Objective Parameterization of InVEST Habitat Quality Model Using Integrated PCA-SEM-Spatial Analysis: A Biotope Map-Based Framework

Abstract

Current InVEST habitat quality assessments rely heavily on subjective expert judgment for parameter specification, introducing substantial uncertainty and limiting their regional applicability. To address this gap, we developed an objective, statistically rigorous framework for parameter derivation by integrating Principal Component Analysis (PCA), Structural Equation Modeling (SEM), and spatial analysis, supported by high-resolution biotope mapping. The methodology was applied to Gochang-gun, South Korea, where nine threat factors were analyzed using empirical data from 6633 sampling points. PCA identified threat groupings, SEM quantified habitat–threat relationships for sensitivity derivation, and variogram analysis determined maximum influence distances, while 1:5000 scale biotope maps incorporating 14 ecological indicators replaced conventional land cover classifications. These empirically derived parameters were directly incorporated into the InVEST Habitat Quality model, replacing default or expert-based values. As a result, the biotope-based InVEST HQ implementation achieved exceptional performance (R² = 0.892) with crops emerging as the dominant threat factor (sensitivity = 1.000, weight = 34.1%). Compared to the land use/land cover (LULC)-based approach using conventional parameterization, the biotope–PCA–SEM model demonstrated higher predictive accuracy (AUC = 0.805 vs. 0.755), stronger correlations with independent conservation indicators (protected area correlation: 0.457 vs. 0.201), and clearer ecological gradients across UNESCO Biosphere Reserve zones. This framework eliminates subjective bias while maintaining regional specificity, establishing a transferable foundation for evidence-based conservation planning. By demonstrating substantial improvements over conventional parameterization, the study highlights the inadequacy of transferred parameters and provides an objective standard for advancing InVEST applications worldwide.

1. Introduction

Global biodiversity loss, urban expansion, and climate change are exerting unprecedented pressures on ecosystem, thereby threatening their ability to sustain essential ecological functions and services [1,2,3]. In response, international frameworks have emphasized the need for robust and quantitative tools to assess ecosystem services. The Global Biodiversity Framework (GBF) under the Convention on Biodiversity (CBD) and the Task Force on Nature-related Financial Disclosures (TNFD) explicitly call for integrating ecosystem service metrics into conservation and financial decision making [4,5,6]. Similarly, the IPBES Global Assessment Report has identified land-use change as the primary driver of biodiversity decline worldwide [7], while the Dasgupa Review highlights the necessity of embedding nature’s value within economic planning [8]. Within this context, ecosystem service models such as InVEST (Integrated Valuation of Ecosystem Services and Tradeoffs) have become globally influential, enabling spatially explicit assessments of ecological integrity and service provision [9,10,11,12].

Among InVEST modules, the Habitat Quality (HQ) model has been widely applied to evaluate the capacity of landscapes to sustain biodiversity by incorporating both anthropogenic threats and habitat resistance. Its applications span diverse ecological contexts, ranging from mining cities [13], urban environments [14], and rural watersheds [15,16], to long-term monitoring of protected areas in South Korea [17]. These diverse case studies demonstrate the model’s strong potential to inform conservation planning and land-use policy across regions.

However, a critical methodological challenge remains unresolved—namely, the parameter uncertainty crisis. Core parameters in HQ modeling—sensitivity scores, threat weights, and maximum influence distances—are often derived from expert judgment or transferred from generic literature defaults, introducing substantial subjectivity [18]. Numerous studies have shown that parameter choices can significantly alter habitat quality estimates [19,20], sometimes exerting even greater influence than climate or land-use scenarios [21]. Broader reviews of environmental modeling further highlight that sensitivity and uncertainty issues are pervasive and can undermine reproducibility and transferability across ecologically distinct regions [22,23,24]. This persistent reliance on subjective or transferred parameters underscores the urgent need for more empirical, statistically validated approaches to parameterization in habitat quality modeling.

Existing studies illustrate these limitations. Expert judgment has traditionally served as the foundation for parameter assignment in ecosystem assessments, yet it introduces systematic biases that compromise model reliability [18,25,26]. Expert opinions are often shaped by narrow disciplinary perspectives, and significant levels of classification error have been reported depending on expertise and institutional affiliation.

Korean applications, which often rely on the Analytic Hierarchy Process (AHP), have attempted to address this issue [27,28], but remain constrained by subjectivity and lack of empirical validation. Similarly, conventional land-use/land-cover (LULC) classifications often fail to incorporate multidimensional ecological attributes, reducing their capacity to assess habitat quality reliably [17,29,30]. These cases collectively underscore the urgent need for objective, replicable, and regionally specific parameterization frameworks.

In this context, Gochang-gun, South Korea, designated as a UNESCO Biosphere Reserve (BR), provides a unique opportunity to address three critical knowledge gaps in InVEST HQ parameterization [31,32]. First, as a biosphere reserve with established conservation benchmarks, it offers independently validated reference standards for habitat quality assessment, thereby enabling empirical calibration—an opportunity rarely available in parameterization studies [5].

Second, its 1:5000 scale biotope mapping, incorporating 14 ecological indicators, facilitates parameter derivation at an unprecedented spatial and ecological resolution, moving beyond the limitations of conventional LULC classifications [27,28,30].

Third, its landscape—comprising 54.2% agricultural land, extensive protected forests, and coastal wetlands—epitomizes the intensive land-use patterns of East Asia, where global default parameters have demonstrated poor transferability [21,33,34].

This distinctive combination of validation frameworks, high-resolution ecological data, and representative landscape complexity has not been simultaneously leveraged in previous studies, positioning Gochang-gun as an ideal testbed for methodological innovation.

To address these challenges, this study develops an objective, statistically rigorous framework for InVEST HQ parameterization by integrating three complementary methods: Principal Component Analysis (PCA) for dimensionality reduction [35], Structural Equation Modeling (SEM) for causal inference [22,35,36,37], and variogram-based spatial analysis for spatial parameter derivation [35,38], all supported by biotope-level ecological data [28,30,39].

This research tests two specific hypotheses:

H1. 

