling and rotation.

4. Next, segmentation is applied for noisy data handling with a background removal algorithm.

5. The augmented images were then fed into baseline models (VGG16 and Resnet 50) separately for training and validation to observe classification results. To further examine the effect of segmentation on classification results, segmented images are also fed into the baseline models. We have recorded disease recognition and classification results in both cases.

6. Finally, we have fed segmented images into the proposed Adaptive multi-feature fusion-based hybrid deep architecture for training and validation.

7. The significances of segmentation and Adaptive multi-feature fusion-based hybrid deep model for rice disease recognition and classification were illustrated by comparing the performance of the proposed model with baseline and existing models.

Dataset collection and description

The rice leaf dataset in this study comprises 5 types of rice leaf images. These 5 leaf image types represent 5 leaf disease classes: Blast infection, Brown-Spot (BS), Bacterial Blight (BB), Tungro (Tun), and Sheath Blight (SB). It is collected from the online open repository Kaggle³⁷. The use of images in the training and test experiments is shown in Table 2.

Table 2.

Dataset description.

Classes	Total samples	Test samples	Training samples
Leaf Blast (LF)	510	153	357
Brown Spot (BS)	510	153	357
Bacterial Blight (BB)	510	153	357
Tungro (Tun)	510	153	357
Sheath Blight (SB)	510	153	357

Training and testing samples are preprocessed and segmented to obtain high-quality images with the background removed. Preprocessed training images are then used to create augmented samples by applying on-the- fly image augmentation techniques during model training at runtime, ensuring that each epoch receives a new set of augmented samples in the training process. Thus, the model is trained with augmented samples. Test samples serve as unseen samples for the model and are kept separate for model testing prior to the augmentation process to avoid the data leakage. Details of the augmentation process and augmented data generation is discussed in the coming section 'Diversifying Data with Image Augmentation.' Class-wise augmented sample counts during the entire training process (6 epochs) are shown in Table 3.

Table 3.

Augmented sample counts.

Classes	Training samples before augmentation	Training samples after augmentation
Leaf Blast (LF)	357	2,142
Brown Spot (BS)	357	2,142
Bacterial Blight (BB)	357	2,142
Tungro (Tun)	357	2,142
Sheath Blight (SB)	357	2,142

Image preprocessing to enhance image quality

The source image is resized to a 224′224′3-pixel square shape while maintaining a consistent aspect ratio, as most CNN models require a square-shaped input image. Next, global contrast stretching is applied to the resized image to improve its visual quality. The contrast of a digital image may vary due to incorrect instrument calibration or poor lighting conditions. Size-reduced image samples are contrast-enhanced to improve their visual appearance. Contrast enhancement techniques expand the image’s brightness range, thereby improving its visual quality. This process typically involves converting RGB images into a format that includes luminosity information. Consequently, the resized images are converted to the Lab* colour space, where the L* component (luminosity) is adjusted to enhance contrast. After modification, the image is converted back to the RGB format. Adjusting the luminance alters pixel intensity while preserving the original colour. The global contrast enhancement of the L* component is computed using Eq. 1.

1

where is normalized luminance, is original luminance, is maximum luminance, and is minimum luminance.

Segmentation for noisy data handling with background removal

This part describes the steps involved in the background noise removal procedure. This preprocessing is required to get the infected leaf portion in the input image. The background removal algorithm is based on Otsu’s global thresholding technique and described in Algorithm 1. Algorithm 1 is applied to all input images. Otsu’s global thresholding technique is specified in Algorithm 2.

Algorithm 1.

Background Removal.

Algorithm 2.

Otsu’s global thresholding.

Although, several advanced-level segmentation algorithms have been discussed in recent literature to analyze plant diseases, such as clustering-based algorithms³⁸ (e.g., K-means and fuzzy C-means), graph-based algorithms (e.g., graph-cut), and deep learning-based algorithms³⁹ (e.g., U-Net and Mask R-CNN). They are useful for handling complex backgrounds and irregular object boundaries, but they often require more computational resources and additional annotated data when used with deep learning methods. In contrast, Otsu’s thresholding is a basic yet efficient global thresholding method that automatically sets the optimal threshold, which depends on the characteristics of the histogram of the images. In the dataset used in this study, leaf areas have sufficient contrast with the background, and so Otsu’s method is appropriate for accurately extracting the foreground. Hence, in our study, Otsu-based segmentation provides the best trade-off in segmentation quality and computational cost.

Diversifying data with image augmentation

Several on-the-fly image augmentation techniques are used to generate diverse images in this research. Augmentation is a set of image transformation techniques that create slightly geometrically altered samples. Image augmentation helps to generate more generic samples. It helps the model learn more complex features from the data and reduces overfitting. The following are the three fundamental changes made to the input dataset.

• i.Translation: Random horizontal and vertical translations are applied to shift the image horizontally and vertically by a random amount in the range of pixel distances of -25 and 25, respectively.
• ii.Scaling: Random horizontal and random vertical scaling are applied by a factor of 0.9 and 1.1, respectively.
• iii.Rotation: Random rotation is applied to rotate the images by a random degree in a specified angle range of − 20◦ to + 20◦.

