An Efficient Convolutional Neural Network Based Deep Learning Framework for Rice Leaf Disease Classification

Full text

Text extracted from the article PDF for reading and search. Tables, figures and equations are in the PDF, which remains the authoritative version.

1. Introduction

Rice represents one of the most important staple crops. It is a critical crop that provides food for billions of people. Moreover, rice cultivation seems important especially in Asian countries, where It plays a key part in food security and agricultural economy. Majority of Indians preferred rice as an staple diet [25]. Rice crop growth depends on numerous components. Plant disease is one of the major important biological components that lowers crop production [15].

Here are some of the diseases of rice leaf: Bacterial leaf blight: Bacterium Xanthomonas oryzae is responsible for causing this disease. It attacks rice plant leaves and spreads rapidly under favorable environment conditions. It is characterized by elongated yellow or pale lesions along leaf edges.

Brown spot: It is caused by Bipolaris oryzae. It looks like as small circular or oval shaped brown lesions on the leaf surface.

Leaf smut: It is a fungal disease. It is caused by Entyloma oryzae. It appears as a narrow black or dark brown streaks on the surface of rice leaves.

These diseases can be detected through visible symptoms on plants leaves. If these diseases are not detected, then they may spread and can affect production. Traditionally, specialists used to detect disease by visual inspection. Those techniques were effective but they had some limitations. They required lots of time and effort. In remote or rural regions, it may not be accessible to farmers. As AI and deep learning technologies have emerged, agricultural practitioners can use image-based analysis tools that can be helpful in earlier disease detection. Thus, deep learning methodologies, especially CNNs, showed superb achievement in visual pattern recognition tasks. From raw images, CNNs can identify relevant image features by themselves and eliminate the need for manual feature extraction. This can improve classification accuracy. Several studies used well known deep learning models like CNN, MobileNetV2, ResNet50 VGG16. Although they performed well but they had some limitations like they require millions of parameters which make them computationally expensive and they are difficult to deploy. Therefore, there is need of a lightweight or efficient model which can classify or identify rice leaf disease accurately. Which also maintains low computational complexity.

Motivated by this need, in this work, a custom CNN-based model is proposed that could classify rice leaf diseases using leaf pictures. Commonly used performance metrics are used to measure or analyze the model and also compared it with transfer learning methods. The statistics suggest that the proposed model holds up on accuracy, making it appropriate for practical implementation in the field.

2. Major contributions of this study

Suggested model shows high classification accuracy and performs better than transfer learning models like MobileNetV2, ResNet50 and VGG16 on the given dataset. It has lightweight architecture, which has only about 424K parameters. It reduces memory usage which makes it appropriate for implementing on gadgets that have confined computational resources. The model also provides faster inference time (~10 ms per image). The proposed framework is a workable and scalable solution for actual agricultural applications since it generally accomplished a fair balance among accuracy, efficiency and complexity.

3. Related work

By adopting approaches of machine learning and deep learning, in past few years, several research studies have examined automated diagnosis of plant disease. Along with traditional classifiers, earlier methods used handcrafted features. While these research achieved success, their performance was strongly depended on feature design. Plant disease classification accuracy has been improved by deep learning models like CNNs. On several crop datasets, researchers applied various models like AlexNet, VGGNet, Inception, and ResNet. Transfer learning models have achieved high accuracy, but they have very high computation cost which makes the not well suitable for deployment in resource-constrained environments. This study differs by developing a custom CNN architecture which achieves competitive accuracy and also maintains lower computational complexity.

For plant disease classification. several studies have analyzed the employment of prelearnt deep learning frameworks like AlexNet, VGGNet, GoogLeNet, ResNet, and DenseNet through transfer learning. These models showed excellent performance, but usually they required a large number of parameters and require high computational resources which makes them less fit for utilization on budget friendly or mobile agricultural appliance [7].

Kurmi et al. projected a deep CNN based schema which combined region-of-interest localization with CNN-based feature extraction for crop leaf disease detection to address computational efficiency and robustness. It removed redundant background information from leaf images, improved classification accuracy and generalization. The framework was examined on the PlantVillage dataset. It achieved higher performance compared to conventional methods of image processing [7].

Priyadharshini et al. achieved 97.89% accuracy with 80:20 train-test split by modifying LeNet architecture for identifying diseases in maze leaf [13]. Jasrotia et al. achieved 96.76% accuracy for four maize disease classes by combining CLAHE preprocessing with custom CNN [22]. Upadhyay and Kumar used CNN with Otsu's thresholding and achieved 99.7% accuracy for rice disease classificatio006E [15].