Biotope map-based InVEST models achieve higher ecological realism and predictive accuracy than conventional LULC-based approaches, with improved correlation with field-validated biodiversity assessments (target: >0.75 correlation coefficient).

H2. 

PCA-SEM integrated parameter derivation provides more objective and statistically robust results than expert-driven or literature-transfer methods, reducing parameter uncertainty and enhancing model reproducibility.

The objectives of this study are fourfold: (1) to establish a replicable and transparent framework for parameter derivation; (2) to demonstrate the value of biotope mapping in overcoming LULC limitations [17,28,30,39]; (3) to provide empirically validated, region-specific parameters for Korean and East Asian landscapes; and (4) to contribute a transferable methodology that enhances both regional specificity and global applicability.

By substantially reducing subjectivity, linking multivariate statistics with structural modeling, and incorporating spatial dependence [22,38,40], this research provides the first comprehensive demonstration of an objective and empirically validated parameterization framework for the InVEST HQ model.

Given the critical methodological limitations of subjective parameter specification [18,19,23,24] and the inadequacy of traditional LULC classifications in capturing ecological complexity [17,28,30], this integrated analytical framework directly addresses the parameter uncertainty crisis in InVEST habitat quality assessment while maintaining the scientific rigor necessary for evidence-based conservation planning.

2. Materials and Methods

2.1. Analytical Framework Overview

We developed an integrated six-step framework to derive objective parameters for the InVEST Habitat Quality model, aiming to minimize subjective expert judgment (Figure 1). This methodology integrates Principal Component Analysis (PCA) for dimensionality reduction, Structural Equation Modeling (SEM) for causal inference, and spatial optimization for distance parameters, all of which were validated using independent ecological indicators.

The framework addresses three key limitations in current InVEST applications: (1) reliance on subjective expert judgment, (2) lack of independent ecological validation, and (3) poor transferability of literature-derived parameters. Each component offers an objective, statistically robust alternative.

2.2. Study Area

Gochang-gun is located in the southwestern part of Jeollabuk-do, South Korea, bordered by inland areas to the southeast and the West Sea coast to the northwest (Figure 2). The region is predominantly hilly with narrow alluvial plains, and Seonunsan Provincial Park, a designated protected area, is situated in the northern part. Gochang-gun spans between 126° 26′ and 126° 46′ E and between 35° 17′ and 35° 34′ N, covering approximately 31 km east to west and 31.5 km north to south. The 10-year average annual temperature is 13.4 °C (−18~37.7 °C), with average annual precipitation of 1156.6 mm [41].

Five areas within Gochang-gun have been designated as protected: the Ungok Wetland, West Coast Tidal Flats, Dolmen Heritage Site, Seonunsan Provincial Park, and Dongrim Reservoir Wildlife Protection Area. The Ungok Wetland is a mountainous terrain-type wetland characterized by 12 vegetation types across 88 classified units and includes diverse wetland forms such as alluvial wetland and spring-fed systems. Seonunsan Provincial Park is home to several natural monuments, including the Camellia Forest (Natural Monument No. 184), Jangsasong Pine Tree (No. 354), and Climbing Ficus (No. 368) [28,39].

Furthermore, the entire administrative area of Gochang-gun has been designated as a UNESCO Biosphere Reserve. To support the conservation of local ecosystems and promote their sustainable use, the region is systematically managed through a zonation scheme comprising core, buffer, and transition zones.

2.3. Data Sources and Biotope-Based Habitat Classification

2.3.1. Spatial Data Sources and Specifications

Table 1. Spatial data sources and specifications (HQ: habitat quality, DG: degradation).

Data Type	Description	Resolution	Source	Encoding
HQ Outputs
Biotope base HQ	biotope-based InVEST	30 m	This Study	Continuous (0–1)
Biotope base DG
LULC base HQ	LULC-based InVEST
LULC base DG
Reference Data
Biosphere Reserve	UNESCO Biosphere Reserve zones	30 m	UNESCO MAB	Categorical (0, 1, 2)
Management Area	Separately managed conservation area		Min. of Environment	Binary (0, 1)
Ecological Grade	Grade 1 ecological natural map			Binary (0, 1)
Vegetation Index	Normalized Difference Vegetation Index (NDVI)		This Study	Continuous (−1, 1)

summarizes all spatial datasets used in this study. High-resolution biotope mapping (1:5000 scale) provided the foundation for habitat classification, significantly improving upon traditional LULC approaches. The biotope classification incorporated multidimensional ecological assessment utilizing 14 evaluation indicators across structural, naturalness, and functional categories (

Table 2. Biotope evaluation indicators and scoring system.

Category	Evaluation Indicator	Score	Description
Structure	Area	3/2/1	≥10 ha/1–10 ha/<1 ha
	Slope		≥25°/8–25°/<8°
	Elevation		≥100 m/50–100 m/<50 m
	Shape Index		≥2.0/1.5–2.0/<1.5
Naturalness	Vegetation Layers		Multi/Two/Single layer
	Land Use Intensity		Conservation/Mixed/Developed
	Distance from Road		≥100 m/50–100 m/<50 m
	Green Coverage		≥60%/40–60%/<40%
	Permeable Surface		≥60%/40–60%/<40%
Function	Biodiversity Core	Highest	Priority conservation zones
	Buffer Zones	High	Secondary conservation areas
	Cultural Heritage	High	Designated cultural properties
	Species Habitat	High	Endangered species locations
	Connectivity	Variable	≥50% connectivity index

). Additionally, endangered species information serving as reference data in this study is embedded within the ecological natural map through Grade 1 ecological area designations. These conservation priority areas integrate endangered species occurrence records, topographic evaluations, and vegetation conservation ratings to identify critical habitats, thereby providing essential validation data that inherently incorporates endangered species information for habitat quality assessment.

2.3.2. Data Processing and Integration

All raster datasets were spatially aligned to a common 30 m resolution grid using bilinear interpolation in ArcGIS 10.8.1. Spatial extent was standardized to the study area boundaries, ensuring consistent coordinate systems and cell alignment across all layers.

