Search Results (1,948)

The 4 km/15 min downward surface shortwave radiation (DSSR) product from Fengyun-4B (FY-4B)/AGRI shows great potential for solar energy assessment in China, but its applicability requires further validation. This study conducts a comprehensive evaluation of the FY-4B DSSR product over Guangxi for 2025, using ground-observed irradiance from a dense network of 101 photovoltaic (PV) power stations. The overall comparison shows a correlation coefficient (R) of 0.84, a root-mean-square error (RMSE) of 161.78 W/m², a relative prediction error (RPE) of 51.02%, and a mean bias error (MBE) of 56.23 W/m², indicating systematic overestimation. Seasonally, the largest discrepancies occur in spring (MBE = 85.27 W/m², RPE = 53.12%) and summer (R = 0.82, RMSE = 184.10 W/m²). Diurnally, retrievals are most reliable around 09:00–13:00 local time, deteriorating notably in the early morning and, especially, the afternoon and evening. Spatially, errors are larger in the hilly, elevated terrain of northwestern Guangxi (e.g., Hechi) than in flatter southern and coastal cities, with RPE rising from roughly 40–60% at lower elevations to around 80% above 600–700 m. Sky-condition classification confirms that data quality follows clear > cloudy > overcast sky, while AOD-binned analysis shows aerosol loading playing a secondary but non-negligible role, especially under high-AOD pollution events. Solar zenith angle (SZA) also strongly affects accuracy: R peaks around 0.75–0.8 in the 30–50° SZA range and drops below 0.4 beyond about 75°. This study offers the most spatially and dimensionally comprehensive validation of FY-4B DSSR over Guangxi to date, characterizing accuracy across seasonal, diurnal, spatial, cloud, aerosol, solar-geometry, and elevation dimensions using a denser ground-truth network than previously available, with direct implications for photovoltaic resource assessment and power forecasting in subtropical hilly regions. Full article

Accurate precipitation data are critical for hydrological and climatic studies in Guangdong Province, where complex terrain and frequent extreme rainfall pose substantial challenges. However, the performance of gridded precipitation products is still not well understood. This study evaluates nine products, including gauge-based (CHM_PRE, CN05.1, GMCP, NOAA CPC), satellite-based (IMERG-E, IMERG-F, TMPA RT, TMPA 3B42), and ERA5 reanalysis against NCDC observations from 2001 to 2019 using metrics including trend significance, correlation (R), root mean square error (RMSE), categorical statistics (POD, FAR, ETS), and relative bias across rainfall intensities. The results indicate that, based on validation against NCDC observations, CHM_PRE performs the best across all temporal scales, capturing significant increasing trends (p < 0.05) and achieving the highest consistency with observations at the annual (R = 0.99), monthly (R = 0.99), and daily (R = 0.89) scales. Using CHM_PRE as the reference, CN05.1 shows the highest spatial consistency with it, especially for extreme events. NOAA CPC exhibits the best performance in monthly event detection (ETS = 0.42; BIAS ≈ 1). Satellite products show acceptable performance at the monthly scale but exhibit intensity-dependent biases and high daily variability, with pronounced “light rain overestimation and heavy rain underestimation.” ERA5 shows limitations, particularly in its severe underestimation of extreme precipitation. CHM_PRE is thus identified as the most suitable dataset for Guangdong based on its agreement with NCDC observations. With CHM_PRE as the reference, CN05.1 provides a reliable alternative for spatial analyses; NOAA CPC performs the best in monthly event detection. Satellite products suit monthly use but require daily-scale caution; ERA5 shows a relatively poor performance. Full article

Canopy Leaf Area Index (CLAI) is a stand attribute containing information on the real-time health and growth potential of managed pine plantations. Current remote sensing techniques for quantifying CLAI rely on simple linear models applied to satellite multispectral imagery, or on techniques based on light detection and ranging (LiDAR) data that are costly and less frequently collected. This study demonstrates a convolutional neural network (CNN) approach to retrieving CLAI from 10 m Sentinel-2 multispectral imagery with a model trained on gridded LiDAR-based CLAI estimates. We demonstrate large gains in accuracy with the CNN compared to traditional linear models based on vegetation indices (e.g., Simple Ratio), but also clear shortfalls in model skill when predicting “blind” in some spatial domains that were completely excluded during model training. Pixel-scale root mean squared error ranged from 0.34 to 0.64 by domain when exposed to CLAI training data from all available spatial domains, but rose to 0.58–1.74 when predicting without prior domain-specific training. Prediction accuracy was consistently lower when applied to completely unobserved USGS LiDAR-based CLAI estimates. Traditional linear models, in contrast, had the advantage of usually lower prediction error across unobserved spatial domains (0.43–1.98), but with lower maximum accuracy. These results demonstrate a potential route for deploying more complex models for LiDAR “mimicry”, e.g., between data acquisitions widely separated in time, but advocate for the development and use of more stable generalized approaches for use in unobserved managed pine stands. Full article

