Accurate prediction of Standard Penetration Test (SPT) blow counts from Cone Penetration Test (CPT) data is critical for reliable geotechnical characterization, particularly when SPT data are scarce or difficult to obtain. This study presents a data-driven framework that employs an Artificial Neural Network (ANN) to estimate the corrected SPT blow number ($N_{60}$) using key CPT parameters. The database was compiled from two construction sites in Nasiriyah, Iraq, comprising cone tip resistance ($q_c$), sleeve friction ($fs$), and effective overburden pressure ($\sigma_{vo}^{\prime}$) as input variables. Multiple ANN architectures were trained and validated, and optimal performance was achieved using one hidden layer with eight neurons, yielding a coefficient of determination ($R^2$) of 0.9967, and two hidden layers with six and sixteen neurons, achieving $R^2$ = 0.9976. Relative importance analysis indicated that cone tip resistance contributed 44% to the model’s predictive strength, followed by sleeve friction and effective overburden pressure, each accounting for approximately 26%. Sensitivity analysis confirmed that $N_{60}$ increases with higher input parameters, consistent with soil behavior principles. The ANN model demonstrated high accuracy and generalization capability across both sandy and clayey soils. Design charts derived from the trained model enable practical estimation of $SPT-N$ from CPT results, providing geotechnical engineers with a rapid and reliable tool for site characterization and preliminary design.

Rapid motorization and insufficient traffic management continue to intensify congestion in major Iraqi cities such as Baghdad, Basra, and Mosul, highlighting the need for intelligent mobility solutions. Traditional shortest-path algorithms, including Dijkstra and Bellman–Ford, remain limited by static edge weights and cannot respond to evolving traffic states. To address this limitation, this study develops a hybrid computational-intelligence framework that integrates a temporal-attention-enhanced recurrent neural network (RNN) for sequential travel-time prediction, an adaptive neuro-fuzzy inference system (ANFIS) for interpretable decision support, and a genetic algorithm (GA) for dynamic route optimization. A synthetic dataset reflecting diverse congestion patterns and diurnal fluctuations across major Iraqi road networks was constructed for evaluation. Experimental results show that the proposed model reduces mean absolute error by up to 32% in travel-time prediction and shortens route travel time by 15% compared with conventional shortest-path algorithms. These findings demonstrate the advantages of coupling predictive modeling with evolutionary optimization for improving urban mobility performance. The proposed framework offers a scalable basis for future intelligent transportation systems in developing urban environments.

Accurate fruit recognition in natural orchard environments remains a major challenge due to heavy occlusion, illumination variation, and dense clustering. Conventional object detectors, even those incorporating attention mechanisms such as YOLOv7 with attribute attention, often fail to preserve fine spatial details and lose robustness under complex visual conditions. To overcome these limitations, this study proposes DeepHarvestNet, a YOLOv8-based hybrid network that jointly learns depth and visual representations for precise apple detection and localization. The architecture integrates three key modules: (1) Efficient Bidirectional Cross-Attention (EBCA) for handling overlapping fruits and contextual dependencies; (2) Focal Modulation (FM) for enhancing visible apple regions under partial occlusion; and (3) KernelWarehouse Convolution (KWConv) for extracting scale-aware features across varying fruit sizes. In addition, a transformer-based AdaBins depth estimation module enables pixel-wise depth inference, effectively separating foreground fruits from the background to support accurate 3D positioning. Experimental results on a drone-captured orchard dataset demonstrate that DeepHarvestNet achieves a precision of 0.94, recall of 0.95, and F1-score of 0.95—surpassing the enhanced YOLOv7 baseline. The integration of depth cues significantly improves detection reliability and facilitates depth-aware decision-making, underscoring the potential of DeepHarvestNet as a foundation for intelligent and autonomous harvesting systems in precision agriculture.

Phenol is a persistent and toxic pollutant in industrial wastewater, demanding efficient and sustainable removal technologies. Conventional treatment methods often suffer from high operational costs, incomplete degradation, and secondary contamination. In this study, ZnO–Fe$_2$O$_3$ nanocomposites were synthesized using pulsed laser ablation in liquid (PLAL)-a clean, surfactant-free, and environmentally benign route—to develop eco-friendly adsorbents for phenol removal. The structural, morphological, and optical characteristics of the as-prepared nanoparticles were examined using X-ray diffraction (XRD), scanning electron microscopy (SEM), UV-visible spectroscopy, and zeta potential analysis. The 50:50 ZnO–Fe$_2$O$_3$ composite demonstrated moderate colloidal stability (-28.54 mV), nanoscale crystallinity, and a heterogeneous surface morphology conducive to adsorption. Batch adsorption experiments at an initial phenol concentration of 100 mg/L revealed a maximum removal efficiency of 68.44% under 600 laser pulses after 50 minutes of contact time. The consistent optical band gap values (2.48-2.50 eV) across all samples indicated structural and electronic stability. The enhanced adsorption efficiency was attributed to synergistic interfacial interactions between ZnO and Fe$_2$O$_3$ within the nanocomposite matrix. Although the present work is limited to batch-scale trials under fixed conditions, future studies will investigate the effects of pH, adsorption kinetics, isotherm behavior, and material reusability. Overall, the findings highlight the potential of PLAL-fabricated ZnO–Fe$_2$O$_3$ nanocomposites as sustainable adsorbents for aqueous phenol remediation.