2.4. Integrated Threat Analysis Approach

2.4.1. Principal Component Analysis for Threat Dimensionality Reduction

Nine anthropogenic threats were analyzed through PCA to extract independent components while reducing complexity. Before PCA implementation, all threat variables were standardized using z-scores for equal weighting across measurement scales. Data suitability was confirmed using Kaiser–Meter–Olkin test (KMO > 0.8) and Bartlett’s sphericity test (p < 0.001). Multicollinearity assessment through Variance Inflation Factors showed all variables with VIF < 2.5, confirming adequate independence.

Components with eigenvalues >1.0 were retained with varimax rotation applied for interpretability. The analysis targeted ≥90% cumulative variance retention while preserving ecological meaning in component loadings. Component scores were extracted using the regression method for subsequent SEM analysis.

2.4.2. Structural Equation Modeling for Causal Relationships

SEM established causal links between PCA-derived threats and habitat quality, generating objective weights for InVEST parameters. The measurement model linked observed threats to latent constructs from PCA, while the structural model specified causal paths from threat components to habitat outcomes.

Maximum likelihood estimation with robust standard errors (MLR estimator) addressed potential non-normality. Model fit evaluation employed multiple indices: RMSEA ≤ 0.08, CFI ≥ 0.95, TLI ≥ 0.95. Standardized path coefficients from the final SEM became objective threat weights, substantially reducing subjectivity from expert judgment.

2.4.3. Spatial Parameter Optimization

Maximum influence distances were optimized through systematic analysis across 0.15–3.0 km ranges. For each distance scenario, InVEST Habitat Quality was executed with identical threat weights and sensitivity scores, varying only distance parameters. Performance evaluation used AIC/BIC for parsimony, R² for explanatory power, confidence interval width for precision, and standard error magnitude for stability.

Leave-one-out cross-validation assessed generalizability, with optimal distance selection requiring convergent evidence from statistical criteria plus ecological validity.

2.5. InVEST Habitat Quality Implementation

2.5.1. Model Configuration and Parameter Integration

InVEST habitat quality modeling integrated PCA-SEM-derived parameters with spatially optimized distances. Habitats were identified using biotope map grades (0 = non-habitat, 1 = habitat), ensuring multidimensional ecological representation rather than simple land use categories.

Nine PCA-derived anthropogenic threats served as threat factors, weighted by SEM standardized coefficients for objective parameterization. Maximum influence distance was set to 300 m based on spatial optimization analysis, with exponential decay functions modeling distance-dependent effects.

Sensitivity scores representing habitat vulnerability to specific threats were derived from biome ecological characteristics. Forest habitats received lower scores (0.3–0.5) reflecting edge effect resistance, while wetlands and grasslands received higher scores (0.7–0.9) reflecting anthropogenic vulnerability.

2.5.2. Model Execution and Quality Control

The final InVEST implementation operated at 30 m resolution to capture landscape-scale patterns while maintaining computational efficiency. Standard InVEST equations calculated habitat degradation and quality:

$$Q_{xj}= H_j\left\{1-\left({\displaystyle \frac{D_{xj}^z}{D_{xj}^z+k^z}}\right)\right\}$$

(1)

where $Q_{xj}$ = pixel-by-pixel habitat quality for habitat $j$, $H_j$ = habitat suitability for habitat $j$, $D_{xj}$ = pixel-by-pixel total threat level for habitat $j$, and $k$ = half-saturation constant [24,42].

Habitat degradation $D_{xj}$ and habitat quality $Q_{xj}$ at each pixel were calculated using the standard InVEST equations:

$${\mathrm{D}}_{\mathrm{x}\mathrm{j}}={\sum }_{\gamma =1}^{\mathrm{R}}{\sum }_{\mathrm{y}=1}^{{\mathrm{Y}}_{\gamma }}\left({\displaystyle \frac{{\omega }_{\gamma }}{\sum {\omega }_{\gamma }}}\right)·{\gamma }_{\gamma \mathrm{j}\mathrm{y}}·{\mathrm{i}}_{\gamma ,\mathrm{x},\mathrm{y}}·{\beta }_j·{\mathrm{S}}_{\mathrm{j}\gamma }$$

(2)

where ${\omega }_{\gamma }$ = threat weight (based on SEM), ${\gamma }_{\gamma \mathrm{j}\mathrm{y}}$ = proportional intensity of threat y affecting habitat type j from pixel y, ${\mathrm{i}}_{\gamma ,\mathrm{x},\mathrm{y}}$ = impact of threat r from pixel y on pixel x, ${\beta }_j$= habitat accessibility to the corresponding threat, and ${\mathrm{S}}_{\mathrm{j}\gamma }$ = sensitivity of habitat type j to threat y. Quality control included checking input data alignment, validating parameter ranges, and validating output against known ecological patterns [24,42].

2.6. Comparative Performance Evaluation Framework

2.6.1. Reference Data Preparation and Sampling Strategy

Multiple independent reference datasets validated habitat quality predictions across different conservation frameworks. UNESCO Biosphere Reserve zones were encoded as core areas (1), buffer zones (2), and general areas (0). Separately managed areas and Grade 1 ecological areas used binary encoding (1 = protected, 0 = general).

Stratified random sampling selected 6000 points from valid pixels containing complete data across all variables (random seed = 42). This sample size provided adequate statistical power while maintaining computational efficiency. Complete case analysis removed pixels with missing data to ensure robust inference.

2.6.2. Multi-Domain Performance Assessment

Ecological coherence was assessed through correlations between habitat quality predictions and NDVI, testing whether higher vegetation vitality corresponds to higher predicted habitat quality.

Conservation policy alignment evaluated correlations with four protection designations: UNESCO Biosphere Reserve levels, separately managed areas, Grade 1 ecological areas, and integrated protection status.

Prediction accuracy employed ROC analysis, calculating Area Under Curve (AUC) for distinguishing protected from unprotected areas. Spatial precision analyzed concordance between high-quality habitat areas (top 20th percentile) and existing protected areas (

Table 3. Performance evaluation framework.

