Size‐, sex‐ and site‐dependent morphological variation in the stone flounder (Kareius bicoloratus) from Korean coastal waters Authors: Eun Jeong Kim, Songhyun Gwon, Ju Hwan Park, In Chul Bang, Yoon Kwon Nam Published: 2026 ABSTRACT Abstract Kareius bicoloratus is a commercially important flatfish in Korean coastal fisheries and an ecologically relevant benthic species in nearshore ecosystems. To evaluate how multiple biological and environmental factors jointly shape flatfish morphology, we analysed 81 morphometric indices and two‐dimensional geometric morphometrics (GMM) in wild K. bicoloratus collected from two Korean coastal regions, incorporating size, sex and sampling site. Classical morphometrics based on a three‐way ANOVA framework identified body size and sex as the primary drivers of shape variation, particularly influencing surface area‐, depth‐ and angular traits, whereas geographic effects were more limited and context dependent. GMM analyses supported these patterns. Procrustes ANOVA detected significant shape effects of size, sex and site, and CVA revealed clear multivariate separation among groups, with Mahalanobis distances of 3.28–5.49 among sex groups and Procrustes distances of 0.032–0.049. Geographic divergence was also evident in GMM space, with Procrustes distances of 0.033–0.036 between sites. Allometry analyses indicated heterogeneous allometric trajectories for size and sex but not for site. Mitochondrial DNA analyses revealed moderate genetic differentiation between regions, and Mantel tests detected a weak but significant association between genetic and morphological distances. These findings provide a robust morphological baseline for K. bicoloratus.Keywords: morphometrics, mtDNA, sex, site, size, stone flounder INTRODUCTION Morphometric analysis has long served as a fundamental tool for understanding biological variation in fish populations. Quantitative assessments of body shape and proportional traits provide critical insights into ecological adaptation, developmental trajectories and population differentiation (Cadrin, 2000; Luo, 2024; Tripathy, 2020). In fisheries science and aquaculture, morphometric characters are routinely applied to stock monitoring, broodstock selection and hybrid identification, thereby linking field‐based ecological observations with practical management and production objectives (He et al., 2020; Ndiwa et al., 2016; Park et al., 2020; Yershov et al., 2022). Flatfishes (Order Pleuronectiformes) represent one of the most morphologically diverse groups of teleosts, characterized by pronounced asymmetry, benthic specialization and complex ontogenetic transitions associated with metamorphosis and habitat shifts (Dierickx et al., 2023; Munroe, 2014; Park, 2021). The stone flounder Kareius bicoloratus (Basilewsky, 1855) is a commercially valuable demersal species widely distributed across the Northwest Pacific, including Korea, Japan and the Russian Far East (Jun & Im, 2004). Although large‐scale aquaculture production of this species has not yet been fully established, interest in K. bicoloratus has increased due to its potential use in selective breeding, hybridization and triploidization studies aimed at improving aquaculture performance traits (Jung et al., 2019; Kim, Park, et al., 2025). A robust understanding of phenotypic variation in natural populations is therefore essential not only for fisheries management but also for guiding future aquaculture development. Recent studies, including Kim, Kim, et al. (2025), have begun to characterize morphological variation in K. bicoloratus along the Korean coastline, primarily through comparisons among local populations. While these studies provide valuable baseline information, they largely rely on univariate or single‐factor approaches and do not explicitly test how multiple biological and environmental factors interact to shape morphology. In flatfishes, intrinsic factors such as ontogenetic allometry, sexual dimorphism and reproductive condition, together with extrinsic drivers such as habitat type and local oceanographic conditions, collectively influence morphometric patterns (Cadrin & Silva, 2005; Currie et al., 2025; Shuai et al., 2018). Because these factors rarely operate independently, analytical frameworks that neglect interaction effects may yield incomplete or potentially misleading interpretations of morphological variation. Despite their importance, multifactorial approaches capable of jointly testing size‐, sex‐ and site‐specific effects—and their interactions—have not yet been applied to K. bicoloratus. This gap limits the establishment of reliable morphological baselines for natural stocks and constrains the broader application of morphometric data in aquaculture, selective breeding programmes and ecological assessments. To address these limitations, we examined stone flounder collected from two coastal regions of Korea—Gadeokdo (Southeastern Sea) and Taean (Yellow Sea)—that differ markedly in oceanographic and ecological characteristics. These contrasts provide a strong environmental gradient for evaluating habitat‐associated morphological divergence. Importantly, robust multifactorial morphometric analyses require strict control of confounding biological factors, including body size distribution, sex composition and seasonal physiological condition. Such standardization is achievable only when sampling is concentrated at a limited number of well‐characterized sites where sufficient numbers of similarly sized individuals can be collected within the same season. The selection of these two regions, therefore, reflects both ecological contrast and methodological necessity. In addition, this study focused on wild‐caught individuals, given that dedicated aquaculture strains or stable hatchery‐produced stocks of K. bicoloratus have not yet been established in Korea. In this study, we integrate classical truss‐based morphometrics with two‐dimensional geometric morphometrics (GMM) to obtain a high‐resolution assessment of multivariate shape variation across biological (size, sex) and environmental (site) factors. Specifically, we aim to (1) identify morphometric traits significantly influenced by total length (TL), sex or sampling site; (2) test interaction effects among these factors using a multifactorial statistical framework; and (3) evaluate shape variation using both linear measurements and landmark‐based GMM, including analyses of allometry and size‐independent shape components. By combining parametric and non‐parametric approaches with residual‐based GMM analyses, this study provides a comprehensive assessment of morphological variation in K. bicoloratus and establishes a robust baseline for future ecological, developmental and aquaculture‐related research. MATERIALS AND METHODS Ethical statement The care and use of experimental animals complied with The Animal Protection Act of the Republic of Korea and the guidelines of the Institutional Animal Care and Use Committee (IACUC), as approved by the IACUC of Pukyong National University (Approval No. PKNUIACUC‐2024‐75). Fish sampling and classification A total of 101 stone flounders (Kareius bicoloratus) were collected from two coastal regions of the Korean peninsula: Site G (Gadeokdo, South Sea; n = 50) and Site T (Taean, Yellow Sea; n = 51) (Table 1; Figure S1). The two sites are geographically separated and differ markedly in long‐term oceanographic and habitat characteristics, including tidal regime, sediment composition and hydrodynamic conditions, based on publicly available environmental datasets and published literature (Table S1). These contrasting settings were selected to evaluate potential geographic structuring of body shape within the natural distribution range of the species. TABLE 1Summary of sampling sites and collection metadata for Kareius bicoloratus.SiteLocationSample size (F/M/U)Collection periodNotesG: GadeokdoN: 34°55′–35°05′ E: 128°40′–128°55′17F/9M/24UNov–Dec 2024South Sea (microtidal regime; coarse‐grained sandy sediment)T: TaeanN: 36°30′–36°50′ E: 126°00′–126°25′30F/21M/‐Nov–Dec 2024Yellow Sea (macrotidal regime; mixed sand‐mud bottom) Note: Geographic coordinates indicate the approximate spatial extent of the sampling areas, based on the sampling location map (Figure S1). Sex categories include female (F), male (M) and underdeveloped individuals (U), classified based on visual inspection of eggs and sperm using a custom‐designed catheter and/or by direct hand‐stripping of gametes. Long‐term environmental characteristics (2020–2024) of the two sampling regions, derived from publicly available databases, are summarized in Table S1 to support ecological interpretation of morphological variation. Summary of sampling sites and collection metadata for Kareius bicoloratus. Note: Geographic coordinates indicate the approximate spatial extent of the sampling areas, based on the sampling location map (Figure S1). Sex categories include female (F), male (M) and underdeveloped individuals (U), classified based on visual inspection of eggs and sperm using a custom‐designed catheter and/or by direct hand‐stripping of gametes. Long‐term environmental characteristics (2020–2024) of the two sampling regions, derived from publicly available databases, are summarized in Table S1 to support ecological interpretation of morphological variation. All specimens were wild‐caught by local commercial fisheries using bottom trawl nets or hook‐and‐line methods during the late autumn to winter period (November–December 2024), thereby minimizing seasonal variation in growth and reproductive condition. Sampling locations represent nearshore demersal habitats typical for K. bicoloratus in each region and encompass a limited spatial extent within each site (Figure S1). Phenotypic sex determination was conducted by visual inspection of eggs and sperm using a custom‐designed catheter and/or by direct hand‐stripping of gametes. Individuals for which reproductive products could not be confirmed were classified as underdeveloped (U). Based on these criteria, 17 females (F), 9 males (M) and 24 underdeveloped individuals were identified at Site G, whereas 30 females and 21 males were identified at Site T, where no underdeveloped individuals were observed (Table 1). To assess ontogenetic (size‐dependent) effects on body shape, preliminary evaluations—including skewness checks, k‐means clustering and coefficient of variation (CV) analyses—were conducted to identify an appropriate size metric. TL was selected as the most robust grouping variable due to its strong correlation with body weight (r = 0.902). TL‐based size classes were defined using percentile‐based quantile binning into four categories: small (S), small‐medium (SM), medium‐large (ML) and large (L), comprising 26, 26, 24 and 25 individuals, respectively. These size classes were treated as a categorical factor in subsequent statistical analyses, together with sex (F, M, U) and sampling site (G, T). Line morphometric analysis Body measurements and data acquisition The body weight of each fish was measured to the nearest 0.01 g using a digital scale. Each fish was photographed from dorsal and ventral perspectives after proper alignment on a flat surface using a Canon EOS‐1D Mark II camera with a 50 mm f/1.4 USM lens (Canon Inc., Tokyo, Japan). Images were taken under a standardized setup (fixed zoom, distance and orientation). All morphometric data were extracted from these images using Fiji (ImageJ v1.54m; https://imagej.net/; National Institutes of Health, Bethesda, MD, USA) (Schneider et al., 2012). For traditional and truss‐based measurements, a total of 17 classical morphometric traits (e.g., [TL], standard length [SL], head length [HL]) were measured (Figure 1a). Truss‐based morphometrics were conducted using 26 anatomical landmarks (12 on the dorsal side, 14 on the ventral side) to form a comprehensive framework of 42 inter‐landmark distances (Figure 1b). These distances were normalized to TL, SL or HL. The definitions of landmarks and the morphometric measurement protocol followed the methodologies of Cadrin and Silva (2005) and Kim, Park, et al. (2025). A detailed summary is provided in Table S2. Morphometric measurement schemes used for Kareius bicoloratus. (a) Traditional morphometric measurements, with abbreviations defined within the figure. (b) Landmark‐based configuration consisting of 26 anatomically homologous dorsal and ventral landmarks used to calculate inter‐landmark distances and angular traits (Ang1–Ang5) for line‐based morphometric analyses, and to construct the geometric morphometric (GMM) dataset used in subsequent shape analyses. Landmark definitions are listed in Table S2. This figure is presented as a schematic illustration of the morphometric framework rather than absolute body size. Adapted from Kim, Park, et al. (2025) with modifications. In addition, four surface area traits were computed: total body surface area (BSA), dorsal body area (DSA), head surface area (HSA) and bony plate surface area (BPSA). Five angular (head, body and caudal angles) traits were also measured (Figure 1b). Each measurement was repeated three times, and the mean value was used for statistical analysis. From the full raw dataset, 81 indices were computed, including the condition factor (CF = 100 × body weight/TL3), as well as 80 other proportion‐, ratio‐ and angle‐based morphometric indices. A complete list of the 81 indices can be referred to Table S3. All measurements were recorded to at least two decimal places and reviewed for consistency before statistical analysis. Statistical analysis All statistical analyses were conducted using IBM SPSS Statistics (version 29.0.2, IBM Corp., Armonk, NY, USA). For all 81 morphometric indices, normality was tested using the Shapiro–Wilk test, and homogeneity of variance was evaluated using Levene's test. Proportion‐ and ratio‐based indices were arcsine square root‐ or log‐transformed as appropriate. For each test, a significance level of p < 0.05 was used. A three‐way analysis of variance (ANOVA) was performed for each index to evaluate the effects of three main factors—size (S, SM, ML, L), sex (F, M, U) and sampling site (G, T)—and their two‐way and three‐way interactions. For indices that passed assumptions and did not show significant interaction effects, statistical subgrouping was performed using post hoc tests (Tukey's HSD or Games–Howell), depending on the results of Levene's test. For indices with significant interaction effects, interpretation was conducted using estimated marginal means (EMMs) and simple effects tests. EMM‐based pairwise comparisons provided subgrouping across size, sex and site, accounting for marginal adjustments. Simple effects tests were run to identify which subgroups drove the interaction, typically by stratifying one factor within levels of another. When conducting simple effects analyses, Levene's test was re‐applied per subgroup to determine appropriate post hoc tests. In cases where parametric assumptions were not met—even after transformation—non‐parametric tests were applied. Specifically, the Kruskal–Wallis test was used for comparisons across three or more groups, followed by pairwise comparisons with Bonferroni correction. For two‐group comparisons, the Mann–Whitney U test or Welch's t‐test was used depending on equality of variance. A standardized decision tree was used to guide the selection of appropriate statistical approaches (see Figure S2), ensuring consistency across all 81 indices. Graphical representations, including colour‐coded statistical subgrouping matrices and interaction plots, were generated using Microsoft Excel 365 (Microsoft Corp., Redmond, WA, USA). GMA Landmark configuration and digitization GMM analyses were based on the same 26 anatomically homologous landmarks (dorsal and ventral sides combined) used for linear morphometric measurements (Figure 1b), ensuring full anatomical correspondence between truss‐ and landmark‐based datasets. All landmarks represented clearly identifiable skeletal or morphological reference points that could be consistently located across individuals. Landmark configurations were analysed following established GMM principles for quantifying biological shape variation, including the use of homologous landmarks to capture spatial relationships independent of size, position and orientation (Zelditch et al., 2012). Digital images were converted to TPS format using tpsUtil32 v1.83 (Stony Brook University, NY, USA), and landmark coordinates were manually digitized using tpsDig2 v2.32 (Stony Brook University, NY, USA). All digitization procedures were conducted by a single observer to minimize inter‐observer variability. Digitizing error assessment and landmark reliability To evaluate the reliability and repeatability of landmark placement, digitizing error was formally assessed prior to downstream shape analyses. A total of 20 individuals representing both sampling sites, sexes and size classes were randomly selected, and each specimen was digitized independently three times for each body side. Digitizing error was quantified using Procrustes ANOVA implemented in MorphoJ v1.08 (Klingenberg, 2011; https://morphometrics.uk/MorphoJ_page.html), with individual identity specified as the main effect and replicate digitizations treated as the error term. This analysis tested whether inter‐individual shape variation exceeded variation attributable to repeated landmark placement. In addition, the magnitude of digitizing error was quantified using root mean square error (RMSE) calculated from pairwise Procrustes distances among repeated digitizations within each individual. RMSE values provided an intuitive measure of within‐individual shape discrepancy in Procrustes shape space. These complementary approaches allowed both statistical and quantitative evaluation of landmark placement precision. Procrustes preprocessing and shape variable extraction Following confirmation of high digitization repeatability, all landmark configurations were imported into MorphoJ v1.08 for GMM analyses. Classifiers for TL size group, sex (female, male, undetermined), sampling site (Gadeokdo vs. Taean) and individual identity were assigned. Generalized Procrustes Analysis (GPA) was performed to remove non‐shape variation attributable to translation, rotation and isometric scaling, resulting in Procrustes‐aligned coordinates and centroid size for each specimen. Outlier inspection based on Procrustes distances detected no aberrant landmark configurations. Consensus configurations, covariance matrices and wireframe models were generated for subsequent multivariate analyses. Principal component analysis (PCA) PCA of the covariance matrix of Procrustes‐aligned coordinates was conducted as an exploratory approach to summarize the major axes of shape variation and to visualize overall morphological patterns in the dataset. PCA was performed separately for dorsal and ventral landmark configurations using MorphoJ v1.08. Scores of individual specimens along principal component axes were examined to assess the distribution of shapes in morphospace with respect to TL size group, sex and sampling site. Shape changes associated with major principal components were visualized using deformation grids and wireframe diagrams generated in MorphoJ, allowing intuitive interpretation of landmark displacements along each axis. PCA results were used to identify dominant trends in shape variation and to guide subsequent multivariate analyses, including CVA and Procrustes ANOVA, rather than as a formal hypothesis‐testing framework. CVA, Procrustes ANOVA and discriminant function analysis (DFA) CVA was performed separately for TL‐size classes, sex categories and sampling sites to quantify and visualize among‐group shape differences. Mahalanobis and Procrustes distances between group means were calculated, and their significance was tested using 10,000 permutation replicates. Global significance of group effects was further assessed by Procrustes ANOVA, which partitions total Procrustes variance into group and residual components and provides F‐statistics and permutation‐based p‐values for each factor. Where appropriate (number of groups >2), DFA with cross‐validation was used to evaluate the classification performance of shape variables. Original and cross‐validated assignment accuracies were calculated to complement CVA by quantifying the predictive strength of group‐level differences. Allometry testing and residual‐shape analysis Potential size‐related shape variation (allometry) was evaluated in MorphoJ using two steps. First, the Compare Vector Directions (CVD) procedure was applied to test whether group‐related shape changes (TL size, sex, site) were aligned with the common allometric trajectory (regression of shape on centroid size). When CVD indicated significant alignment or heterogeneity among regression vectors (p < 0.05), group differences were interpreted as size‐associated, and residual analyses were not pursued further for that factor. For factors in which CVD indicated no significant difference in allometric trajectories between groups (p > 0.05), Procrustes coordinates were regressed on centroid size in a pooled within‐group design, and the resulting residuals were extracted as size‐independent shape variables. These residual coordinates were then analysed by PCA, CVA and Procrustes ANOVA to test whether group differences persisted after removal of allometric effects. This procedure was applied to the sampling‐site factor; TL‐size and sex were retained in their original Procrustes form because their shape variation included substantial and heterogeneous allometric components. Genetic analysis mitochondrial DNA (mtDNA) sequencing and sequence generation For mtDNA analysis, 60 individuals were randomly selected, comprising 30 specimens from Site G and 30 specimens from Site T. Three protein‐coding mitochondrial genes—cytochrome c oxidase subunit I (COX1), cytochrome b (CYTB) and NADH dehydrogenase subunit 2 (ND2)—were targeted for sequencing. Primer pairs were designed based on the complete mitochondrial genome of Kareius bicoloratus (GenBank accession no. NC_003176.1) to amplify each gene as a single fragment encompassing the full open reading frame (ORF). The primer sequences were as follows: COX1 (forward 5′‐TTGCACCTCTGTCTATGGGT‐3′, reverse 5′‐TGCCTTGACAAGGCAGTGTT‐3′), CYTB (forward 5′‐CTATGCCAGCTACCACAACT‐3′, reverse 5′‐GCTAGAATCCTAGCTTTGGG‐3′) and ND2 (forward 5′‐GATCGAACCCAACCTGAAGA‐3′, reverse 5′‐TGCACGTAGAAGATGTGGGA‐3′). Polymerase chain reaction (PCR) amplifications were performed using AccuPower® ProFi Taq PCR PreMix (Bioneer, Daejeon, Republic of Korea) under the following cycling conditions: an initial denaturation at 94°C for 2 min, followed by 32 cycles of denaturation at 94°C for 30 s, annealing at 58–60°C for 30 s and extension at 72°C for 30 s. Each target gene was amplified as a single PCR product. Amplified products were purified prior to sequencing and subjected to Sanger sequencing using a primer‐walking strategy to ensure complete coverage of the target regions. All sequences were obtained in both forward and reverse directions, with each region read at least twice. Raw sequence reads were trimmed and assembled using Sequencher (GeneCodes, Ann Arbor, MI, USA). Forward and reverse reads were assembled to generate consensus sequences, and sequence quality was verified to minimize ambiguous base calls. The integrity of each coding region was confirmed by comparison with the reference mitochondrial genome (NC_003176.1), ensuring complete ORF coverage without frame shifts or premature stop codons. Sequence alignment and dataset construction Sequence alignments were performed separately for each mitochondrial gene (COX1, CYTB and ND2) using MEGA v12 (https://www.megasoftware.net/). Nucleotide sequences