Multi-scale environmental drivers of heron and egret colony assemblages in Korea using self-organizing map clustering Authors: SeonYong Lee, Han-I Choi, Jae-Woong Hwang, Ji-Yeon Lee, Dong-Won Kim, Yu-Seong Choi, Hyung-Kyu Nam Published: 2026 ABSTRACT AbstractColonial herons and egrets breed within highly heterogeneous landscapes where environmental conditions vary across multiple spatial scales. Although landscape structure is expected to shape colony-level species assemblages, explicit multi-scale empirical assessments remain scarce, particularly at national extents. We sought to (1) identify emergent species-assemblage types among heron and egret colonies across South Korea, (2) determine how these assemblages correspond to landscape heterogeneity across nested spatial scales, and (3) quantify the scale-dependent importance of environmental predictors influencing colony identity. We compiled species-specific nest counts from 176 breeding colonies and applied Self-Organizing Maps (SOM) to classify colony assemblages. Landscape variables were extracted from ESA CCI Plant Functional Type (PFT) datasets within five buffer radii: 500 m, 1 km, 3 km, 5 km, and 10 km. The relative importance of land-cover predictors at each spatial scale was evaluated using Random Forest models. SOM analysis revealed six ecologically distinct colony groups differing in species dominance, richness, and geographic distribution. Large-bodied species predominated in inland forested regions, whereas medium- and small-bodied species clustered in coastal and urban–rural mosaics. Environmental predictors showed scale-dependent patterns: built areas, managed herbaceous vegetation, and water-related cover dominated at fine scales; natural grasslands, bare land, and tree-related PFTs emerged at intermediate scales; and broader land-use variables remained important at 10 km. Colony assemblages arise from hierarchical interactions among fine-scale habitat features, river-corridor landscapes, and broader-scale land-use regimes. This hierarchical multi-scale perspective provides a landscape-ecological framework for anticipating how land-use change may alter the multi-scale landscape contexts supporting colonial wading birds. Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1038/s41598-026-57885-5.Keywords: Landscape heterogeneity, Multi-scale analysis, Colony assemblage, Self-organizing map (SOM), Colonial wading birds, ArdeidaeSubject terms: Ecology, Ecology, Environmental sciences Introduction Colonial wading birds such as herons and egrets depend on spatial mosaics of foraging and breeding habitats embedded within heterogeneous landscapes1,2. Mixed-species colonies—often comprising species with differing ecological requirements—emerge from interactions among local habitat conditions, regional resource availability, and broader-scale landscape structure2,3. Understanding how these colonies assemble across space therefore requires an integrative framework that links ecological processes operating across multiple, nested spatial scales, a core principle of landscape ecology4,5. Because landscape structure constrains movement, regulates access to food resources, shapes exposure to disturbance, and mediates interspecific interactions, colonies can be viewed as spatially discrete breeding units whose species composition reflects the surrounding landscape mosaic and its multi-scale context2,6. In this study, we define a colony as a geographically discrete breeding site that may include one or more heron and egret species, and use the term “colony assemblage” to refer to the species-specific nest composition within each colony. A multi-scale landscape-ecological perspective is particularly important for understanding colony assemblages because breeding colonies are influenced by environmental conditions operating at different spatial extents, from immediate nesting surroundings to broader landscape contexts7–9. Fine-scale surroundings may determine immediate nesting conditions, disturbance exposure, and access to nearby foraging habitats, whereas broader landscape contexts may reflect river corridors, agricultural mosaics, coastal–inland gradients, and regional land-use intensity5,10,11. Therefore, colony assemblages are unlikely to be explained by a single spatial scale alone, and a nested multi-scale framework is needed to evaluate how local habitat features and broader landscape structure jointly shape colony-level species composition. Across East Asia, rapid land-use change has substantially altered both the distribution and spatial configuration of habitats used by Ardeidae species12,13. Key foraging environments—including rice paddies, coastal lowlands, river corridors, and forest–agriculture ecotones—have undergone pronounced modification driven by urban expansion, agricultural intensification, and river engineering14,15. These changes affect not only the quantity of foraging habitat but also the spatial arrangement and configuration of habitat patches, thereby