Plant disease represents one of the most severe threats to agricultural production. Deep learning has emerged as a promising solution for automating the recognition of these diseases, leading to a richness of disease recognition applications based on deep learning. However, most existing applications do not address the challenge of simultaneous multi-disease detection from the same leaf. In this study, we introduce a deep learning-based model designed to detect and recognize multiple diseases from the same leaf simultaneously. Our method enables the recognition of each disease’s symptoms separately from small leaf regions, independent of other diseases or specific crop types, through an isolation method. This approach also allows the model to generalize disease detection to new crops not encountered during training. Additionally, our method calculates the prevalence rate of each disease on the leaf and determines the overall extent of all diseases present. To evaluate the effectiveness of our approach, we applied it to the widely recognized PlantVillage dataset, creating a new version for training and testing with three CNN models: Small Inception, MiniVGGNet, and LeNet5. The results demonstrate that the Small Inception architecture outperformed the other two CNNs in terms of classification performance. Despite some class imbalances in the new dataset, which were addressed through the use of class weights, this approach significantly enhanced the model’s performance. Furthermore, while the proposed method demonstrates high performance in controlled environments, though its consistency under real field conditions still warrants deeper investigation. Overall, the findings underscore the effectiveness of our method and highlight its potential as an efficient solution applicable across diverse agricultural contexts.

The Smart Grid, a prominent IoT application, is experiencing rapid growth driven by the proliferation of connected embedded devices. This evolution has resulted in an exponential increase in time-series data, emphasizing the need for efficient data storage and retrieval mechanisms, particularly for real-time IoT environments. Existing indexing structures primarily focus on either time-based or consumption-based organization, often overlooking the interdependence between these dimensions, which limits their query efficiency. To address this limitation, this paper introduces a novel Temporal-Consumption Binary Tree (TCB-Tree), a hybrid tree-based indexing structure that jointly exploits temporal and consumption attributes for efficient data retrieval. The proposed method operates in three main phases: (i) horizontal segmentation, which applies clustering to identify key consumption levels; (ii) vertical segmentation, which groups temporally successive data within the same consumption range; and (iii) hybrid index construction, where internal nodes index time while leaf nodes index consumption patterns. Experimental evaluation using three real-world datasets demonstrates that the TCB-Tree achieves rapid construction times (under 0.20 s) and efficient hybrid query execution (under 0.9 s) on large datasets, while maintaining minimal storage overhead (below 18%). These results confirm the scalability, efficiency, and suitability of the proposed structure for Smart Grid and real-time IoT applications.

Field-service operations for utility companies require routing technicians across multiple depots while guaranteeing same-day response to critical infrastructure customers, a constraint that standard multi-depot routing methods cannot structurally enforce. We introduce the MDVRP with Priority Requests (MDVRP-PR), formalised as a lexicographic optimisation problem that guarantees service to priority customers before maximising coverage and minimising route duration. A three-stage pipeline is proposed: hybrid DBSCAN-Hierarchical clustering for topology-aware depot assignment, an Enhanced Max-Min Ant System (MMAS) with priority-driven construction, lexicographic solution selection, and repair, and a Boundary Relocate post-optimisation stage with global cross-depot recovery. The approach is evaluated on a real-world applied case study from Algérie Télécom (Guelma, Algeria), comprising a single four-depot field-service instance scaled to three sizes (55, 90, and 150 customers) and assessed over 2135 controlled runs. On this case study, the proposed clustering method outperforms the MDVRP-adapted Sweep baseline by 22.9 percentage points on the largest instance (n = 150; Friedman p < 0.001). The priority mechanisms sustain 100% feasibility across all configurations, compared to complete collapse without them (0/10 seeds at ≥40% priority), at a route-time overhead below 5%. Relative to the company’s current manual practice, the framework improves customer coverage by 16.1 percentage points within 28 s, confirming its practical utility for daily deployment in this capacity-constrained, priority-sensitive routing context.

