Search Results (1,043)

OCR for low-resource languages is still held back by the same small number of issues: too little labeled image-text data, too few benchmarks, and thin language-specific tooling. Kurdish Sorani is a particularly awkward case. It is written in a modified Arabic script, runs right to left, and has orthographic habits that standard Arabic OCR engines handle poorly. This paper describes a transformer OCR system for Sorani trained almost entirely on synthetic data, meaning line images rendered on the fly from a text corpus rather than manually transcribed scans. The pipeline has three parts: corpus-driven line synthesis, a deterministic script-aware normalization step based on character-level transliteration, and a TrOCR encoder–decoder recognizer. Text lines are rendered with randomly sampled fonts and sizes, then passed through stochastic augmentation to mimic realistic distortions. The system is evaluated twice. On an in-distribution synthetic set of 200 rendered lines, the best model reaches a character error rate of 0.0434, a word error rate of 0.1246, and 64.0% exact matches. More importantly, on a real-world test set of 19 scanned Kurdish documents (467 lines, 28,468 characters) processed end-to-end through detection and recognition, it reaches a character error rate of 0.0305 and a word error rate of 0.1770, beating both Arabic and Kurdish Tesseract baselines and an existing Kurdish TrOCR model while being considerably smaller than the latter. A controlled ablation, in which eight variants are trained under one shared budget and scored on identical images, then isolates what each design choice contributes. The label space is the largest design effect, and the reason is concrete: the decoder’s pre-trained tokenizer has no representation for seven common Sorani graphemes, which cover 14.7% of the corpus and place a floor under any model trained on native-script labels. Corpus size dominates overall and behaves as a threshold, font diversity helps with diminishing returns, and stochastic augmentation buys robustness at a small cost in in-distribution accuracy. Aligning detected lines against the transcribed ones further shows that line detection contributes under 1% of the reported character error on this material. The broader point, at least for Sorani, is that the synthetic training data and the label space in which the model predicts have to be designed together: a compact recognizer built that way outperforms a substantially larger released Kurdish model on genuine document images. Full article

Whole-slide image (WSI) compression is a fundamental requirement in digital pathology. Yet, current validation metrics, such as PSNR and pathologist agreement, were developed before the emergence of pathology foundation models (PFMs). PFMs capture fine-grained tissue details in high-dimensional spaces that can be disrupted by compression artifacts invisible to the human eye, potentially affecting downstream tasks. To the best of our knowledge, we present a systematic study of model- and dataset-specific embedding-drift indicators using representation drift ($\Delta E$), a cosine-based similarity measure, across clinically motivated compression ratios. We evaluate three vision transformer models of identical ViT-Large architecture but different training data: DINOv2 (general vision, 142 million natural images), UNI (pathology, over 100,000 clinical WSIs), and Phikon-v2 (a second PFM). We use 4000 TCGA-BRCA H&E tumor tiles from two WSIs at six compression ratios, ranging from lossless to 80:1. The results demonstrate that UNI first exceeds the predefined $\Delta E>0.01$ sentinel criterion at the tested CR 10:1 operating point (PSNR $=37.76$ dB, usually considered excellent). DINOv2 first exceeds the same sentinel criterion at CR 20:1. At CR 20:1, UNI’s $\Delta E=0.109$ while DINOv2’s is $0.012$, a nearly tenfold difference at the same image quality. At CR 80:1, UNI’s cosine similarity drops to 0.316 while DINOv2 retains 0.895. Phikon-v2 first exceeds the same sentinel criterion at CR 10:1 and tracks UNI far more closely than DINOv2 ($\Delta E=0.311$ versus $0.037$ at CR 40:1). The two pathology-pretrained models exhibit substantially greater drift than DINOv2, a pattern consistent with an association between pathology-domain training and increased compression sensitivity. However, checkpoint-specific factors such as training objectives, preprocessing, learned invariances, embedding normalization, and training data may also contribute. Additional tests on a balanced 500-tile TCGA-LUAD pilot set, denser compression ratios, and an alternative JPEG2000 encoder show similar model-dependent drift patterns; however, the empirical drift values and operating criteria should be interpreted as model- and dataset-specific rather than as generalizable thresholds across tissues, institutions, scanners, or staining protocols. We introduce the Rate Distortion Representation (RDR) curve as a model-aware evaluation tool that reveals this representation blind spot: a compression range where PSNR remains conventionally acceptable while embedding geometry is substantially altered. These results are model- and dataset-specific and should not be interpreted as universal compression thresholds or clinical safety boundaries. Overall, the findings support the use of representation-level robustness assessment as a complement to image-level fidelity metrics in AI-oriented digital pathology workflows. Full article