were aligned based on their coding regions, and alignments were visually inspected to confirm reading‐frame consistency and the absence of insertions or deletions (indels). ORFs were rechecked after alignment to ensure intact coding regions for all sequences. The final aligned sequence lengths were 1650 bp for COX1, 1141 bp for CYTB and 1045 bp for ND2. In addition to gene‐wise datasets, a concatenated mitochondrial dataset was constructed by sequentially combining the three genes in the order COX1–CYTB–ND2, yielding a total alignment length of 3746 bp. This concatenated dataset was used exclusively for analyses requiring combined genetic information, including individual‐level genetic distance estimation for Mantel tests. Population genetic analyses Population genetic analyses were conducted separately for each mitochondrial gene to assess genetic diversity and differentiation between the two sampling sites. Genetic diversity indices, including the number of haplotypes (Nh), haplotype diversity (Hd), nucleotide diversity (π) and the number of segregating sites (S), were calculated for each site using DnaSP v6.12 (https://www.ub.edu/dnasp). Genetic differentiation between Site G and Site T was evaluated using pairwise fixation indices (F_ST) and exact tests of population differentiation implemented in Arlequin v3.5.2.2 (https://cmpg.unibe.ch/software/arlequin35/). The statistical significance of F_ST values was assessed based on 10,000 permutations. Analysis of molecular variance (AMOVA) was additionally conducted to partition genetic variation within and between sampling sites for each gene. Haplotype relationships were visualized using median‐joining haplotype networks constructed in PopArt v1.7 (https://popart.maths.otago.ac.nz). Networks were generated separately for each mitochondrial gene to illustrate genealogical relationships among haplotypes and their distribution between the two sampling sites. Mantel test between genetic and morphological distances Because Mantel tests are generally most informative when multiple populations are available, an individual‐level Mantel framework was adopted to evaluate the association between genetic and morphological divergence in this study, which involved only two sampling sites. A total of 60 individuals with matched morphometric and mtDNA data (30 from Site G and 30 from Site T) were included, and individual identities were strictly harmonized across datasets to ensure consistent ordering of distance matrices. To avoid potential bias associated with differences in sample composition relative to the full morphometric dataset, GMM analyses were repeated using only the 60 individuals included in the genetic analysis. Procrustes superimposition was performed separately for dorsal and ventral landmark configurations, and individual‐by‐individual morphological distance matrices (60 × 60) were calculated as Procrustes distances derived from Procrustes coordinates using the R package geomorph implemented in R (R Core Team, 2024; https://www.r-project.org/). For genetic distances, pairwise genetic distance matrices (60 × 60) were calculated from the concatenated mtDNA dataset (3746 bp) using the p‐distance model with pairwise deletion implemented in MEGA v12. Mantel tests were conducted between the morphological and genetic distance matrices using the mantel function in the R package vegan, with Mantel's correlation coefficient (r) and permutation‐based p‐values estimated from 9999 permutations. Analyses were performed separately for dorsal and ventral datasets. RESULTS Variation in raw morphometric indices The overall distribution of 81 morphometric indices prior to statistical grouping is shown in Figure S3. Each index is plotted as a mean ± standard deviation across all individuals, categorized by measurement type (relative to TL, SL, HL, other ratio‐based indices and angular measurements). Although no group‐level trends are presented, the extent of variation differed across indices, with some exhibiting notably wider standard deviation ranges, as seen in indices such as UBD/LBD, MALBP/BPAL and BSA/BW. Figure S4 summarizes the coefficient of variation (CV%) for all indices in descending order to provide a baseline understanding of variability prior to statistical testing and helping to contextualize the significance results described in the next section. As shown in the Figure S4, the majority of indices displayed %CV values ranging from 5% to 20%. However, several traits exhibited considerably higher %CV values (greater than 25%), which included surface area‐related indices (e.g., BSA/BW, BSA/SL, DSA/SL and HSA/HL) and bony plate‐related indices (e.g., BPSA/DSA and MSLBP/BPA). In contrast to the greater inter‐individual variability in these indices, lower CV% values (lower than 5%) were typically observed in certain fin‐related and angular measurements, as exemplified by DFBL/SL(TL), AFBL/SL(TL), SL/TL (related to caudal fin length), Ang3 and Ang4. Line morphometric analysis Three‐way ANOVA: Main effects The results of the three‐way ANOVA for all 81 morphometric indices are summarized in Figure 2 as heatmaps of significance (i.e., log‐transformed p‐values for the main effects of size, sex and sampling site), and effect size estimates (min‐max normalized partial η 2), providing an intuitive index‐wise comparison of how strongly each main factor contributes to morphometric variation. Full statistical outputs are provided in Table S4. Overall, 44 indices (54.3%) showed statistically significant differences (p < 0.05) for at least one main factor (Figure 2a). Among the three main factors, sex accounted for the largest number of significant indices (n = 29), followed by size (n = 19) and site (n = 18). Of the 44 significant indices, 24 showed significance for one factor, 18 for two factors and only 2 indices (DSA/SL and eye ball diameter [EBD]/HL) were significant across all three factors. Although these 44 indices were distributed broadly across measurement types, higher incidences of significant effects were observed among surface area‐, depth‐ and bony plate‐related ratios, as well as angular indices, compared to linear proportions measured relative to TL or SL. Heatmaps summarizing three‐way ANOVA results for 81 morphometric indices of Kareius bicoloratus. This provides a heatmap‐based overview that enables intuitive, index‐wise comparison of which main factors (size, sex and sampling site) exert statistically significant and biologically meaningful effects on each morphometric index, and which factors do not, based on significance and effect size estimates. (a) Significance of main effects (size, sex, site) shown as –log10(p) values (scale: 0.0–3.0). Dark blue indicates higher significance. Actual p‐values are annotated on the scale bar. Yellow cells denote indices with significant two‐way interactions. (b) Effect sizes (partial η 2) shown as min–max normalized values (scale: 0.0–1.0), with dark green indicating stronger explanatory power. Actual η 2 values are annotated on the scale bar. See Table S4 for full numerical results. Patterns of effect size did not precisely mirror those of statistical significance, although the overall heatmap gradients were generally comparable (Figure 2b). The highest partial η 2 values (up to 0.60) were observed in sex‐related effects, particularly for UBD/BL and UBD/LD. The number of indices with partial η 2 values exceeding 0.1 (a commonly accepted threshold for large effects) was similar between size and sex factors. In contrast, the site factor was associated with far fewer indices exhibiting large effect sizes (η 2 > 0.1), as detailed in Table S4. Thirteen indices exhibited significant interaction effects among the three main factors. Upper lip length (ULL)/HL showed both size × sex and sex × site interactions, while other indices showed specific two‐way interactions: size × sex (e.g., CL/SL, EBD/HL, d1‐2/HL), size × site (SFBPD/SL, d2‐5/SL, d5‐10/SL) and sex × site (e.g., HL/SL) interaction (see Figure 2a). These interaction effects are addressed in more detail in the following sections. Post‐hoc group contrasts and effect‐size pattern The colour‐coded matrix in Figure 3 shows statistical subgroupings of all 81 morphometric indices across size, sex and site, derived using a decision‐tree workflow that incorporated both parametric and non‐parametric post hoc comparisons as appropriate (see Figure S2). Among the 81 indices examined, 70 exhibited statistically significant differences in at least one of the three grouping factors: size, sex and sampling site. In terms of the number of significantly affected indices, size and sex groups were associated with slightly more indices (63 and 62, respectively) than the site group (56 indices). Colour‐coded matrix summarizing the results of post hoc comparisons across 81 morphometric indices, showing statistical subgrouping for each of the three main factors: Size, sex and site. Shading intensity indicates subgroup differences based on multiple post hoc procedures selected via the decision‐tree workflow (see Figure S2). Subgrouping is displayed across size group (S = small, SM = small‐medium, ML = medium‐large, L = large), sex group (F = female, M = male, U = underdeveloped) and sampling site (G = Gadeokdo, T = Taean). Identical shades indicate no significant differences within a factor, while different shades represent distinct statistical subgroups. For statistical groupings, alphabetical labels (a, ab, b, etc.) denote homogenous subsets. Indices with significant interaction effects are labelled on the right; grey cells indicate non‐significant results or post hoc exclusion due to interaction terms. Regarding patterns of group‐level differentiation, many indices showed clear statistical separation between the smallest and largest size classes. In particular, several indices demonstrated sharp size‐dependent effects, with all four size classes (S, SM, ML, L) forming distinct statistical groups. These included traits that were either positively associated with fish size (e.g., Ang3, HSA/HL, DSA/SL, BSA/SL, d6–10/SL) or inversely related (e.g., HL/TL, d2–5/SL, d8–9/SL). For the sex factor, the most frequent differences occurred between underdeveloped individuals and either mature males or females, although sex‐specific differentiation between adult females and males was also observed in several indices, particularly UBD/BD, UBD/LBD and DSA/BSA. For sampling site, group‐level differentiation was more frequently observed among indices measured relative to SL (30/38 = 78.9%) and among depth, surface area and angular traits (11/15 = 73.3%) than among TL‐based (4/11 = 36.4%) or HL‐based (9/17 = 52.9%) indices. The overall distribution of these factor‐specific effects is summarized in Figure 4. As shown in the Venn diagram, 44 indices were influenced by all three grouping factors simultaneously. Among these, nine indices also exhibited significant two‐way interaction effects: size × sex (n = 5), size × site (n = 3) and sex × site (n = 1). An additional 14 indices—including two interaction‐involved indices (d2–3/HL in size × sex, and d5–10/SL in size × site)—were jointly affected by size and sex. Five indices were jointly affected by TL and site, and three by sex and site without involvement of the third factor. Among these three, the ULL/HL index exhibited interaction effects in both size × sex and sex × site. One index was influenced exclusively by sex, and three indices exclusively by site. Venn diagram summarizing the number of morphometric indices showing significant differences by size (n = 63), sex (n = 62) and site (n = 56), including their overlap. Central overlaps indicate indices affected by two or all three main factors, while peripheral regions represent factor‐specific effects. Text labels within overlaps specify the number of indices with significant interaction effects between factors. A total of 11 indices showing no significant