restructuring the ecological template upon which colonial breeding sites are established13; Schweizer et al., 2020). Despite the pronounced spatial heterogeneity of these landscapes, however, little is known about how colony-level species assemblages respond to landscape structure across spatial scales. Most previous studies have focused on habitat selection at individual breeding sites or on single-species responses, providing limited insight into the multi-scale processes underlying the formation and persistence of mixed-species colonies. The Korean Peninsula offers an ideal setting for addressing these questions. The region encompasses strong environmental gradients from coastal to inland areas, extensive river systems forming distinct riparian corridors, and mosaic agricultural landscapes that support the primary foraging activities of herons and egrets (NGII 2013a, 2013b). Colonial breeding sites occur across this diverse landscape, ranging from estuarine rice fields and coastal plains to inland forested valleys, providing a unique opportunity to examine how landscape heterogeneity shapes colony-level species assemblages16,17. Nationwide surveys further reveal substantial variation among colonies in nest numbers, species richness, and dominant species combinations (NIER 2012), suggesting that spatial configurations of landscape elements may function as ecological “filters” influencing colony identity17. Nevertheless, no study to date has evaluated these patterns at a national scale using spatially explicit, multi-scale landscape metrics. A landscape-ecological perspective provides several key insights into colony assembly. First, it recognizes that ecological patterns arise from hierarchical spatial processes, allowing colony structure to be interpreted within nested spatial contexts, ranging from immediate foraging environments within hundreds of meters to broader-scale land-use regimes spanning several kilometers8,9. Second, it accounts for the possibility that species assemblages are shaped by nonlinear interactions among water availability, agricultural matrices, forest cover, and other land-cover components18,19. Third, it emphasizes the importance of spatial configuration and connectivity—factors that are particularly critical for highly mobile foragers whose reproductive success depends on repeated movements between breeding sites and multiple foraging habitats20; Teitelbaum et al., 2021). In this study, we integrate nationwide breeding colony surveys with Self-Organizing Map (SOM) analysis, an unsupervised nonlinear ordination method that arranges observations according to similarity in multivariate composition (Fig. S1), to identify species-composition patterns and emergent colony assemblage groups across heron and egret breeding sites in South Korea. We then assess how these colony assemblage groups correspond to landscape gradients and structural features measured across five nested spatial scales: 500 m, 1 km, 3 km, 5 km, and 10 km7,9. Using ESA CCI Plant Functional Type (PFT) layers, we quantify key land-cover components—including water bodies, managed and natural herbaceous cover, forest structure, bare land, and urban areas—to evaluate colony responses not only to local habitat conditions but also to landscape mosaics at broader spatial scales (ESA CCI21. By combining unsupervised species-assemblage modeling with multi-scale landscape analysis, this study provides the first nationwide assessment of how hierarchical landscape structure shapes colony-level species composition in herons and egrets. Methods Breeding colony dataset This study examined the breeding distribution of egret and heron species in the Republic of Korea using nationwide survey data collected in 2011–2012 (NIER 2012). These baseline data were supplemented with additional colony records identified through researcher field observations, news reports, and online sources such as blogs, which occasionally document newly established or previously unrecorded colonies. Based on this integrated dataset, all candidate locations were visited twice during the main breeding season from March to July in 2018 and 2019 to verify the presence of active breeding. For each confirmed colony, geographic coordinates (longitude and latitude) and species-specific nest counts were recorded. When a colony was surveyed twice within a breeding season, the maximum species-specific nest count recorded across the two visits was used for analysis to reduce the risk of underestimating active nests due to asynchronous breeding phenology or temporary differences in detectability. Self-organizing map (SOM) analysis A Self-Organizing Map (SOM) was used to identify latent ecological structure in the species composition of breeding colonies. SOM is an unsupervised neural-network algorithm that projects high-dimensional data onto a lower-dimensional lattice while preserving topological relationships among observations22,23. This approach is useful for ecological