Background: Miscarriage is a significant global health concern, affecting approximately 10–20% of recognized pregnancies and resulting in substantial physical and psychological consequences. Chronic inflammatory conditions such as periodontitis may contribute to pregnancy loss through systemic inflammatory and immune-mediated pathways. However, the combined predictive value of periodontal destruction and hematological inflammatory markers for miscarriage risk remains insufficiently understood. This study aimed to evaluate the potential of periodontal and hematological parameters for predicting miscarriage risk using artificial intelligence-based models. Methods: A total of 82 participants (41 women who experienced miscarriage and 41 healthy pregnant controls) were included in this study. Fourteen clinical variables, including maternal age, periodontal indices, and hematological inflammatory markers, were analyzed. Artificial Neural Network (ANN) and eXtreme Gradient Boosting (XGBoost) models were developed to predict miscarriage risk. Model performance was assessed using the coefficient of determination (R²), mean squared error (MSE), root mean square error (RMSE), mean absolute error (MAE), and Willmott’s Index of Agreement. Results: Clinical attachment loss (CAL) was identified as a primary predictor of miscarriage risk within the study cohort. Both machine learning models exhibited promising predictive potential; however, the XGBoost model showed enhanced performance compared to the ANN. XGBoost achieved an R² of 0.92088 and an MSE of 0.0235, indicating consistent predictive capabilities for this preliminary dataset. Conclusions: The results highlight a significant relationship between oral health, systemic inflammation, and adverse pregnancy outcomes. The integration of periodontal and hematological biomarkers within machine learning frameworks provides a non-invasive, cost-effective, and preliminary proof-of-concept approach for miscarriage risk assessment. These findings support the development of personalized risk prediction strategies and may facilitate earlier clinical interventions aimed at improving maternal and fetal health outcomes. Full article

Achieving high-accuracy path following remains challenging for an unmanned surface vehicle (USV) in narrow waterways with time-varying curvature and straight–curve transitions; fixed-parameter line-of-sight (LOS) guidance can cause delayed response, overshoot, and steady-state cross-track error. This paper proposes a curvature-coupled adaptive vector-field integral LOS (AVFILOS) guidance law. It incorporates curvature-adaptive guidance: a lookahead distance regulated by curvature and cross-track error and a field-source radius that contracts with curvature to strengthen centripetal correction in high-curvature regions. A fuzzy adaptive proportional–integral–derivative (PID) controller tracks surge speed and heading. A stability analysis establishes local exponential stability for straight and constant-curvature paths and local ISS with local uniform ultimate boundedness for time-varying curvature under a bounded-rate condition. Across six elliptical and sinusoidal cases, AVFILOS achieved an average root mean square error (RMSE($y_e$)) of 0.1325 m, reducing RMSE($y_e$) by 90.6%, 63.6%, and 37.2% compared with LOS, time-varying LOS (TLOS), and vector-field integral LOS (VFILOS), respectively. Its average maximum absolute cross-track error (Max($|y_e|$)) was 0.3478 m, with reductions of 88.2%, 48.2%, and 30.7%. The ablation and sensitivity results indicate that coupled adaptive mechanisms improve curved-path tracking and reduce overshoot. The simulations indicate that AVFILOS is promising for cross-track-error-sensitive USV navigation. Full article

Photon-counting constitute a promising technology for ultra-weak signal satellite communications. Considering the Poisson shot noise impairment, atmospheric turbulence fading, and multi-user interference, in this paper, a meta-learning-driven photon-counting multi-user single-input multiple-output (MU-SIMO) scheme is developed and analyzed. Referred to as meta-learning-driven signal detection (Meta-SD), this scheme can achieve rapid convergence with limited samples and significantly improve system detection performance. Simulation results demonstrate that the proposed meta-learning scheme outperforms the mean square error based signal detection (MSE-SD) baseline, in terms of detection accuracy, robustness to signal-dependent Poisson shot noise, convergence speed, and generalization to few-shot detection tasks with previously untrained signal classes. Specifically, Meta-SD achieves nearly a tenfold reduction in BER, compared with the derived MSE-SD benchmark, in a $4\times 8$ MU-SIMO scenario at $E_{\mathrm{s}}=-140$ dBJ. Full article