Modern dairy farming and increasing cattle herd sizes are economic challenges that require cost-effective and accurate solutions that ensure high fertility. This can be achieved by early and reliable detection of heat phenomena in the estrous cycle. Traditional observation-based methods present significant challenges such as observer fatigue and variability in observer expertise. In addition, recent methods proposed for heat detection are often insufficient due to high costs or limited reliability. To overcome these issues, we proposed a new system called CowStrus that overcomes the limitations of existing methods by exploiting the overhead view using surveillance cameras fixed to the ceiling of the barn or surveillance drones. The captured videos are processed using an enhanced version of YOLOv8, fine-tuned through transfer learning to optimize performance in detecting cows and accurately identifying their heads. This information is used to extract spatiotemporal features, which are then used by a decision tree-based classifier to ensure proper detection of cows in heat. Decision making is performed based on a cumulative score of the classifications of sexual activities observed in cows. Evaluation of the system on real images showed promising results highlighting the potential of CowStrus as a reliable and economically viable solution, achieving high scores in four classes of sexual activity: F-Score of 0.965 for the Cajoling class, 0.897 for Sniffing, 0.857 for Resting_the_chin and 1.000 for Mounting.

Food supply chains have increasingly adopted blockchain technology due to its critical role in ensuring food safety and providing consumers with traceability and transparency of food products. Existing blockchain-based traceability systems often rely on complementary technologies such as QR codes, RFID, and barcodes to facilitate tracking and data retrieval. However, these systems face significant challenges in managing large-scale, complex traceability queries. Moreover, their query time performance tends to degrade as the volume of data and the size of the blockchain ledger grow, since the search must be performed sequentially across blocks especially in complex food supply chains. To address these limitations, we propose a novel blockchain-based traceability system that integrates a natural language processing (NLP) model and B+ tree indexing technique. This integration ensures accurate data responses and optimizes traceability queries within the blockchain ledger. The system was simulated using the hyperledger fabric framework, with various scenarios tested based on different data and block sizes, and the experimental results demonstrate a significant improvement in complex traceability range queries, achieving an average reduction of 68.32

Accurate detection of wheat grain quality is crucial for ensuring food safety, economic efficiency, and post-harvest value preservation. Traditional classification methods often require extensive annotated datasets and struggle to generalize across diverse impurity types. In this paper, we propose an object-level contrastive self-supervised learning approach to address these limitations. The proposed method first pretrains a feature encoder using SimCLR on unlabeled grain instances to learn rich visual representations, then fine-tunes it on labeled samples for classifying healthy, damaged, and impure grains. By operating at the instance level, the model captures fine-grained visual features crucial for quality assessment. Experiments conducted on a publicly available wheat dataset demonstrate that our method outperforms both supervised ResNet-50 and YOLOv8 baselines in precision and F1-score, particularly for the underrepresented impurity class. The results validate the effectiveness of self-supervised pretraining in enhancing classification accuracy while reducing dependency on large annotated datasets. This study lays the groundwork for scalable and annotation-efficient wheat quality inspection systems, with potential applications in automated sorting and grain certification pipelines.

The growing need for personalised advertising has intensified interest in intelligent recommendation systems capable of adapting to user preferences. This paper introduces a collaborative advertisement recommender that integrates demographic, geographic and behavioural data to enhance ad relevance. Two clustering strategies are examined: a centric approach based on the K-means algorithm and a hierarchical approach employing the KD-tree algorithm. Experiments were conducted using data gathered from Hazmit, a purpose-built social platform for evaluating advertising recommendations. The comparative analysis - covering accuracy, precision, recall, F-score and execution time - demonstrated that KD-tree achieved superior precision (0.75) and overall accuracy (0.65), whereas K-means obtained the highest recall (0.98). KD-tree produced outstanding results in food-related advertisements, while K-means yielded stronger performance in technology and clothing categories. Both methods showed limited effectiveness for news advertisements, reflecting the unpredictability of user interests in that domain. With average runtimes below 1.2 s, both algorithms proved efficient for real-time deployment. Overall, the findings indicate that KD-tree offers more targeted and accurate recommendations, while K-means ensures broader user coverage, making each approach advantageous under specific advertising contexts.

