PeerJ Computer Science最新文献

Background: The occurrence of pandemics in the last 20 years highlighted the unpreparedness of healthcare systems. There is a worldwide increased trend in the vector borne diseases. Ticks are one of the most common organisms that play a vital role in global ecosystem as well as being vectors of diseases affecting human and livestock. They are able to carry infectious agents that might cause illnesses including paralysis and to some certain extend death. Therefore, it is crucial to identify different genera of ticks to track infectious agents. Conventionally, tick classification is done by acarologists who are experts in the field. For this reason, the identification process is carried out in a difficult and time-consuming manner.

Method: The aim of the study was to develop a web-based application by using artificial intelligence-based algorithms to easily identify Hyalomma and Rhipicephalus ticks, which are the most abundant genera in Northern Cyprus, with high sensitivity and accuracy. The experimental procedure is structured based on five phases. Phase 1 revolves around data collection in which pictures of 35 identified ticks are taken by experienced acarologists and the curation of non-tick images (spiders, beetles, mites, mosquitos and scorpions). Phase 2 revolves around pre-processing steps and data split. Phase 3 involves training and testing custom Convolutional Neural Network (CNN), Visual Geometry Group 16 (VGG16), Residual Network 50 (ResNet-50) using 6,972 images (3,486 images for each class) for discrimination between ticks and non-ticks and 9,556 images (4,778 images for each class) for the discrimination between Hyalomma and Rhipicephalus. Phase 4 revolves around performance evaluation. Phase 5 is characterized by development of a web-based application (I-TickNet), created to enable a widespread use of the tick classifier.

Results: The performance evaluation and comparison of the model performance has shown that ResNet50 achieved the best result for binary classification of tick and non-tick (experiment A) with accuracy of 100% and Area Under the Curve (AUC) score of 100%. Moreover, VGG16 achieved the best result for binary classification of ticks (experiment B) with an accuracy of 96.97% and AUC score of 99.55% respectively. All the three models were employed for the development of artificial intelligence/Internet of Things (AI/IoT) framework known as I-TickNet for real-time and on-spot classification of tick images. In conclusion, this study provided a web-based application that can identify two distinct tick genera with high accuracy and sensitivity. The application developed enabled a user-friendly interface to identify genera without requiring any expertise.

Accurate crop yield prediction is vital for ensuring food security and informing agricultural policy, particularly in wheat-dependent regions like Pakistan where manual estimation methods are labor-intensive and imprecise. This study introduces a novel algorithmic framework, MS-YieldStackNet, to predict wheat yield with high spatial resolution by integrating multispectral satellite imagery, in-situ soil analytics, and meteorological variables. A unified feature space is constructed using Normalized Difference Vegetation Index (NDVI) and Difference Vegetation Index (DVI), soil physicochemical attributes, and temporal climate patterns, processed through a stacked ensemble neural architecture (MS-YieldStackNet) combining three parallel feed-forward neural networks (FFNNs) and a Random Forest meta-learner. The model achieved robust performance with an R-squared of 0.81, Mean Squared Error (MSE) of 6,114.30 kg/ha, root mean squared error (RMSE) of 78.19 kg/ha, mean absolute error (MAE) of 59.07 kg/ha, and mean absolute percentage error (MAPE) of 3.55%, demonstrating its potential for precise and scalable crop yield forecasting.

Coffee, the world's most traded tropical crop, is vital to the economies of many producing countries. However, coffee leaf diseases pose a serious threat to coffee quality and sustainable production. Deep learning has shown strong performance in plant disease identification through automatic image classification. Nevertheless, reliance on a single convolutional neural networks (CNNs) architecture restricts feature variability and real-world generalization. Moreover, limited work has systematically combined feature selection/reduction with CNNs, which constrains the advancement of hybrid models capable of capturing complementary features while ensuring computational efficiency without accuracy loss. This article presents an enhanced deep learning-based framework for coffee disease classification incorporating a hybrid strategy that integrates CNNs and advanced feature selection algorithms. GoogLeNet and ResNet18 are paired for complementary feature extraction, Principal Component Analysis (PCA) and Singular Value Decomposition (SVD) are employed for dimensionality reduction, and ANOVA and Chi-square are applied to select the most informative features. An Adam optimizer (learning rate = 0.001, batch size = 20, epochs = 50) with early stopping is used for training. Experiments on the BRACOL dataset achieved 99.78% accuracy, with precision, recall, and F1-score all exceeding 99% across classes. To the best of our knowledge, this study systematically integrates GoogLeNet and ResNet18 with PCA/SVD dimensionality reduction and analysis of variance (ANOVA)/Chi-square feature selection, for coffee disease classification, thereby addressing a key gap in prior research.

