In India, agriculture remains the backbone of the economy, with a substantial percentage of
In [1], the key elements of CNN architecture, such as convolutional layers, ReLU activation, pooling layers, and dropout layers, were employed in the Caffe framework. In [2], CNN architecture was used with layers such as Convo2D, flatten, max, pooling etc. on the Plant Village dataset with accuracy of 88%. In [3], the five different architectures are compared which include VGG16, ResNet50, InceptionV3, InceptionResNet, and DenseNet169, achieving the best result from ResNet50. [4] evaluates multiple deep learning models such as GoogleNet, ResNet101, ResNet50, InceptionV3, AlexNet, InceptionResNetV2, SqueezeNet, VGG16, VGG19. For object recognition, conventional techniques like LBP, HOG, colour features and GLCM for are also assessed. Bi-CNN employs pre-trained VGG and ResNet models for feature extraction followed by ADAM optimizer on Plant Village Dataset [5].
In [6], CNN is achieved through a re-parametrization method and a dynamic pruning gate to manage computational complexity, optimizing the feature extraction network. In [7], three classifiers LeafNet, SVM and MLP are evaluated to detect diseases in tea plants. LeafNet performed best with accuracy of 90.16%. In [8], the recognition of diseases in tomato leaves is done by S-CNN in which the model is trained using segmented images. In [9], on tomato, potato and pepper crops in the Plant Village dataset, CNN with image preprocessing is done and it achieves 98.029% of accuracy.
In [10], CNN incorporated a fully connected layer for classification, convolutional and pooling layer for feature extraction on approx. 35,000 images of Plant Village dataset. In [11], max pooling layers come after the convolutional layers, and the final layer includes an Adam optimizer and softmax activation to lower the loss function. In [12], using CNN with Raspberry Pi kit to anticipate crop diseases in advance. With a suggested activation function and two convolutional layers, it achieves 95% of system accuracy. Fertilizer optimization is aided by K-means clustering-based image segmentation.
In [13], convolution is used to identify patterns and edges, while pooling serves to reduce the image dimensions. CNN architectures which are applied simple CNN, VGG and InceptionV3. In [14], a diverse dataset is captured through various sensors. Subsequently, transfer learning is employed to leverage a pre-trained GoogLeNet CNN, facilitating detection and classification tasks. The dataset is expanded (XDB) through manual subdivision of images into smaller regions for optimal results.
In [15], The synthesis of three different CNN models (VGG-16, Google Net, ResNet 50) used with the application of two different classifiers (SVM and KNN). In [16], CNN is being used where the feature extraction is done by DWT, GLCM giving an accuracy of 98.12%. In [17], three approaches are used, a customized CNN, transfer learning with INCEPTIONv3, and visual transformers (small and large). Training involves Adam and RM-Sprop optimization, categorical cross-entropy loss, and callbacks and appropriate learning rates.
[18] explores ML and DL techniques, like random forest, SVM, and CNN like VGG-16, VGG-19, and Inception-V3, to accurately detect and classify citrus leaf diseases based on a manually curated dataset. (Evaluation involves area under the curve (AUC), precision, F1-score, recall and accuracy, comparing the performance of ML and DL methods, with DL demonstrating higher overall effectiveness). [19] Utilizing the Plant Village dataset for supervised learning, applying pixel-based operations, and employing CNNs for image classification.
In [20], utilization of advanced deep learning meta-architectures including RFCN, SSD and Faster RCNN, SSD with Inception-v2 and the highest mean average precision (73.07%) was achieved and optimization with Adam significantly improves accuracy, particularly for specific disease classes. In [21], while the model is trained, its process has included 160 images of the papaya leaves. There are numerous machine learning algorithms, such as KNN, Naive Bayes, Random Forest, Support Vector Machine, CART, and Logistic Regression, these all have been applied. Out of which, random forest performed the best with an accuracy of 70.14%.
In [22], detection of plant infections relies on K Means clustering and GLCM technique. Accuracy achieved was 98.27%. In [23], through the introduction of a rice plant disease recognition system, the ML algorithms such as KNN, Logistic Regression, Naive Bayes and Decision Tree are introduced. The Decision Tree algorithm gave best results by achieving an accuracy of a perfect 97.9167%. This dataset consisted of three different disease classes wherein each class has 40 images.
In [24], this paper's primary objective was to suggest enhancements to the existing machine-learning based classification methods which are for plant disease detection, supported by a comparison of the KNN classifier and SVM classifier. The outcomes demonstrated that the suggested algorithm has achieved a good accuracy of 98.56%, which also surpassed the 97.6% accuracy of the old/existing system. In [25], the suggested approach detects the plant diseases with an average accuracy of 93% by using the Random Forest Classifier as well as the digital Image processing technique.
In [26], Transfer learning is implemented with five pre-trained deep neural network architectures: VGG16, DenseNet169, InceptionV3, ResNet50, and Xception. Following model training, images representing different corn diseases from various datasets are employed as test data to evaluate the models' generalization capabilities. The DenseNet169 model demonstrated superior performance. The highest generalization accuracy of 81.60% was achieved when training the DenseNet169 model using (RGBA) images from the CD&S corn disease dataset, with backgrounds removed. In [27], the study compares 4 deep neural models such as fasterRCNN, EfficientDET, YoloV5 and YoloV6. Amongst all, YoloV5 model, which was trained with 93% accuracy on pre-trained hyper parameters, produced the best result. [28] achieves a detection accuracy of 98.26% by using the EfficientNetV2 model for cardamom plant disease detection and the U2-Net for background removal.
