This paper presents a hybrid model for medium- and long-term electricity demand forecasting, developed with the aim of achieving accurate forecasts on horizons ranging from a few weeks to several months. Univariate deep models show a significant drop in performance at higher time horizons and are unable to adequately model long-term nonlinearities and changes in electricity demand. To address these limitations, an approach is proposed that combines deterministic time-series decomposition, residual modeling using deep dilated convolutional networks (TitanResNet), and additional residual correction using the LightGBM model. In addition, the meta-model integrates two complementary components in order to achieve greater robustness and stability on long horizons.The proposed approach achieves strong long-horizon forecasting performance on the investigated dataset, especially on horizons from 720 to 2160 h, where MAPE≈4% and R2>0.95 under the adopted chronological protocol. By using climatological averages to generate future features, the framework preserves causality and remains applicable when true future meteorological measurements are unavailable. The results indicate that residual learning, deep convolutions, tree-based correction, and calibration-stage fusion provide a robust forecasting framework, while the broader generality of the approach requires validation on additional datasets.

In the development of Global Navigation Satellite System (GNSS) Positioning, Navigation, and Timing (PNT), short-lived, fast-growing ionospheric storms (lasting up to three days) were assumed to constitute a single, uniform class of space weather events. To challenge this premise, here is presented a circular analysis of single-frequency commercial-grade Global Positioning System (GPS) horizontal positioning errors in the mid-latitude region during four observed storms (one in 2015 and three in 2017), revealing significant heterogeneity. Maximum positioning errors reached 7.495 m in March 2015, but peaked at 4.261 m, 4.387 m, and 3.948 m during storms in May, early September, and late September 2017, respectively. By demonstrating that short-duration, fast-growing storms require subclassification, in this study is offered a new perspective on disturbances.

This paper presents a class-specific U-Net-based segmentation framework designed for high-precision segmentation across diverse image domains, including underwater inspection, surface crack detection, and medical imaging. Unlike traditional multi-class models, our approach trains a separate U-Net for each class, enabling domain-specific tuning and threshold optimization. The proposed framework includes a Dice (F1 score) loss function tailored for multi-image performance, domain-customized preprocessing, and a stratified training pipeline. Evaluations across three distinct datasets, including underwater concrete inspection, the Roboflow surface crack dataset, and clinical cardiac magnetic resonance imaging (MRI), demonstrate consistent performance improvements over state-of-the-art methods, such as the You Only Look Once (YOLO) model and the baseline U-Net. Our results show that this modular segmentation strategy achieves superior Dice coefficients, Intersection over Union (IoU), and class precision, particularly for small or ambiguous structures. The computational trade-offs and the method's potential for real-world deployment in healthcare and infrastructure monitoring are discussed to prove the model's applicability.

Vessel detection and multi-object tracking (MOT) in marinas remain challenging because of occlusions, small targets, cluttered backgrounds, and changing illumination. This paper presents a controlled, feasibility-focused comparative evaluation of YOLO-driven vessel detection and tracking in a fixed-camera marina environment, comparing YOLOv11 with the more attention-oriented YOLOv12 detector family. A dataset of 3546 annotated images was collected using a static ground-level camera at a marina in the northern Adriatic, Croatia. YOLOv11s, YOLOv11m, YOLOv12s, and YOLOv12m were fine-tuned using transfer learning, evaluated across five random seeds, and integrated into a common tracking-by-detection pipeline with Kalman filter (KF) and extended Kalman filter (EKF) motion models. Detection was assessed using COCO-style mAP metrics, while tracking was evaluated using MOT and HOTA-based metrics. YOLOv11m achieved the highest mean test-set mAP@50–95 (0.7458 ± 0.0027) and significantly outperformed YOLOv11s and YOLOv12s. Although YOLOv12m did not achieve the highest frame-level mAP, it obtained the highest mean AssA and IDF1 when averaged across the two motion models. KF achieved higher mean HOTA, DetA, AssA, MOTA, and IDF1 than EKF, whereas EKF achieved only marginally higher MOTP; the KF advantage was statistically significant for IDF1. Overall, the results show that frame-level detection accuracy does not necessarily determine downstream tracking performance, thereby supporting the joint evaluation of detector and motion-model choices.