Technological developments in recent years led to the emergence of increasingly sophisticated recommender systems to support multi-day travel itineraries that fall under the Tourist Trip Design Problem (TTDP). Various problem analogies are widely used to solve TTDP, such as Traveling Salesman Problem (TSP), Vehicle Routing Problem (VRP), Orienteering Problem (OP), and Team Orienteering Problem with Time Windows (TOPTW). For multi-day route recommendation, TOPTW is suitable as a problem analogy since there is a per-day travel duration constraint. So far, TTDP with TOPTW does not consider the weighting (priority level of users) for each requirement attribute in a multi-attribute-based TOPTW to ensure personalized recommendations. In addition, running time remains a challenge in many studies in the TOPTW area. Many metaheuristic algorithms have been adopted to TOPTW for generating a time-efficient approach. Komodo Mlipir Algorithm (KMA) emerges as a new algorithm that promises good scalability. Therefore, we propose KomoTrip, a method that adopts the discrete version of KMA and Multi-Attribute Utility Theory (MAUT) to recommend optimal travel routes per day by accommodating the multi-attribute preferences of users. We perform three evaluation scenarios, i.e., general performance, Degree of Interest (DOI) combinations, and varying numbers of Points of Interest (POI), consistently demonstrating that KomoTrip outperforms several benchmark algorithms in terms of computational time efficiency and also exhibits robust fitness values across different problem dimension scales. Thus, KomoTrip can be regarded as an efficient algorithm to recommend optimal multi-day tour routes, effectively incorporating weighted multi-attribute preferences into its optimization process. We further benchmarked KomoTrip against state-of-the-art TOPTW heuristics on the public Solomon dataset, where it demonstrated competitive profit values, particularly for a larger number of days (tours), and consistently achieved superior runtime performance.

Target detection in remote sensing is essential for applications such as law enforcement, military surveillance, and search-and-rescue. With advancements in computational power, deep learning methods have excelled in processing unimodal aerial imagery. The availability of diverse imaging modalities including, infrared, hyperspectral, multispectral, synthetic aperture radar, and Light Detection and Ranging (LiDAR) allows researchers to leverage complementary data sources. Integrating these multi-modal datasets has significantly enhanced detection performance, making these technologies more effective in real-world scenarios. In this work, we propose a novel approach that employs a deep learning-based attention mechanism to generate depth maps from aerial images. These depth maps are fused with RGB images to achieve enhanced feature representation. For image segmentation, we use Markov Random Fields (MRF), and for object detection, we adopt the You Only Look Once (YOLOv4) framework. Furthermore, we introduce a hybrid feature extraction technique that combines Histogram of Oriented Gradients (HOG) and Binary Robust Invariant Scalable Keypoints (BRISK) descriptors within the Vision Transformer (ViT) framework. Finally, a Residual Network with 18 layers (ResNet-18) is used for classification. Our model is evaluated on three benchmark datasets Roundabout Aerial, AU-Air, and Vehicle Aerial Imagery Dataset (VAID) achieving precision scores of 98.4%, 96.2%, and 97.4%, respectively, for object detection. Experimental results demonstrate that our approach outperforms existing state-of-the-art methods in vehicle detection and classification for aerial imagery.

We propose a self-supervised point cloud representation learning framework CrossAlignNet based on cross-modal mask alignment strategy, to solve the problems of imbalance between global semantic and local geometric feature learning, as well as cross-modal information asymmetry in existing methods. A geometrically consistent mask region is established between the point cloud patches and the corresponding image patches through a synchronized mask alignment strategy to ensure cross-modal information symmetry. A dual-task learning framework is designed: the global semantic alignment task enhances the cross-modal semantic consistency through contrastive learning, and the local mask reconstruction task fuses the image cues using the cross-attention mechanism to recover the local geometric structure of the masked point cloud. In addition, the ShapeNet3D-CMA dataset is constructed to provide accurate point cloud-image spatial mapping relations to support cross-modal learning. Our framework shows superior or comparative results against existing methods on three point cloud understanding tasks including object classification, few-shot classification, and part segmentation.

This work introduces a method for the semantic similarity of long documents using sentence transformers and large language models. The method detects relevant information from a pair of long texts by exploiting sentence transformers and large language models. The degree of similarity is obtained with an analytical fuzzy strategy that enables selective iterative retrieval under noisy conditions. The method discards the least similar pairs of sentences and selects the most similar. The preprocessing consists of splitting texts into sentences. The analytical strategy classifies pairs of texts by a degree of similarity without prior training on a dataset of long documents. Instead, it uses pre-trained models with any token capacity, a set of fuzzy parameters is tuned based on a few assessment iterations, and the parameters are updated based on criteria to detect four classes of similarity: identical, same topic, concept related, and non-related. This method can be employed in both small sentence transformers and large language models to detect similarity between pairs of documents of random sizes and avoid truncation of texts by testing pairs of sentences. A dataset of long texts in English from Wikipedia and other public sources, jointly with its gold standard, is provided and reviewed to test the method's performance. The method's performance is tested with small-token-size sentence transformers, large language models (LLMs), and text pairs split into sentences. Results prove that smaller sentence transformers are reliable for obtaining the similarity on long texts and indicate this method is an economical alternative to the increasing need for larger language models to find the degree of similarity between two long texts and extract the relevant information. Code and datasets are available at: https://github.com/omarzatarain/long-texts-similarity. Results of the adjustment of parameters can be found at https://doi.org/10.6084/m9.figshare.29082791.