classification because colonies with similar multivariate species-composition profiles are arranged close to one another on the SOM lattice. In this study, each colony was represented by a vector of log-transformed species-specific nest counts. During SOM training, each colony vector was matched to the most similar neuron on the SOM lattice, referred to as the best-matching unit (BMU). The BMU and its neighboring neurons were iteratively updated so that colonies with similar species-composition profiles were arranged close to one another on the two-dimensional lattice. Breeding colony data are typically right-skewed, with a small number of large colonies exerting disproportionate influence on the distribution. To reduce this bias and improve numerical stability in the SOM analysis, all species-specific nest counts were log-transformed using log(n + 1) before SOM training. This transformation reduced the influence of extremely large colonies while retaining variability among smaller colonies and improving the suitability of abundance data for multivariate pattern analysis24. Model performance was assessed using two diagnostic metrics: quantization error and topographic error. Quantization error was defined as the mean distance between each observation and its best-matching unit, where lower values indicate better representation of the input space. Topographic error was defined as the proportion of observations for which the first and second best-matching units were not adjacent on the SOM lattice, where lower values indicate stronger preservation of topological structure. These metrics are commonly used to evaluate representation accuracy and topology preservation in SOM outputs23,25. Using the selected grid, the final SOM was trained for 200 iterations with a learning rate decreasing linearly from 0.05 to 0.01. Similarity among neurons in standardized species-composition space was quantified using Euclidean distance. After training, QE and TE were recalculated to confirm adequate representation of the input structure. Codebook vectors for all neurons were extracted for subsequent cluster analysis. Because SOM does not inherently define discrete clusters, hierarchical clustering was applied to the final codebook vectors. Euclidean distance matrices were computed, and Ward’s D2 linkage method was used to minimize within-cluster variance26. Candidate cluster solutions from k = 4 to k = 7 were evaluated using both quantitative and ecological criteria. Quantitative diagnostics included cluster-size distribution, cluster-size coefficient of variation, mean within-cluster distance, mean silhouette width, Calinski–Harabasz index, and Davies–Bouldin index27–29. In addition, SOM-lattice continuity and ecological interpretability of species-composition patterns were considered to avoid over-partitioning ecologically similar assemblages. These results are summarized in Supplementary Table S1. The final SOM was trained on a 6 × 7 hexagonal lattice for 200 iterations and showed acceptable representation of the input species-composition data, with a quantization error of 0.609 and a topographic error of 0.006. The six-cluster solution was selected because it provided the best balance between statistical separation, SOM-lattice continuity, and ecological interpretability. Although the seven-cluster solution showed slightly greater compactness in some diagnostics, it further divided ecologically similar assemblages and was therefore considered over-partitioned. Each colony was assigned to a cluster according to the cluster membership of its best-matching unit. Cluster boundaries were visualized on the SOM grid to evaluate cluster separation and topological continuity. Kernel density estimation of breeding colony distributions To quantify spatial concentration patterns of breeding colonies within each SOM-derived cluster, kernel density estimation (KDE) was applied separately to each cluster. KDE was used only as a descriptive visualization of the spatial concentration of observed colonies, not as a predictive model of colony occurrence or colony-group suitability across unsampled areas. To account for differences in colony importance, KDE was weighted by the number of nests at each breeding site, such that colonies with larger nest numbers contributed more strongly to the estimated density surface. Weighted KDE was implemented using a bivariate Gaussian kernel, with bandwidth matrices automatically selected for each cluster using the plug-in estimator (Hpi), as implemented in the ks R package30. To ensure numerical stability and comparability among clusters, kernel weights were rescaled prior to density estimation so that their sum equaled the sample size within each cluster. Because the present study did not include an explicit non-colony or background sampling framework, we did not extrapolate SOM-derived colony-group probabilities to the entire study area. Environmental variable extraction and landscape analysis To assess whether SOM-derived