For apple disease classification by developing a custom CNN Vishnoi et al. achieved 98% accuracy [9]. Despite significant progress, many models performed well on controlled datasets but shown reduced accuracy on field-acquired images. This shows the need for lightweight yet accurate CNN architectures which can generalize well across diverse crop species and environmental conditions. For automated crop leaf disease classification, the reviewed literature shows that CNN based models can offer an effective and scalable solution. However, further study is necessary to develop optimized CNN architectures that balance accuracy, computational efficiency, and real-world applicability. This forms the inspiration for the proposed work, which attempts to design a CNN framework for accurate identification of diseases in crop leaf images.

4. Description of The Dataset

In the current work, rice-leaf dataset [26] and A dataset of rice leaf diseases [27] are used. In Rice leaf dataset, it consists of 16000 images sorted into four different categories or classes (each class contains 4000 images): different diseases classifications (bacterial leaf blight, brown spot and leaf smut) and healthy leaves. From a dataset of rice leaf disease only brown spot class’s images (139 images) are used to add variation in the dataset. These were collected from kaggle. Duplicate images are removed and dataset had total 16058 images.

Table 1: Number of images per class Class name Number of images Healthy 4000 Bacterial leaf blight 4084 Brown spot 4009 Leaf smut 3965 Every image is sorted to a respective class and the dataset was divided into training set, test set andvalidation set (70%: 15%: 15% respectively).

Fig 1: Sample rice leaf images from different disease categories used in this study [26, 27]: (a) Bacterial leaf blight (b) Brown spot (c) Leaf smut (d) healthy

4.1. Dataset statistics

Dataset contains four classes. For every class, dataset in independently divided as shown in table II.

Table 2: Class distribution Dataset split Total sample images Objective Training set 11238 Model learning Validation set 2409 Hyper parameter tuning Test set 2411 Final evaluation Preprocessing leads training, images are rescaled to 224×224 pixel dimensions and pixel value is normalized to ensure compatibility with deep learning techniques. Thus, the training set experience data augmentation techniques like rotation, zooming and allows model to learn meaningful feature representations from images. During training, to optimize model hyper parameters and detect overfitting validation set is employed. Test set is utilized to evaluate the model. It contains images that are not seen by the model before.

5. Methodology

5.1. Framework Used

The proposed model is implemented using Jupyter Notebook, open-source web-based execution platform. Tensorflow and Keras are utilized to construct and train the model. Code execution, documentation, and visualization are all made easier using Jupyter.

5.2. Image Preprocessing and Data Augmentation

For improving model performance, image preprocessing plays an important role. Images are rescaled to 224×224 pixel dimensions and value of pixels is normalized to range 0-1 through dividing by 255 to ensure compatibility with deep learning techniques.

In the dataset to include variability, data augmentation is used which avoids overfitting. The augmentation techniques used in this study include rotation, flipping, zooming and contrast adjustment. So that it learns better.

6. Proposed Model Architecture

Fig 2: Proposed methodology In this study, for automatic rice leaf disease categorization, a custom convolutional neural network (CNN) is developed. From leaf images the model extracts hierarchical features while maintaining relatively low computational complexity. The input images which it processes are of size 224x224x3. It includes four convolutional layers, five batch normalization layers, four pooling layers and two fully connected layers.

Before performing final categorization, the network progressively learns increasingly abstract representation of the input image. It contains approximately 424,262 parameters which make it lighter than deeper architectures such as VGG16 and ResNet50.

Fig 3: Proposed model’s fundamental architecture Table 3: Proposed framework’s Layer wise architecture Sr. No. Layer Output Size 1. Input layer 224×224×3 14. Global average pooling 256 15. Dropout 256 16. Dense 128 17. Batch normalization 128 18. Dropout 128 19. Output layer 4

6.1. Input layer

Here RGB images are accepted. Images are scaled to 224 × 224 pixels. Each image is represented as a 3-D tensor (224×224×3). The image height and width is 224. Here 3 represents RGB color channels.

Prior to feature extraction, pixel values are normalized using a rescaling layer. The input values i.e. [0,255] are rescaled to [0,1].

6.2. Convolutional layers

In convolutional neural networks, convolutional layer is an essential part. Spatial feature from input image are extracted by these layers. They do this by using trainable filters. These filters scan across the input image’s spatial dimensions (height and width). An element wise multiplication and summation is performed by each filter and creates a feature map. Patterns like edges, textures and disease lesions are recognized by it.