Generative AI (GenAI) adoption in higher education (HE) raises significant ethical concerns. The focus is shifting from rules- or outcomes-based learning environments towards the development of moral character, personality traits, integrity, and practical wisdom (phronesis). However, most existing AI ethics validation instruments are predominantly shaped by Global North paradigms. In Global South HE contexts, in particular, open distance e-learning (ODEL) HE institutions (HEIs) characterised by limited direct supervision and a digital divide, validation remains scant. Ethical risks are intensified by the adoption and integration of GenAI tools, such as large language models (LLMs), to enhance teaching, learning, research, and student support, thus recognising the need to develop and validate virtue ethics scales. The current paper attempts to address this gap by validating the Virtue Ethics Measurement Scale (VEMS) within South Africa’s largest comprehensive ODEL institution. Guided by the positivist paradigm, a 36-item cross-sectional survey of 503 undergraduate and postgraduate students measured six virtue dimensions (justice, honesty, responsibility, care, prudence, and fortitude). Confirmatory factor analysis (CFA) compared four competing models. The single-factor model showed poor fit, rejecting unidimensionality. A second-order hierarchical model demonstrated an acceptable fit (χ²/df = 2.992, CFI = 0.933, RMSEA (Root Mean Square Error of Approximation) = 0.063, SRMR (Standardized Root Mean Squared Residual) = 0.043) with subscale reliabilities ranging from Cronbach’s α = 0.84 to 0.90, supporting a multidimensional yet hierarchical virtue structure. The VEMS offers a psychometrically sound instrument for evaluating ethical AI use in ODEL institutions. This aligns with virtue ethics theory, which emphasises that moral character is a constellation of dispositions (e.g., honesty, care, prudence) rather than a single trait. The VEMS thus enables HEIs to assess students’ virtues, design targeted ethics capacity-development programmes, and inform policy reform for responsible GenAI adoption in under-researched Global South HE settings. Full article

Electrospinning is a widely used nanofabrication technique capable of producing fibers with a range of diameters, morphologies, and porosities through the adjustment of experimental parameters. However, achieving reliable fiber diameter tuning remains challenging because of the complex, nonlinear interdependence among multiple electrospinning variables. In this study, a data-driven methodology is proposed that integrates predictive machine learning (ML) modeling with evolutionary algorithm-based optimization, specifically employing a genetic algorithm (GA), to predict fiber diameter and guide electrospinning parameter selection across nano- and microscale ranges. A curated dataset comprising 388 data points from 30 scientific publications was developed, focusing exclusively on polyacrylonitrile (PAN) dissolved in dimethylformamide (DMF). Multiple ML models were trained and tested to predict fiber diameter as a function of key electrospinning parameters. Among the evaluated ML models, the eXtreme gradient boosting (XGB) model achieved the highest predictive performance, yielding a coefficient of determination (R²) value of 0.93 with low prediction errors (root mean square error [RMSE]: 127.76 nm, mean absolute error [MAE]: 56.27 nm) on the test set. Experimental validation was performed by fabricating electrospun PAN nanofibers under one independent set of conditions, with scanning electron microscopy (SEM) showing close agreement between predicted and actual fiber diameters. The trained XGB model was subsequently integrated with a GA to identify electrospinning parameter sets for user-defined target fiber diameters ranging from 100 to 2000 nm. The evolutionary optimization process exhibited rapid convergence with low fitness error when evaluated using the trained predictive model. Overall, this study demonstrates the potential of a data-driven methodology to generate model-guided candidate conditions for target-driven PAN-DMF electrospinning, subject to broader experimental validation. Full article