Medical big data analysis is crucial to advancing healthcare research and improving patient care. However, missing data in these datasets creates a significant challenge for understanding patient profiles and disease patterns. The absence of critical information caused by factors such as incomplete patient records or unreported variables introduces uncertainties that can compromise the accuracy and reliability of analytical results. Resolving the problem of missing data is therefore paramount to ensuring the efficiency of healthcare analyses and improving the overall quality of medical research. This study examines and reviews different methods of imputing missing data in the context of medical big data. Traditional imputation techniques, advanced statistical approaches, machine learning-based models and deep learning-based models are evaluated with regard to their relevance, scalability and performance. Particular emphasis is placed on the unique challenges presented by medical big data, such as high dimensionality and data heterogeneity, as well as the incorporation of domain-specific knowledge. This study further investigates the role of big data technologies in enabling efficient imputation for large-scale medical datasets and evaluates metrics designed to assess imputation quality. Finally, future directions are discussed with the aim of improving missing data imputation strategies and, by extension, data-driven decision-making in the era of medical big data.

Diabetes is a time-transcending sickness that happens when your body can't make or use insulin well. This can harm many body parts like your heart, blood vessels, eyes, kidneys, and nerves. Finding diabetes early is crucial to treat it fast and stop it from getting worse. But regular computer programs don't do a great job at predicting what will happen with diabetes because they struggle with some data issues. In our research, we looked at a special kind of computer program called stacking-bagging, which combines different types of programs to make predictions. We used it along with Support Vector Machine (SVM) on a famous dataset called Pima Indian diabetes dataset. After training and testing our models a lot, we found out that our new method worked really well. It got 90.25% of predictions right and was very good at being precise and remembering things too. In fact, it did better than other fancy computer programs people use for this problem.

As the Internet of Medical Things (IoMT) expands, securing healthcare data has become increasingly vital due to the inherent vulnerabilities associated with wireless communication. Traditional security measures often fall short in effectively mitigating sophisticated cyber threats that target IoMT networks. This study proposes a Deep Reinforcement Learning (DRL)-based Intrusion Detection System (IDS) specifically designed to enhance transmission-level security for IoMT applications. The proposed system utilizes an advanced Deep Q-learning algorithm to develop optimal strategies for dynamic decision-making, allowing for real-time adaptation to emerging threats. Through comprehensive experiments and comparative analyses across IoMT datasets, the proposed DRL-based IDS demonstrates significant improvements in detection accuracy, precision, recall, and response time when compared to conventional IDS methodologies. The findings highlight the potential of employing DRL techniques to substantially strengthen the security posture of IoMT ecosystems, thereby ensuring the integrity and safety of sensitive healthcare data.

Sentiment analysis has become an essential tool in understanding global narratives across social, economic, political, and commercial sectors. As social media platforms increasingly produce vast amounts of uncontrolled textual data, the need to analyze content in regional languages has grown significantly. This article focuses on the unique challenges posed by sentiment analysis in Arabic and dialectal Arabic, with a special emphasis on the Algerian Dialects often referred to as Darija, which is characterized by its linguistic diversity and multilingual nature. One of the primary issues addressed is the lack of annotated datasets for this dialect and the complexities of accurately interpreting this diverse language. To tackle these challenges, we present DZDialect, a new dataset comprising 117,569 annotated comments, and we explore various advanced methodologies for sentiment analysis. Furthermore, the study compares the performance of machine learning algorithms (SVM, NBM, and KNN), deep learning algorithms (LSTM and CNN) using wor2vec as word embedding tool, and transformer-based classifiers (AraBERT Base, AraBERT Mini, AraBERT Medium, DistilBERT Multilingual, and AraGPT-2) in classifying social media posts into positive or negative sentiment categories. In addition, we introduce an innovative ensemble architecture that combines multiple pre-trained models for enhanced performance. The system leverages DistilBERT and AraBERT Base as encoders, while AraGPT-2 serves as the decoder. These components work in concert through a sophisticated stacking and voting mechanism. The results reveal promising accuracy rates, with the AraBERT Base model achieving 87.9%, the LSTM model 85%, and the SVM classifier 82%. The stacking model attained an accuracy of 91.1%, while the majority voting model reached 90%. This research contributes valuable insights into sentiment analysis for dialectal Arabic, with practical implications for real-world applications across various sectors.

