In a recent publication in the journal Scientific Reports, researchers proposed a novel framework for early plant disease diagnosis, integrating fog computing, edge environments, and IoT sensor technology. They utilized pre-trained convolutional neural network (CNN) models for feature extraction and coupled them with an enhanced grey wolf optimization (GWO) algorithm for disease identification.
The emergence of agri-technology as a pioneering scientific discipline signifies a shift toward data-intensive methodologies, fostering agricultural productivity while mitigating environmental repercussions. Contemporary agricultural practices rely on sensor-generated data, offering comprehensive insights into operational variables such as weather conditions, crop and soil dynamics, and machinery-related data, facilitating informed decision-making.
The adoption of conservation agriculture, recognized for its efficacy and ecological benefits in enhancing yields, prompts an exploration of its holistic impact on crop output improvement. Amidst these considerations, the critical research domains of plant disease and pest identification within the field of machine learning come to the fore.
Employing machine vision tools to discern diseases or pests in plant images represents a pivotal facet of agricultural innovation. Automating precise and early-stage plant disease detection becomes imperative, especially considering the challenges faced by non-expert farmers and the logistical hurdles associated with consulting experts.
Machine vision applications, leveraging artificial intelligence (AI) and machine learning, are gradually making inroads in agriculture, particularly in disease and pest identification. Deep learning advancements hold promise for heightened accuracy in plant disease detection, surpassing traditional machine learning and image processing approaches. The synergy of Internet of Things (IoT) sensors, deep learning models, fog computing, and a novel GWO algorithm constitutes a comprehensive approach to address these challenges.
Several studies employ deep learning techniques for plant disease diagnosis. For instance, DenseNet121 for tomato leaf categorization, CNN, and transfer learning for the classification of insect species. For rice disease detection, Faster Recurrent-CNN (FR-CNN) and the k-means algorithm (FCM-KM) are used, and for rice leaf illnesses, residual network (ResNet50) with support vector machines (SVM). There are deep-learning models developed for plant disease recognition. Various studies contribute to the advancement of plant disease prediction.
Methodology for enhanced plant disease detection
The authors employ transfer learning with CNN-based models such as AlexNet and GoogleNet, integrating support vector machines (SVMs) for classification. The study introduces the GWO algorithm for feature selection. Fog computing and the IoT contribute to real-time data processing and decision-making in smart agriculture, addressing challenges such as energy consumption and user proximity associated with cloud computing. The proposed solution involves a three-fold approach: applying deep learning models for feature extraction, using the GWO algorithm for feature optimization, and employing SVM for image classification. The architecture encompasses both fog and cloud computing for efficient processing.
The proposed methodology incorporates transfer learning, pretraining methods, and an optimization algorithm leveraging IoT sensors. The architecture comprises three layers: IoT devices, image processing in a fog environment, and resource utilization from cloud computing. The model, rooted in deep learning, transfer learning, and shallow machine learning, employs a data acquisition layer, preprocessing, feature extraction, feature subset selection using a modified GWO (MGWO), and SVM training.
The MGWO introduces adjustments to overcome the limitations of the traditional GWO, enhancing its efficiency in searching for optimal values. The proposed model represents an innovative approach to plant disease detection, adaptable to both fog and cloud computing environments.
Experiments and results
Performance metrics delineate true positives, true negatives, false positives, and false negatives, facilitating a detailed examination of classification proportions. Addressing unbalanced data collection's impact on accuracy, the text emphasizes the significance of sensitivity and specificity. Researchers performed two experiments to evaluate the models. The first Experiment evaluates a modified GWO for feature selection, which is crucial in machine learning model development.
The experiment uses fifteen standard feature selection datasets, employing a chaotic map for initialization and addressing the multi-objective issue. The process involves transforming binary values and utilizing cross-validation for robustness. Results highlight MGWO's superiority in feature reduction and classification accuracy among the compared algorithms.
The second Experiment builds on MGWO's efficacy, applying it as a wrapper feature selection algorithm for plant disease classification. The feature extraction process uses a pre-trained AlexNet CNN, followed by MGWO feature selection and SVM training. Datasets involving ten different plants and healthy and diseased images. The proposed model, integrating AlexNet, MGWO, and SVM, outperforms AlexNet, GoogleNet, and SVM individually. A comparison between SVMs trained directly on extracted features and those selected by MGWO. The receiver operating characteristic (ROC) curve for the proposed model SVM on the test set showcases its robust performance.
In summary, researchers proposed a hybrid approach for plant disease identification, utilizing SVM, AlexNet, and Google Net-based transfer training on edge servers. The proposed model, incorporating modified GWO, operates efficiently on IoT devices, integrating fog and cloud computing. Experimental results demonstrate accurate disease detection with minimal computational resources. Future enhancements include leveraging blockchain for improved fog environments and developing deep learning-supported apps for plant disease detection in smart agriculture.