The convolution operation:

Table 4: Architecture of convolution blocks Layer Filters Output size The output size of each convolutional layer is represented in the form of height × width × number of feature maps. The spatial dimensions (height and width) decrease progressively due to pooling operations, while the number of filters increases to capture more complex and abstract features at deeper layers of the network. As the depth grows, more complex visual features can be learnt by the network by adding additional filters as the depth increases.

6.3. Batch normalization

After each convolution layer, for balancing process of learning, batch normalization layer is deployed. During training, it lowers the internal covariate shift or variation in the distribution of layer inputs. Model can learn more effectively with the help of batch normalization as it preserves a uniform distribution.

Each layer’s output is normalized by applying the following equation:

�̂=??????−μ √σ2+ϵ here: x = input to the layer, σ2 = batch’s variance, μ = batch’s mean, ϵ = a small constant for numerical stability It improves training speed and lowers the risk of disappearing or exploding gradients.

6.4. Max pooling layer

After convolutional blocks max pooling layers come. It reduces spatial resolution as well as preserves essential features. it reduces the number of parameters and network activities which helps in reducing computational complexity and managing overfitting.

From a small region of the feature map the maximum value is selected by the max pooling operation:

here x = input features map.

Using 2 × 2 pooling windows the feature maps are down sampled in this architecture. It retains the most prominent features while discarding less significant information. The feature map dimensions are reduced as in table V.

Table 5: Feature map dimension’s architecture Stage Output size The output size of each pooling layer is represented in the form of height × width × number of feature maps.

6.5. Global pooling average

Global average pooling is used instead of using traditional flattening layers in this architecture. This layer reduces spatial dimensions to single vector by calculating average value of every feature map. 14 × 14 × 256 feature map’s dimensions are converted to 256 by global average pooling. It decreases the parameter count and helps avoid overfitting.

6.6. Fully connected layer

For performing classification, through fully connected layers the extracted feature vector is passed. First dense layer contains 128 neurons which learn high-level feature interactions. By randomly disabling neurons during training dropout regularization is applied to reduce overfitting. Prior to classification, for feature learning, 128 neurons are utilized in fully connected dense layer; but the number of features produced after global average pooling is 256. The dropout operation can be represented as:

6.7. Output layer

There are 4 neurons in the final output layer which corresponding to the four classes: healthy, leaf smut, bacterial leaf blight and brown spot.

For converting the output values into class probabilities a Softmax activation function is applied:

P(yi)=ezi ∑ezjkj=1 here: �??????= output score for class i, k = number of classes. Class having highest probability is chosen as predicted class.

6.8. Model complexity

Proposed model’s complexity is analyzed with regards to its parameters counts. Proposed CNN architecture contains 424,262 total parameters. Out of total parameters 423,044 are trainable parameters which are actively updated during backpropagation. These trainable parameters include convolutional kernel weights and biases in fully connected layers, which are optimized using the Adam optimizer. Model is of size approximately 1.62 MB. The proposed model is significantly lighter while still achieving competitive classification performance as Compared to deeper architectures such as VGG16, MobileNetV2 and ResNet50. It lowers complexity which makes the model appropriate for resource-constrained environments and real time agricultural implementations.

6.9. Evaluation metrics

Standard classification metrics such as accuracy, precision recall and F1 score are applied to analyze model’s performance. They also help to check classification quality. For class-wise predictions analysis and to recognize patterns of misclassification, confusion matrix is used. overall classification accuracy appears defined as: Accuracy shows proportions of predictions model got correct.

Precision could examine positive predictions. Furthermore, precision checks how many were actually right.

Recall may represent model’s capability to correctly detect all relevant samples:

The significant F1-score could provide important measure between precision and recall:

7. Experimental Findings and Discussions

7.1. Qualitative Performance Analysis

The proposed model is trained for 25 epochs. The model is trained using 70-15-15 dataset split (11238 training, 2409 validation and 2411 test images). Using accuracy and loss curves, the learning behavior of the model is analyzed.

Fig 4: Graph of model accuracy Fig. 4 shows that on the validation and test sets, around 98% - 99% training accuracy is attained by the trained CNN model and it hits around 99% validation accuracy. It shows that proposed model effectively identified the essential features and there is absence of significant overfitting.

Fig 5: Model loss graph Fig. 5 reveals that training loss decreased and reached nearly 0.03-0.05. Validation loss is also decreased from 2 to 0.05. This figure confirms there is stable convergence, high confidence predictions and effective feature learning.