Deep learning has demonstrated exceptional learning capabilities, leading to the development various deep unfolding networks for image reconstruction. However, current deep unfolding networks often replace certain steps of traditional optimization algorithms with neural networks, thereby compromising the interpretability of the optimization algorithms. Additionally, each iteration in the unfolding process may result in certain image information loss, negatively impacting image reconstruction quality. This paper proposes a deep unfolding Alternating Direction Method of Multipliers (ADMM) network named LSRA-CSNet for compressive sensing image reconstruction, incorporating a long-short term residual optimization mechanism. The LSRA-CSNet is constructed by stacking multiple stages, with each stage consisting of a Fast ADMM Block (FAB) and a Residual Optimization Block (ROB). In FAB, inspired by the Woodbury matrix identity, we propose a fast version of the ADMM algorithm. Meanwhile, instead of replacing certain steps of the ADMM with neural networks, we leverage CNNs to replace some matrix operations. ROB consists of the Short-Term Residual Refinement Module (SRRM) and the Long-Term Residual Feedback Module (LRFM), which optimize the reconstruction details by leveraging inter-stage image residuals and multi-stage measurement residuals, respectively. Experiments on four datasets show the effectiveness of LSRA-CSNet, demonstrating superior reconstruction accuracy compared to existing CS image reconstruction networks.

Collision course detection plays a vital role in enhancing safety in personal watercraft rental services, particularly in tourist regions where inexperienced operators frequently navigate busy waterways. Collision course detection is defined here as forecasting personal watercraft trajectories up to 30 seconds ahead and identifying potential intersections that may result in collisions. This work builds on previous research by the authors, introducing deep learning approaches for personal watercraft trajectory forecasting on real-world data with RNN (Recurrent Neural Network), LSTM (Long Short-Term Memory), and GRU (Gated Recurrent Unit) architectures, an attention-based model, and the foundation model Unified Time Series (UniTS) for time-series data. Trajectory forecasting is integrated into collision detection, replacing the previous Intelligent Distance Control (IDC) system by analyzing data from 527 locations across continents collected over several years, and analyzing key variables such as longitude, latitude, speed, and heading. The attention model using longitude and latitude offsets, with the longest forecasting horizon of thirty seconds, increased the true negative rate to 95.45%, which is a significant improvement compared to the non-machine learning approach. This study applies forecasting in real-time safety systems, replacing static proximity-based rules with advanced learning for adaptive and accurate collision prevention. The GRU attention model achieves inference under 12 milliseconds per sequence, enabling real-time deployment.

Elliptic Curve Cryptography (ECC) is widely used for secure communication, digital signatures, and key exchange. While its performance is well-studied in standard cryptographic contexts, less is known about its behavior on structured sensitive data at the attribute level in databases. This paper presents a hybrid experimental analysis combining cryptographic benchmarking with lightweight machine learning to evaluate ECC performance across diverse attribute types. Encryption and decryption times were measured using secp256r1 and brainpoolP256r1 curves, with X25519 included for reference key exchange. Data-level features such as length, entropy, and character composition were analyzed using Random Forest regression to assess their influence on execution time, with the machine learning component used for interpretability rather than predictive deployment. Results show that curve choice significantly affects performance, encryption scales linearly with dataset size, and data-level characteristics have minimal impact. These findings support simplified selective encryption strategies in database systems, reducing the need for data-aware cryptographic tuning while maintaining secure and predictable performance.