Blockchain technology has emerged as an innovative solution to enhance food safety in the food supply chain process by ensuring effective distributed data management and providing consumers with traceability and transparency of food products. However, the scalability of the blockchain reduces the time efficiency of traceability queries, as the search must be performed sequentially across blocks. This paper addresses this issue and proposes integrating the B-tree indexing technique with blockchain technology to optimize traceability queries and improve the time efficiency of inter-block searches. Initially, we designed a new blockchain network with a modified transaction data structure that includes fabrication time as a new attribute. Then, we created a function to compute the minimum and maximum fabrication times within each block, thereby establishing a range [FabTime_MIN, FabTime_MAX] specific to the block’s transactions. These ranges are stored in the B-tree structure along with their corresponding block numbers and are continuously updated with each block creation until they encompass all the ranges and associated block numbers within the blockchain network. This process allows us to locate relevant blocks containing traceability data and then perform the traceability query only in these identified blocks. The simulation was conducted using the Hyperledger Fabric framework, which supports food supply chain scenarios. The obtained results show that the proposed solution improves search time without altering quality compared to the standard method.

Smart grid systems, which integrate traditional energy infrastructure with modern communication technologies, face significant cybersecurity challenges due to their dynamic architecture and continuous data exchange. The diversity and interconnection of devices increase vulnerability to malicious intrusions, highlighting the need for advanced and scalable detection methods. This study aims to develop an intrusion detection system (IDS) for smart grids by leveraging recent advances in deep learning, specifically enhanced variants of Long Short-Term Memory (LSTM)—xLSTM, sLSTM, and mLSTM. These sequence modeling architectures were adapted and fine-tuned within our IDS framework to capture complex spatio-temporal patterns and handle heterogeneous, high-dimensional data effectively. A comprehensive evaluation on two benchmark datasets, NSL-KDD and DNP3, demonstrates the robustness of the proposed approach. On the NSL- KDD, xLSTM, sLSTM, and mLSTM achieved accuracies of 98.16%, 98.55%, and 98.54%. On the more modern, protocol-specific DNP3 dataset, which represents real-world SCADA-focused attacks, the models maintained their superior performance, achieving accuracies of 99.50%, 99.33%, and 99.42%, respectively. The high and consistent accuracy across both datasets demonstrates the models' dependability and adaptability for intrusion detection in smart grid infrastructures. The study's targeted enhancement of LSTM-based architectures contributes a novel and effective approach to protecting critical intelligent systems from emerging cyber threats.

The characteristics of Food Supply Chains (FSC) make them hard to manage properly, and many efforts have been conducted to alleviate the difficulties related to their management, especially when it comes to integrating the latest Information and Communications Technologies. The Internet of Things (IoT) has shown to be very beneficial in providing a holistic and real-time vision of FSCs. Blockchain, with its decentralization and immutability, is another promising technology, that is showing a great potential in managing FSCs. A lot of research has been carried out to prove the advantages of each of these technologies on its own. However, the research investigating their adoption together is still not enough. Our paper presents a study of recent advances in the integration of IoT and Blockchain in Food Supply Chain Management (FSCM) over the past five years. We identify key research trends, analyze the benefits and limitations of IoT–blockchain integration, and highlight major challenges hindering large-scale adoption. Finally, we propose future research directions to address these challenges and improve the adoption of IoT–blockchain solutions in FSCs. This study aims to serve as a reference for researchers and practitioners seeking to understand and advance the integration of these emerging technologies in FSCM.