difference for any main factor and are noted on the right. Meanwhile, 11 indices did not show significant differences in response to any of the three main effects. These included fin‐related indices (e.g., AFBL/TL, AFBL/SL, DFBL/SL and v15–16/HL), caudal shape‐related indices (e.g., d10–12/TL, d10–11/SL and Ang5) and ventral head shape traits (v13–14/HL and v13–18/HL), suggesting that these are relatively invariant morphological features in this flounder species (Figures 3 and 4). Interaction effects TL × sex interactions Figure 5 presents the simple effects of size within each sex for the seven morphometric indices that exhibited significant size × sex interaction effects. These interactions reflect cases where size‐related morphometric patterns differ between males and females. The index CL/SL (Figure 5a) showed a clear interaction, with males displaying a consistent upward trend across size groups, while females showed no significant variation across size classes. ULL/HL (Figure 5b) values were stable across size classes in both sexes, with no significant within‐sex differences. For d1–2/HL (Figure 5c), both males and females exhibited a decline in value with increasing size, but the trend was more pronounced in males, resulting in non‐parallel trajectories. In EBD/HL (Figure 5d), size‐related decline was significant in both sexes, although the rate of reduction was steeper in females. In d2–3/HL (Figure 5e), males showed a peak at intermediate size groups, while females remained flat, contributing to a diverging pattern. CD/BD (Figure 5f) exhibited a statistically significant size effect only in females, even though males appeared to show a peak at SM; the interaction likely reflects sex‐specific differences in trend strength or variability. Lastly, Ang1 (Figure 5g) showed clear size‐dependent changes in females, with a hump‐shaped pattern peaking at SM and ML, while males showed no significant differences across TL groups. Collectively, these results demonstrate that size × sex interactions arise from both qualitative and quantitative differences in size‐related morphological changes between males and females. Interaction plots illustrating size × sex effects for seven morphometric indices: (a) CL/SL, (b) ULL/HL, (c) d1–2/HL, (d) CD/BD, (e) EBD/HL, (f) d2–3/HL and (g) Ang1. Each plot shows mean values ± SE for each TL‐based size group (S, SM, ML, L) separated by sex. Different letters indicate statistically significant differences within sex groups based on post hoc comparisons (p < 0.05). BD, body depth; HL, head length; L, large; ML, medium‐large; S, small SE, standard error; SL, standard length; SM, small‐medium. TL × site interactions Figure 6 illustrates the simple effects of size group across sampling sites (panels a–e) and sex group across sites (panels f–g) for the morphometric indices that showed significant interaction effects between size × site or sex × site in the three‐way ANOVA. For the size × site interactions, five indices were examined. Among them, d2–5/SL, d5–10/SL and Ang3 (Figure 6b,c,e) showed size‐dependent variation that was significant only within Site G. In d2–5/SL, values decreased sharply from S to L size groups at Site G, with three distinct size groups, while values at Site T remained statistically constant. A similar site‐specific trend was observed in d5–10/SL, where Site G displayed moderate differences between size groups, while Site T showed no significant changes. In Ang3, Site G exhibited a strong increasing trend with size, whereas Site T remained flat across size groups. In contrast, SFPBD/SL and d6–7/SL (Figure 6a,d) showed diverging patterns between the two sites, but without statistically significant size‐based subgrouping within either site, indicating apparent trends that were not statistically significant. Interaction plots illustrating size × site effects for six morphometric indices: (a) SFBPD/SL, (b) d2–5/SL, (c) d5–10/SL, (d) d6–7/SL, (e) Ang3 and sex × site effects for two indices: (f) HL/SL and (g) ULL/HL. Each plot shows mean values ± SD for TL‐based size groups (S, SM, ML, L) or sex groups (F, M, U). Different letters indicate statistically significant differences within site groups based on post hoc comparisons (p < 0.05). HL, head length; L, large; ML, medium‐large; S, small SE, standard error; SFBPD, snout to first bony plate distance; SL, standard length; SM, small‐medium. Sex × site interactions For the sex × site interactions, HL/SL (Figure 6f) showed sex‐based differences only in Site G, where the underdeveloped group (U) was significantly different from females and males, while all sex groups at Site T were statistically similar. In contrast, ULL/HL (Figure 6g) displayed statistically significant sex‐based differences in both Site G and Site T, though the relative magnitude of differences varied by site. These findings confirm that interaction effects in these indices stem from site‐specific amplification or suppression of group‐level morphometric patterns. The full results of the simple effects analyses for all indices exhibiting significant interaction effects are summarized in Table 2, including within‐group significance patterns across size, sex and site levels. TABLE 2Summary of statistical results for morphometric indices with significant interaction effects from the three‐way ANOVA.IndexInteraction typeSignificance p a Effect size b [observed power c ]Simple effects test summary d Post hoc in simple effects testsRelated figureCL/SLSize × Sex0.0440.068 [0.604]Size dependency significant only in malesTukey‐HSDFigure 5a ULL/HLSize × Sex0.0360.072 [0.635]No significant simple effects in sexesn/aFigure 5b EBD/HLSize × Sex0.0070.106 [0.824]Size dependency significant in both sexesTukey‐HSDFigure 5c d1‐2/HLSize × Sex0.0290.077 [0.665]Size dependency significant in both sexesTukey‐HSDFigure 5d d2‐3/HLSize × Sex0.0030.121 [0.878]Size dependency significant only in malesTukey‐HSDFigure 5e CD/BDSize × Sex0.0490.065 [0.586]Size dependency significant only in femalesTukey‐HSDFigure 5f Ang1Size × Sex0.0380.071 [0.624]Size dependency significant only in femalesTukey‐HSDFigure 5g SFPBD/SLSize × Site0.0270.054 [0.605]No significant size dependency within sitesn/aFigure 6a d2‐5/SLSize × Site0.0440.045 [0.523]Size dependency significant only in the G siteTukey‐HSDFigure 6b d5‐10/SLSize × Site0.0330.050 [0.575]Size dependency significant only in thr G siteGames‐HowellFigure 6c d6‐7/SLSize × Site0.0370.048 [0.555]No significant size dependency within sitesn/aFigure 6d Ang3Size × Site0.0330.050 [0.573]Size dependency significant only in the G siteTukey‐HSDFigure 6e HL/SLSex × Site0.0380.048 [0.550]Sex dependency significant only in the G siteTukey‐HSDFigure 6f ULL/HLSex × Site0.0220.058 [0.638]Sex dependency significant in both G and T sitesTukey‐HSDFigure 6g Note: For each index, the interaction type, significance (p), effect size (η 2) and observed power are shown. Simple effects test results indicate where size‐, sex‐ or site‐based differences were significant within levels of the interacting factor. Post hoc tests were performed only when simple effects were statistically significant. n/a = Post hoc not applied due to lack of significant simple effects. a Significance of interaction term from three‐way ANOVA. b Effect size η 2 = partial eta squared. c Observed power computed at α = 0.05. d Simple effects interpreted from EMM‐based tests within interaction levels.Abbreviations: HL, head length; SL, standard length. Summary of statistical results for morphometric indices with significant interaction effects from the three‐way ANOVA. Note: For each index, the interaction type, significance (p), effect size (η 2) and observed power are shown. Simple effects test results indicate where size‐, sex‐ or site‐based differences were significant within levels of the interacting factor. Post hoc tests were performed only when simple effects were statistically significant. n/a = Post hoc not applied due to lack of significant simple effects. Significance of interaction term from three‐way ANOVA. Effect size η 2 = partial eta squared. Observed power computed at α = 0.05. Simple effects interpreted from EMM‐based tests within interaction levels. Abbreviations: HL, head length; SL, standard length. GMM analysis Digitizing error assessment Procrustes ANOVA revealed highly significant inter‐individual shape variation relative to digitizing error for both dorsal (F = 47.20, p < 0.0001) and ventral sides (F = 24.38, p < 0.0001). Mean RMSE values were low (dorsal: 0.0132 ± 0.0067; ventral: 0.0180 ± 0.0048), confirming high repeatability of landmark placement (Table S5). PCA PCA was conducted separately for dorsal and ventral landmark datasets. In the dorsal view, PC1 and PC2 explained 38.9% and 19.4% of total shape variance (cumulative 58.3%), whereas in the ventral view, PC1 and PC2 accounted for 28.7% and 21.3% (cumulative 50.0%). Thus, a few leading components captured most of the variance in both views. Deformation patterns indicated that PC1 mainly represented variation in head expansion, dorsal–ventral body depth and overall body outline, while PC2 reflected changes in posterior body tapering and ventral curvature. PC3 contributed additional, more localized deformation around cranial and caudal regions (Figure S5). Across TL‐size, sex and site, PCA scatter plots showed broad overlap of groups in morphospace. TL‐size classes exhibited a shallow gradient along PC1 consistent with general growth‐related expansion, but with extensive overlap among classes. Sex groups also overlapped widely in dorsal and ventral views, indicating limited sex‐related separation in the unsupervised PCA. The two sampling sites (G and T) showed slightly shifted but overlapping score distributions. These patterns motivated the use of supervised methods (CVA, Procrustes ANOVA, DFA) to detect and test more subtle group‐level shape differences. CVA TL‐size variation CVA identified TL‐size as a major determinant of shape in both views. In the dorsal dataset, CV1 and CV2 explained 76.2% and 13.7% of the among‐group variance; in the ventral dataset, 70.6% and 16.1% were explained by the first two canonical variates. Multivariate tests indicated highly significant effects of TL‐size (dorsal: Goodall's F = 8.971, Pillai's trace = 1.4756; ventral: Goodall's F = 8.9156, Pillai's trace = 1.6686; all p < 0.0001). Most pairwise comparisons among the four TL‐size classes were significant in both datasets. An exception occurred for the intermediate ML–SM comparison, which was non‐significant in the dorsal view (Mahalanobis p = 0.0673; Procrustes p = 0.1072) but significant in the ventral view (Mahalanobis p = 0.0023; Procrustes p = 0.0190). CVA scores showed a graded ordering of the four size classes along CV1, consistent with a continuous allometric trajectory (Figures 7a and S6). Dorsal wireframes highlighted coordinated cranial expansion, body deepening and posterior narrowing with increasing size, whereas ventral deformation emphasized abdominal widening and ventral curvature changes (Figure S7). Canonical variate analysis (CVA) of dorsal‐side geometric morphometric shape variation in Kareius bicoloratus. CVA based on Procrustes‐aligned landmark coordinates illustrating multivariate shape differentiation among groups. (a) CVA scatter plot showing size‐class separation, with individuals grouped into large (L), medium‐large (ML), small‐medium (SM) and small (S) size categories. (b) CVA plot depicting sexual dimorphism, including female (F), male (M) and underdeveloped (U) individuals. (c) Frequency distributions of canonical variate 1 (CV1) scores for the two sampling sites (Site T and Site G), highlighting site‐dependent differences in overall shape structure. Each point represents an individual specimen, and ellipses indicate group dispersion around centroids. Corresponding CVA results for the ventral side are provided in Figure S6. Sex variation Sex‐related