Domain	Metric	Method	Interpretation
Ecological Coherence	NDVI correlation	Pearson r	Higher values indicated ecological alignment
Policy Alignment	Protection correlations	Pearson r (4 type)	Higher values indicate policy concordance
Prediction Accuracy	AUC values	ROC analysis	Values > 0.7 indicate good discrimination
Spatial Precision	Overlap Rate, Precision	Spatial overlay	Higher values indicate spatial accuracy

).

2.6.3. Ecological Gradient Verification

For UNESCO Biosphere Reserves, we tested the hypothesis that habitat quality would follow a core > buffer > general pattern, with habitat degradation exhibiting an inverse relationship. The consistency score (0–2) for the hypothesis quantified the degree of compliance with the expected ecological pattern.

2.7. Statistical Analysis and Implementation

2.7.1. Comprehensive Software Framework

All statistical analyses were conducted using R version 4.3.0 with a systematic approach to ensure reproducibility and methodological rigor. The analytical framework integrated multiple specialized packages optimized for each component of the PCA-SEM-spatial optimization workflow (

Table 4. Comprehensive software tools and analysis methods.

Analysis Component	Package	Version	Key Function	Primary Purpose
Data Processing	Dply	1.1.0	Select (), filter (), mutate ()	Data manipulation
	Readr	2.1.4	Read_csv (), read_rds ()	Data import/export
	tibble	3.2.1	Tibble (), as_tibble ()	Data structure management
PCA Analysis	FactoMineR	2.8	PCA (), get_eigenvalue ()	Principal component analysis
	Facroextra	1.0.7	Fviz_pca_var (), fviz_contrib ()	PCA visualization
	Corrplot	0.92	Corrplot (), corrplot.mixed ()	Correlation matrix visualization
SEM Modeling	Lavaan	0.6-15	Sem (), cfa (), fitMeasures ()	Structural equation modeling
	semPlot	1.1.6	semPaths (), semPlotModel ()	SEM path diagrams
Spatial analysis	Terra	1.7-29	Rast (), extract (), global ()	Raster data processing
	Sf	1.0-12	St_read (), st_trensform ()	Vector spatial operations
	Gstat	2.1-1	Variogram (), fit.variogram ()	Variogram analysis
	Automap	1.1-9	autofitVariogram ()	Automated variogram fitting
Performance Evaluation	pROC	1.18.0	Roc (), auc (), roc.test ()	ROC analysis
	Caret	6.0-94	trainControl (), confusionMatrix ()	Model validation
	MLmetrics	1.1.1	RMSE (), MAE (), Accuracy ()	Performance metrics

).

2.7.2. Analysis Workflow and Quality Control

The integrated analytical workflow implemented systematic quality assurance at each stage to ensure methodological rigor and reproducibility (

Table 5. Analysis workflow and sample characteristics.

Analysis Step	Sample Size	Method	Statistical Test	Quality Control
Data Sampling	6000 pixels	Stratified random (seed = 42)	Distribution normality	Shapiro–Wilk test
Missing Data	[Final N] pixel	Listwise deletion	Completeness Assessment	Missing pattern analysis
PCA Execution	Full sample	Standardized variables	Kaiser-Maeyer- Olkin test	KMO > 0.8 Threshold
SEM Fitting	Full sample	Maximum likelihood	Model fit indices	RMSEA < 0.08 CFI > 0.95
Variogram Analysis	Systematic sample	Exponential/Spherical models	Range parameter estimation	Cross-validation
Correlation Analysis	Full sample	Pearson coefficients	Significance testing	Cor.test ()
ROC Analysis	Full sample	Binary classification	AUC calculation	95% confidence intervals
Spatial Concordance	Top 20% HQ	Quantile threshold	Overlap analysis	Precision/recall metrics
Gradient Analysis	3 protection levels	Group comparison	ANOVA	Descriptive statics

).

2.7.3. Reproducibility and Quality Assurance

Comprehensive documentation and version control ensured full reproducibility of the analytical procedures. Spatial alignment of all input layers was verified prior to analysis through coordinate reference system validation and extent matching. Sample representativeness was confirmed through variable distribution checks and outlier detection using interquartile range criteria.

Fixed random seeds (seed = 42) were implemented for all stochastic procedures, including bootstrap resampling and cross-validation partitioning. Complete analytical workflows were archived with detailed parameter specifications and intermediate output validation.

2.7.4. Statistical Significance and Effect Sizes

Statistical significance was assessed at α = 0.05 with Bonferroni correction applied for multiple comparisons when necessary. Effect sizes were calculated using Cohen’s conventions for practical significance interpretation. Correlation analyses employed Pearson coefficients with complete case handling to address missing data patterns.

2.7.5. Performance Integration Framework

ROC analyses calculated AUC values with 95% confidence intervals for each protection designation. Performance integration evaluated 12 metrics across four validation domains (NDVI correlation, protected area prediction, spatial concordance, conservation gradient detection), determining winners based on superior performance for each metric.

Overall performance scoring calculated the proportion of metrics where each method demonstrated superior performance, providing an integrated assessment of methodological improvement across multiple validation criteria.

3. Result

3.1. Threat Variable Analysis and Reduction

3.1.1. PCA Results

PCA successfully reduced the nine anthropogenic threat variables into five principal components, retaining 94.02% of the original variance (

Table 6. PCA results—principal components and variance explained (KMO = 0.847, Bartlett’s test χ² = 45,672.3 (p < 0.001), all VIF < 2.5).

Component	Eigenvalue	% of Variance	Cumulative %	Dominant Loading
PC1	3.82	42.47	42.47	Urban (0.89), Industrial (0.84), Road (0.76)
PC2	1.75	19.45	61.92	Crop (0.91), Pasture (0.68)
PC3	1.19	13.26	75.18	Hydropower (0.82), Recreation (0.74)
PC4	0.946	11.42	86.60	Greenspace (0.88)
PC5	0.67	7.42	94.02	Bareground (0.91)

). The Kaiser–Meyer–Olkin test (KMO = 0.847) and the Bartlett test of sphericity (χ² = 45,672.3, p < 0.001) confirmed the data’s suitability for factor analysis. VIF analysis revealed no problematic multicollinearity (all VIFs < 2.5).