Real-time face detection in crowded scenes remains challenging due to small-scale facial regions, heavy occlusion, and complex illumination, which often degrade detection accuracy and computational efficiency. This study presents an enhanced detection framework that integrates Slicing-Aided Hyper Inference (SAHI) with the YOLOv11 architecture to improve small-face recognition under diverse visual conditions. While YOLOv11 provides a high-speed single-stage detection backbone, it tends to lose fine spatial information through downsampling, limiting its sensitivity to tiny faces. SAHI addresses this limitation by partitioning high-resolution images into overlapping slices, enabling localized inference that preserves structural detail and strengthens feature representation for small targets. The proposed YOLOv11–SAHI system was trained and evaluated on the WIDER Face dataset across Easy, Medium, and Hard difficulty levels. Experimental results demonstrate that the integrated framework achieves Average Precision (AP) scores of 96.33%, 95.87%, and 90.81% for the respective subsets—outperforming YOLOv7, YOLOv5, and other lightweight detectors, and closely approaching RetinaFace accuracy. Detailed error analysis reveals that the combined model substantially enhances small-face detection in dense crowds but remains sensitive to severe occlusion, motion blur, and extreme pose variations. Overall, YOLOv11 coupled with SAHI offers a robust and computationally efficient solution for real-time face detection in complex environments, establishing a foundation for future work on pose-invariant feature learning and adaptive slicing optimization.

Sediment deposition in Iraqi water channels represents a persistent constraint on agricultural irrigation and industrial water supply systems. Existing predictive models often neglect the unique hydraulic and sedimentological conditions of arid-region channels, limiting their applicability. This study integrates controlled laboratory experiments with statistical modeling to establish an empirical equation that quantifies sediment deposition mass (D) as a function of flow velocity (V), sediment concentration (C), and channel slope (S). A series of 54 experiments were conducted in a recirculating flume under precisely monitored conditions, including triplicate trials to ensure statistical robustness. The resulting power-law model, D=0.024·V-1.32·C0.89·S-0.75, exhibited strong predictive capability with R²=0.93, identifying flow velocity as the dominant governing parameter (56% influence). Optimal channel slopes between 5° and 7° were found to minimize deposition. Field validation within the Al-Diwaniyah irrigation network confirmed the model’s reliability, achieving 89% agreement between predicted and observed deposition values. These findings provide a practical and region-specific framework for improving channel design and maintenance strategies in arid environments. Future extensions will incorporate computational fluid dynamics (CFD) simulations and IoT-based monitoring to support adaptive sediment management.

The integration of heterogeneous medical data remains a major challenge for clinical decision support systems (CDSS). Most existing deep learning (DL) approaches rely primarily on imaging modalities, overlooking the complementary diagnostic value of electronic health records (EHR) and physiological signals such as electrocardiograms (ECG). This study introduces MIMIC-EYE, a secure and explainable multi-modal framework that fuses ECG, chest X-ray (CXR), and MIMIC-III EHR data to enhance diagnostic performance and interpretability. The framework employs a rigorous preprocessing pipeline combining min–max scaling, multiple imputation by chained equations (MICE), Hidden Markov Models (HMMs), Deep Kalman Filters (DKF), and denoising autoencoders to extract robust latent representations. Multi-modal features are fused through concatenation and optimized using a Hybrid Slime Mould–Moth Flame (HSMMF) strategy for feature selection. The predictive module integrates ensemble DL architectures with attention mechanisms and skip connections to capture complex inter-modal dependencies. Model explainability is achieved through Local Interpretable Model-agnostic Explanations (LIME) and Shapley Additive Explanations (SHAP), enabling transparent clinical reasoning. Experimental results demonstrate superior performance, achieving 98.41% accuracy, 98.99% precision, and 98.0% sensitivity—outperforming state-of-the-art baselines. The proposed MIMIC-EYE framework establishes a secure, interpretable, and generalizable foundation for trustworthy AI-driven decision support in critical care environments.

Drones have a problem with command transmission under Ultra-Reliable Low Latency Communication (URLLC) requirements. This paper discusses minimizing Packet Error Rate (PER) in an Unmanned Aerial Vehicle (UAV) relay system that transmits commands under Ultra-Reliable Low Latency Communication requirements. The problem is solved through joint optimization of block-length allocation and UAV placement. To tackle these challenges, the optimization problem was split into two sub-problems to analyze the convexity and monotonicity of each. An iterative optimization algorithm for PER minimization was then formulated, combining the Alternating Direction Method of the Multipliers algorithm (ADMM) with the bisection search method through a perturbation-based iterative approach. Simulation results confirm that the proposed algorithm achieves up to 16.42% improvement in computation time and up to 57.14% in convergence speed compared to the algorithm using the bisection method alone for both problems, and it gives the same performance as that of the exhaustive search method.