On-the-fly augmentation is applied using MATLAB’s augmentedImageDatastore; it is applied at each epoch during training, so the model sees new augmented versions of the data with 357 samples per class. A total of 10,710 augmented samples are created and used during the 6-epoch training process. The augmentation process aids in generalization of the model. The number of augmented samples generated per class, presented epoch-wise, is shown in Table 4.

Table 4.

Epoch-wise augmented samples generation.

Epoch	Blast	Brown Spot	Blight	Sheath Blight	Tungro	Total Augmented Samples (Cumulative)
1	357	357	357	357	357	1,785
2	357	357	357	357	357	3,570
3	357	357	357	357	357	5,355
4	357	357	357	357	357	7,140
5	357	357	357	357	357	8,925
6	357	357	357	357	357	10,710

Baseline deep networks

We have utilized two pretrained deep CNNs, namely VGG16 and ResNet50, as baseline models for our study. In this section, we first describe basic CNN architectures and then discuss the structures of the pretrained deep architectures used as baseline deep networks with transfer learning (complete parameter training) for our research.

CNN

A subclass of neural networks, known as CNNs, is designed to mimic the human visual cortex. Basic components like convolutional layers apply filters to input data to identify patterns, and pooling layers help maximize computational efficiency while preserving consistency across spatial scales. Fully connected layers then combine the detected characteristics to address challenges involving regression or classification. CNNs have revolutionized computer vision applications such as object detection, facial recognition, and picture classification. The key components of CNN are discussed as follows:

• i)Convolution Operation by Convolution Layer
  A convolutional layer is a key component of a CNN that performs convolutional operations to extract relevant features. Convolution is the core operation in CNNs, used to obtain feature maps from given images. The convolution operation can be represented by Eq. 2:

  2

  Where : (x, y) position output in the filter, : The input image patch, : filter weight, bk ​: filter bias, and M,N: Size of the kernel (e.g., 3 × 3 or 5 × 5).

• ii)Activation Function
  Non-linearity is introduced into the network using the ReLU activation (Eq. 3).

  3
• iii)Pooling
  Pooling is utilized either by max pooling (Eq. 4) or average pooling (Eq. 5) for spatial dimensions reduction of feature maps by keeping significant features.

  4

  5
• iv)Fully Connected Layers
  The FC layer accepts the outputs of the previous layers. It transforms the same into an n (number of disease classes) feature vector using Eq. 6, which is then used as input to the final softmax layer for classification. Class forecasting is usually done using the Softmax function on the outputs of the previous layers.

  6

  where: ​: neuron’s Outcome, ​: Weight connecting the input to the neuron, : neuron’ bias, and : Input to the neuron.
• v)Loss Functions
  Loss functions quantify the error between predictions and true labels. Generally, Categorical Cross-Entropy is used as loss function (Eq. 7) for multi-class classification.

  7

  where: C: Number of classes,​: True label (1 if the sample belongs to class i, otherwise 0), ​: Predicted probability for class i.

VGG16 Architecture

The VGG16 architecture consists of sequential convolution layers, pooling (max-pooling), and fully connected layers.

• i)Convolutional Layers:13 convolutional layers with 3 × 3filters.Each convolutional layer uses ReLU activation. Filters are doubled as the depth increases (e.g., 64 → 128 → 256 → 512).
• ii)Pooling Layers:2 × 2 max-pooling operations cause spatial dimensionality reduction by half.
• iii)Fully Connected Layers:2 dense layers with 4096 neurons each, followed by ReLU. Final dense layer with softmax for classification.

ResNet50 architecture

ResNet50 proposes and utilizes residual connections to address the diminishing gradient problem, allowing a deeper network architecture to train effectively. Here is the breakdown:

• i)Convolutional Layers: The first convolution layer uses a 7*7 filter, followed by ReLU, batch norm and a 3*3 max-pooling layer. The first convolutional layer is followed by four stages of residual blocks, each consisting of several bottleneck layers. Bottleneck layers use 1 × 1,3 × 3, and 1 × 1 convolutions.
• ii)
  Residual Connections: Residual connections in deep networks allow learning residual mappings via the introduction of shortcut connections to enhance gradient flow, training stability, and performance. The residual block output is computed using Eq. 8:

  8

  where is the learned residual mapping, and x is the input.
• iii)Global Average Pooling: The feature maps are reduced to a single vector using global average pooling.

Baseline deep networks with Otsu’s background removal

We have used two pretrained deep CNN models, VGG16 and ResNet50, as baseline networks for our work. Otsu’s method is applied to the noisy input dataset to segment the leaf portion of the image by removing the background. Background removed segmented image is given to each baseline deep learning model separately to classify the diseases. Integrating Otsu’s method with baseline deep networks, VGG16 and ResNet50, improves the feature extraction by removing the noisy background. Otsu’s method, a global thresholding technique, automatically determines an optimal threshold to separate the foreground (e.g., plant leaves) from the background, improving model robustness against environmental variations. By applying Otsu’s background removal before feeding images into deep learning models, we ensure the network focuses solely on essential disease-related patterns rather than distractions such as soil, sky, or other plants. This preprocessing step significantly reduces computational complexity, accelerates training, and improves classification accuracy.