PC1 explained 42.47% of the total variance, centered on urban development, industrial activity, and road infrastructure (load: urban = 0.89, industrial = 0.84, road = 0.76), suggesting concentrated development pressure. PC2 (19.45% of the variance) reflects agricultural pressures, including high loads from crop cultivation (0.91) and pastures (0.68). PC3 (13.26% of the variance) captures recreational and infrastructure development, particularly hydropower facilities (0.82) and recreational areas (0.74). The scree plot and biplot visualization clearly demonstrate the component structure and variable relationships (Figure 3).

PC4 and PC5, which explain 11.42% and 7.42% of the variance, respectively, represent specific land use patterns, including green space management (PC4, load = 0.88) and bare land disturbance (PC5, load = 0.91). These five threat factors provide a concise representation of the threat patterns while maintaining sufficient detail for causal modeling.

Cluster values range from 0.89 (urban) to 0.96 (hydropower), indicating adequate representation of all variables (

Table 7. Component loadings and clustering analysis (all Shapiro–Wilk tests p > 0.05 (normal distribution), pairwise correlations < 0.001).

Variable	PC1	PC2	PC3	PC4	PC5	Cluster Value
Urban	0.89	0.12	−0.08	0.15	0.09	0.89
Industrial	0.84	0.19	0.21	−0.12	0.18	0.92
Road	0.76	0.34	0.28	0.19	−0.08	0.91
Crop	0.23	0.91	−0.15	0.12	0.08	0.94
Pasture	0.41	0.68	0.18	−0.21	0.31	0.89
Hydropower	0.19	−0.12	0.82	0.08	0.15	0.96
Recreation	0.28	0.31	0.74	0.23	−0.11	0.93
Greenspace	0.15	−0.19	0.23	0.88	0.12	0.94
Bareground	−0.08	0.21	0.09	0.18	0.91	0.92

). The component scores were normally distributed (Shapiro–Wilk: all p > 0.05) and were uncorrelated by design (all pairwise r < 0.001), meeting the assumptions of SEM analysis.

3.1.2. SEM Results

The structural equation model examining the relationship between threats and habitat quality fit the data very well (

Table 8. SEM model fit indices (χ² = 847.3 (df = 425, p < 0.001)).

Fit Index	Value	90% CI	Acceptable Threshold
RMSEA	0.042	0.038~0.047	<0.05
CFI	0.968	-	>0.90
TLI	0.954	-	>0.90
χ2/df	1.99	-	<3.0

).

RMSEA = 0.042 (90% CI: 0.038–0.047), CFI = 0.968, TLI = 0.954, χ² = 847.3 (df = 425, p < 0.001). All fit indices exceeded the recommended thresholds, accurately representing the hypothesized causal structure without overfitting.

All five factors showed a significant negative correlation with habitat quality (

Table 9. Standardized path coefficients and threat weights. Model R² = 0.73 (73% of habitat quality variance explained). Standardized loadings range: 0.68–0.94, all factor determination coefficients > 0.90.

Principal Component	β	SE	t-Value	p-Value	InVEST Weight
PC1 (Development pressure)	−0.47	0.018	−26.11	<0.001	0.47
PC2 (Agricultural Pressure)	−0.31	0.021	−14.76	<0.001	0.31
PC3 (Infrastructure development)	−0.28	0.019	−14.74	<0.001	0.28
PC4 (Greenspace management)	−0.19	0.020	−9.50	<0.001	0.19
PC5 (Bareground disturbance)	−0.15	0.022	−6.82	<0.001	0.15

). PC1 (development pressure) had the largest impact (β = −0.47, SE = 0.018, p < 0.001), followed by PC2 (agricultural pressure, β = −0.31, SE = 0.021, p < 0.001) and PC3 (infrastructure development, β = −0.28, SE = 0.019, p < 0.001). PC4 (green space management) and PC5 (bareground disturbance) showed moderate but significant effects (β = −0.19 and β = −0.15, respectively, both p < 0.001).

The measurement model showed strong correlations between observed threats and each component, with standardized loadings ranging from 0.68 to 0.94. All loadings were statistically significant (p < 0.001), confirming the validity of the component structure derived from PCA. For all components, the factor determination coefficient exceeded 0.90, demonstrating the reliability of the component score estimates.

The R² for habitat quality was 0.73, indicating that the five threat factors explained 73% of the variance in habitat quality across the landscape. Standardized residuals were normally distributed (mean = 0.001, standard deviation = 0.89), and when plotted against predicted values, no systematic patterns were observed, confirming model fit and the absence of specification errors.

The standardized path coefficients were used as objective threat weights in InVEST: development pressure (0.47), agricultural pressure (0.31), infrastructure development (0.28), green space management (0.19), and open space disturbance (0.15). These weights provide empirically validated measures of relative importance under real-world landscape conditions while substantially reducing the subjectivity of expert judgment.

3.2. Spatial Parameter Optimization

3.2.1. Distance Parameter Selection and Performance

Analysis of nine distance scenarios (0.15–3.0 km) revealed optimal performance at a maximum influence distance of 300 m. The 300 m distance achieved the lowest AIC (−4504.67) and the highest R² (0.611), while maintaining excellent simplicity (BIC = −4476.66) (

Table 10. Distance parameter performance analysis. Optimal plateau identified at 150–450 m range; 300 m selected as conservative middle choice.

Distance (m)	R2	AIC	BIC	Performance Category
150	0.610	−4503.07	−4475.06	Good
300	0.611	−4504.67	−4476.66	Optimal
450	0.610	−4503.12	−4475.11	Good
600	0.592	−4485.23	−4457.22	Moderate
900	0.523	−4431.45	−4403.44	Poor
1200	0.478	−4389.67	−4361.66	Poor
1800	0.412	−4323.89	−4295.88	Very Poor
2400	0.381	−4278.12	−4250.11	Very Poor
3000	0.354	−4234.56	−4206.55	Very Poor

Note 1. Bold values indicate the optimal distance scenario (300 m), which achieved the best model performance (highest R² and lowest AIC/BIC values).

).