differences were clearly detected by CVA. In the dorsal dataset, CV1 and CV2 accounted for 57.99% and 42.01% of the among‐group variance; the corresponding values for the ventral dataset were 53.59% and 46.41%, so that two axes captured virtually all sex‐associated variation. Global tests indicated strong sex effects (dorsal: Goodall's F = 17.39, Pillai's trace = 1.4015; ventral: Goodall's F = 16.33, Pillai's trace = 1.3972; all p < 0.0001). All pairwise comparisons among F, M and U were significant in both views, with Mahalanobis distances of 3.28–4.26 (dorsal) and 3.59–4.22 (ventral), and Procrustes distances of 0.032–0.049 (all permutation p < 0.0001). Group centroids formed a triangular configuration in CVA plots, and confidence ellipses showed only partial overlap. Wireframe changes associated with canonical variates indicated that sex‐related differences were concentrated in head shape, body depth and abdominal outline rather than the caudal region (Figures 7b and S6 and S7). Site variation For the sampling‐site factor, a single canonical variate (CV1) captured 100% of the among‐site variance in both views. Multivariate tests supported a strong site effect (dorsal: Goodall's F = 20.06, Pillai's trace = 0.7616; ventral: Goodall's F = 18.37, Pillai's trace = 0.7678; all p < 0.0001). Mahalanobis distances between G and T were 3.54 (dorsal) and 3.60 (ventral), and corresponding Procrustes distances were 0.0355 and 0.0326 (all permutation p < 0.0001). CVA scores showed largely distinct G and T clusters with limited overlap. Wireframes indicated relatively deeper bodies with more pronounced cranial and abdominal expansion in fish from site G, whereas fish from site T exhibited a more streamlined, flattened outline. Similar deformation patterns in dorsal and ventral views indicated consistent site‐related shape differences across perspectives (Figures 7c, S6 and S7). Procrustes ANOVA and discrimination analysis Procrustes ANOVA performed separately for dorsal and ventral views showed significant effects of TL size, sex and site on centroid size and Procrustes shape coordinates (Table 3). TL‐size groups differed strongly in centroid size (dorsal: F = 343.66; ventral: F = 351.61; both p < 0.0001) and size‐standardized shapes remained significantly different (dorsal: F = 8.97, Pillai's trace = 1.48; ventral: F = 8.92, Pillai's trace = 1.67; both p < 0.0001). Sex exerted marked effects on centroid size (dorsal: F = 90.60; ventral: F = 88.78; both p < 0.0001) and on shape (dorsal: F = 17.39; ventral: F = 16.33; both p < 0.0001), indicating that sexual dimorphism in shape is not reducible to size alone. Site differences were associated with large centroid‐size contrasts (dorsal: F = 115.47; ventral: F = 121.02; both p < 0.0001) and produced the strongest shape effects among the three factors (dorsal: F = 20.06, Pillai's trace = 0.76; ventral: F = 18.37, Pillai's trace = 0.77; both p < 0.0001). Patterns were closely matched between dorsal and ventral datasets. TABLE 3Summary of Procrustes ANOVA and discriminant analysis based on geometric morphometrics.Procrustes ANOVA: Centroid size and shape a ViewFactorCentroid size (F)Shape (F)Pillai's trace p‐valueDorsalSize343.668.971.48<0.0001Sex90.6017.391.40<0.0001Site115.4720.060.76<0.0001VentralSize351.618.921.67<0.0001Sex88.7816.331.40<0.0001Site121.0218.370.77<0.0001Discriminant analysis: Distances and classification accuracy b ViewFactorProcrustes distance (range)Mahalanobis distance (range)Classification accuracy (%)Cross‐validation (%)DorsalSize0.0149–0.05112.2115–6.059886–10060–98Sex0.0325–0.04933.3428–4.638392–10082–93VentralSize0.0175–0.04483.5322–6.614998–10070–88Sex0.0324–0.04533.6631–5.486294–10083–93Abbreviations: ML, medium‐length; SM, small‐medium. a Shape effects were evaluated using Procrustes ANOVA with Pillai's trace. Pillai's trace values refer to multivariate shape effects only and are not applicable to centroid size. All tests were significant after permutation procedures. b All pairwise comparisons were significant (p < 0.001 or p < 0.0001), except the ML–SM size comparison in the dorsal dataset, which was non‐significant in permutation tests. Classification accuracies were based on leave‐one‐out cross‐validation. Distance and accuracy values for each pairwise comparison are provided in Table S5. Summary of Procrustes ANOVA and discriminant analysis based on geometric morphometrics. Abbreviations: ML, medium‐length; SM, small‐medium. Shape effects were evaluated using Procrustes ANOVA with Pillai's trace. Pillai's trace values refer to multivariate shape effects only and are not applicable to centroid size. All tests were significant after permutation procedures. All pairwise comparisons were significant (p < 0.001 or p < 0.0001), except the ML–SM size comparison in the dorsal dataset, which was non‐significant in permutation tests. Classification accuracies were based on leave‐one‐out cross‐validation. Distance and accuracy values for each pairwise comparison are provided in Table S5. Discriminant analysis of dorsal landmarks indicated increasing morphological separation with greater size contrast (Table 3). Pairwise Procrustes distances ranged from 0.0149 to 0.0511, with the largest values between the smallest (S) and largest (L) groups. Most pairwise comparisons were highly significant in parametric and permutation tests (p < 0.001), except ML versus SM, which did not reach significance in the permutation test. Classification accuracy was 86%–100%, and cross‐validation indicated robust discrimination, especially for contrasts involving the extreme size classes. In the ventral dataset, all pairwise comparisons, including ML versus SM, were significant in both testing frameworks. Procrustes and Mahalanobis distances were generally larger than in the dorsal view, and classification accuracy reached 98–100% with cross‐validation accuracies of 70%–88%. The ventral dataset showed size‐associated shape variation primarily involving ventral cranial breadth, the branchial region and the relative positions of the pelvic and anal fin bases. DA of dorsal landmarks detected significant separation among F, M and U. Procrustes distances ranged from 0.0325 (F vs. M) to 0.0493 (M vs. U), and Mahalanobis distances from 3.34 to 4.64; all pairwise Hotelling's T 2 and permutation tests gave p < 0.0001. Classification accuracy ranged from 92.2% to 100%, with cross‐validation scores of 81.8%–93.0%. In the ventral dataset, Procrustes distances were comparable to dorsal values, while Mahalanobis distances increased (3.66–5.49), particularly for M versus U. Again, all pairwise tests were significant (p < 0.0001). Discriminant accuracy remained high (93.5%–100%), and cross‐validation scores were slightly higher than in the dorsal view (83.1%–92.6%) (Table S6). In the ventral dataset, sex‐related shape differences were mainly associated with variation in ventral cranial width, branchial contour and pelvic‐fin base position. Allometry testing and residual analysis Regression of Procrustes coordinates on centroid size within each factor level, followed by CVD, showed that TL‐size classes and sex groups exhibited significant heterogeneity in allometric trajectories. For TL‐size, pairwise angular comparisons among the four groups yielded relatively large angles with consistently small p‐values (all p < 0.05; data not shown). For sex, dorsal angular differences were 37.1° (F–M), 47.6° (F–U) and 54.7° (M–U), all with p < 0.01. In the ventral dataset, the F–M comparison remained significant (45.0°, p = 0.00004), whereas F–U and M–U were not (87.7° and 87.0°, p = 0.423 and 0.401). These patterns indicate that a single homogeneous allometric trajectory is not appropriate for the sex factor. In contrast, the site factor showed no evidence of differing size–shape trajectories: angles between G and T vectors were 115.9° (dorsal) and 112.0° (ventral), with large p‐values (0.976 and 0.967). Centroid‐size effects were therefore considered homogeneous between sites, and residual‐based analyses were restricted to this factor. Pooled within‐site regressions of shape on centroid size explained 8.9% of within‐site variance in the dorsal dataset and 6.4% in the ventral dataset (permutation tests, both p < 0.0001). The largest regression coefficients occurred along the mid‐ to posterior ventral and ventrolateral body contour, whereas snout landmarks showed negligible size‐related displacement. The small magnitude of coefficients and nearly overlapping regression wireframes indicate that, although statistically detectable, allometric effects for the site factor were weak in magnitude (Table S7). PCA of regression residuals showed that the first components captured a substantial proportion of remaining variance (dorsal: PC1 = 36.6%, PC2 = 17.6%; ventral: PC1 = 32.0%, PC2 = 14.8%). Residual PCA scatter plots revealed partial but consistent separation of G and T along major axes, with some overlap. Deformation grids for PC1 and PC2 indicated size‐independent differences mainly in the anterior head and mid‐body configuration (Figure S8). Residual CVA yielded a single canonical axis (CV1) that separated G and T in both views (eigenvalues: dorsal = 4.75; ventral = 3.72), with permutation tests confirming highly significant group effects (Goodall's F: dorsal = 16.829, ventral = 13.935; all p < 0.0001; Pillai's trace also p < 0.0001). Mahalanobis distances were 4.316 (dorsal) and 3.819 (ventral), and Procrustes distances 0.0310 and 0.0275, respectively (all permutation p < 0.0001). Histograms of canonical scores showed minimal overlap between sites (Figure S9). Procrustes ANOVA on the residual datasets detected no centroid‐size differences between sites (dorsal: F = 1.70, p = 0.1951; ventral: F = 0.90, p = 0.3440), confirming effective removal of size effects. In contrast, the shape term remained strongly significant (dorsal: F = 16.84, Pillai's trace = 0.83; ventral: F = 13.94, Pillai's trace = 0.79; both p < 0.0001). Thus, statistically robust site‐related shape differences persisted after allometric correction, indicating genuine site‐specific divergence in body shape that is independent of centroid size. Genetic analysis Genetic diversity based on mtDNA sequences Genetic diversity indices were estimated separately for COX1, ND2 and CYTB using mtDNA sequences obtained from Site G and Site T. Across all three genes, the Site T population consistently exhibited equal or higher levels of genetic diversity than Site G. Haplotype diversity (Hd) ranged from moderate to high in both populations, with higher values generally observed in Site T, particularly for ND2 (Hd = 0.897 in Site T vs. 0.671 in Site G). Nucleotide diversity (π) also showed a consistent pattern, with Site T displaying higher π values than Site G for all genes. Among the three markers, ND2 exhibited the highest levels of nucleotide diversity and the largest number of haplotypes, whereas COX1 showed comparatively lower variability. CYTB displayed intermediate levels of genetic diversity, characterized by high haplotype diversity but moderate nucleotide diversity. These patterns indicate gene‐specific differences in variability, while consistently suggesting greater mitochondrial genetic diversity in the Site T population (Table 4). TABLE 4Population genetic parameters of mitochondrial DNA markers (COX1, ND2 and CYTB) in Kareius bicoloratus from two sampling sites (G and T).GeneSite N No. of haplotypes (h)Haplotype diversity (Hd)Nucleotide diversity (π)Pairwise F_STAMOVA (% among pops.) COX1 G3080.7200.00170.101* 10.14T30120.7660.0025 ND2 G30100.6710.00390.118** 11.81T30160.8970.0064 CYTB G30120.8690.00330.072* 7.20T30130.8670.0045* p < 0.05;** p < 0.01. Population genetic parameters of mitochondrial DNA markers (COX1, ND2 and CYTB) in Kareius bicoloratus from two sampling sites (G and T). p < 0.05; p < 0.01. Genetic differentiation between Site G and Site T Genetic differentiation between Site