Proposed hybrid deep network with adaptive multi-feature fusion and Otsu’s background removal

The proposed hybrid model combines feature maps from modified VGG16 and modified ResNet50. This combination is beneficial because the modified VGG16 (VGG16 path) focuses on shallow, fine-grained texture features, while the modified ResNet50 (ResNet50 path) captures hierarchical, deep, and contextual information.

It is important to note that the suggested hybrid model does not use the full architecture of VGG16 and ResNet50. Rather, truncated and altered versions of both networks are used to reduce computational cost. In the case of modified VGG16, only 6 out of the original 13 convolutional layers are retained. In particular, the architecture has changed to consist of 1 convolutional layer with 64 filters in Block 1, 1 with 128 filters in Block 2, 2 with 256 filters in Block 3, and 2 with 512 filters in Block 4 and Block 5 is completely removed. Also, the three fully connected layers, which account for a large share of the total parameters, are removed and replaced with a single Dense layer comprising 512 neurons with ReLU activation. The architectural details of the modified VGG16 are shown in Table 5, under the column ‘VGG16 path’. On the same note, in ResNet50, only a few residual blocks are retained, and the last classification layers are excluded. The architectural details of the modified ResNet50 are shown in Table 5, under the ‘ResNet50 path’ column. Moreover, compact feature representations are produced by means of a global average pooling layer. These architectural changes reduce the number of parameters to be trained whilst maintaining the strong feature extraction capacity.

Table 5.

Proposed model’s architecture.

Layer	VGG16 Path	ResNet50 Path	Hybrid Model (Fusion Path)
Input	Input: 224 × 224 × 3	Input: 224 × 224 × 3	Two inputs, one for each model
Convolutional Layer 1	Conv (3 × 3, 64 filters, ReLU)	Conv (7 × 7, 64 filters, stride 2, ReLU)	–
MaxPooling Layer	MaxPool (2 × 2, stride 2)	MaxPool (3 × 3, stride 2)	–
Convolutional Layer 2	Conv (3 × 3, 128 filters, ReLU)	Residual Block (Conv 1 × 1 → Conv 3 × 3 → Conv 1 × 1)	–
MaxPooling Layer	MaxPool (2 × 2, stride 2)	MaxPool (3 × 3, stride 2)	–
Convolutional Layer 3	Conv (3 × 3, 256 filters, ReLU)	Residual Block (Conv 1 × 1 → Conv 3 × 3 → Conv 1 × 1)	–
Convolutional Layer 4	Conv (3 × 3, 256 filters, ReLU)	Residual Block (Conv 1 × 1 → Conv 3 × 3 → Conv 1 × 1)	–
MaxPooling Layer	MaxPool (2 × 2, stride 2)	MaxPool (3 × 3, stride 2)	–
Convolutional Layer 5	Conv (3 × 3, 512 filters, ReLU)	Residual Block (Conv 1 × 1 → Conv 3 × 3 → Conv 1 × 1)	–
Convolutional Layer 6	Conv (3 × 3, 512 filters, ReLU)	Residual Block (Conv 1 × 1 → Conv 3 × 3 → Conv 1 × 1)	–
MaxPooling Layer	MaxPool (2 × 2, stride 2)	Global Average Pooling (Flatten output)	–
Flatten Layer	Flatten (14 × 14 × 512 = 100,352 features)	Output → 1024 features	–
Dense Projection	Dense (512 neurons, ReLU)	–	–
Attention Fusion	–	–	VGG Attention: Sigmoid applied to 512 features → element-wise multiplication ResNet Attention: Sigmoid applied to 1024 features → element-wise multiplication
Fusion Output			Concatenate weighted features (512 + 1024 = 1536)
Dense Layer 1	–	–	Dense (512 neurons, ReLU)
Dropout Layer	–	–	Dropout (Rate = 0.5)
Output Layer	–	–	Dense (5, Softmax activation)

The proposed hybrid approach is an advanced approach designed to improve the results of image analysis and computer vision tasks. This method integrates multiple feature spaces, such as fine-grained texture features and deep contextual information, into a unified deep network, enabling the model to extract richer, more discriminative information from the input data. By leveraging adaptive multi-feature fusion, the network can effectively integrate complementary information from different feature spaces, improving its robustness and accuracy in complex scenarios. Additionally, the Otsu method is applied to preprocess the input images, automatically separating the leaf portion from the images using intensity thresholding. This step minimizes noise and irrelevant data, allowing the model to focus on the most salient regions of interest. The architecture of the proposed model is shown in Table 5. Suggested, Adaptive multi-feature fusion-based hybrid network (AMFF-Hybrid Net) with background removal is not a heavyweight model. The trainable parameters in the proposed model are shown in Table 4. Below is the architecture of the combined model:

Table 6 indicates that the total number of parameters of the proposed model (57.85 million) is much lower than that of the standalone VGG16 architecture (138 million). This reduction is mainly due to the elimination of the 3 fully connected layers of VGG16, which account for a large percentage of its parameters in the proposed design. Further, truncated backbone networks and global average pooling also reduce parameter growth. Consequently, even though the two feature-extraction branches are combined in the model, the model remains computationally efficient.