Reliable navigation in GPS-denied environments remains a critical challenge for autonomous vehicles (AVs), particularly in complex indoor and urban settings. GPS-based localization systems often fail under these conditions, highlighting the need for resilient multimodal solutions. In this article, we present a radar-assisted tracking system that integrates LiDAR and inertial measurements within a sensor-fusion architecture to achieve robust navigation. The principal methodological contribution is a unified tracking and prediction framework that combines Bayesian state estimation with learning-based temporal prediction, enabling accurate tracking while continuously forecasting the slave robot’s short-term future state from mapping observations generated by the master robot, with a typical end-to-end perception-to-action latency of 20–60 ms. The communication and prediction forecasting module operates with an update interval below 35 ms, enabling real-time cooperative robotic operation. Sensor data are fused through a pipeline incorporating Gaussian Mixture Models (GMMs) for post-processing, which helps mitigate the limitations associated with individual sensors during edge processing. Moreover, Kalman filtering is employed to mitigate sensor noise and drift, thereby improving state estimation accuracy through trajectory smoothing. The fused spatiotemporal information is subsequently exploited by a Convolutional Recurrent Neural Network (CRNN) coupled with a Nonlinear Autoregressive model with eXogenous Inputs (NARX) to model the robot’s motion dynamics and provide short-horizon state prediction. Through simulations and real-world indoor experiments conducted in GPS-denied environments, we validate the system’s ability to provide accurate and continuous pose estimation with low localization errors. Experimental results show that the proposed framework achieves root-mean-square errors of 0.12 m, 0.15 m, and 0.28 m along the X, Y, and Z axes, respectively, while maintaining sub-meter maximum position deviations throughout the evaluated trajectories. These results confirm that the proposed framework provides reliable localization and predictive state estimation for cooperative robotic navigation in indoor GPS-denied environments. Future work will investigate outdoor validation and extend the framework to additional data-driven decision-making models for future robotic services. Full article

Temperature–time profiles obtained through thermistor-based monitoring provide a rich but noise-sensitive source of information for estimating kinetic and thermal parameters in exothermic batch reactions. Conventional workflows typically combine deterministic smoothing with numerical differentiation, an approach that can amplify measurement noise and fail to propagate preprocessing uncertainty into the resulting reaction-rate and parameter estimates. To address these limitations, this study presents an uncertainty-aware Bayesian machine-learning framework that integrates scalable random-Fourier-feature Gaussian-process (RFF–GP) smoothing, analytical differentiation, temperature-derived apparent conversion, Bayesian parameter inference, posterior validation, predictive calibration, model comparison, ablation, sensitivity analysis, probabilistic benchmarking, simulation of thermal nonideality, and endpoint diagnostics. The framework was applied to $379,631$ cleaned thermistor observations. The production RFF–GP achieved a validation root-mean-square error of $0.04805\mathrm{K}$, yielding a stable latent temperature trajectory and an uncertainty-aware estimate of $dT/dt$. On a smaller matched subset, exact Gaussian-process regression achieved the highest predictive accuracy and the best probabilistic scores, whereas the RFF–GP reduced central-processing-unit runtime by approximately $4.1$-fold and remained applicable to the larger production fit. A Monte Carlo dropout neural comparator produced larger prediction errors and substantially wider predictive intervals. Six apparent thermokinetic structures were evaluated using mean-field variational inference, after which the nth-order and autocatalytic structures were validated using the No-U-Turn Sampler (NUTS). Under mean-field variational inference, the apparent autocatalytic structure achieved the lowest point estimate of the widely applicable information criterion (WAIC), the lowest derivative-domain error, and the lowest full-profile temperature-reconstruction root-mean-square error of $0.2920\mathrm{K}$. Its posterior obtained using NUTS yielded $E_a=40.98\mathrm{kJ}{\mathrm{mol}}^{-1}$, $k_{\mathrm{ref}}=0.005815{\min }^{-1}$, $\Delta T_{\mathrm{ad}}=56.11\mathrm{K}$, $m=0.1694$, and $n=1.0784$. The sampling diagnostics indicated satisfactory convergence, large effective sample sizes, and no divergent transitions. Although the MFVI posterior means and NUTS posterior medians were similar, variational inference produced narrower uncertainty intervals for several correlated parameters. Moving-block bootstrap intervals did not establish a decisive separation in WAIC among the leading structures. Expanded sensitivity, ablation, imperfect-insulation simulation, and endpoint-holdout analyses further showed that the apparent parameter estimates were sensitive to optimization, thermal nonideality, sensor response, and Gaussian-process boundary behavior. The autocatalytic formulation should therefore be interpreted as the best-performing apparent structure among the candidates tested rather than as evidence of a unique chemical mechanism. Overall, the framework extracted physically plausible apparent thermokinetic information from noisy temperature-only measurements while explicitly quantifying uncertainty arising from prediction, parameter estimation, model form, computation, thermal nonideality, and boundary behavior. Full article