Pandemics present critical challenges to global health systems, economies, and societal structures, necessitating the development of accurate forecasting models for effective intervention and resource allocation. Classical statistical models such as the autoregressive integrated moving average (ARIMA) have been widely employed in epidemiological forecasting; however, they struggle to capture the nonlinear trends and dynamic fluctuations inherent in pandemic data. Conversely, deep learning models such as long short-term memory (LSTM) networks demonstrate strong capabilities in modeling complex dependencies but often require substantial data and computational resources. To boost forecasting precision, hybrid models such as ARIMA-LSTM integrate the advantages of traditional and deep learning methods. This study evaluates and compares the performance of ARIMA, LSTM, and hybrid ARIMA-LSTM models in predicting pandemic trends, using COVID-19 data from the Malaysian Ministry of Health as a case study. The dataset covers the period from 4 January 2021 to 18 September 2021, and model performance is evaluated using key metrics, including mean squared error (MSE), mean absolute error (MAE), mean absolute percentage error (MAPE), root mean squared error (RMSE), relative root mean squared error (RRMSE), normalized root mean squared error (NRMSE), and the coefficient of determination (R²). The results demonstrate that ARIMA performs poorly in capturing pandemic trends, while LSTM improves forecasting accuracy. However, the hybrid ARIMA-LSTM model consistently achieves the lowest error rates, confirming the advantage of integrating statistical and deep learning methodologies. All findings support the adoption of hybrid modeling approaches for pandemic forecasting, contributing to more accurate and reliable predictive analytics in epidemiology. Future research should investigate the generalizability of hybrid models across various infectious diseases and integrate additional real-time external variables to improve forecasting reliability.

The proliferation of social media has intensified concerns about the societal and psychological impacts of hate speech, particularly in Arabic-speaking communities, where dialectal diversity, morphological complexity, and sociopolitical factors complicate detection. Despite platform efforts, the automated detection of Arabic hate speech remains challenging owing to limited annotated datasets and linguistic nuances. This survey reviews the advances (2020-2024) in deep learning approaches, including convolutional neural networks (CNNs), recurrent neural networks (RNNs), transformer-based models (e.g., bidirectional encoder representations from transformers (BERT) and AraBERT), and hybrid architectures for Arabic hate speech detection. It further examines the dataset constraints involving dialectal variation, annotation inconsistencies, and scarcity. The analysis identified critical research gaps and proposed future directions: expanding multilingual datasets, enhancing contextual modeling, and developing ethically grounded frameworks. This review consolidates state-of-the-art methodologies to guide effective countermeasures against Arabic online hate speech.

Background: Medication non-adherence remains a significant global healthcare challenge, resulting in inadequate disease management, increased hospitalisations, and higher healthcare costs. Mobile health (mHealth) applications have emerged as promising digital health tools for enhancing medication adherence through real-time monitoring, personalised reminders, artificial intelligence (AI)-driven interventions, and improved patient engagement.

Objectives: This scoping review examines the effectiveness, key features, and challenges of mHealth applications in promoting medication adherence across diverse patient populations and healthcare settings. It also seeks to identify research gaps and inform future development and implementation strategies for digital therapeutics.

Eligibility criteria: Studies published between 2020 and 2024 were included if they investigated the use of mHealth applications to improve medication adherence and reported outcomes related to adherence rates, patient health indicators, or user engagement. Only studies with empirical data, including randomised controlled trials, observational studies, or mixed-methods research, were considered.

Sources of evidence: A comprehensive search was conducted across Scopus, Web of Science, PubMed/MEDLINE, Google Scholar, and CINAHL databases. In total, 319 studies met the inclusion criteria following a systematic screening process based on Preferred Reporting Items for Systematic reviews and Meta-Analyses extension for Scoping Reviews (PRISMA-ScR) guidelines.

Charting methods: Data were extracted on study design, app functionalities, patient demographics, adherence outcomes, and barriers to adoption. The charted data were thematically synthesised to identify trends, success factors, and limitations.

Results: Among the included studies, 85% reported improved medication adherence associated with features such as personalised medication reminders, real-time health tracking, and AI-powered adherence prediction. Clinical outcomes were also frequently observed, including improved blood pressure, glucose control, and patient-reported quality of life. Key barriers to adoption included limited digital literacy, concerns about data privacy, socioeconomic disparities, and a lack of integration with electronic health records (EHRs).

Conclusions: mHealth applications show significant potential to improve medication adherence and health outcomes, particularly in the management of chronic diseases. However, inclusive design, robust data privacy frameworks, and evidence-based implementation strategies are essential for scalability and sustained impact. Future research should focus on long-term effectiveness, cost-efficiency, and integration of mHealth tools within broader healthcare systems.