Storm-surge residuals represent one of the most destructive marine-coastal hazards, and reliable short-term surge residual prediction is critical for coastal disaster preparedness. Conventional empirical forecasting approaches suffer from limited cross-regional generalization, while high-fidelity physics-based hydrodynamic models such as ADCIRC-SWAN can reproduce complete storm-surge physical processes but demand substantial computational resources. In this study, a three-layer back-propagation neural network (BPNN) for storm-surge residual forecasting is constructed, which is driven by output datasets from the validated ADCIRC-SWAN coupled hydrodynamic model. Wind speed, significant wave height, sea-surface atmospheric pressure, and the simulated current-time storm-surge residual are selected as input predictors. Simulation-derived samples are pre-processed via data cleaning and Min-Max normalization, and two different dataset partitioning strategies (random mesh-point-based partition and time-sequential partition) are implemented for comparative experiments. After hyperparameter sensitivity tests, the optimal network configuration with 30 hidden-layer neurons is determined. Model predictive performance is quantitatively evaluated via multi-station time-series comparison and universal statistical metrics including R, NSE, and RMSE. The results show that the BPNN achieves satisfactory performance under random mesh-point-oriented partitioning, yet obvious performance degradation occurs under time-sequential temporal extrapolation, with prominent underestimation of surge peaks. On the basis of BPNN-predicted spatial surge residual fields, storm-surge intensity grading is carried out following the Chinese national standard GB/T 39418-2020. Statistical comparisons between the full computational domain and the Pearl River Estuary sub-region reveal strong spatial aggregation of high-intensity storm-surge grids within the estuary driven by funnel-shaped topographic amplification. This work demonstrates the feasibility of using a BPNN as a surrogate emulator for hydrodynamic outputs under a given typhoon condition; however, limitations in temporal extrapolation performance still need to be addressed before this approach can be practically used in operational early-warning applications. Full article

Continuous radar surface velocity records require purpose-specific conditioning because the temporal information available for processing differs between real-time monitoring and retrospective analysis. Short-period fluctuations and spikes can obscure stage-related flow responses under both settings. We evaluated a stepwise quality control framework that separates these two processing roles using 10 min records from six monitoring sites in South Korea. Causal preprocessing combined Huber-weighted recursive least squares, fuzzy correction, and a trailing Hampel filter to generate a provisional series. Retrospective processing applied a centered Hampel filter followed by criterion-based zero-phase moving average smoothing. Causal preprocessing reduced the standard deviation of successive velocity increments by 14.1–53.4%, with a further 1.4–7.8% reduction observed after centered filtering. A three-point moving-average window was selected at all sites, retaining 98.4–99.9% of the peak velocity and a velocity sum ratio of 1.000. Three sites satisfied all the selection criteria, two satisfied the shape retention criteria, and one was retained under a flagged fallback because the increment variance and slope criteria were not met. Postprocessed index-velocity-method-derived hydrographs showed a lower RMSE and higher R² when compared to the operational stage–discharge benchmark at all sites, while signed biases varied by site. The proposed framework provides a traceable pathway from observation availability and data status to shape assessment and downstream discharge evaluation. Full article

Red mud, an alkaline industrial by-product of alumina refining generated in enormous volumes worldwide, poses a persistent environmental disposal challenge; using it as a partial cement replacement offers a promising route toward more sustainable concrete, but the resulting compressive strength is governed by complex, nonlinear interactions among the mix constituents that conventional empirical and regression-based models struggle to capture accurately. To address this challenge, the present study develops and compares four machine learning and deep learning models, namely the Deep Gradient Boosting Machine (DGBM), the Differentiable Neural Decision Tree (DNDT), Long Short-Term Memory (LSTM), and the Monte Carlo Dropout Neural Network (MCDNN), for the accurate and uncertainty-aware prediction of the compressive strength of red mud concrete. A dataset of 183 data points, compiled from the literature and supplemented with experimental results, was used to capture the influence of red mud content, curing period, and other mix parameters, including cement dosage, water content, and admixture proportions. The data were pre-processed prior to model training, and predictive performance was evaluated using R², RMSE, MAE, and WMAPE, among other indicators. The results show that the deep learning models outperformed the tree-based models: LSTM achieved the highest accuracy (R² = 0.942 on the testing dataset), while MCDNN additionally provided reliable uncertainty estimates alongside comparable prediction accuracy; DNDT and DGBM were comparatively less effective. Global sensitivity analysis identified fly ash and water content as the most influential contributors to strength development. By combining rigorous data-driven modeling with sensitivity and uncertainty analysis, this study contributes to the literature a validated, uncertainty-aware deep learning framework for sustainable concrete strength prediction, and offers practical value to the construction industry by providing engineers with a reliable, data-driven tool for optimizing red mud content in concrete mix design, thereby supporting the safe and wider industrial utilization of this problematic waste stream. Full article