Sea-state estimation (SSE) supports safe, efficient, and autonomous maritime operation. Conventional sources, including wave buoys, satellites, radar systems, and metocean products, are valuable but cannot provide continuous, local estimates. Ship-motion-based SSE offers a complementary solution by using the vessel as a wave-sensing platform through the wave buoy analogy (WBA). This review examines machine-learning (ML) approaches to SSE from ship-motion responses published since 2018. The studies are synthesized by learning objective, data source, representation, and modeling strategy, covering regression, classification, directional wave-spectrum estimation, wave-elevation reconstruction, transfer learning, domain adaptation, and hybrid physics-ML methods. The synthesis shows that ML-based SSE has evolved from feasibility studies toward pipelines using simulated, experimental, and in-service data. Under controlled conditions, wave height and wave-period quantities are estimated more reliably than direction, spectral-shape parameters, secondary systems, and directional spectra. Full-scale studies demonstrate promise but reveal limitations from proxy labels, uneven coverage, missing vessel-state information, sensor heterogeneity, and temporal domain shift. Transfer learning and hybrid physics-ML can improve data efficiency and trustworthiness when domains are physically compatible and estimates are checked against WBA-based response consistency. Future priorities include richer operational datasets, deployment-oriented evaluation, uncertainty and explainability mechanisms, collaborative learning, multimodal representations, and foundation-model-inspired approaches grounded in vessel physics.

Maritime navigation safety is the basis for global logistics and marine ecosystems. The increasing availability of Automatic Identification System (AIS) data has opened new avenues for forecasting vessel trajectories with higher precision and robustness. This review presents a comprehensive survey of recent data-driven approaches, including probabilistic models, classical statistical methods, and deep learning (DL) architectures. Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Recurrent Neural Network (RNN), Transformer-based models, and hybrid systems are evaluated for maritime trajectory prediction. By categorizing studies based on their input features (speed, course, and position) and computational methods, this review highlights methodological trends and performance benchmarks. Additionally, an overview of dataset accessibility and contextual variable use, including weather and wave state, is presented to support future reproducibility. We conclude by identifying gaps and future directions toward explainable and adaptive maritime forecasting systems.

Nature-inspired metaheuristic algorithms are commonly applied to complex combinatorial optimization problems where exact methods are computationally impractical. Tourist route optimization is a representative multi-objective problem characterized by realistic constraints such as travel time, cost, opening hours, and transportation modes. Although Mushroom Reproduction Optimization is computationally efficient, it often experiences premature convergence in complex search spaces. This paper proposes a novel hybrid algorithm, Mushroom-Weed Hybrid Reproduction Optimization (MWHRO), which integrates the colony-based local search of the Mushroom Reproduction algorithm with the fitness-proportional reproduction and competitive elimination mechanisms of Invasive Weed Optimization. Hybridization enhances population diversity and global exploration while preserving fast convergence. The proposed algorithm is evaluated based on a realistic tourist route optimization problem using real-world data from Zagreb, Croatia, across multiple transportation modes and objective-weight scenarios. Performance is compared against Ant Colony Optimization, Invasive Weed Optimization, Particle Swarm Optimization, and standard Mushroom Reproduction Optimization under equal evaluation budgets. Experimental results demonstrate that the proposed MWHRO algorithm consistently achieves high-quality solutions with significantly lower execution times, particularly in constrained and multimodal scenarios. Statistical analysis confirms the robustness and practical suitability of the proposed approach for real-world route optimization.

Identifying kinematic anomalies is critical for maritime safety, especially in high-traffic areas. This paper presents a novel, lightweight, scalable method for trajectory clustering and kinematic anomaly detection using inflection-point sequences across varying trajectory lengths and environments. By encoding trajectory shape changes into compact fingerprints, the technique circumvents the computational burdens of deep learning. Results indicate that the proposed fingerprinting method can differentiate anomalous driving patterns, offering a viable solution for real-time anomaly detection. This study presents a simple yet robust trajectory analysis framework applicable to personal watercraft monitoring and to larger maritime vessels, paving the way for improved safety and real-time decision-support systems. Trajectories are segmented, inflection points are extracted, and clustering is performed using trajectory fingerprints. Human evaluation experiments validate the approach aligned with Intelligent Distance Control (IDC), and a larger window size improves accuracy to 85% or higher when aligned with the rate of turn (ROT).