Table 6.

Trainable parameters in the proposed model.

Proposed Model components	Architecture basis	Trainable parameters
Modified VGG16 (VGG16 path)	6 conv blocks (3 × 3, 64 → 512) + Dense (512 neurons, ReLU)	55,881,408
Modified ResNet50 (ResNet 50 path)	7 × 7 conv + 5 residual blocks + GAP	1,818,304
Attention	Element-wise sigmoid	0
Fusion Dense head	concat features → Dense (512) → Dense (5)	789,893
Proposed AMFF-Hybrid Net + Otsu	VGG16 path + ResNet path + fusion Dense head	57,851,301≈ 57.85 M

Adaptive multi-feature fusion strategy

The proposed hybrid model uses an adaptive multi-feature fusion approach to integrate complementary features from the VGG16 and ResNet50 paths effectively. In contrast to traditional fusion approaches that use simple static concatenation, the proposed approach enables the network to learn feature importance during training via an attention mechanism, thereby improving its discriminative power and overall classification performance.

In particular, the feature vector at the VGG16 path (512 dimensions) is passed through a sigmoid activation function to obtain attention weights, which are then multiplied element-wise with the feature vector to highlight informative and suppress redundant features. In the same manner, the feature vector of ResNet50 (1024 dimensions) undergoes the same procedure to generate weighted deep contextual features.

Let be the feature vector extracted from the final dense projection of the VGG16 path, and be the global average-pooled feature vector from the ResNet50 path. The attention-weight vector of the VGG16 path ( is computed using Eq. (9).

9

Adaptive feature of VGG16 path ( is computed by Eq. (10) using element-wise multiplication of the attention vector and feature vector.

10

Similarly, the attention weights vector of the ResNet50 path ( is computed by using Eq. (11), and the adaptive feature of the VGG16 path ( is computed by Eq. (12).

11

12

Finally, adaptive vectors are concatenated to obtain the fusion vector () as in Eq. (13).

13

Proposed model’s novelty

Even though hybrid architectures based on transfer learning have been discussed in earlier studies, the proposed framework offers several important architectural and methodological improvements that set it apart from traditional models.

First, instead of integrating complete pre-trained models, truncated and modified versions of VGG16 and ResNet50 are used, in which computationally expensive blocks such as fully connected layers and more fully formed residual steps are eliminated. This truncation yields an efficient architecture with fewer complex parameters.

Second, the proposed model incorporates an adaptive multi-feature fusion technique, in which the concatenated feature representations are fed through a trainable fusion layer. This mechanism dynamically weights feature contributions, unlike standard methods that use fixed concatenation, enabling the model to focus on the most discriminative features of each backbone network.

Third, an Otsu threshold is applied to segment the leaf portion from the noisy background. This preprocessing is more useful for improving feature quality before classification, something that is not usually considered in existing hybrid structures.

Lastly, the proposed architecture aims to achieve the desired trade-off between performance and computational efficiency by a lightweight backbone change and an adaptive fusion strategy.

All this contributes to the proposed model being unique among existing methods of transfer learning-based hybrid methods and helps it to be better at its classification tasks.

Experimental procedure and findings

This section describes the experimental setup and the research implementation, using a baseline and a proposed hybrid network architecture to develop a practical algorithm for paddy leaf disease recognition and identification. Findings of experiments, including the outcomes of each experiment and an evaluation of these results, are briefly described in this section. Experimental findings are presented in a variety of figures and charts.

Experimental framework

The experimental framework in deep learning refers to the hardware, software, libraries, and configurations used to train, validate, and test deep learning models. All experiments are conducted on a well-equipped Dell i7 laptop, with specifications listed in Table 7.

Table 7.

Experimental environment.

Category	Specification
Hardware
System	Dell Laptop
CPU	Core i7 processor@ 3.60 GHz
GPU	NVIDIA RTX 3080
RAM	16 GB DDRAM
STORAGE	1 TB with 512 GB SSD
Software
Implementation Language	MATLAB 2019a
Library	Tensor flow
Operating System	Window 10
Dataset
Source	Kaggle
Dataset Name	Rice Leaf Disease
Samples Count	2,550
Training- Testing Ratio	70:30

Model performance metrics

Assessment of a deep network architecture is done by various performance metrics depending on the task, such as classification, regression, or natural language processing (NLP). The proposed system is based on a classification task and assessed by the classification evaluation metrics described in Table 8. Model evaluation Terms are described in Table 9.

Table 8.

Model evaluation Metrics.

Metric	Description	Formula
Accuracy (A)	The fraction of accurately identified samples out of the whole samples	A = (TN + TP)/( TN + TP + FN + FP)
Precision (P)	The fraction of true positives of a class out of the total predicted positives for that class	P = TP/(TP + FP)
Recall (R)	The proportion of true positives of a class to the whole actual positives for that class	R = TP/(FN + TP)
F1-Score (F)	Harmonic-mean of Recall and Precision	F = 2⋅ R⋅ P/( R + P)
Overall Accuracy	Average of all class’s true positives	Overall Accuracy  = Sum of True Positives of All Classes/ Number of all tested samples

Table 9.