colony assemblage groups reflected meaningful differences in landscape structure rather than statistical artifacts, environmental variables were quantified within multiple spatial buffers surrounding each colony. Circular buffers with radii of 500 m, 1 km, 3 km, 5 km, and 10 km were generated. These five nested spatial scales were selected to represent ecological contexts relevant to colonial herons and egrets and to follow a multi-scale habitat framework commonly used in landscape ecology7–9, rather than exact species-specific foraging ranges. The 500 m and 1 km buffers were used to characterize immediate colony surroundings, local nesting conditions, disturbance exposure, and nearby foraging habitats. The 3 km and 5 km buffers represented local landscape mosaics that may influence routine movements between colonies and surrounding foraging habitats, whereas the 10 km buffer captured broader land-use, coastal–inland, riverine, agricultural, and geomorphological contexts. Land-cover composition was quantified as both the area and proportional cover of each ESA CCI Plant Functional Type (PFT) class within each buffer. Proportional cover was calculated as the area of each PFT class divided by the total buffer area. Land-cover data were obtained from the ESA CCI Plant Functional Type (PFT) dataset. The PFT layer used in this study corresponded to 2019 and was therefore temporally aligned with the 2018–2019 breeding colony survey period. A total of 16 PFT classes were used in the analysis31; https://maps.elie.ucl.ac.be/CCI/viewer/, accessed 21 November 2025). The ecological definitions of the 16 PFT classes are summarized in Table 1, and their proportional cover across colony buffers is provided in Supplementary Table S3. Table 1Summary of the Plant Functional Type (PFT) classes derived from the ESA CCI Land Cover dataset. Each class represents the fractional contribution of a specific vegetation or surface type within a pixel. Definitions provided here outline the ecological meaning and classification criteria for all PFT variables incorporated into this study.CodeLong nameDescriptionWATERWaterGeneral water mask. Indicates the presence or proportion of all water surfaces, including both inland and oceanic water.WATER_INLANDInland waterProportion of inland water bodies such as lakes, rivers, and reservoirs within a pixel; excludes oceanic water.WATER_OCEANOceanic waterProportion of marine or coastal ocean water within a pixel.LANDLandLand–water mask used to distinguish terrestrial areas from aquatic surfaces; often used as a reference layer for PFT interpretation.BAREBareProportion of non-vegetated bare surfaces such as rock, soil, gravel, sand, and deserts.BUILTBuiltProportion of artificial/urban surfaces, including buildings, roads, paved areas, and other human-made structures.GRASS_MANManaged grassesProportion of managed herbaceous vegetation, including pasture, hayfields, and herbaceous croplands (e.g., cereals, rice, forage crops).GRASS_NATNatural grassesProportion of natural herbaceous vegetation, including wild grasslands, natural meadows, and herbaceous wetland vegetation.SHRUBS_BDBroadleaved deciduous shrubsProportion of broadleaved deciduous shrubs, which shed leaves seasonally.SHRUBS_BEBroadleaved evergreen shrubsProportion of broadleaved evergreen shrubs, which retain leaves year-round.SHRUBS_NDNeedleleaved deciduous shrubsProportion of needle-leaved deciduous shrubs, typically small larch-type shrubs that drop needles seasonally.SHRUBS_NENeedleleaved evergreen shrubsProportion of needle-leaved evergreen shrubs, low-stature coniferous shrubs that maintain needles year-round.TREES_BDBroadleaved deciduous treesProportion of broadleaved deciduous trees, typical of temperate deciduous forests (e.g., oaks, maples, birches).TREES_BEBroadleaved evergreen treesProportion of broadleaved evergreen trees, common in subtropical and warm-temperate evergreen forests (e.g., evergreen oaks, laurels).TREES_NDNeedleleaved deciduous treesProportion of needle-leaved deciduous trees, such as larch (Larix spp.), which lose needles in winter.TREES_NENeedleleaved evergreen treesProportion of needle-leaved evergreen trees, including pines, firs, and spruces that retain needles throughout the year. Summary of the Plant Functional Type (PFT) classes derived from the ESA CCI Land Cover dataset. Each class represents the fractional contribution of a specific vegetation or surface type within a pixel. Definitions provided here outline the ecological meaning and classification criteria for all PFT variables incorporated into this study. To evaluate the environmental predictors associated with the SOM-derived colony assemblage groups, Random Forest classification models were applied because this method can handle nonlinear relationships and interactions among predictors and is relatively robust to multicollinearity among environmental variables32. For each buffer radius, the response variable was SOM-derived colony-group membership, which refers