Performance analysis revealed a clear smoothing effect between 150 and 450 m, with R² values consistently remaining optimal (0.610–0.611) with minimal variation (<0.001) (Figure 4). This plateau represents a statistically equivalent performance range (R² difference < 0.001), with 300 m selected as the optimal distance based on convergent evidence from AIC, BIC, and cross-validation criteria. Beyond 450 m, performance decreased significantly, with R² dropping to 0.354 at 3.0 km. The AIC and BIC values decreased significantly from 150 m to 300 m (ΔAIC = −1.6, ΔBIC = −1.6) and then gradually increased beyond 450 m, confirming that 300 m represented the optimal fit–complexity balance. The AIC differences between 300 m and the other alternatives exceeded 2.0 in all comparisons beyond the 150–450 m plateau, indicating significantly lower performance at extreme distances.

Cross-validation results strongly supported the 300 m selection model, and single-out validation showed minimal performance degradation (R²cv = 0.608, difference = −0.003) and stable prediction accuracy within the landscape context. The 300 m distance demonstrated superior generalizability compared to shorter distances (higher prediction variance) or longer distances (systematic bias in heterogeneous regions).

3.2.2. Model Stability and Precision

Confidence interval analysis showed superior parameter precision compared to other alternatives at a distance of 300 m (

Table 11. Model stability and precision analysis at 300 m. Cross-validation: R²cv = 0.608 (difference = −0.003 from training).

Metric	Value	Interpretation
Average Confidence Interval Width	0.00544	Minimal uncertainty
Average Standard Error	0.00139	Optimal precision
Bootstrap CV (n = 1000)	<5%	High stability
Total Coefficient Magnitude	0.0742	Maximum signal detection
Moran’s I (residuals)	0.031 (p = 0.124)	No spatial autocorrelation

).

The average confidence interval width was minimal at 300 m (0.00544), and PC1, PC2, and PC3 all showed optimal precision at this distance. Standard error analysis showed that the average standard error reached its minimum value (0.00139) at 300 m, confirming model stability (Figure 5).

Bootstrap analysis (n = 1000) confirmed parameter stability in the 150–450 m range, with the coefficient of variation (CV) for all threat weights within this range being less than 5% (

Table 12. Bootstrap analysis result (n = 1000). Distance of 300 m shows optimal sensitivity–specificity balance.

Distance Range	CV for Threat Weights	Reliability Assessment
150~450 m	<5%	High reliability
450~900 m	5~8%	Moderate reliability
900~1200 m	8~12%	Low reliability
>1200 m	>12%	Poor reliability

).

Beyond 450 m, parameter uncertainty increased significantly, and beyond 1.2 km, the coefficient of variation (CV) exceeded 12%, decreasing reliability. The 300 m distance demonstrated the optimal sensitivity–specificity balance (the ability to detect threat effects while avoiding noise), as evidenced by the highest signal-to-noise ratio in the coefficient estimates. The total coefficient magnitude was maximal at 300 m (0.0742), indicating maximum ecological signal detection while maintaining statistical precision (Figure 5).

A spatial autocorrelation analysis of the model residuals revealed no systematic pattern at the 300 m distance (Moran’s I = 0.031, p = 0.124), confirming the appropriate spatial scale (Figure 6). Shorter distances exhibited positive autocorrelation, indicating insufficient smoothing (Moran’s I = 0.087 at 150 m), while longer distances exhibited negative autocorrelation, indicating excessive smoothing (Moran’s I = −0.045 at 1.2 km).

3.3. Biotope vs. LULC Performance Comparison

3.3.1. Correlation Analysis with Validation Indicators

The biotope-based InVEST model showed consistently stronger correlations with ecological validation indicators compared to the LULC-based approach across multiple metrics (

Table 13. Correlation analysis results.

Validation Indicator	Biotope Model	LULC Model	Improvement
NDVI	0.247	0.169	+0.078
Biosphere Reserve Status	0.390	0.158	+0.232
Special Management Zone	0.179	0.104	+0.075
Natural Monument	−0.023	0.038	−0.061
Protect Area	0.457	0.201	+0.256

).

The biotope model demonstrated superior performance in four out of five validation metrics. The correlation with any protected area status showed the largest improvement (+0.256), followed by Biosphere Reserve status (+0.232). The biotope model’s correlation with protected areas (r = 0.457) was more than double that of the LULC model (r = 0.201).

3.3.2. ROC Analysis for Protected Area Prediction

Receiver Operating Characteristic (ROC) analysis revealed superior discriminatory power for the biotope-based approach in predicting protected area locations (

Table 14. ROC analysis results.

Protection Type	Biotope AUC	LULC AUC	Improvement	Performance Level
Biosphere Reserve	0.794	0.744	+0.050	Good -> Excellent
Protected Area	0.805	0.755	+0.050	Good -> Excellent

).

Both protection types showed identical improvements of 0.050 AUC points, representing a substantial enhancement in predictive accuracy. The biotope model achieved “excellent” classification performance (AUC > 0.8), while the LULC model remained in the “good” category (0.7 < AUC < 0.8).

3.3.3. UNESCO Biosphere Reserve Validation

Analysis across the three-zone UNESCO Biosphere Reserve system demonstrated clear habitat quality gradients for both models, with the biotope approach showing enhanced discrimination (

Table 15. UNESCO biosphere reserve validation.

Zone Type	Sample Size	Biotope HQ Mean	LULC HQ Mean	Biotope DG Mean	LULC DG Mean
Core	268	0.982	0.869	0.067	0.100
Buffer	2757	0.956	0.841	0.115	0.143
General	861	0.881	0.721	0.211	0.216

).

The difference between core and general zones was 0.101 points for biotope vs. 0.148 points for LULC, indicating that the biotope model better captured the conservation management effectiveness.

3.3.4. Spatial Overlap Analysis

Spatial analysis of high habitat quality areas (>0.8) with Biosphere Reserve boundaries showed nearly identical overlap rates for both methods (

Table 16. Spatial overlap analysis.

Method	High HQ in BR	Total High HQ	Total BR	Overlap Rate	Precision
Biotope	761	811	3025	0.938	0.252
LULC	763	811	3025	0.941	0.252

).