Model evaluation Terms description.

Term	Definition
True_Positive (TP)	Samples accurately identified as positive (Expected)
False_Positive (FP)	Samples inaccurately identified as positive (Unexpected)
True_Negative (TN)	Samples accurately identified as negative (Expected)
False_Negative (FN)	Samples inaccurately identified as negative (Unexpected)

Model training and hyperparameter setup

Five diseased leaf classes- Blast infection, Brown-Spot (BS), Bacterial Blight (BB), Tungro, and Sheath Blight are available in the dataset. The training and testing datasets are separated into a 70:30 ratio from the dataset. Images from the entire 2,550-sample dataset are partitioned into two segments. The training dataset contains 1,750 images, and the testing dataset contains 765 images. Two baseline models (VGG16 and ResNet50) and one proposed hybrid network are trained and tested using the same hyperparameters specified in Table 10.

Table 10.

Models hyperparameters.

Category	Specification
Batch size	30
Learning Rate	0.0001
Optimizer	Stochastic Gradient Descent with Momentum
Training Epoch	6
Iteration per epoch	49
Activation Function	ReLu
Validation frequency	10
Loss Function	Categorical Cross-Entropy

Results discussion

We have done three experiments: the first two for baseline models and the third for the proposed hybrid model. In the first experiment, Rice disease prediction was performed using the VGG16 baseline model to observe its behaviour on the given dataset. In the second experiment, rice disease classification was performed using the ResNet50 baseline model to assess its nature and classification performance on the given dataset. In the third experiment, a complete discussion of the results of the proposed hybrid framework was conducted. Classification results are presented as a confusion matrix. True positive rates (Recall) of each disease are shown in the bottom row of the confusion matrix, whereas the Precision of each class is shown in the last column of the confusion matrix. The model’s overall accuracy is shown in the bottom-right cell.

Baseline deep network model

The first experiment is implemented with 2 baseline deep networks: VGG16 and ResNet50. These two pretrained models are fed the preprocessed rice leaf dataset separately. Then, fine-tuning is applied to the baseline networks, and finally, the best hyperparameter combination is selected based on improved classification results. The selected hyperparameters are specified in Table 10. The learning process of the baseline network, VGG16, is shown in Fig. 3a, whereas rice disease classification results using the VGG16 model are shown in Fig. 3b. Baseline Vgg16 achieved validation accuracy of 86.10%. Similarly, the learning phase of the baseline network, ResNet50, is shown in Fig. 4a, whereas the classification results of the second baseline network are shown in Fig. 4b. The ResNet50 model achieved a validation accuracy of 89.50%.

Fig. 3.

Simulation results of baseline VGG16 (a) Training and validation graph of baseline VGG16 (b) Classification result of baseline VGG16.

Fig. 4.

Simulation results of baseline ResNet50 (a) Training and validation graph of baseline ResNet50 (b) Classification result of baseline ResNet50.

ResNet50 shows slightly better training and validation performance because it is deeper than VGG16 and also uses residual connections during training. The accuracy of ResNet is slightly higher than that of VGG16  as shown in Fig. 5.

Fig. 5.

Background removal process for tungro infected sample (a) RGB to Gray conversion of tungro infected leaf (b) Gray to binary Mask by Otsu’s method (c) Binary Mask to target image with removed background.

Baseline deep networks with Otsu’s background removal

Size-reduced and contrast-stretched image samples undergo Otsu’s background-removal process to isolate the leaf portion with diseased lesions. To assess the significance of segmentation and background-removal techniques on the input data, the second experiment is implemented by feeding the background-removed input dataset to the same baseline deep networks: VGG16 and ResNet50.In the results, we observed that this method performs well on the given input data and improves the performance of baseline networks. For demonstration purposes, the background removal result of a tungro-infected leaf is shown in Fig. 5. Similarly, the segmentation results for Blast, Brown spot, Blight, and Sheath blight are shown in Figs. 6, 7, 8, and 9, respectively.

Fig. 6.

Background removal process for Blast infected sample (a) Blast infected sample (b) Grayscale image (c) Background removed sample.

Fig. 7.

Background removal process for Brown spot infected sample (a) Brown spot infected sample (b) Grayscale image (c) Background removed sample.

Fig. 8.

Background removal process for Blight-infected sample (a) Blight-infected sample (b) Grayscale image (c) Background-removed sample.

Fig. 9.

Background removal process for Sheath Blight infected sample (a) Sheath Blight infected sample (b) Grayscale image (c) Background removed sample.

The training and validation progress of the baseline network, VGG16 with background removal, is shown in Fig. 10a. The classification results of this model are shown in Fig. 10b. Vgg16 with background removal achieved a validation accuracy of 90.30%. In contrast, the training/validation progress of the second baseline network, ResNet50 with background removal, is shown in Fig. 11a, and the Classification results of the second baseline network with background removal are depicted in Fig. 11b. ResNet50 with background removal achieved a validation accuracy of 90.70%.

Fig. 10.

Simulation results of baseline VGG16 with Otsu’s method (a) Training and validation graph of VGG16 with Otsu’s method (b) Classification result of VGG16 with Otsu’s method.