to the categorical cluster identity assigned to each confirmed breeding colony based on species-specific nest composition, not the number of nests. Predictor variables were the 16 ESA CCI Plant Functional Type variables measured within each buffer. For model validation, colonies were divided into training and testing datasets using a 70:30 stratified random split that preserved the proportional representation of the six SOM-derived colony assemblage groups. Each Random Forest model was built using 1,000 trees. Predictive performance was evaluated using out-of-bag accuracy, test-set accuracy, balanced accuracy, and Cohen’s kappa33, and the results are provided in Supplementary Table S2. Variable importance was assessed based on mean decrease in accuracy derived from out-of-bag samples. For cross-scale comparison, the top five predictors at each radius were rescaled to a 0–100% range relative to the most influential variable within that scale. Software environment All analyses were conducted in R version 4.5.1. Key packages included kohonen for SOM analysis23, ks for weighted kernel density estimation30, sf34 and geodata35 for spatial data processing, dplyr36 and tidyr37 for data preprocessing, and rstatix for non-parametric statistical analyses38. Results Species composition across 176 heron and egret colonies Analysis of 176 breeding colonies revealed that Grey Heron and Great Egret were the predominant species nationwide, accounting for the largest numbers of nests in most colonies (Fig. 1). These large-bodied species were consistently present across southern, central, and eastern regions and were particularly dominant at inland sites. In contrast, medium- and small-bodied species—including Little Egret, Cattle Egret, Intermediate Egret, and Black-crowned Night Heron—showed more spatially restricted distributions. These species were concentrated primarily in southern provinces and lowland plains along the west coast, where several colonies formed regionally distinct mixed-species assemblages (Fig. 1). Spatial distribution of nests for the six heron and egret species breeding in Korea during 2018–2019. Maps show the geographic locations of breeding colonies for Grey Heron (N = 166), Great Egret (N = 134), Intermediate Egret (N = 54), Cattle Egret (N = 40), Little Egret (N = 46), and Black-crowned Night Heron (N = 45). Circle size is proportional to the number of nests at each colony, illustrating spatial variation in colony size and regional differences in breeding distribution across the Korean Peninsula. Regional variation in colony structure Clear geographic differentiation in colony structure was evident across regions (Fig. 1). Colonies in southern inland and southeastern areas were dominated by Grey Heron and Great Egret, forming large mixed aggregations characterized by high nest abundance. In contrast, colonies in the metropolitan region and western coastal plains frequently supported co-occurrence of Little Egret, Cattle Egret, and Black-crowned Night Heron, resulting in more diverse multi-species assemblages. Colonies in central inland areas and parts of the east coast tended to be smaller and simpler, often comprising one or two species and frequently dominated by either Grey Heron or Great Egret. SOM-based colony grouping Self-Organizing Map (SOM) analysis identified six distinct colony assemblage groups (k = 6) based on similarities in species-specific nest composition. These SOM-derived assemblage groups represent emergent assemblage types rather than predefined habitat categories or simple nest-abundance classes. This six-cluster solution was selected after comparing candidate solutions from k = 4 to k = 7 using quantitative diagnostics and ecological interpretability criteria (Supplementary Table S1). The final SOM showed acceptable representation of the high-dimensional species-composition data, with a quantization error of 0.609 and a topographic error of 0.006. The resulting clusters differed consistently in both species composition and spatial distribution (Fig. 2). SOM-derived colony assemblage groups and codebook dendrogram. (a) SOM lattice displaying six colony assemblage groups (k = 6) identified from species-composition patterns of breeding colonies. Labels shown within SOM units indicate neuron identifiers used for visualization and do not represent additional cluster categories. Cluster boundaries indicate the six SOM-derived colony assemblage groups. (b) Hierarchical clustering dendrogram of SOM codebook vectors using Ward’s D2 method, showing the corresponding six-group structure. Cluster I was concentrated in mid-western coastal and adjacent inland regions and consisted of mixed-species colonies in which Grey Heron, Great Egret, and smaller species co-occurred (Figs. 3 and 4). Cluster II occurred primarily in inland agricultural landscapes and was dominated by Grey Heron and Great Egret, forming relatively simplified assemblages. Cluster III was distributed mainly in forested and riverine areas