G and Site T was evaluated using pairwise F_ST and AMOVA for each mtDNA gene. Pairwise F_ST values indicated significant genetic differentiation between the two populations for all three genes (COX1: F_ST = 0.101, p < 0.05; ND2: F_ST = 0.118, p < 0.01; CYTB: F_ST = 0.072, p < 0.05). Among them, ND2 showed the highest level of genetic differentiation, whereas CYTB exhibited the lowest. AMOVA results further supported these findings, revealing that 7.20%–11.81% of the total genetic variation was attributable to differences between populations, depending on the gene examined. The highest among‐population variance was observed for ND2 (11.81%), followed by COX1 (10.14%) and CYTB (7.20%), while the majority of genetic variation resided within populations. Together, these results indicate statistically significant but moderate genetic differentiation between Site G and Site T (Table 4). Haplotype network analysis Median‐joining haplotype networks constructed for each mtDNA gene revealed substantial haplotype sharing between Site G and Site T, accompanied by varying degrees of site‐associated structure. In the COX1 network, a common central haplotype was shared by individuals from both sites, with multiple low‐frequency haplotypes radiating outward, some of which were site‐specific. This pattern suggests limited genealogical separation between populations for this gene. The CYTB network showed a star‐like topology centered on shared haplotypes and additional site‐associated variants distributed across the network. In contrast, the ND2 network exhibited a more complex structure, characterized by multiple haplotype clusters and a greater number of derived haplotypes, particularly in Site T. Several haplotypes were predominantly associated with one site, reflecting stronger genealogical differentiation than observed for the other genes. Overall, haplotype networks indicated incomplete lineage separation between Site G and Site T, while revealing gene‐dependent differences in the extent of haplotype structuring (Figure 8). Haplotype networks of mitochondrial DNA sequences from Kareius bicoloratus collected at two sampling sites. Median‐joining haplotype networks constructed from mitochondrial DNA sequences showing the genetic relationships among haplotypes detected at Site G and Site T. Panels (a–c) represent haplotype networks based on individual mitochondrial gene regions (a, COX1; b, CYTB; c, ND2). Each circle represents a unique haplotype, with circle size proportional to haplotype frequency. Colours indicate sampling sites (cyan, Site G; blue, Site T), and pie charts denote the relative contribution of individuals from each site within shared haplotypes. Small black nodes represent median vectors (hypothetical unsampled haplotypes). Hatch marks on branches indicate the number of mutational steps separating haplotypes. Association between genetic and morphological distances The relationship between genetic and morphological divergence was evaluated using individual‐level Mantel tests based on pairwise genetic distances derived from concatenated mtDNA sequences and Procrustes distances from GMM data. For dorsal landmark configurations, a weak but statistically significant positive correlation was detected between genetic and morphological distances (Mantel r = 0.147, p = 0.012). A similar pattern was observed for ventral landmark configurations (Mantel r = 0.131, p = 0.013) (Figure S10). Although the strength of the correlations was low, the consistent significance across both dorsal and ventral datasets indicates that individuals with greater genetic divergence tend to exhibit greater morphological divergence. These results demonstrate a statistically detectable, albeit weak, association between mitochondrial genetic variation and body shape differences at the individual level. DISCUSSION Overview of main findings This study provides an integrated understanding of morphological variation in Kareius bicoloratus by combining classical truss‐based indices with GMM analyses. Although these approaches differ in how they quantify form, both consistently identified size and sex as dominant drivers of shape variation, with geographic origin contributing additional but more moderate differences. Strong concordance between key truss indices and landmark deformation patterns further demonstrates that linear and geometric methods capture complementary dimensions of the same phenotypic structure, strengthening the biological interpretation of observed shape differences. In addition to morphological analyses, mitochondrial DNA data revealed significant but moderate genetic differentiation between sampling sites, with gene‐specific patterns of variation. Importantly, individual‐level Mantel tests detected a weak yet statistically significant association between genetic and morphological distances, indicating that genetic divergence contributes, at least in part, to phenotypic differentiation among individuals. Together, these findings highlight the multifactorial nature of morphological variation in this species, shaped by the combined influences of size, sex, geographic context and underlying genetic structure. Size‐ and sex‐related shape variation Growth was accompanied by pronounced reorganization of body form. Larger individuals showed marked expansion of the dorsal and total body‐surface areas relative to body length (reflected in higher DSA/SL, BSA/SL and HSA/HL) and a relative shortening of the head and anterior trunk (evident in declining HL/TL and d2–5/SL). These directional shifts align closely with GMM patterns that documented cranial compression (d1–d4), increased depth and breadth of the midbody (d5, d8, v21, v22), and tapering of the posterior body (d10–d11, v23–v25). The broadening of the midbody region was particularly evident, with both truss measurements (e.g., widening of d6–10/SL) and landmark‐based deformation suggesting coordinated lateral expansion. Together, these concordant patterns indicate that allometric growth in K. bicoloratus involves coordinated restructuring across the cranial, dorsal and caudal regions, offering functional advantages such as greater stability and manoeuvrability on the benthic substrate (Loy et al., 2000; Oliveira et al., 2023; Onyekwelu et al., 2021; Pepe‐Victoriano et al., 2021). Sexual dimorphism was also clear and involved multiple anatomical regions. Females exhibited deeper abdominal profiles and broader surface‐area development (e.g., elevated BD/SL and DSA/BSA), while males tended to show greater upper‐body depth relative to total depth (UBD/BD, UBD/LBD) and a more compact dorsal outline. These observations align with previous studies in other flatfish species where females exhibit larger heads and deeper abdomens compared to males (Cadrin & Silva, 2005). These proportional differences were mirrored in GMM, with females showing pronounced outward displacement of abdominal landmarks (v18–v22) and broader cranial contours, and males showing a more contained dorsal margin. Such concordance across methods reinforces the view that sex‐related differences are not limited to reproductive traits but manifest as whole‐body changes linked to energy allocation, reproductive investment or locomotive specialization (Jacquemin & Pyron, 2013; Ryu et al., 2020). Underdeveloped individuals, which often displayed intermediate or distinct profiles, appeared to reflect interactions between size and maturation stage in shaping morphology. Geographic divergence and ecological interpretation Geographic differences, although less extensive than those associated with size or sex, were consistently detected across both line‐based and GMM analyses. In the line morphometric dataset, ImageJ‐derived surface area and proportion indices further supported these patterns. Fish from Gadeokdo exhibited a characteristically deeper body profile with fuller cranial–abdominal contours (e.g., higher BD/SL and BD/TL), whereas individuals from Taean showed a more streamlined body form accompanied by relatively greater dorsal surface‐area proportions (e.g., higher DSA/SL). These trends were congruent with GMM deformation patterns, which revealed outward landmark displacements around the dorsal and ventral midbody regions in Gadeokdo specimens (d5, d8, v21, v22), and anterior shifts associated with a more elongate head and snout region in Taean specimens (d1–d3, v13–v15), consistent with higher SNL/HL values. These site‐specific morphological patterns align with well‐documented ecological contrasts between the two regions (Rebstock & Kang, 2003). The hydrodynamically energetic and structurally heterogeneous rocky–sandy habitats of Gadeokdo in the Southeast Sea may favour deeper‐bodied forms with enhanced lateral stability, whereas the macrotidal, fine‐sediment environments of Taean in the Yellow Sea may select for more streamlined profiles that facilitate efficient benthic contact and movement over softer substrates. Although direct causal inference remains limited in the absence of long‐term environmental datasets or controlled experimental validation, the concordance between observed morphological differences and regional habitat characteristics suggests that local environmental conditions contribute to the geographic structuring of body shape (Santos et al., 2019; Špelić et al., 2021; Spinner et al., 2016). Comparable site‐associated morphological divergence has been reported in other benthic fishes inhabiting Korean coastal waters, including flatfishes such as Paraplagusia japonica and Platichthys stellatus, as well as other demersal species, where regional habitat characteristics and limited effective dispersal were suggested to contribute to population‐level differentiation (Jeong et al., 2014; Myoung et al., 2016; Park & Kim, 2022). These regional comparisons provide a broader ecological context for the patterns observed in Kareius bicoloratus, while not implying direct equivalence in the underlying mechanisms or selective pressures. In addition, life‐history traits common to benthic coastal fishes—such as settlement to demersal habitats and relatively restricted post‐settlement movement—may facilitate the maintenance of site‐associated morphological differences once established, as suggested in regional studies of Korean coastal species (Myoung et al., 2016; Park & Kim, 2022). Hydrographic features, including regional current systems, have also been shown to influence effective connectivity in coastal fishes, potentially reinforcing spatial structuring even over relatively short geographic distances (Yu & Kim, 2018). Importantly, mitochondrial DNA analyses revealed significant but moderate genetic differentiation between the two regions, and individual‐level Mantel tests indicated a weak yet statistically significant association between genetic and morphological distances. While this association does not imply a direct genetic determination of body shape, it suggests that the observed geographic patterning of morphology is not solely environmentally driven, but likely reflects the combined effects of local environmental conditions and population‐level genetic structuring. Such environmentally mediated shape variation, operating in concert with genetic differentiation, may influence swimming stability, bottom adhesion and manoeuvrability under conditions of strong currents, low visibility or unstable sediments (Mohaddasi et al., 2013; Yeung et al., 2021). Collectively, these findings underscore the importance of integrating spatial, ecological and genetic contexts when interpreting morphometric variation in coastal demersal species such as Kareius bicoloratus. Interaction effects and stable traits Morphological development in K. bicoloratus did not operate independently across biological and environmental factors. Rather, multiple indices exhibited significant interaction effects between factors, highlighting the non‐additive and context‐dependent nature of