Fig. 11.

Simulation results of baseline ResNet50 with Otsu’s method (a) Training and validation graph of ResNet50 with Otsu’s method (b) Classification result of ResNet50 with Otsu’s method.

Here, we observed that the training and validation curves for both models on the augmented and segmented datasets are similar and stable, indicating good generalization. Here, the gap between the training and validation curves is smaller than in the first experiment with baseline networks, suggesting the model is not overfitting.

Adaptive multi-feature fusion-based hybrid network (AMFF-Hybrid Net) with background removal

Single CNN models may not capture all important features, such as fine details (texture, shape, and pattern) and important global contextual features (spatial hierarchy). The third experiment is implemented to get multi-fused features by applying an adaptive fusion technique at the concatenation layer of the suggested hybrid deep networks using the same baseline models. Background-removed samples are fed into the hybrid model to observe the outcomes. Training and validation progress of the hybrid network with background removal is shown in Fig. 12a. The classification results for the suggested network are shown in Fig. 12b. Proposed adaptive multi-feature fusion-based hybrid network (AMFF-Hybrid Net) with background removal approach achieved a validation accuracy of 98.50%.

Fig. 12.

Simulation results of proposed AMFF-Hybrid Net with Otsu’s method (a) Training and validation graph of proposed model with Otsu’s method (b) Rice disease classification result of AMFF-Hybrid Net with Otsu’s method.

At the beginning of training, the model achieves an accuracy close to 100% and remains steady, demonstrating the model’s highly optimized and effective learning ability. Training accuracy (blue) and validation accuracy (black dashed) are almost identical, indicating excellent generalization. Classification results are shown in a confusion matrix. True positive rates (Recall) of each disease are shown in the bottom row of the matrix. The Precision of each class is shown in the last column of the confusion matrix. Overall accuracy is shown in the bottom-right cell. The high per-class accuracy confirms the model’s reliability and robustness. The slight misclassification of blight and sheath blight in the sample is probably due to their visual overlap and similar symptoms.

It has been observed that the proposed Adaptive multi-feature fusion-based hybrid network (AMFF-Hybrid Net) with background removal achieves 98.5% accuracy, which is better than the rest, as illustrated in Fig. 12. The proposed framework outperforms related plant disease identification solutions outlined in the “Performance comparison with the related works” section.

Ablation study

The ablation study shows that Otsu segmentation and the AMFF-Hybrid architecture are important elements that should be optimized to enhance performance. The simple VGG16 and ResNet50 baselines achieve average F1 scores of 85.8% and 89.5%, respectively, and show limited robustness to complex field backgrounds. These are boosted to 90.1 per cent (VGG16 + Otsu) and 90.6 per cent (ResNet50 + Otsu) with the addition of Otsu segmentation, demonstrating that background removal benefits network training by enabling them to focus on lesion regions without requiring additional trainable parameters. The suggested AMFF-Hybrid Net with Otsu, however, offers 98.5% average F1-score with almost perfect class-wise scores (e.g., BS 99.6, LB 99.0, BB 96.4, SB 97.7 and more details in Tun 100.0 as depicted in Table 11 ) indicating that adaptive fusion of VGG16 path texture features with ResNet50 path deep contextual features on Otsu-refined images provides a much more discriminatory representation than any single backbone. The average recall, Precision, and F1-score for the studied components are shown in Fig. 13, whereas the overall accuracies are depicted in Fig. 14. This performance of the proposed system is obtained with a moderate 57.85 M trainable parameters and a reasonable inference time of 22.9 ms per image on a Dell i7 + RTX 3080 system, as shown in Table 12, which is an acceptable trade-off between accuracy and efficiency for real-time, field-deployable rice disease diagnosis.

Table 11.

Performance of the proposed system.

classes	Baseline VGG16			Baseline ResNet50			VGG16 with the Otsu method			ResNet50 with the Otsu method			Proposed AMFF-Hybrid Net with Otsu method
	R	P	F	R	P	F	R	P	F	R	P	F	R	P	F
BS	88.2	84.9	86.5	85.6	96.3	90.6	94.8	89.0	91.8	90.2	83.6	86.8	99.3	100	99.6
LB	71.9	84.0	77.5	80.4	90.4	85.1	81.7	91.2	86.2	77.8	93.0	84.7	98.0	100	99.0
BB	73.9	80.7	77.2	88.9	86.6	87.7	75.2	89.1	81.6	88.2	89.4	88.8	97.4	95.5	96.4
SB	97.4	91.4	94.3	92.8	95.9	94.3	100	89.5	94.5	99.3	91.0	95.0	98.0	97.4	97.7
Tun	99.3	88.4	93.5	100	81.4	89.7	100	92.7	96.2	98.0	97.4	97.7	100	100	100
Avg	86.1	85.9	85.8	89.5	90.1	89.5	90.3	90.3	90.1	90.7	90.9	90.6	98.5	98.5	98.5

Fig. 13.

Performance comparison of the proposed system with its baseline components.

Fig. 14.

Accuracy comparison of the proposed system with its baseline components.

Table 12.

Efficiency of the proposed system.