of Gangwon-do and Gyeongsangbuk-do provinces and exhibited strong dominance of large-bodied species. Cluster IV occurred in mosaic landscapes of forest and farmland and showed relatively high proportions of medium- and small-bodied species, particularly Cattle Egret and Little Egret, with Grey Heron and Great Egret present at moderate levels. Cluster V was associated with urban–rural transitional zones and was characterized by a high contribution of medium- and small-bodied species, resulting in the most compositionally diverse assemblages. Cluster VI was found in forested and wetland-dominated landscapes with high naturalness and consisted largely of Grey Heron, forming near single-species assemblages. Spatial organization of SOM-derived colony assemblage groups across South Korea. Weighted kernel density estimation (KDE) maps illustrate the spatial concentration of colonial breeding sites for each self-organizing map (SOM)-derived colony assemblage group (Groups I–VI). KDE was performed using a bivariate Gaussian kernel, with colony locations weighted by nest numbers; bandwidth matrices (Hpi) were individually selected for each cluster, and weights were rescaled to allow numerical comparability among clusters. Warmer colors indicate higher relative colony density. For geographic context, additional reference panels include (upper left) provincial administrative boundaries labeled using abbreviations (GW: Gangwon-do, GG: Gyeonggi-do, CB: Chungcheongbuk-do, CN: Chungcheongnam-do, JB: Jeonbuk Special Self-Governing Province, JN: Jeollanam-do, GB: Gyeongsangbuk-do, GN: Gyeongsangnam-do) and (lower left) a relief map with major rivers (Han, Geum, Nakdong, Seomjin) overlaid. Species composition of heron and egret colonies across the six SOM-derived colony assemblage groups. Relative contributions of six species (Grey Heron, Great Egret, Intermediate Egret, Cattle Egret, Little Egret, and Black-crowned Night Heron) are shown for each cluster, illustrating distinct assemblage structures associated with the SOM-defined groups. Differences among clusters in species composition Comparison of species composition among clusters revealed marked differences in relative nest abundance despite the consistent presence of Grey Heron and Great Egret across all groups (Fig. 4). Clusters III and VI were strongly dominated by large-bodied species, which accounted for more than 90% of total nests. In contrast, Cluster V showed the highest contribution of medium- and small-bodied species, comprising over 80% of total nests. Clusters I and IV exhibited intermediate, mixed-species structures with variable proportional contributions of multiple species. Scale-dependent importance of land-cover variables in Random Forest models Random Forest models indicated that SOM-derived colony assemblage groups were associated with land-cover variables across all spatial scales, with predictor importance varying by buffer radius (Fig. 5). Model-validation results showed modest classification performance across buffer radii, with test-set accuracy ranging from 0.302 to 0.396, balanced accuracy from 0.150 to 0.247, and Cohen’s kappa from − 0.005 to 0.138 (Supplementary Table S2). Therefore, we interpret these models primarily as exploratory tools for assessing the relative importance of land-cover predictors rather than as high-accuracy predictive models of colony-group identity. At the 500 m scale, artificial or urban surfaces (BUILT), coastal water (WATER_OCEAN), managed herbaceous and agricultural vegetation (GRASS_MAN), water (WATER), and broadleaved deciduous shrublands (SHRUBS_BD) were the most influential. At 1,000 m, urban surfaces, managed vegetation, broadleaved deciduous forests (TREES_BD), natural grasslands (GRASS_NAT), and needle-leaved evergreen forests (TREES_NE) ranked highest. At 3,000 m, bare surfaces (BARE), natural grasslands, broadleaved deciduous forests, managed vegetation, and urban surfaces emerged as key predictors. At 5,000 m, urban surfaces, bare surfaces, coastal water, managed vegetation, and natural grasslands showed the greatest importance. At the broadest 10 km scale, urban surfaces, bare surfaces, managed vegetation, coastal water, and needle-leaved evergreen forests were the most influential variables distinguishing SOM-derived colony assemblage groups. Random Forest variable importance for the top five land-cover predictors associated with SOM-derived heron and egret colony assemblage groups at five buffer radii (500 m, 1 km, 3 km, 5 km, and 10 km). Importance values were derived from the mean decrease in classification accuracy based on out-of-bag samples and rescaled to 0–100% within each radius to allow comparison across spatial scales. Discussion Self-organizing map (SOM) analysis revealed that heron and egret breeding colonies in Korea form distinct colony assemblage groups that differ in species composition, dominance structure, and co-occurrence patterns. The main contribution of this study is its integration of colony-level