One of the most serious threats to global food security is plant diseases compromising agricultural productivity and threatening the livelihoods of millions. These diseases can decimate crops, disrupt food supply chains, and escalate the risk of food shortages, underscoring the urgency of implementing robust strategies to safeguard the world’s food sources. Deep learning methods have revolutionized the field of plant disease detection, offering advanced and accurate solutions for early identification and management. However, a recurring problem in deep learning models is their susceptibility to a lack of robustness and generalization when facing novel crop and disease types that were not included in the training dataset. In this paper, we address this issue by proposing a novel deep learning-based system capable of recognizing diseased and healthy leaves across different crops, even if the system was not trained on them. The key idea is to focus on recognizing the diseased small leaf regions rather than the overall appearance of the diseased leaf, along with determining the disease’s prevalence rate on the entire leaf. For efficient classification and to leverage the excellence of the Inception model in disease recognition, we employ a small Inception model architecture, which is suitable for processing small regions without compromising performance. To confirm the effectiveness of our method, we trained and tested it using the widely acclaimed PlantVillage dataset, recognized as the most utilized dataset for its comprehensive and diverse coverage. Our method achieved an accuracy rate of 94.04%. Furthermore, when tested on new datasets, it achieved an accuracy rate of 97.13%. This innovative approach not only enhances the accuracy of plant disease detection but also addresses the critical challenge of model generalization to diverse crops and diseases. In addition, it outperformed the existing methods in its ability to identify any disease across any crop type, showcasing its potential for broad applicability and contribution to global food security initiatives.

In response to escalating cyber threats and privacy issues within the Industrial Internet of Things (IIoT), this research presents FedGenID, an advanced Federated Generative Intrusion Detection System, to safeguard IIoT networks. Our approach introduces a three-model framework: 1) a federated generative model, incorporating a Conditional Generative Adversarial Network (cGANs) for data augmentation, emphasizing only generator model updates to be shared among clients. This model uses a Wasserstein loss function with Gradient Penalty to amplify sample diversity, indicative of varying cyber threats. Concurrently, we address the issues of imbalanced and distributed data and deploy a data curation technique to align generated data within specific constraints. 2) A secondary model fine-tunes local Critics for enhanced resilience and detection of various adversarial attacks. 3) The third model focuses on precise cyber threat identification, leveraging augmented data for improved training under a synthetic federated learning schema, bolstering detection capability, especially against zero-day threats. Our evaluation of FedGenID, utilizing a novel industrial cybersecurity dataset, highlights its efficacy in non-IID, multi-class cyber threat detection and its resilience to adversarial attacks. Furthermore, we demonstrate how FedGenID can mitigate the negative impact of differential privacy-enhanced FL on model performance. The findings underscore FedGenID’s proficiency in detection accuracy, surpassing traditional FedID by 10% in the presence of zero-day attacks and high privacy regimes.

Dataset bias is a critical issue in medical machine learning models that can significantly impact the efficacy of disease prediction and diminish the ability to generate relevant clinical decisions. To address this problem, oversampling the underrepresented class presents a promising solution. This study investigates the impact of oversampling techniques, specifically SVM-SMOTE and BorderlineSMOTE, on both machine learning using K-Nearest Neighbors (KNN) algorithm and Ensemble Learning using Gradient Boosting Machine (GBM). we assessed changes in accuracy, precision, recall, F1 score, Positive Predictive Value (PPV), Equal Opportunity Difference (EOD), Disparate Impact (DI), and Imbalance Ratio (IR) across two medical datasets. In heart disease prediction using Gradient Boosting Machine (GBM), both SVM-SMOTE and BorderlineSMOTE significantly enhanced key metrics. BorderlineSMOTE maintained high accuracy at 0.85, with a notable increase in recall to 0.92., indicating improved identification of positive cases. Precision and F1 score also showed robust improvements, reflecting the models' enhanced ability to accurately classify instances across all classes. Also, the introduction of synthetic samples notably reduced EOD and DI values, indicating a reduction in bias within model predictions.

This paper proposes an arrhythmia classification system that combines 1D Convolutional Neural Network with three selected temporal features: QRS duration, and the current and next RR intervals. Our aim is to leverage the advantages of deep learning, which allows for automatic feature extraction, while also utilizing some crucial hand-crafted features. We tested the effectiveness of this model with the MIT-BIH Arrhythmia Dataset, adhering to the guidelines provided by the Association for the Advancement of Medical Instrumentation. The proposed hybrid approach demonstrates good performance in terms of accuracy, sensitivity, and specificity, while maintaining a simpler structure.