Model	Trainable parameters	Training time (6 epochs, 49 iters/epoch)	Inference time	Notes/setup
VGG16 (Baseline)	≈ 134 M	28.1 min	27 ms	Dell i7 @ 3.60 GHz, RTX 3080, 16 GB RAM, MATLAB 2019a + TensorFlow, Windows 10
ResNet50 (Baseline)	≈ 25.6 M	17.8 min	20.6 ms	Same setup
Proposed AMFF-Hybrid Net + Otsu	≈ 57.85 M	25.2 min	22.9 ms	Same setup; Otsu preprocessing adds < 1 ms/image on CPU

Hyperparameter sensitivity analysis

To evaluate the strength of the suggested model, a hyperparameter sensitivity analysis was carried out to measure the fixation of the learning rate, batch size, dropout rate, and training epochs. The model has proven to perform well with varying settings with a minimum accuracy of 97 percent as seen in Table 13. Optimal results (98.7) are obtained at a dropout rate of 0.6 and 8 training epochs, which demonstrates a better generalization with a slightly increased level of regularization and longer training time. Changes in batch size and learning rate have only slight modifications to the accuracy, which confirms that the suggested model is not very sensitive to the variations of the hyperparameters. In general, the findings confirm the stability and strength of the proposed framework in various training environments.

Table 13.

Hyperparameter Sensitivity Analysis.

Learning Rate	Batch Size	Dropout	Epochs	Accuracy of proposed model (%)
0.0001	30	0.5	6	98.5
0.0005	30	0.5	6	98.3
0.0001	16	0.5	6	98.1
0.0001	64	0.5	6	97.2
0.0001	30	0.3	6	98.0
0.0001	30	0.6	8	98.7
0.0001	30	0.5	10	98.6

Performance comparison with the State of the Art (SOTA) model

SOTA comparison makes it very clear that the proposed AMFF-Hybrid Net + Otsu is a far more potent solution than traditional CNN baselines and current lightweight models on rice disease classification. Although VGG16 and ResNet50 are considered classic and achieve decent accuracies of 86.1% and 89.5%%, they are limited in their ability to model global context and lack explicit mechanisms to curb background complexity. InceptionV3, EfficientNet-B0, and DenseNet121 are sequentially enhanced in accuracy through multi-branch convolutions, compound scaling, and dense feature reuse, but either fail on highly similar disease patterns or are more costly than other models. By comparison, the AMFF- Hybrid Net + Otsu has an accuracy of 98.5% based on a combination of modified VGG16 and ResNet50 backbones built into an adaptive multi-feature fusion model: VGG16 provides fine-grained local texture information, ResNet50 also provides deeper global contextual data, and Otsu-based segmentation eliminates background noise such that the fused representation only targets the portions of the image that are directly related to the disease. A performance comparison with the State of the Art (SOTA) model is shown in Table 14. This synergy leads to almost complete separation of classes. It significantly reduces misclassification, demonstrating that selective segmentation and complementary local–global feature fusion are more efficient than simply scaling a single backbone.

Table 14.

Performance comparison with the State of the Art (SOTA) model.

SOTA Model	Architecture	Accuracy (%)	Remark
VGG16	Deep CNN with stacked 3 × 3 conv layers	86.1	Strong baseline, but limited global context modelling and heavier fully connected layers
SqueezeNet	Lightweight CNN with Fire modules	92.6	Very compact model; good efficiency, but weaker at capturing fine-grained disease patterns
ResNet50	Deep residual network	89.5	Robust representation via residual blocks, but lacks explicit attention for context
InceptionV3	Inception-style multi-branch CNN	94.3	Better multi-scale feature extraction; still relatively computationally intensive
EfficientNet‑B0	Lightweight compound-scaled CNN	91.3	Good local feature extraction, but struggles with highly similar disease classes
DenseNet121	Dense connections with feature reuse	96.1	Performs well with dense feature reuse, but is computationally heavier
Proposed AMFF-Hybrid Net + Otsu	Modified VGG16 + Modified ResNet50 with adaptive multi-feature fusion and Otsu segmentation	98.5	Exploits VGG16 for rich local texture details and ResNet50 for deeper global contextual cues. At the same time, Otsu-based segmentation removes background clutter, enabling the fusion to focus on disease-relevant regions and yielding near-perfect class separation

Performance comparison with the related works

Table 15 shows a comparison of the proposed model and some of the recent related state-of-art methods selected in the literature. As it can be seen, despite the high classification accuracy of existing methods with the help of transfer learning and generative augmentation, as well as multi-scale fusion techniques, they usually suffer with limitations such as high computational complexity, high reliance on large datasets, or lack of adaptability in feature fusion. The suggested hybrid model does not use the full architecture of VGG16 and ResNet50. Rather, truncated and altered versions of both networks are used to lower the amount of computation. Further, the proposed model, in turn, presents an adaptive multi-feature fusion scheme in conjunction with segmentation-based preprocessing that allows successfully incorporating complementary features with the preservation of computational efficiency. This leads to competitive performance as opposed to related methods.

Table 15.

Performance comparison with the related works.