This study evaluates the performance of YOLOv7 with modified mosaic augmentation, which resizes images without cutting, against newer YOLO versions (YOLOv8, YOLOv9, YOLOv10, and YOLOv11) when trained with an unbalanced dataset in the context of sewage inspection. The modified mosaic augmentation helps preserve small, underrepresented objects in unbalanced datasets which are often cut off with original mosaic augmentation. Results show that YOLOv7 with modified mosaic augmentation outperforms YOLOv7 and YOLOv8 in precision and recall metrics, while YOLOv9 slightly surpasses it. However, YOLOv10 and YOLOv11 offer the best overall performance, in terms of both precision and recall. These findings show the importance of model selection based on application-specific needs and highlight the potential of modified mosaic augmentation for handling class imbalances. Future research should validate these findings across different datasets.

We address the challenge of abandoned luggage detection by adapting a YOLOv11 model and validating it on the IITP20 dataset. Our methodology involves optimizing our pre-trained YOLOv11 model for the domain adaptation to specific characteristics of the IITP20 dataset. Our initial analysis revealed significant semantic inconsistencies within the dataset’s annotations; a large portion of “abandoned” luggage instances were simply items in lockers or storage areas, which do not constitute security threats. To create a more realistic benchmark, we systematically corrected these inaccuracies. A comparative analysis reveals that our person and baggage detection model, along with the proposed algorithm, achieves state-of-the-art results, demonstrating exceptional performance in scenarios relevant to security surveillance. These findings underscore the robustness of our detection approach and emphasize the importance of high-quality, semantically accurate annotations for developing reliable abandoned object detection systems in real-world applications.

In this paper, we propose a system for the real-time automatic detection of abandoned luggage in an airport recorded by surveillance cameras. To do this, we use an adapted YOLOv11-s model and a proposed algorithm for detecting unattended luggage. The system uses the OpenCV library for the video processing of the recorded footage, a detector, and an algorithm that analyzes the movement of a person and their luggage and evaluates their spatial and temporal relationships to determine whether the luggage is truly abandoned. We used several popular deep convolutional neural network architectures for object detection, e.g., Yolov8, Yolov11, and DETR encoder–decoder transformer with a ResNet-50 deep convolutional backbone, we fine-tuned them on our dataset, and compared their performance in detecting people and luggage in surveillance scenes recorded by an airport surveillance camera. The fine-tuned model significantly improved the detection of people and luggage captured by the airport surveillance camera in our custom dataset. The fine-tuned YOLOv8 and YOLOv11 models achieved excellent real-time results on a challenging dataset consisting only of small and medium-sized objects. They achieved real-time precision (mAP) of over 88%, while their precision for medium-sized objects was over 96%. However, the YOLOv11-s model achieved the highest precision in detecting small objects, corresponding to 85.8%, which is why we selected it as a component of the abandoned luggage detection system. The abandoned luggage detection algorithm was tested in various scenarios where luggage may be left behind and in situations that may be potentially suspicious and showed promising results.

Monitoring and predicting personal watercraft trajectories is a novel and largely unexplored research area where any development is valuable for various rental services. Unlike existing work focused on specific maritime routes, this study introduces a location-agnostic deep-learning approach capable of generalizing across diverse environments. This is achieved by using an innovative preprocessing approach including offset and scaling. A novel Long Short-Term Memory (LSTM) bidirectional and convolutional model developed by the authors for binary peptide classification was adapted to accommodate the forecasting of continuous values. By leveraging Recurrent Neural Network (RNN) architectures with LSTM and Gated Recurrent Unit (GRU) layers, and cutting-edge attention-based and foundation models, we benchmark trajectory forecasting performance using real-world data from 1282 rental sites worldwide. Most notably, we extend the applicability of the listed models and the original LSTM bidirectional and convolutional models to maritime trajectory estimation, eliminating the need for training separate models for different locations while achieving superior predictive accuracy. Our results demonstrate that foundation models and attention mechanisms significantly outperform traditional methods, offering a robust and scalable solution for watercraft trajectory forecasting. These findings pave the way for intelligent monitoring systems that enhance maritime safety and operational efficiency.