species-composition clustering with multi-scale land-cover analysis at a national extent. This approach extends previous work that focused primarily on single-species habitat associations or local breeding-site characteristics by identifying emergent colony assemblage groups and evaluating how they are embedded within nested landscape contexts. These clusters closely corresponded to broad geographic gradients, including coastal–inland transitions, contrasts between major riverine lowlands and forested uplands, and variation in urban–rural land-use intensity. Colonies located in coastal plains and estuarine lowlands tended to form large mixed-species aggregations with high nest densities, whereas inland mountainous or hilly regions generally supported smaller colonies dominated by one or two species. This spatial organization indicates that colony-level assemblages are shaped not only by local habitat conditions but also by broader biogeographic context and multi-scale landscape heterogeneity across the Korean Peninsula. Random Forest analysis identified land-cover variables associated with SOM-derived colony assemblage groups, although classification performance was modest and relative importance varied across nested spatial scales. At fine spatial scales (500–1,000 m), the most influential predictors were artificial or urban surfaces (BUILT), managed herbaceous and agricultural vegetation (GRASS_MAN), and freshwater or coastal water habitats (WATER and WATER_OCEAN). These habitat types represent immediate colony surroundings, including human-modified landscapes, rice fields or other managed open vegetation, and nearby aquatic foraging areas. Their importance suggests that fine-scale colony assemblage differences are closely associated with local nesting surroundings, disturbance exposure, and access to shallow-water or agricultural foraging habitats10,11,16,39–42. This interpretation is consistent with previous studies showing that herons and egrets frequently use rice paddies, wetlands, and other shallow-water habitats during the breeding season10,43. In addition, colony placement itself appears shaped by human-associated landscape features: Shin et al.44, documented that approximately 12.46% of potential breeding habitat for herons and egrets occurs within ~ 50 m of residential areas in Korea, suggesting that many colonies are established adjacent to villages or settlement edges where agricultural landscapes, shallow-water habitats, and disturbance-mediated foraging opportunities co-occur. The prominence of BUILT at this spatial scale therefore likely reflects not only disturbance pressures but also the long-recognized tendency of some Ardeidae colonies to form near human settlements, leveraging the proximity to rice-paddy foraging grounds and aquatic feeding habitats. At intermediate spatial scales (3–5 km), the most influential predictors shifted toward variables representing landscape mosaics at broader spatial extents, including bare land (BARE), natural grasslands (GRASS_NAT), and tree-related PFTs (TREES_BD/NE). These land-cover types are closely associated with the structural components of river corridors in Korea45. Natural grasslands and riparian tree cover typically occur along river embankments, floodplains, and buffer zones, forming linear landscape elements composed of semi-natural vegetation and sparsely vegetated open areas46; Lim and Lee 2022). The prominence of these variables therefore suggests that many breeding colonies are embedded within the ecological influence zones of river corridors, and that these linear landscape structures play a key role in differentiating SOM clusters. Colonies associated with major river systems or extensive alluvial lowlands tended to coincide with higher proportions of open or semi-natural land cover, whereas colonies in interior mountainous regions corresponded more closely with forest-dominated landscapes2,47,48. These findings highlight river corridors as key landscape axes shaping foraging opportunities, movement connectivity, and interspecific interactions at intermediate spatial scales. At the broadest spatial extent (10 km), artificial or urban surfaces again emerged as an important predictor, suggesting that broader-scale land-use regimes may be associated with differences in colony assemblage structure17. The continued importance of bare or sparsely vegetated land and managed herbaceous or agricultural vegetation further emphasizes the role of extensive open-land systems, while coastal water and tree-related PFT variables may reflect broader coastal–inland, geomorphological, and elevational gradients4,49. However, tree-related variables should be interpreted cautiously because they may represent not only forest cover but also mountainous terrain and topographic constraints, particularly in northeastern Korea. Taken together, the combined SOM and scale-dependent Random Forest analyses demonstrate that the species composition of heron and egret colonies in