morphometric variation (de la Hoz Franco & Budy, 2005). Several cranial and dorsal features changed with size in different ways between males and females (e.g., CL/SL, d1–2/HL), indicating sex‐specific allometric patterns. Other traits varied with size only in particular environments, such as certain anterior‐to‐midbody spans that showed size sensitivity at Gadeokdo but not Taean (e.g., d2–5/SL or d5–10/SL), suggesting that ecological context modulates the expression of growth‐related shape changes (Claramunt & Wahl, 2000; Horta‐Lacueva et al., 2021; Traverso et al., 2024). Likewise, some cranial and dorsal proportions differed between sexes at one site but not the other (e.g., ULL/HL), indicating that sexual dimorphism may be enhanced or diminished depending on local habitat. GMM analyses confirmed these interactions, revealing nonparallel deformation patterns across size–sex–site combinations, with differences concentrated in cranial, dorsal and abdominal landmark regions. These findings highlight the complexity of morphological variation, reflecting overlapping influences of development, physiology and environment (Claramunt & Wahl, 2000; de la Hoz Franco & Budy, 2005; Horta‐Lacueva et al., 2021; Linløkken & Hesthagen, 2011). Recognizing such interactions is critical for understanding the developmental, ecological and potentially genetic underpinnings of morphological plasticity in fish (Mir et al., 2014; Mulyasari et al., 2023). Despite this complexity, certain aspects of body form remained remarkably stable. Measurements associated with fin‐base lengths (AFBL/SL, DFBL/SL) and specific cranial or caudal dimensions (e.g., v13–14/HL, Ang5) showed minimal variation across groups, a pattern mirrored in GMM by limited landmark displacement in these areas. Such invariance suggests underlying developmental or functional constraints in regions essential for propulsion, balance or cranial support, and these traits may serve as reliable structural references for comparative studies or aquaculture evaluation (Kim, Park, et al., 2025; Lal et al., 2024; Zenger et al., 2019). Possible implications for fisheries management and aquaculture Although primarily analytical, the findings offer several potential applications to fisheries and aquaculture. Morphological traits that clearly differentiate size and sex groups, such as BD/SL, UBD/BD and DSA/SL, may serve as phenotypic indicators for broodstock evaluation or developmental quality assessment in hatcheries or farms (Jun et al., 2009; Jung et al., 2019; Kim, Park, et al., 2025; Mojekwu & Anumudu, 2015). The site‐related differences in body depth, snout projection and dorsal robustness (e.g., BD/TL, SNL/HL, BPSA/DSA) may assist in identifying population structure or evaluating habitat suitability for stock enhancement or relocation (Daga et al., 2012; Ndiwa et al., 2016). Interaction‐sensitive indices such as ULL/HL and EBD/HL highlight the value of considering combined effects when evaluating trait plasticity, especially in ecologically variable systems. Because several shape differences likely stem from environmental influences, morphometric assessments may also act as supplementary indicators of ecological conditions in coastal systems. Establishing natural morphological baselines for wild populations provides an important reference for evaluating the performance and developmental stability of hybrids, triploids or selectively bred lines (Shuai et al., 2018; Zenger et al., 2019). Limitations and future considerations This study was based on two sampling sites and a single collection period, and the lack of temporally resolved environmental data limits the strength of ecological inferences. Although classical and GMM analyses provided consistent insights into the biological and spatial drivers of shape variation, integrating ecological and experimental approaches under artificial rearing conditions would enable more precise identification of underlying mechanisms (Kang et al., 2012; Kim et al., 2013; Yu et al., 2023). In addition, genetic inferences were based on mitochondrial markers, which may not fully capture genome‐wide patterns underlying complex morphological traits. Future studies incorporating multi‐year sampling, controlled rearing experiments and genomic analyses will help clarify the relative contributions of phenotypic plasticity, local adaptation and demographic history to morphological variability in K. bicoloratus. CONCLUSION This study provides an integrated characterization of morphological variation in wild Kareius bicoloratus by combining classical truss‐based measurements with GMM analyses. Both approaches consistently identified body size and sex as the primary drivers of shape variation, while geographic origin contributed additional but more moderate differentiation. The strong correspondence between key truss indices and landmark‐based deformation patterns demonstrates the complementary value of linear and geometric methods in capturing phenotypic structure. Shape variation was most pronounced in body depth, surface expansion and cranial–abdominal contours, whereas fin‐base lengths and portions of the cranial and caudal regions remained comparatively stable, and interaction effects highlighted context‐dependent morphological development. Mitochondrial DNA analyses revealed significant but moderate genetic differentiation between regions, and individual‐level comparisons indicated a weak yet significant association between genetic and morphological divergence, suggesting that observed shape variation reflects the combined influences of biological, spatial and genetic factors. Collectively, these findings establish a coherent morphological reference for K. bicoloratus and underscore the utility of multifactorial morphometric frameworks for describing shape diversity in flatfishes. AUTHOR CONTRIBUTIONS Eun Jeong Kim: Investigation; methodology and visualization. Songhyun Gwon: Methodology; validation and investigation. Ju Hwan Park: Methodology; investigation and data curation. In Chul Bang: Conceptualization and validation. Yoon Kwon Nam: Conceptualization; data curation; validation; writing—original and revised draft. Supporting information Table S1. Summary of long‐term environmental characteristics of the two sampling regions (2020–2024), based on monthly averaged data. Table S2. Nomenclature and anatomical descriptions of the morphological landmarks used in this study. Landmark definitions are based on external body features identifiable in both dorsal and ventral views. This classification refers to the previously published anatomical framework by Kim, Park, et al. (2025). Table S3. Mean ± SDs (untransformed raw means in proportions, ratios or degrees) of morphological indices based on factor grouping. Table S4. Summary of statistical results from the three‐way ANOVA for each morphometric index, including main effect significance, effect size (η 2), observed power and assumption checks. Decision route refers to the final statistical method applied based on normality and variance tests. Table S5. Results of digitizing error assessment based on Procrustes ANOVA and root mean square error (RMSE) calculated from repeated landmark digitizations (three replicates per individual, n = 20) for dorsal and ventral landmark configurations. Table S6. Pairwise discriminant analysis results for TL‐size classes and sex groups based on geometric morphometric data. Table S7. Summary of allometry tests and residual‐based analyses for sex and site factors. Figure S1. Map showing the geographic locations of the two sampling sites used in this study. Site T is located on the western coast of the Korean Peninsula (Yellow Sea, Taean Peninsula), whereas Site G is located on the southern coast (South Sea, Gadeokdo area). Enlarged panels illustrate detailed sampling locations within each site, with purple dots indicating individual sampling points (MOS#1–MOS#4). Latitude and longitude grids are shown to provide geographic reference, and scale bars indicate distances of 10 km. The coordinates of nearby marine observation stations used for environmental and oceanographic context are listed below the map, including tidal‐ and wave‐observation stations associated with each sampling region. Figure S2. Workflow for statistical decision‐making in three‐way ANOVA and post hoc comparisons. This decision tree outlines the selection of appropriate post hoc or non‐parametric tests based on the outcomes of normality testing, Levene's test for equal variances, and interaction effects among main factors. EMM (estimated marginal means) were used for post hoc pairwise comparisons within the GLM framework. Interaction effects were further assessed using simple effects tests within subgroups. The non‐parametric route was used when the assumption of normality was violated. Figure S3. Raw means (±SD) of all 81 morphometric indices across all samples prior to grouping by size, sex or site. Indices are grouped and colour‐coded by type: relative to total length (TL), standard length (SL), head length (HL), other ratio‐based indices and angular measurements (degrees). For visualization purposes, indices with large values—including angular indices and several ratio‐based indices (e.g., CF, BSA/SL, HSA/HL, etc.)—are plotted using the right Y‐axis, while the rest are plotted against the left Y‐axis. Figure S4. Coefficient of variation (%CV) for all 81 morphometric indices calculated from the ungrouped dataset (prior to size, sex or site classification). Indices are arranged in descending order of CV. Dot colours correspond to the significance categories determined from multiple post hoc comparisons (as shown in Figure 3). Dashed horizontal lines at 5% and 25% CV are shown for visual reference. Figure S5. Principal component analysis (PCA) of dorsal‐ and ventral‐side geometric morphometric data. Principal component analysis (PCA) of Procrustes‐aligned landmark coordinates showing patterns of shape variation for the dorsal side (left panels) and ventral side (right panels). The upper panels display the percentage of variance explained by successive principal components. Shape changes associated with the first principal component (PC1) are visualized using thin‐plate spline deformation grids and corresponding wireframe representations, illustrating the major directions of shape variation captured by PC1. The lower panels show scatter plots of individual specimens projected onto the PC1–PC2 morphospace for dorsal and ventral datasets, respectively. Landmark numbers correspond to those defined in the geometric morphometric configuration. Figure S6. Canonical variate analysis (CVA) of ventral‐side geometric morphometric shape variation in Kareius bicoloratus. CVA based on Procrustes‐aligned landmark coordinates of the ventral side, presented as a supplementary comparison to the dorsal‐side analysis shown in Figure 7. CVA scatter plots illustrate shape variation according to size class (large, L; medium‐large, ML; small‐medium, SM; small, S) and sex (female, F; male, M; underdeveloped, U). Frequency distributions of canonical variate 1 (CV1) scores are shown for the two sampling sites (Site T and Site G). Each point represents an individual specimen, and ellipses indicate group dispersion around centroids. Figure S7. Wireframe visualizations of canonical shape changes along CV1 and CV2 axes derived from CVA. Wireframe diagrams illustrating canonical shape changes associated with canonical variates 1 (CV1) and 2 (CV2) from the canonical variate analysis (CVA). Shape deformations are shown separately for the dorsal side (upper panels) and ventral side (lower panels), and are presented according to the major grouping factors: size class (Size), sex (Sex) and sampling site (Site). For each factor, wireframes represent shape changes along the positive and negative extremes of the corresponding canonical axis, highlighting the principal landmarks contributing