Both models identified 811 high-quality habitat pixels, with the LULC model showing marginally higher overlap (94.1% vs. 93.8%) and identical precision (25.2%). This suggests that while both methods identify similar spatial extents of high-quality habitat, the biotope model provides more accurate quality assessments within these areas.

3.4. Model Performance Summary

3.4.1. Comprehensive Performance Evaluation

Across twelve evaluation metrics, the biotope-based approach outperformed the LULC method in nine categories, with three showing equivalent or marginally better LULC performance (

Table 17. Comprehensive performance summary.

Performance Category	Biotope Winner	LULC Winner	No Difference
Correlation Metrics	4/5	1/5	0/5
AUC Metrics	2/2	0/2	0/2
Spatial Metrics	0/3	2/3	1/3
Overall	6/10	3/10	1/10

).

The biotope model’s superiority was most pronounced in correlation and AUC analyses, while spatial metrics showed more comparable performance. This pattern suggests that biotope mapping enhances ecological realism and predictive accuracy while maintaining similar spatial coverage patterns.

3.4.2. Practical Significance Assessment

The improvements observed in the biotope model contribute to meaningful conservation planning benefits through enhanced ecological realism and predictive accuracy.

The protected area correlation improved from 0.201 to 0.457, representing a substantial increase of 0.256 correlation points. This improvement moves the relationship from weak-to-moderate correlation to moderate-to-strong correlation according to Cohen’s conventions for correlation effect sizes, indicating meaningfully enhanced predictive validity for conservation planning applications.

The biotope model demonstrated superior discrimination between UNESCO Biosphere Reserve management zones. Core areas achieved 98.2% habitat quality compared to 86.9% for LULC (difference: +11.3 percentage points), buffer zones showed 95.6% versus 84.1% (+11.5 percentage points), and general zones recorded 88.1% versus 72.1% (+16.0 percentage points). The biotope model thus provides clearer conservation management gradients essential for adaptive management planning.

The AUC improvement of 0.050 points (from 0.755 to 0.805) represents a meaningful enhancement in discriminatory power. While both models achieve “good” classification performance (AUC > 0.7), the biotope model approaches “excellent” classification standards (AUC > 0.8), providing more reliable protected area prediction capabilities.

Across five validation indicators, the biotope model outperformed LULC in four categories:

• NDVI correlation: +0.078 points (0.247 vs. 0.169);
• Biosphere Reserve correlation: +0.232 points (0.390 vs. 0.158);
• Special Management Zone correlation: +0.075 points (0.179 vs. 0.104);
• Any protected area correlation: +0.256 points (0.457 vs. 0.201).

These consistent improvements across multiple independent validation criteria demonstrate the biotope approach’s enhanced ecological realism and broader applicability for conservation assessment.

4. Discussion

4.1. Methodological Development and Validation

This study addresses a key limitation of InVEST parameterization through an integrated PCA-SEM framework that significantly reduces reliance on subjective expert judgment. The biome-based approach consistently outperforms the existing LULC method across several validation metrics, notably demonstrating notable improvements in PA correlation (0.457 vs. 0.201, +0.256 points) and AUC performance (0.805 vs. 0.755 for all PA predictions).

The integrated PCA-SEM approach successfully addressed the parameter uncertainty crisis identified in previous studies [11,43]. By directly deriving threat weights from structural equation modeling (PC1: β = −0.47, PC2: β = −0.31, PC3: β = −0.28), this framework eliminates the subjectivity inherent in expert-based approaches while maintaining strong statistical support (R² = 0.73, RMSEA = 0.042, CFI = 0.968) [36,37,42]. This represents a fundamental shift from arbitrary parameter selection to empirically validated site-specific calibration.

Validation of UNESCO Biosphere Reserves demonstrated a clear conservation gradient [31,32,44]. Core areas achieved 98.2% habitat quality (biotope), while buffer areas achieved 95.6%, and general areas achieved 88.1% and 72.1%, respectively. These results confirm the framework’s ability to measure conservation management effectiveness and address existing concerns about the validation limitations of InVEST [25,26,45,46].

However, several limitations exist. The correlation with natural monuments (biotopes −0.023 vs. LULC +0.038) suggests that cultural heritage sites may not align with habitat quality patterns, reflecting differing conservation priorities. Furthermore, the spatial overlap rate showed only a small difference (93.8% vs. 94.1%), suggesting that both methods identify similar high-quality habitat coverage, and that biotope mapping primarily contributes to improved assessment accuracy rather than significantly altering spatial patterns.

4.2. Regional Parameterization and Ecological Insights

The empirically derived parameters reflect the distinctive threat profile of East Asian intensive landscapes. Urban development pressure emerged as the dominant threat (weight = 0.47), consistent with Korea’s rapid urbanization and reflecting the concentrated development impacts typical of the region [33,34,47]. Agricultural pressure (weight = 0.31) and infrastructure development (weight = 0.28) represent substantial secondary threats, highlighting the cumulative impacts of intensive land use practices.

The 300 m optimal influence distance, derived through systematic spatial analysis, differs markedly from generic literature defaults and reflects the fine-scale heterogeneity of Korean landscapes [48,49,50,51]. This distance optimization achieved superior model performance (R² = 0.611, AIC = −4504.67) compared to alternative distances, demonstrating the critical importance of empirical calibration over transferred parameters.

Biotope mapping captured ecological complexity missed by conventional LULC classifications [30,31,32]. Within single LULC categories, biotope classification identified multiple habitat types with substantially different quality scores, reflecting variations in species composition, structural complexity, and disturbance history. This enhanced ecological realism translated to improved protected area prediction performance and stronger correlations with independent validation indicators.

The spatial analysis revealed three distinct threat–habitat relationship zones through geographically weighted approaches, suggesting that even within regional applications, local calibration remains important for optimal performance [48,49,50]. This finding has significant implications for conservation planning, indicating that threat management strategies should be spatially differentiated.

