Journal of Intelligent Systems and Internet of Things

Turmeric is a rhizomatous crop recognized for its medicinal effects which requires significant observation to ensure appropriate growth and progression. Turmeric plant diseases cause yield losses impacting food production systems and causing economic losses. Early prevention of these diseases is crucial for improving agricultural productivity. For this reason, The Improved YOLOV3-Tiny Model (IY3TM) was developed using Cycle-GAN and Convolutional Neural Network (CNN) with residual network for the early turmeric plant disease detection. However, this model leads to the omission of vital details along with the exact positioning of key attributes, thereby decreasing prediction accuracy. To resolve this, Convolutional and Vision Transformer model for Turmeric Diseases Detection (ConViT-TDD) is proposed for the prediction of turmeric plant diseases. ConViT-TDD is integrated into IY3TM with a self-attention mechanism and CNN-based global perspective to enhance the performance of the model A ConViT-TDD block involves the input channel transformation, the channel as well as spatial attention mechanism and global-minded transformers. The input channel transformation utilizes a convolutional layer to minimize the dimension of input channel and reduces the computational complexity. Global-minded transformers generate a feature vector based on the input channel transformation that is then transmitted to the encoder component. By collecting channel weights and spatial weights, respectively, the channel and spatial attention modules enhance the model's sensitivity to certain channel attributes and spatial locations, hence altering the feature representation of those channels and spatial locations. The attention module can adaptively change the weights of channel and spatial features for improved feature extraction and fusion. Once the initial attributes are reformed, the IY3TM detects and classifies the turmeric plant diseases. The test outcomes reveal that the ConViT-TDD model accomplishes an overall accuracy of 93.16% on the collected turmeric plant diseases images which is contrasted with the classical CNN models.

The following study investigates the role and impact of IoT and Al technologies on operational efficiency, sustainability, and cost optimization of freight forwarding companies. Their goals are to measure the effects of these technologies on logistics performance, assess sustainability improvements like decreased carbon emissions and waste, and identify cost-saving drivers for AI and IoT integration. H1: The operational efficiency of IoT and AI should enhance information sharing, route planning, and warehouse management significantly H2 claims that it will contribute to the reduction of carbon emissions and waste production by allowing real-time tracking, optimizing the usage of materials throughout the production cycle. H3- Cost Reduction in Logistics Operations through AI-based Automation, Predictive analytics and Improved Asset Management The approach was a quantitative research design, and data were obtained from 240 respondents from five large freight forwarders (companies): DHL Global Forwarding; Kuehne + Nagel; DB Schenker; XPO Logistics; and CEVA Logistics. Objective: Improvements after adoption are analyzed using structured questionnaires to measure key performance indicators (KPI) and frequency analysis and percentage calculation methods. The results confirm the transformative role of IoT and AI in freight logistics, increasing operational efficiency, sustainability, and cost efficiency. Logistics performance must be further optimized through continued investment in digital innovation.

Selecting the most relevant feature subset for a task is demanded and recommended for high accuracy and reduced model training time. Ensemble learning has shown superior results in classification; hence, we propose an ensemble method for feature selection and shown stability analysis for the selected feature set. The research question being investigated is whether ensemble methods are effective at selecting informative features in a dataset and if the selected features are stable compared to other feature selection methods. This paper presented a tree-based ensemble learning approach for feature selection. Our approach for ensemble feature selection includes function perturbation with the voting ensemble, an ensemble with a fixed number of features, and an ensemble with a contiguous number of features. Ensemble learning is found to be superior to other traditional feature selection algorithms. Ensemble learning algorithms are implemented on two high-dimensional microarray biomedical datasets. From our experimental study, it is observed that the voting ensemble outperforms other ensemble techniques, thereby reducing feature subset size and achieving higher accuracy. Stability analysis of all the algorithms has been studied and it is found that all ensemble techniques have higher stability than the traditional feature selection methods. Thus, ensemble learning proves to be a superior technique for feature selection. Our results demonstrate that the proposed method is effective in identifying relevant features and stable features and can improve the performance of machine learning models.

Pedestrian detection using object detection and deep learning has been found to be effective method for identifying pedestrians in video frames or images accurately. It is more commonly used in many real-time applications, such as security observing systems, autonomous driving systems, and robotics. The combination of deep learning techniques and object detection algorithms allows efficient and robust detection of pedestrians in several real-time scenarios. However, it is necessary to improve the detection efficacy for complex environments such as cases with worse visibility due to weather or daytime, crowd scenes, and rare pose samples. Continuous improvement and research in DL algorithms, dataset collection, and TRA models contribute to accelerating the robustness and acc of pedestrian detection systems. Therefore, this research models a novel marine predator algorithm with DL-based pedestrian detection and classification (MPADLB-PDC) method. The objective of the MPADLB-PDC system lies in the accurate recognition and identification of pedestrians. To achieve this, the MPADLB-PDC technique involves two major processes, namely object detection and classification. In the first stage, the MPADLB-PDC technique uses an improved YOLOv7 object detector for the recognition of the objects in the frame. Next, in the second stage, the ensemble classifier comprises three classifiers such as deep feed-forward neural networks (DFFNNs), extreme learning machine (ELM), and long short-term memory (LSTM). To improve the recognition performance of the ensemble classifier, the MPA is used to optimally select the parameters related to it. The simulation outcome of the MPADLB-PDC technique was authorized on the pedestrian database, and the outcome can be studied in terms of various aspects. The experimentation values validated the better outcome of the MPADLB-PDC approach compared to other approaches.

Face recognition technology plays a vital role in modern educational systems by enabling efficient and accurate student identification. The growing demand for efficient and accurate student identification systems has highlighted the limitations of conventional face recognition methods, particularly in handling variations in pose, lighting, and occlusions. To address this, our Precision-Optimized Human Recognition Model builds an Adaptive Information Retrieval System utilizing a Histogram of Oriented Gradients (HOG)-based detector for face detection and a ResNet-34-based Deep Metric Learning Model for face recognition. The system encodes facial features and performs identity verification using Euclidean distance for precise and reliable student identification. By integrating these techniques, the model ensures real-time data retrieval with high accuracy and adaptability to diverse conditions. The proposed approach enhances computational efficiency while maintaining robust recognition performance, making it a scalable and practical solution for identity verification in educational institutions.

Through its integration with the Federated Learning (FL) and Digital Twin (DT) technology, Internet of Things (IoT) based smart livestock farming is revolutionized toward real-time health monitoring and predictive analytics combined with secure decision-making. Privacy risks, inefficient models, large computational overheads, and heterogeneous data remain prominent in existing frameworks. This work introduces a “Privacy-Enhanced Digital Twin Livestock Optimization (PEDLO)” system, combining several adaptive and AI-driven components, including IntelliSense-Livestock Monitoring Framework (ISLMF) for multi-sensor data fusion, Privacy-Preserving Hybrid Aggregation (PPHA) Algorithm for secure federated learning, and Digital Twin-Augmented Reinforcement Learning (DTARL) for simulation-based decision-making. The PEDLO system optimizes disease prediction and anomaly detection, aims to reduce false alarms, and ensures data privacy for enhanced livestock welfare. Experimental results show 0.94 of accuracy, 0.93 of anomaly detection sensitivity, and a 40-second model convergence time, which outperform state-of-the-art techniques by a wide margin. The proposed system will enable scalable, efficient, and secure livestock management, marking a transformative shift toward sustainable precision farming.