to group separation. Landmark numbers correspond to those defined in the geometric morphometric configuration. These visualizations complement the CVA scatter plots shown in Figure 7 (dorsal side) and Figure S6 (ventral side) by illustrating the morphological basis of canonical differentiation. Figure S8. Principal component analysis (PCA) of size‐corrected residual shape data following allometry testing. PCA performed on size‐corrected residuals of Procrustes‐aligned landmark coordinates to evaluate shape variation independent of allometric effects. Residual datasets were obtained after regression of shape on centroid size as part of the allometry test. Results are shown separately for the dorsal side (left panels) and ventral side (right panels). The upper panels display the percentage of variance explained by successive principal components. Shape changes associated with the first principal component (PC1) are visualized using thin‐plate spline deformation grids. The lower panels show scatter plots of individual specimens projected onto the PC1–PC2 morphospace based on residual shape data, illustrating overall dispersion after removal of size‐related shape variation. Figure S9. Canonical variate analysis (CVA) of size‐corrected residual shape data following allometry testing. CVA was performed on size‐corrected residuals of Procrustes‐aligned landmark coordinates to evaluate group differentiation independent of allometric effects. Residual datasets were obtained after regression of shape on centroid size as part of the allometry test. Results are shown separately for the dorsal side (left panels) and ventral side (right panels). Upper panels present wireframe visualizations of canonical shape changes along the primary canonical axis, illustrating deformation patterns after removal of size‐related shape variation. Lower panels show frequency distributions of canonical variate 1 (CV1) scores for the two sampling sites (Site T and Site G), highlighting site‐associated differences in residual shape structure. Figure S10. Mantel test assessing the relationship between genetic and morphological distances. Scatter plots showing the relationship between pairwise genetic distances (mitochondrial DNA p‐distance) and morphological distances (Procrustes distance) based on geometric morphometric data. Analyses were conducted separately for the dorsal side (left panel) and ventral side (right panel). Each point represents a pairwise comparison between individuals. Red lines indicate linear regression trends for visualization purposes. Mantel test results are shown within each panel, including Mantel's r statistic and associated p‐value, indicating a weak but statistically significant correlation between genetic and morphological distances. ACKNOWLEDGMENTS This research was financially supported by the Ministry of Small and Medium‐sized Enterprises (SMEs) and Startups (MSS), Korea, under the ‘Industry‐Academia‐Research Collaborative R&D program (project no. RS‐2023‐00224997)’ supervised by the Korea Technology and Information Promotion Agency for SMEs. DATA AVAILABILITY STATEMENT The data that support the findings of this study are available from the corresponding author upon reasonable request. REFERENCES Associated Data Supplementary Materials Table S1. Summary of long‐term environmental characteristics of the two sampling regions (2020–2024), based on monthly averaged data. Table S2. Nomenclature and anatomical descriptions of the morphological landmarks used in this study. Landmark definitions are based on external body features identifiable in both dorsal and ventral views. This classification refers to the previously published anatomical framework by Kim, Park, et al. (2025). Table S3. Mean ± SDs (untransformed raw means in proportions, ratios or degrees) of morphological indices based on factor grouping. Table S4. Summary of statistical results from the three‐way ANOVA for each morphometric index, including main effect significance, effect size (η 2), observed power and assumption checks. Decision route refers to the final statistical method applied based on normality and variance tests. Table S5. Results of digitizing error assessment based on Procrustes ANOVA and root mean square error (RMSE) calculated from repeated landmark digitizations (three replicates per individual, n = 20) for dorsal and ventral landmark configurations. Table S6. Pairwise discriminant analysis results for TL‐size classes and sex groups based on geometric morphometric data. Table S7. Summary of allometry tests and residual‐based analyses for sex and site factors. Figure S1. Map showing the geographic locations of the two sampling sites used in this study. Site T is located on the western coast of the Korean Peninsula (Yellow Sea, Taean Peninsula), whereas Site G is located on the southern coast (South Sea, Gadeokdo area). Enlarged panels illustrate detailed sampling locations within each site, with purple dots indicating individual sampling points (MOS#1–MOS#4). Latitude and longitude grids are shown to provide geographic reference, and scale bars indicate distances of 10 km. The coordinates of nearby marine observation stations used for environmental and oceanographic context are listed below the map, including tidal‐ and wave‐observation stations associated with each sampling region. Figure S2. Workflow for statistical decision‐making in three‐way ANOVA and post hoc comparisons. This decision tree outlines the selection of appropriate post hoc or non‐parametric tests based on the outcomes of normality testing, Levene's test for equal variances, and interaction effects among main factors. EMM (estimated marginal means) were used for post hoc pairwise comparisons within the GLM framework. Interaction effects were further assessed using simple effects tests within subgroups. The non‐parametric route was used when the assumption of normality was violated. Figure S3. Raw means (±SD) of all 81 morphometric indices across all samples prior to grouping by size, sex or site. Indices are grouped and colour‐coded by type: relative to total length (TL), standard length (SL), head length (HL), other ratio‐based indices and angular measurements (degrees). For visualization purposes, indices with large values—including angular indices and several ratio‐based indices (e.g., CF, BSA/SL, HSA/HL, etc.)—are plotted using the right Y‐axis, while the rest are plotted against the left Y‐axis. Figure S4. Coefficient of variation (%CV) for all 81 morphometric indices calculated from the ungrouped dataset (prior to size, sex or site classification). Indices are arranged in descending order of CV. Dot colours correspond to the significance categories determined from multiple post hoc comparisons (as shown in Figure 3). Dashed horizontal lines at 5% and 25% CV are shown for visual reference. Figure S5. Principal component analysis (PCA) of dorsal‐ and ventral‐side geometric morphometric data. Principal component analysis (PCA) of Procrustes‐aligned landmark coordinates showing patterns of shape variation for the dorsal side (left panels) and ventral side (right panels). The upper panels display the percentage of variance explained by successive principal components. Shape changes associated with the first principal component (PC1) are visualized using thin‐plate spline deformation grids and corresponding wireframe representations, illustrating the major directions of shape variation captured by PC1. The lower panels show scatter plots of individual specimens projected onto the PC1–PC2 morphospace for dorsal and ventral datasets, respectively. Landmark numbers correspond to those defined in the geometric morphometric configuration. Figure S6. Canonical variate analysis (CVA) of ventral‐side geometric morphometric shape variation in Kareius bicoloratus. CVA based on Procrustes‐aligned landmark coordinates of the ventral side, presented as a supplementary comparison to the dorsal‐side analysis shown in Figure 7. CVA scatter plots illustrate shape variation according to size class (large, L; medium‐large, ML; small‐medium, SM; small, S) and sex (female, F; male, M; underdeveloped, U). Frequency distributions of canonical variate 1 (CV1) scores are shown for the two sampling sites (Site T and Site G). Each point represents an individual specimen, and ellipses indicate group dispersion around centroids. Figure S7. Wireframe visualizations of canonical shape changes along CV1 and CV2 axes derived from CVA. Wireframe diagrams illustrating canonical shape changes associated with canonical variates 1 (CV1) and 2 (CV2) from the canonical variate analysis (CVA). Shape deformations are shown separately for the dorsal side (upper panels) and ventral side (lower panels), and are presented according to the major grouping factors: size class (Size), sex (Sex) and sampling site (Site). For each factor, wireframes represent shape changes along the positive and negative extremes of the corresponding canonical axis, highlighting the principal landmarks contributing to group separation. Landmark numbers correspond to those defined in the geometric morphometric configuration. These visualizations complement the CVA scatter plots shown in Figure 7 (dorsal side) and Figure S6 (ventral side) by illustrating the morphological basis of canonical differentiation. Figure S8. Principal component analysis (PCA) of size‐corrected residual shape data following allometry testing. PCA performed on size‐corrected residuals of Procrustes‐aligned landmark coordinates to evaluate shape variation independent of allometric effects. Residual datasets were obtained after regression of shape on centroid size as part of the allometry test. Results are shown separately for the dorsal side (left panels) and ventral side (right panels). The upper panels display the percentage of variance explained by successive principal components. Shape changes associated with the first principal component (PC1) are visualized using thin‐plate spline deformation grids. The lower panels show scatter plots of individual specimens projected onto the PC1–PC2 morphospace based on residual shape data, illustrating overall dispersion after removal of size‐related shape variation. Figure S9. Canonical variate analysis (CVA) of size‐corrected residual shape data following allometry testing. CVA was performed on size‐corrected residuals of Procrustes‐aligned landmark coordinates to evaluate group differentiation independent of allometric effects. Residual datasets were obtained after regression of shape on centroid size as part of the allometry test. Results are shown separately for the dorsal side (left panels) and ventral side (right panels). Upper panels present wireframe visualizations of canonical shape changes along the primary canonical axis, illustrating deformation patterns after removal of size‐related shape variation. Lower panels show frequency distributions of canonical variate 1 (CV1) scores for the two sampling sites (Site T and Site G), highlighting site‐associated differences in residual shape structure. Figure S10. Mantel test assessing the relationship between genetic and morphological distances. Scatter plots showing the relationship between pairwise genetic distances (mitochondrial DNA p‐distance) and morphological distances (Procrustes distance) based on geometric morphometric data. Analyses were conducted separately for the dorsal side (left panel) and ventral side (right panel). Each point represents a pairwise comparison between individuals. Red lines indicate linear regression trends for visualization purposes. Mantel test results are shown within each panel, including Mantel's r statistic and associated p‐value, indicating a weak but statistically significant correlation between genetic and morphological distances. Data Availability Statement The data that support the findings of this study are available from the corresponding author upon reasonable request.