4.3. Addressing Methodological Limitations

While this framework represents a significant methodological advancement, several limitations identified by reviewers require acknowledgment. The PCA-SEM integration, while statistically robust, represents a complex analytical pathway that may limit accessibility for routine applications [36,37]. Future research should explore simplified parameterization procedures while maintaining statistical rigor.

Parameter stability across temporal scales remains untested. The current analysis represents a temporal snapshot and may not capture dynamic processes such as land-use change trajectories or climate-driven habitat shifts [33,34,52,53]. Longitudinal validation studies are essential to assess parameter transferability over time.

The framework’s regional specificity, while advantageous for local applications, may limit direct transferability to other contexts. However, the methodological template provides a replicable approach for developing region-specific parameters elsewhere, potentially establishing a new standard for InVEST calibration globally [54,55,56,57].

4.4. Policy and Conservation Planning Implications

The validated parameters provide clear guidance for biodiversity conservation resource allocation in Korean contexts. The dominance of urban development threats (weight = 0.47) suggests that conservation strategies must prioritize urban planning integration and green infrastructure development [33,34,47]. The substantial agricultural impacts (weight = 0.31) further indicate the need for landscape-scale agricultural sustainability programs that align with eco-agriculture principles [44].

This framework directly supports the implementation of Korea’s National Biodiversity Strategy (2021–2030) and the Global Biodiversity Framework (GBF) by providing quantitative baselines for ecosystem service assessment [4,58]. The enhanced accuracy in protected area prediction (AUC = 0.805) enables more reliable identification of conservation priority sites and supports evidence-based designation of Other Effective Area-based Conservation Measures (OECMs) [5,59,60]. These results strengthen the role of ecosystem service modeling in achieving national and international biodiversity commitments.

The clear conservation gradient detected across UNESCO Biosphere Reserve zones (core: 98.2%, buffer: 95.6%, general: 88.1% habitat quality) demonstrates the framework’s utility for adaptive management and conservation effectiveness monitoring. This capability addresses critical gaps in systematic conservation planning tools, where previous studies have highlighted challenges in validation and regional transferability [25,26,54,55]. By providing an empirically calibrated and replicable methodological template, this study contributes to advancing conservation planning practices and bridging the gap between model-based predictions and policy implementation.

4.5. Future Research Directions

Priority research directions include:

(1)

Temporal validation studies to assess parameter stability under land-use change scenarios [22,33,34], and to evaluate how parameters evolve under unresolved processes and changing landscape properties [61].

(2)

Broader geographical testing is needed to evaluate methodological transferability across diverse ecological and socio-economic contexts [54,55,56]. This aligns with recent calls for expanding systematic conservation planning frameworks to new regions and policy targets [62,63].

(3)

Integration with climate change projections will enable scenario-based conservation planning that accounts for both bioclimatic shifts and species responses [52,53]. In particular, linking bioclimate and population models offers improved forecasts for extinction risk, while iterative scenario testing may strengthen adaptive planning [52,53].

(4)

Development of automated parameter updating procedures is essential for operational applications. Advances in iterative near-term ecological forecasting [46] and Bayesian updating methods [64,65] provide promising pathways, and recent workflow frameworks demonstrate how automation can enhance reproducibility and transparency [66,67].

Future research should also extend the framework to incorporate species-specific habitat requirements and population dynamics, moving beyond generic habitat quality to population viability assessments [32,52]. Moreover, integration with other ecosystem service models would enable comprehensive landscape-scale planning that simultaneously addresses biodiversity, water, agriculture, and climate objectives [11,55]. Yet, caution is warranted, as ecosystem models still face fundamental limitations in predicting the outcomes of conservation decisions [68]. These challenges underscore the need for iterative refinement, model comparison, and stakeholder engagement to ensure robust decision-support tools for conservation.

5. Conclusions

5.1. Research Accomplishments

This research developed and validated a statistically rigorous framework for InVEST Habitat Quality parameterization that reduces subjective expert judgment and enhances ecological realism. The integrated PCA-SEM approach showed consistent superiority over conventional LULC methods across multiple validation criteria, establishing a new methodological standard for ecosystem service modeling. Similar integrative, data-driven approaches have been emerging in recent habitat quality research [69] which support the trend toward empirically based parameterization.

5.2. Scientific and Methodological Contributions

The framework addresses critical limitations in ecosystem service modeling by providing empirically derived, region-specific parameters that reflect local threat profiles and ecological conditions. Statistical validation confirmed robust model performance and reliable parameterization that enhances both scientific rigor and practical applicability for conservation planning. Innovations include:

• Objective parameter derivation substantially reducing expert judgment bias;
• Enhanced ecological realism through detailed biotope mapping;
• Comprehensive validation across independent ecological indicators;
• Transferable methodology applicable to diverse regional contexts.

These innovations align with developments in systematic conservation planning (SCP), which increasingly integrate ecosystem service models, algorithmic advances, and scalability approaches [62].

5.3. Policy and Conservation Implications

The validated framework provides evidence-based tools for Korean conservation policy implementation, particularly supporting the National Biodiversity Strategy (2021–2030) through reliable habitat quality assessment and conservation priority identification. The enhanced discrimination of conservation areas enables more effective resource allocation and systematic conservation planning. The methodology’s emphasis on regional calibration and empirical validation establishes principles that can guide ecosystem service modeling globally, potentially transforming practice from expert-dependent to evidence-based approaches. Moreover, coupling habitat models with multi-model frameworks (e.g., remote sensing + ecological models) has shown promise in recent studies, further bolstering the utility of integrative approaches [70].

5.4. Future Research and Applications

This research establishes a foundation for more reliable, scientifically defensible ecosystem service assessments essential for effective biodiversity conservation planning. Priority extensions include temporal validation studies, broader geographical testing, and integration with climate change scenarios for adaptive conservation planning. The framework’s demonstrated accuracy in conservation assessment provides tools that are increasingly demanded by international frameworks and national policies. As ecosystem service valuation becomes integral to environmental decision making, objective parameterization methods will become essential for maintaining scientific credibility and policy relevance. Yet, while models like this push forward methodological rigor, challenges remain; for example, predicting conservation outcomes under complex decision regimes remains difficult, as highlighted in recent critiques of ecosystem model predictability [68].