When a water treatment plant, power grid, or pipeline is compromised, the consequences extend beyond data loss: a manipulated sensor reading can trigger physical damage, and a disabled safety interlock can endanger lives. Cyber–physical systems (CPSs) sit at this intersection of digital control and physical process, yet existing security reviews treat threats and defences in separate silos, leaving practitioners without a clear picture of which defences fail against which attacks, and why. This paper fills that gap with a structured narrative review of 82 sources (70 from the primary window January 2020 to April 2026, plus 12 foundational pre-2020 works), organised through the CPS Defence-Gap Taxonomy (CPS-DGT)—a framework that classifies 14 attack mechanisms by architectural layer and physical impact, evaluates six defensive technology categories against documented failure modes, and maps five governance dimensions to the institutional conditions required for deployment. Across five intrusion detection system (IDS) studies that differ in dataset, attack selection, training regime and evaluation scope, reported F1 scores lie between 0.796 and 0.969 under each study’s own standard conditions; these values are not a controlled comparison and are reported descriptively. For the one architecture evaluated under adversarial evasion, F1 falls by 37.4 percentage points in absolute terms, a relative reduction of 38.6%. The defence-gap matrix identifies seven entries with insufficient coverage. Five of the 14 attack mechanisms are uncovered: A03, A09, A10, A11 and A14. Two further mechanisms, A01 and A12, have only partial defences. Adversarial evasion of learned detectors is reported separately as a transversal failure mode of one defensive category rather than as an attack mechanism. The uncovered mechanisms cluster at the cyber–physical boundary and in supply-chain channels. We conclude with five concrete research challenges, each with a direct path from the identified gap to a tractable research agenda. Full article

Deep learning for Alzheimer’s disease (AD) detection from structural magnetic resonance imaging (MRI) needs large, labelled datasets, yet many cohorts hold only a few hundred participants, for which conventional augmentation adds little anatomical diversity. In a two-stage pipeline, a variational autoencoder compressed 256 × 256 coronal slices to a 32 × 32 × 8 latent space, and a class-conditional latent diffusion model under classifier-free guidance generated AD and cognitively normal (CN) images using 295 participants from the Alzheimer’s Disease Neuroimaging Initiative (ADNI). The pipeline reached a Kernel Inception Distance (KID) of 0.030 ± 0.002 and a bias-corrected Fréchet Inception Distance (FID_∞) of 43.82. A controlled ablation varying preprocessing alone improved KID by 0.0147 and precision by 0.069, both with 95% intervals excluding zero. FID did not separate the configurations. A ResNet-18 trained only on synthetic slices and tested on 44 held-out real participants (18 AD, 26 CN), each scored as the mean probability over twenty slices, reached an area under the curve of 0.779 ± 0.031 against 0.869 ± 0.027 for real data; the difference was not distinguishable at this sample size. No instance memorisation was found among 880 samples, and a size-matched control exposed a 27.7-percentage-point inflation in the standard memorisation metric. Preprocessing, therefore, measurably affects synthesis quality at the small-cohort scale, though not on every measure. Full article