[29] employs transfer learning with six different CNN architectures, including VGG16, InceptionV3, Xception, Resnet50, MobileNet, and DenseNet121, for multi-class classification of plant diseases using 11,333 images from the PlantVillage dataset, with DenseNet121 achieving the highest accuracy at 95.48%. [30] proposes a rice plant disease diagnosis method using DenseNet169-MLP, combining DenseNet169 as a feature extractor and a multilayer perceptron for classification along with fuzzy c-means (FCM) based segmentation for identifying diseased portions, achieving an accuracy of 97.68%. [31] uses a hyperparameter-optimized Deep Convolutional Neural Network with data augmentation to achieve an accuracy of 98.41%.
We have opted for the PlantVillage dataset, a compilation of images encompassing diverse plant species and diseases. Originally comprising 38 labelled classes, we refined the dataset to 25 classes. This curation, focused on specific plant species and diseases, establishes a controlled framework for the purpose of recognizing and classifying plant diseases. The resulting dataset, presented in Table 1, optimizes precision by concentrating on classes crucial to our research.
Before model construction, thorough data pre-processing procedures were conducted to guarantee the quality and relevance of the dataset. shifts to enhance robustness of our dataset and a scaling function to normalize pixel values. To balance efficiency and information preservation, we resize the images to (224, 224) pixels, aligning with the EfficientNetB3 architecture, which utilizes three color channels (RGB). • Batch Size Selection: For both training and testing, we chose a batch of 40. This action aimed to achieve equilibrium between computational efficiency and model convergence.
Once the data was pre-processed, we constructed the model architecture.
With the model architecture in place, we trained the model using specific parameters and evaluated its performance.
• Epochs: The model underwent training for a total of five epochs.
• Verbose Setting (Verbosity): The training progress was displayed with verbosity set to 1, for real time updates on metrics like loss and accuracy. 2. Validation for performance evaluation: To gauge the performance of our model and confirm that it can be extended to unfamiliar data, a validation dataset was employed during the training process. After each epoch, the model was evaluated on this independent dataset, providing insights into its capacity to extrapolate beyond the provided training dataset.
The classification process determines whether a plant leaf from the Plant Village Dataset is contaminated or not. It further distinguishes the class of plant infection and recognizes the specific plant variety.
We have distributed our dataset into 80% training set, 10% validation set and 10% testing set. We prepared a classification report that provides a detailed assessment of the plant disease recognition system's performance across various diseases affecting plants.
We compared our results with several existing research papers to contextualize the success of our method. Following is the comparison evaluation of performance with existing research papers.
Our approach, concentrating on 25 carefully selected classes, distinguishes itself from studies like [14], which employed a more diverse dataset captured through numerous different sensors widely available. The deliberate emphasis on specificity enhances the precision of disease classification in our model. The stratified split, image processing, and augmentation techniques contributed to the robustness of our dataset.
In comparison to [10], which utilized CNN with a fully connected layer for classification and convolutional and pooling layers for feature extraction, our preprocessing techniques align with the specific requirements of the EfficientNetB3 architecture, ensuring efficiency and preservation of information. Leveraging transfer learning with EfficientNetB3, we introduced batch normalization, densely connected layers with regularization, and dropout for generalization.
Compared to [17], which explored multiple approaches including a customized CNN and transfer learning, our use of EfficientNetB3 with tailored modifications ensures an effective balance between complexity and accuracy. Using categorical cross-entropy as the loss function, five epochs were conducted during the training phase at a learning rate of 0.001.
In comparison to [6], where a reparameterization method and dynamic pruning gate were used to manage computational complexity. Our approach achieves competitive accuracy without resorting to complex computational optimization techniques.
Our model achieved an outstanding accuracy of 98.93% surpassing the performance of [20], wherein deep learning meta-architectures with a mean average precision of 73.07% attained the highest score. The specificity of our model in detecting diverse plant diseases across various species is reflected in the precision, recall, and F1-score metrics, as illustrated in our classification report.
So, our research showcases competitive performance when compared to existing research papers. The specificity and efficiency of our approach position it as a noteworthy contribution to the field of plant disease detection and classification.
While our project has made significant strides, it is crucial to acknowledge areas where can continue to grow and improve.
1. Larger dataset: The project acknowledges that the model's performance is contingent on the size and diversity of the dataset. For the model to be even more effective at generalizing to a wider variety of plant diseases, a bigger dataset might be needed. 2. Additional factors: One aspect worth noting is our model's current focus on visual cues in images, leaving out contextual information such as soil conditions or weather patterns. Integrating these factors could provide a more comprehensive analysis of plant health.
Real-time updates: The project does not explicitly address real-time or frequent updates. Changes in the dataset or emerging diseases may necessitate periodic model updates for sustained effectiveness. Periodic updates could be necessary to keep the model relevant and effective.
Our plant disease recognition system has accomplished prominent results using the plant village dataset when evaluated. The model has achieved test accuracy of 98.93% which shows that the model is highly trained at classifying the different classes of plant diseases accurately. We have also provided a comprehensive classification report containing evaluation parameters such as positive prediction value, sensitivity, and F-measure of different plant disease classes which again declares the efficiency of the model. In this paperwork, transfer learning has been employed with EfficientNetB3, integrated image preprocessing techniques, applied batch normalization, used ReLU activation and dropout layers making sure the model remains efficient and robust in all kinds of situations. In order to efficiently train and test the model, we have divided the dataset in strategic ways. Also, image augmentation is incorporated to enhance model's robustness. The data is fed into the model by dividing it into batches. These techniques contribute to enhancing the efficacy of the model and capability to identify the different plant diseases precisely.