Korea does not arise from a single spatial scale but instead emerges from hierarchical interactions among environmental predictors operating across fine, intermediate, and broader spatial scales. Fine-scale environments capture immediate foraging access and disturbance pressures; intermediate scales reflect landscape mosaics structured around river corridors that influence habitat availability and movement opportunities50; González del Tánago and García de Jalón51,48 and broader scales represent land-use regimes that define the environmental context in which colonies are embedded52; Cho et al., 2025). This multi-scale perspective helps explain why breeding colonies should be interpreted not as isolated sites, but as spatially discrete units embedded within broader landscape contexts. The consistent correspondence between SOM-derived colony assemblage groups and land-cover predictors reinforces the view that colonial breeding systems emerge from nested ecological structures shaped jointly by habitat selection and anthropogenic landscape modification across multiple spatial scales. By integrating species composition, geographic gradients, and land-cover characteristics—including the functional role of river corridors—this study provides a comprehensive framework for understanding the ecological organization of heron and egret breeding colonies in Korea. This hierarchical multi-scale perspective not only clarifies the mechanisms underlying spatial variation in colony assemblages but also offers a foundation for monitoring and anticipating future changes under ongoing land-use transformation. Several limitations should be considered when interpreting our findings. First, KDE maps were used only to visualize the spatial concentration of observed colonies and should not be interpreted as predictive maps of colony occurrence or colony-group suitability. Model-based spatial prediction across the entire study area would require non-colony or background data and independent validation of colony occurrence or absence. Second, although Random Forest models identified land-cover variables associated with SOM-derived colony assemblage groups, these associations should not be interpreted as direct causal effects. Some PFT variables, particularly tree-related classes, may also reflect broader topographic or geomorphological gradients in Korea. The restricted distribution of medium- and small-bodied species in northeastern Korea may therefore be influenced by mountainous terrain, elevation, and reduced availability of lowland foraging habitats, rather than land-cover composition alone. Future studies incorporating elevation, slope, terrain ruggedness, and floodplain structure would allow stronger separation of land-cover effects from topographic constraints. Conservation and management implications From a conservation and management perspective, our results suggest that heron and egret breeding colonies should be managed within hierarchical multi-scale landscape contexts rather than as isolated sites. Fine-scale protection of breeding stands and nearby foraging habitats is necessary but insufficient if broader landscape contexts are degraded. In particular, the strong influence of river corridors and agricultural mosaics at intermediate spatial scales underscores the importance of maintaining riparian buffers, floodplain connectivity, and semi-natural open habitats along major river systems. At broader spatial scales, regional land-use planning that limits excessive urban expansion and maintains mosaics of wetlands, rice paddies, and open grasslands will be critical for sustaining diverse colony assemblages. By identifying landscape features that consistently differentiate colony assemblage groups across nested spatial scales, this study provides a spatially explicit framework for prioritizing conservation actions and for integrating wading-bird conservation into river management, agricultural policy, and regional land-use planning across the Korean Peninsula53–57. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 2 Supplementary Material 3 Supplementary Material 4 Author contributions JL, JH, DK, YC, and HN conceived and designed the study. Data was curated by JL, JH, and DK. HN and SL carried out the analysis, and the original draft of the manuscript were written by SL and HC. YC and HN reviewed and edited the final draft. All authors read and approved the final manuscript. Funding This research was supported by grants from the National Institute of Biological Resource (NIBR), funded by the Ministry of Environment, Republic of Korea (NIBR202617102). Data availability The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Yu-Seong Choi, Email: yschoi0321@korea.kr. Hyung-Kyu Nam, Email: namhk2703@korea.kr. References Associated Data Supplementary Materials Supplementary Material 1 Supplementary Material 2 Supplementary Material 3 Supplementary Material 4 Data Availability Statement The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.