Hyperspectral image (HSI) classification benefits from rich spectral information; however, high dimensionality of HSI data increases computational cost, noise sensitivity, and the risk of overfitting when labeled samples are limited. Most pretrained computer vision networks are designed for three-channel inputs, making direct application to hyperspectral cubes difficult. Conventional principal component analysis (PCA) ranks components by total variance without distinguishing useful signal variance from noise-related variance, which can reduce the reliability of the resulting representation when only a few components are retained. This paper proposes a data-augmented Noise-Adjusted Principal Component Analysis (DA-NAPCA) framework for deep learning-based HSI classification. By accounting for estimated noise covariance, NAPCA orders the transformed components by signal-to-noise ratio rather than total variance, while data augmentation mitigates the overfitting risk when labeled samples are limited. Unlike typical NAPCA/MNF applications, which select the number of retained components empirically, DA-NAPCA deliberately retains three noise-adjusted components to form a compact three-channel representation, enabling pretrained models designed for three-channel inputs to be fine-tuned without modifying their input layers. The framework is evaluated using a 3D convolutional neural network (3D-CNN) for spatial–spectral feature learning and a pretrained EfficientNet-B0 model for lightweight transfer learning. Although this paper uses 3D-CNN and EfficientNet-B0 as illustrative examples, the proposed DA-NAPCA framework is a representation-level preprocessing approach and does not require architecture-specific modification. Experiments conducted on the Indian Pines, University of Pavia, and Salinas datasets compare DA-NAPCA with RGB, band selection, PCA-based dimensionality reduction, and ablation variants. Across the three datasets, DA-NAPCA achieved mean overall accuracies of 93.11–94.71% with 3D-CNN and 95.93–97.44% with EfficientNet-B0. Compared with the second-best baseline method, DA-NAPCA improved overall accuracy by 2.75–7.58 percentage points with 3D-CNN and 1.28–2.12 percentage points with EfficientNet-B0. These results demonstrate that combining a compact noise-adjusted representation with spatial augmentation provides an effective input representation for deep learning-based HSI classification. Full article

Predicting academic performance in face-to-face computing and software engineering courses is hindered by the limited availability of fine-grained process data. This study proposes STEAP, a camera-based static–temporal fusion framework that integrates longitudinal classroom behavior sensing with conventional educational records. Classroom videos from 375 undergraduates enrolled in four computing-related courses were collected over nine teaching weeks. Camera-derived observable behaviors were organized into student-level weekly sequences and transformed into outcome-independent longitudinal representations. Multiple machine-learning classifiers were subsequently applied to predict students’ academic performance. Checkpoint-specific predictions were conducted at Weeks 3, 6, and 9, with each prediction using only the classroom behavioral information available up to the corresponding time point. Using the complete nine-week Temporal representation together with the pre-course Background variables, XGBoost achieved the strongest classification performance among the evaluated models, with an Accuracy of 0.867, a Macro F1 of 0.862, and an At-risk Recall of 0.924. The checkpoint analyses further indicated that classroom behavioral information collected during the early course stage already provided useful predictive information without incorporating behavioral observations from subsequent weeks. After further integrating pre-course background variables and regular assessment information, the final fusion model achieved an Accuracy of 0.896 and a Macro F1 of 0.895. Overall, longitudinal camera-derived classroom behavior provides complementary predictive information beyond conventional educational information and supports the feasibility of earlier academic-risk identification at different course checkpoints. Full article

Background/Objectives: Pediatric uroflowmetry is volume dependent, and low-volume voids can yield recordings that require repetition or cannot be interpreted confidently. To evaluate whether documented bladder volume-guided timing was associated with first-attempt pediatric uroflowmetry process adequacy in children undergoing evaluation for suspected non-neurogenic lower urinary tract dysfunction. Methods: This single-center retrospective workflow cohort included 110 toilet-trained children aged 5–12 years who underwent uroflowmetry for suspected non-neurogenic lower urinary tract dysfunction. The exposure was classified from contemporaneous pre-test documentation as bladder volume-guided timing (n = 55) or standard urge-based timing (n = 55). Expected bladder capacity (EBC) was calculated as (age + 1) × 30 mL. The primary process outcome was first-attempt voided volume ≥ 50% EBC. Repetition after an inadequate first attempt was treated as a deterministic workflow consequence rather than an independent endpoint. Analyses were observational and effect estimates were interpreted as associations. Results: Adequate first-attempt voided volume was documented in 50/55 children (90.9%) with bladder volume-guided timing and 39/55 (70.9%) with standard urge-based timing (unadjusted risk ratio 1.28, 95% confidence interval [CI] 1.06–1.55; Newcombe risk difference 20.0 percentage points, 95% CI 5.3–34.0). After adjustment for age, baseline urgency score, and time since last void, the association remained (adjusted risk ratio 1.27, 95% CI 1.06–1.53; p = 0.011), and the model converged without numerical warnings. Using the age-specific lowest acceptable voided volume, adequacy occurred in 53/55 children (96.4%) versus 46/55 (83.6%) (adjusted risk ratio 1.15, 95% CI 1.01–1.31; p = 0.040). Repetition after an inadequate first attempt occurred in 9.1% versus 29.1% and was treated as a direct consequence of primary-threshold failure; it was not tested independently. Workflow-time measures were exploratory. Conclusions: Documented bladder volume-guided timing was associated with greater first-attempt process adequacy. Adjustment for age, baseline urgency score, and time since last void did not materially change the estimate. The association was attenuated but remained directionally consistent when the age-specific lowest acceptable voided volume was used. Because the timing rule deliberately targeted the same volume construct as the primary outcome, this finding does not establish improved diagnostic accuracy, clinical decision making, or patient outcomes. Retrospective exposure classification and routine documentation further preclude causal interpretation. Documented bladder volume-guided timing was associated with higher first-attempt achievement of a prespecified voided-volume threshold than standard urge-based timing. Because the pathway targeted the same volume construct used to define the primary outcome and was non-randomized, these findings are interpreted as hypothesis-generating workflow data rather than evidence of improved diagnostic accuracy or downstream clinical benefit. Full article