Incomplete signals defined on graphs present a significant challenge for their analysis and processing. This paper presents a method for missing graph signal values reconstruction on directed graphs, whose adjacency matrices are often not diagonalizable. We perform modification of the adjacency matrix, i.e., adding new edges, with the aim of achieving diagonalizability, allowing the application of spectral analysis - i.e. graph Fourier transform as well as advanced reconstruction algorithms. The presented results confirm the effectiveness of the proposed method.

In this work, vessel position is estimated using a Bayesian approach based on heading, speed, time intervals, and offsets of latitude and longitude. An additional approach using a Markov chain is presented. The trajectory data comes from a cloud-based marine watercraft tracking system that enables remote control of the vessels. Wave height and meteorological reports were used to evaluate the impact of weather on personal watercraft trajectories. One proposed approach to trajectory estimation uses the longitude and latitude offsets, while another uses the speed, heading, and actual time intervals. A long-term forecasting window of up to ten seconds is achieved by dividing trajectories into segments that do not overlap. The limitation this method faces in long-term forecasting inspires more sophisticated machine-learning approaches. The most successful estimation method used one or two previous actual values and a Bayesian approach, proving that using previously predicted values in a chain accumulates errors. Considering environmental variables did not improve the model, highlighting that small watercrafts operate well even in unstable sea states. This occurs because they generate and ride waves, having a larger impact than oceanic currents.

The challenges of managing water resources in tidal rivers, exacerbated by climate change and anthropogenic impacts, require innovative approaches for accurate estimation of hydrological parameters. In tidal rivers and estuaries, water levels depend primarily on river discharge and tidal dynamics. Microtidal estuaries are particularly complex due to the strong stratification and two-layer structure, which also affect the water level. This study investigates the potential of machine learning (ML) models for estimating discharge in the Neretva River, Croatia, using only water level data from multiple stations. Comparative analyzes were performed between simple supervised models - Decision Tree (DT), Random Forest (RF), Support Vector Regression (SVR), Light Gradient Boosting Machine (LGBM), Extreme Gradient Boosting (XGB) - and time series models - Long Short-Term Memory (LSTM) and LSTM-Attention. Both simulated and measured data sets were used for this purpose. The results show that time series models perform satisfactorily in the assessment of discharge and overcome the challenges faced by simple supervised models, especially under high flow scenarios. Overall, LSTM-Attention proves to be the best model when analyzing all error metrics with superior performance over the entire range of discharge values. It surpasses the overall LSTM model performance, with a percentage increase of above 9% in RMSE and MAE, above 0.2% in NSE, and above 0.1% in R for both simulated and measured datasets.

In the field of muscle fatigue models (MFMs), the prior research has demonstrated success in fitting data in specific contexts, but it falls short in addressing the diverse efforts and rapid changes in exertion typical of soccer matches. This study builds upon the existing model, aiming to enhance its applicability and robustness to dynamic demand shifts. The objective is to encapsulate the complexities of soccer dynamics with a streamlined set of parameters. Our refined model achieved a slight improvement in the R2 score in the maximum hand-grip test, increasing from 0.87 to 0.89 compared to the existing model. It also demonstrated dynamic change robustness in a soccer-specific 1 min drill and 15 min treadmill protocol extracted from the literature. Through individualized fitting on a 10-repetition 80 m sprint test for a soccer player, the model exhibited R2 scores between 0.62 and 0.80. Furthermore, when tested with actual soccer match data, it maintained a robust performance, with the average R2 scores ranging from 0.70 to 0.72. The proposed approach holds the potential to advance the understanding of tactical decisions by correlating them with real-time physical performance, offering opportunities for more informed strategies and ultimately enhancing team performance.