Table 6: Proposed model’s class-wise performance results Name of the class Precision Recall F1 score Support Bacterial leaf blight 1.00 0.98 0.99 613 Brown spot 1.00 0.99 0.99 602 Healthy 0.98 1.00 0.99 601 Leaf smut 0.99 1.00 1.00 595 Across all disease classes, strong performance is shown by the model and attained nearly perfect precision, recall and F1 score. Model has attained perfect or near perfect precision and recall for all classes. Under the given dataset conditions, high reliable and robust classification capability is shown by the model.

7.2. Test set evaluation

To analyze model’s performance on new data, it was tested on novel test set which was excluded from validation and training. On the test data, the model sustained good performance and attained accuracy around 98.75%. No evidence of train-test data leakage was observed. Thus, unseen rice leaf images can be handled efficiently by the proposed CNN model.

7.3. Confusion matrix analysis

Fig 6: Proposed model’s confusion matrix Proposed model’s confusion matrix of indicates that across all classes, majority of the samples are classified accurately with minor classification failure that occurs between bacterial leaf blight and brown spot and between healthy and brown spot. This may be due to same visual patterns such as distribution of colors and lesion texture. This is common observation among prior research. This error states that in future enhancements could focus on fine-grained feature learning.

7.4. Comparison with transfer learning models

To see how well proposed model performed, it is compared with a transfer learning based frameworks like MobileNetV2, ResNet50 and Vgg16. The transfer learning models, which are already trained on ImageNet, are examined using the same dataset and experimental settings.

Among them comparison is done based on accuracy, model size, time per image and number of parameters. Table VII shows the comparison among them.

Table 7: Comparison among proposed model, MobileNetV2, ResNet50 and VGG16 Model Accuracy (%) Model size (MB) Time per image (ms) Number of parameters Overall suitability Proposed model 99.33 1.618 9.512 424260 Best MobileNetV2 97.17 8.633 14.141 2263108 Good alternative ResNet50 46.93 90.01 64.661 23595908 Not suitable Vgg16 79.75 56.13 193.830 14716740 Heavy From table VII it is clear that proposed model have better accuracy, time per image despite of having lower parameters and model size than others. So overall, the proposed model is achieved best tradeoff between accuracy, speed and model size which makes it more suitable for real life disease detection.

7.5. Discussion

Experimental results showed that superior classification performance is achieved by the proposed framework by achieving 99.33% accuracy on the test set. It shows models ability to learn visual patterns from images of rice leaves. Its hierarchical feature extraction capability is one of the main reason for its performance. In the early convolutional layers, low-level features present on rice leaf surfaces are learnt by the network. These characteristics can be helpful in identifying differences in background patterns and leaf structure. As the input moves through deeper convolutional layers more complex and disease-specific patterns, like lesion shapes, discoloration regions, spot distributions and texture anomalies that define various rice illnesses, are captured by the model. The network's deeper layers are able to identify disease patterns by merging several low-level features into higher-level representations. The training process is stabilized by batch normalization which also enhances convergence. By reducing spatial dimensionality max-pooling layers permit the model to aim on the most prominent features. By requiring the network to summarize feature maps rather than memorize spatial patterns, number of parameters are reduced and overfitting is minimized by global pooling average. To enhance generalization during training, dropout regularization helps in preventing the model from relying too heavily on specific neurons. As a result, while sustaining a fairly compact model size and computational complexity the proposed CNN is able to effectively learn discriminative features from rice leaf images. In opposite, VGG16 achieved 79.75% accuracy, MobileNetV2 achieved 97.17% accuracy while ResNet50 achieved nearly 46% under same conditions. Despite of having lower parameters than other models, proposed model performed well. This shows that when dataset is domain specific, well-designed task -specific architecture can perform well then deeper pre-trained models. Overall the robustness, effectiveness and efficiency of the presented approach are validated by results for identification or categorization of rice leaf disease.

7.6. Conclusion

In this study, for automated disease categorization of rice leaf using image data, a CNN based framework is presented. While maintaining lower computational requirements, the proposed model hits 99.33% prediction accuracy along with strong recall, precision and F1 score which make it appropriate for real world agricultural implementation. Proposed model has several advantages like it is lightweight architecture, has low computational cost, fast inference time and it is appropriate for mobile deployment. A comparison analysis between proposed model and transfer learning models like MobileNetV2, VGG16 and ResNet50 shows that although already trained models perform well, but better disease classification and greater efficiency can be provided by the proposed CNN model which is task-specific. The results show that for faster and better decisions in real- world, the suggested CNN model can be helpful to detect rice leaf diseases accurately. In future, dataset can be expanded with additional disease classes. In real-time mobile or web-based systems, the model can be deployed.