Granite residual and slope-wash soils are important disaster-prone geological bodies in the hilly and mountainous areas of Hunan Province, China. Their engineering classification has long relied on manual visual inspection and laboratory testing, which is inefficient and subjective. In this study, an automatic classification method based on deep learning image recognition is proposed for granite residual and slope-wash soils in the mountainous areas of Hunan Province. First, a three-class primary classification scheme was established, comprising residual clay (RNC), residual sandy clay (RNSC), and residual gravelly clay (RNGC), based primarily on the gravel content of particles larger than 2 mm (RNC < 5%, RNSC 5–20%, RNGC > 20%). Second, 7678 geotechnical test records from 21 counties in Hunan Province were collected, and classification labels were assigned through a strategy combining manual verification and automatic inference using Random Forest (5-fold cross-validation macro F1 = 0.913). From approximately 10,096 original field images, 3380 pure soil image patches were retained after segmentation and screening. A dataset of 43,940 samples was then generated through two-stage preprocessing (including denoising and illumination correction) and 13-fold data augmentation. A CNN image classification model was constructed based on a ResNet18 backbone network pre-trained on ImageNet. On the independent test set (6591 images), the primary classification accuracy reached 91.46%, with a macro F1-score of 0.9128; the per-class F1-scores for RNC, RNSC, and RNGC were 0.921, 0.885, and 0.933, respectively. Grad-CAM visualization analysis demonstrated that the model’s attention was primarily focused on soil particle distribution regions rather than non-soil background areas, confirming the effective learning of mixed-grain features. The study shows that the combined application of transfer learning and 13-fold data augmentation can significantly improve classification performance under limited sample size conditions (an improvement of 15.33 percentage points compared to the baseline of 76.13%), demonstrating promising potential for engineering applications. Full article

The booming data volume of modern time-domain surveys demands fast, robust early classification of sparsely sampled light curves, as newly discovered transients typically have only a handful of observations. We compare classifiers for extremely sparse light curves (3–30 observations) on a benchmark of approximately 1.72 million segments spanning seven astrophysical classes from ZTF and ATLAS. Methods include handcrafted-feature approaches (XGBoost, feature-based Transformers) and end-to-end LSTM and Transformer models. A pre-trained end-to-end Transformer achieves test accuracy of 0.946 (macro F1 0.950), exceeding 90% accuracy with only seven observations, but falls to 0.513 without pre-training. XGBoost-Reduced (38 features, excluding LS descriptors) reaches 0.922, while XGBoost-Full (56 features) reaches 0.913. Reliability diagnostics confirm LS periods and false-alarm probabilities are unreliable on 3–30-point segments; restricting training and evaluation to ≥15 points does not reverse the full-scale preference for the Reduced catalog. On CPU, XGBoost runtime is dominated by feature extraction (ratio ≈ 16:1); adding LS descriptors increases total processing time by ∼7.8% (feature extraction by ∼7.0%) without a commensurate accuracy gain. A lightweight LSTM attains 0.847 accuracy with 0.2 M parameters. These results offer practical guidance for model selection in real-time survey pipelines. Full article