In this study, we assessed the impact of prior-induced regularization using six real datasets from wheat, rice, and potato, spanning 107–758 genotypes, 2–12 environments, 1–18 traits, and 2744–108,024 molecular markers. Two modeling scenarios were evaluated: (i) unimodal genomic prediction based solely on marker information and (ii) multimodal (multi-component) prediction integrating genomic, environmental, and genotype-by-environment (G × E) effects. Predictive performance was evaluated using Pearson’s correlation (COR) and normalized root mean squared error (NRMSE) under 10 repeated random 50% training–50% testing partitions, representing prediction of untested lines in tested environments. Bayesian genomic prediction (BGP) relies on prior distributions to regulate shrinkage and stabilize inference in high-dimensional settings. We evaluated whether predictive performance was driven primarily by the type of Bayesian prior or by the presence of effective prior-induced regularization. Across most datasets, regularized Bayesian models achieved higher predictive correlations and markedly lower NRMSE than the weakly regularized or unregularized baseline. Differences among regularized prior families were generally modest, whereas weakening or removing regularization frequently produced unstable estimates and inflated prediction error. Predictive results were obtained for both winter-wheat datasets as well as for the rice, potato, and DMario datasets. In multimodal analyses, models with coherent regularization across genomic, environmental, and genotype-by-environment components were generally more accurate and stable than configurations in which regularization was absent or weakened in key components. Rice_Kim_2020 was an informative exception in which the baseline remained competitive. These results show that the principal empirical contrast is the presence versus absence of effective prior-induced regularization, rather than a universal ranking of Bayesian prior families. Appropriate regularization should therefore be treated as a central model-design decision in genomic prediction. Full article

Recommender systems are widely used in e-commerce to model user preferences and deliver personalized services. However, conventional interaction-based recommendation approaches fail to fully capture users’ underlying motivations and emotional tendencies. User-generated reviews offer complementary sentiment information, yet integrating sentiment semantics with interaction modeling remains challenging due to feature heterogeneity and sparsity. This paper proposes a hybrid recommendation framework (HAIRec) that jointly models review sentiment and interaction information within a unified architecture. The model extracts sentiment representations, latent rating factors, user preference features, and item attributes through multi-source feature learning. To address the semantic heterogeneity between review sentiment and interaction behavior, a hierarchical heterogeneous feature fusion architecture is proposed. Specifically, a dual-level attention mechanism is designed to jointly model global feature importance and fine-grained semantic interactions, designed to facilitate multi-granularity preference learning across heterogeneous domains. In addition, a Deep & Cross Network V2 (DCN-V2) is adopted to explicitly and implicitly model high-order feature interactions. Extensive experiments are conducted on five real-world subcategories of the Amazon Product Dataset (including Movies and TV, Digital Music, Musical Instruments, Toys and Games, and Home and Kitchen). The experimental results demonstrate that the proposed HAIRec framework consistently outperforms several representative baselines (such as DeepCoNN, NARRE, and LightGCN). Notably, compared with the strong LightGCN benchmark, HAIRec achieves significant rating prediction improvements, reducing the Mean Squared Error (MSE) to 0.884 on Movies and TV, 0.882 on Digital Music, and 0.863 on Toys and Games, which corresponds to maximum performance gains of up to 7.14% in MSE reduction. Full article