Ref	Methodology	Dataset	Accuracy (%)	Key Contribution	Limitation
28	Transfer Learning Ensemble (PDDNet-AE, PDDNet-LVE)	PlantVillage (54,305 images)	97.79	Robust multi-CNN feature extraction, improved generalization	Requires large labeled dataset, limited real-world validation
30	VGG-based Transfer Learning	1509 rice leaf images	92.46	Fine-tuned VGGNet for disease classification	Lower performance with limited dataset
32	DCGAN + ResNet	PlantVillage dataset	95.99	Generative augmentation for improved feature learning	High computational overhead
34	CNN-based real-time system	Plant dataset	98.32	Fast inference (42.6 ms/image), real-time detection	Sensitive to illumination variations and occlusions
35	MobileNetV3 + GKSO + CatBoost	Paddy dataset	98.52	Hybrid optimization and classification framework	Complex architecture and multi-stage processing
36	Hybrid framework with multi features fusion (ResNet50 + VGG16 +  + InceptionV3)	Paddy dataset	96.8	Multi-feature fusion with PCA dimensionality reduction	High computational cost, not suitable for real-time deployment
Proposed	Adaptive Multi-Feature Fusion (VGG16 + ResNet50)	Rice dataset	98.5	Adaptive fusion + segmentation + lightweight design	Balanced performance and efficiency

Discussion on performance improvement and model effectiveness

The integration of the adaptive multi-feature fusion, effective preprocessing and effective backbone design is said to be the reason of the superior performance of the proposed model. The adaptive fusion model improves the classification with the use of a dynamic learning process on the relative importance of the features obtained by modified VGG16 and ResNet50, which allows the model to prioritize the discriminative features and promote a reduced redundancy level. Moreover, preprocessing with the help of thresholding developed by Otsu eliminates background noise and isolates the area of the leaf, and the model can concentrate on disease-specific features and enhance the accuracy of the representation. The use of truncated backbone architectures further reduces computational complexity and mitigates overfitting, while preserving essential feature extraction capability. On the fly data augmentation also implies a better generalization since variations in input conditions are introduced.

The proposed model has high-performance, but the dataset in this research is quite small and found in an online source, which can limit its application in farming conditions in the real world. Although data augmentation and preprocessing measures were used to enhance robustness, additional validation should be done on large field and diverse datasets. In addition, further studies will be aimed at adding the statistical significance analysis and testing the model on independent datasets to further confirm the reliability. Calculational efficiency in this paper has shown that the given model can be implemented in real-time and used in practice in agricultural arrangements.

Conclusion and future scope

Smart agriculture is vital worldwide for improving crop yields and the country’s economy. Plant diseases are a serious threat to the agricultural sector, resulting in massive losses in output and the quality of harvests. As a result, this has necessitated the use of efficient, prompt plant disease detection mechanisms that facilitate early detection and accurate identification, allowing farmers to act in time to curb the transmission and spread of infections. Computer vision and Deep learning can be effective solutions for automated plant disease detection systems that aid early detection. This study was able to come up with a hybrid classification model to recognize and classify five common diseases of rice leaf: Blight, brown spot, sheath blight, tungro, and leaf blast, using an automated model. The proposed methodology effectively integrates global contrast enhancement as a preprocessing step, Otsu-based segmentation, and a multi-feature extraction strategy that leverages transfer learning through modified VGG16 and modified ResNet50 networks. This adaptive integration of network features (fine-grained texture and deep contextual features) proved very effective at highlighting the most discriminatory and informative representations, thereby enabling excellent disease classification. The model has made tremendous progress in automated diagnosis of agricultural diseases, providing a powerful and effective solution for early, specific diagnosis of rice leaf diseases. The proposed AMFF-Hybrid Net model may help farmers improve agricultural productivity and minimize economic losses.

Future research can explore additional approaches to improve the model’s performance and applicability further. To begin with, the data should be further diversified by incorporating a broader range of rice cultivars, growth stages, and environmental conditions to improve generalization. Second, exploring other deep learning architectures or transformer-based models to extract features may yield even greater classification accuracy. Also, it would be a good next step to investigate the real-time application of the model on edge devices to detect diseases in the field and further, statistical validation will be incorporated into future work to achieve additional insight and validation of model generalization and reliability. Lastly, a severity assessment module could be included that goes beyond classification, providing more detailed information to farmers so they can make better decisions about disease management. Severity assessment may entail dividing the diseased regions and measuring the surface area of affected leaves.

Author contributions

Santosh Kumar Upadhyay contributed greatly to the conceptualization of the study, design of the methodology, execution of software, analysis of data as well as writing the original manuscript. Rajesh Prasad helped in the investigation, validation, and literature review and helped in the revision of the manuscript. Vikas had the role of preprocessing data, evaluation of the experiment and managing resources. The study was supervised by Prashant Vats who reviewed and edited the manuscript and approved it. Final manuscript was read and approved by all the authors.

Funding

Open access funding provided by Manipal University Jaipur.

Data availability

The dataset employed in this study is accessible online at https://www.kaggle.com/datasets/rajeshbhattacharjee/rice-diseases-using-cnn-and-svm.

Declarations

Competing interests

The authors declare no competing interests.