Wind direction time series exhibit angular periodic discontinuity, multi-scale non-stationary fluctuations and abrupt wind shifts, which hinder the precision of short-term forecasting for wind turbine yaw control. In this paper, a dual-branch forecasting framework based on Variational Mode Decomposition (VMD) and Transformer is developed to address the above drawbacks, with a hysteresis gating and zoning residual compensation module embedded for targeted error correction. First, sine–cosine encoding is adopted to eliminate the numerical discontinuity between 0° and 360° for wind direction angular data, and valid meteorological input features are screened to discard redundant covariates. Second, the sine–cosine-encoded wind direction sequence is decomposed into multiple band-limited intrinsic mode functions (IMFs) via VMD, extracting frequency-specific features that reduce non-stationarity and facilitate subsequent Transformer modeling. The standard Transformer encoder serves as the normal branch to capture long-range temporal dependencies across the whole time series, while a lightweight multilayer perceptron (MLP) constitutes the compensation branch to learn prediction deviations between baseline predictions and ground-truth values. The hysteresis gating unit activates residual compensation based on historical prediction errors and angular variation, without requiring access to the current ground-truth value, and compensation intensity is adaptively adjusted via the zoning strategy; relevant coefficients are optimized by random search. Verified on a real wind farm dataset consisting of 10,421 15 min sampling points, the proposed model achieves the lowest MAE of 9.64° among six benchmark models. For the improved genuine mutation samples (angle change > 70°), the model achieves a mean improvement of 6.02°. Ablation experiments verify that VMD preprocessing, the MLP compensation branch, and the hysteresis gating mechanism play indispensable roles in forecasting performance. The proposed framework can support accurate yaw control of wind turbines, and the decomposition–compensation workflow can also be generalized to other periodic non-stationary forecasting tasks. Full article

This study examines patterns of ridership recovery across Seoul subway stations following the COVID-19 pandemic and identifies the factors associated with recovery stagnation. To this end, we apply cluster analysis, Markov transition analysis, and binary logistic regression using subway ridership data from 2019 to 2025. We classify station-level recovery into four types: Entrenched Low-Recovery, Partial Recovery, Improved Recovery, and Over-Recovery. The results reveal substantial variation in recovery trajectories across stations. Entrenched Low-Recovery stations display strong path dependence and state persistence, whereas Improved Recovery and Over-Recovery stations exhibit lower state stability and higher transition probabilities. In addition, recovery in morning commuting and evening travel, restaurant density, and average sales per store reduce the likelihood of recovery stagnation, whereas stations with stronger transfer functions are more likely to remain in a low-recovery state. These findings indicate that post-pandemic urban rail recovery is not a uniform return to pre-pandemic ridership levels, but a heterogeneous and dynamic process shaped by station functions and surrounding urban conditions. From a policy perspective, this heterogeneity highlights the need for station-specific recovery strategies that account for differences in transportation and local characteristics. More broadly, by revealing the persistence and transitions of station-level recovery states, this study provides a dynamic perspective on post-pandemic urban rail recovery beyond point-in-time comparisons of ridership change. Full article

Spectral wave models such as SWAN provide reliable harbor wave fields but require repeated, computationally expensive runs during preliminary design. This study develops a domain-knowledge-guided, U-Net-based surrogate model that rapidly approximates two-dimensional wave fields from SWAN simulations. The model takes water depth, incident significant wave height, incident wave direction, and a signed distance field (SDF) encoding structure geometry as inputs, and it combines a dynamic-resolution preprocessing scheme that preserves the terrain aspect ratio with a shadow-weighted loss function that emphasizes the sheltered zone behind breakwaters. Trained on 240 SWAN scenarios that combine 8 idealized geometries with 5 incident directions and 6 wave heights (about 15.7 million grid points), the model reproduced the SWAN significant wave height with a coefficient of determination of R² = 0.8807 and a mean absolute error of 0.07 m over the whole domain under an untrained but within-range (interpolated) wave height condition, with comparable accuracy in the region of interest behind breakwaters (ROI R² = 0.85). In a real sea application to Sokcho Harbor benchmarked against SWAN rather than field data, the model reproduced the offshore to harbor wave height pattern; along entrance and propagation-axis transects, it followed the SWAN profiles with high correlation (R ≈ 0.94–0.95), while overestimating the innermost calm zone waves by a non-negligible margin. The model is thus suited to rapid wave-field screening in preliminary harbor design. Full article