This study presents a comprehensive global accuracy assessment of the medium-range (up to 15 days) forecast pressure-level data generated by the physics-based Integrated Forecast System (IFS) and the newly operational, data-driven Artificial Intelligence Forecasting System (AIFS) from the European Centre for Medium-Range Weather Forecasts (ECMWF) based on the radio wave signal delays during propagation through the troposphere. Troposphere signal path delays are calculated using the software package Ankara Ray-tracing Tools (ART). This newly developed troposphere ray-tracing software package integrates hydrostatic and wet refractivities along the ray path of a radio wave signal using an approximation of a two-dimensional piecewise-linear ray path. Along with the IFS and AIFS pressure-level data, the AIFS/IFS combination generated and appended in this study is used to compute 62 forecast runs for each of the January and August 2025 monthly periods. Each run includes 6-hourly forecast steps over 15 days and is initialized twice daily at 0 and 12 UT throughout January and August 2025. These forecasts cover 52 globally distributed Global Navigation Satellite Systems (GNSS) stations operated by the International GNSS Service (IGS). The forecast zenith delay accuracies were systematically evaluated using the root mean square (RMS) and bias error metrics with respect to the IGS troposphere product and the ECMWF Operational Analysis data as robust validation benchmarks. In addition to the Vienna Mapping Functions 3 (VMF3) troposphere delay product, the empirical troposphere delay models Global Pressure and Temperature 3 (GPT3) and the model utilized by satellite-based augmentation systems (SBAS, e.g., WAAS and EGNOS) GNSS receivers are incorporated into the assessments. Both IFS and AIFS models exhibit exceptional short-range capabilities, keeping global zenith total delay errors (RMS relative to IGS) below 2 cm up to a 2-day lead time. However, a critical performance crossover occurs between the 10-day and 11-day forecasting horizons, where the forecast accuracy of both IFS and AIFS declines below the threshold of the GPT3 model, whose zenith total delay RMS across all stations with respect to the IGS troposphere product is found to be about 4 cm. The findings of this study offer crucial insights for improving the accuracy of real-time satellite navigation, climate monitoring, and satellite-based high-precision positioning applications. Full article

Irrigation water management in arid systems requires spatially distributed estimates of crop evapotranspiration (ET) that fixed crop coefficients cannot provide. The actual ET of flood-irrigated rice (Oryza sativa L.) on the arid northern coast of Peru was mapped with the METRIC surface energy balance model (Mapping EvapoTranspiration at high Resolution with Internalized Calibration) on Google Earth Engine (GEE). Ten cloud-free Landsat 8/9 scenes (January–July 2022) were processed over 113 ha at Ferreñafe (Lambayeque) on the 30 m product grid, onto which the 100 m native thermal observation was resampled, with internal calibration based on automatic anchor-pixel selection and hourly ERA5-Land data. Daily field-mean ET ranged from 4.2 to 8.1 mm d⁻¹, peaking during flooding and establishment and declining towards harvest. Because the same reference $E{T}_o$ underlies the METRIC internal calibration and the FAO-56 estimate, this is a comparison between two modelling approaches rather than an independent validation. Against the FAO-56 reference ET, METRIC showed a positive bias of $+0.65$ mm d⁻¹ (percent bias (PBIAS) $=+13\%$; root mean square error (RMSE) $=1.23$ mm d⁻¹; $r^2=0.57$; $n=9$, after excluding one date with anomalous reanalysis forcing), concentrated during flooding and after harvest, whereas at full canopy cover the two estimates converged. Two global ET products that share neither the METRIC formulation nor the ERA5-Land forcing reproduce the same seasonal decline once the canopy closes ($r=0.63$ and $0.91$) but stay far below in magnitude, as expected from their 500 m pixel. ET did not differ between sowing methods and varied only slightly among cultivars (∼0.3 mm d⁻¹), against marked intra-field variability. The METRIC–GEE workflow offers a low-cost, high-resolution tool for monitoring water use in data-scarce arid rice systems. Full article

The peak height of the F2 layer (hmF2) is a key parameter describing the vertical structure of the ionosphere. It is important for high-frequency radio communication planning and space-weather background assessment, particularly at high latitudes. To improve long-term hmF2 prediction, an empirical model-guided CatBoost model is developed. Predictions from the SHU and E-CHAIM empirical models are incorporated as prior predictors, while spatiotemporal periodicity, solar-activity, and geomagnetic-activity indices are jointly considered to represent the primary drivers of hmF2 variability. A two-stage feature selection procedure, combining stability-based selection and correlation-based redundancy pruning, is employed to identify informative and nonredundant features. The proposed model achieves consistently lower errors than both empirical models. Relative to SHU and E-CHAIM, the proposed model achieves relative root mean square error (RMSE) reductions of 22.67% and 17.70%, respectively, and relative mean relative error (MRE) reductions of 23.21% and 17.79%, respectively. Consistent improvements are observed across different time periods, seasons, and solar-activity conditions. The largest performance gains occur during spring and years of high solar activity. These results demonstrate that integrating empirical-model information with machine learning effectively improves the representation of hmF2 variability at high latitudes. The proposed model provides an effective empirical model-guided approach for long-term spatiotemporal prediction of hmF2 in high-latitude regions. Full article