Machine learning-based assessment of offshore wind farm impacts on soft-bottom benthic communities in the Shandong Peninsula

Abstract

Offshore wind power, as a crucial component of clean energy, is rapidly expanding globally; however, the long-term impact mechanisms on marine benthic ecosystems remain unclear. This study focuses on four offshore wind farms (South 3, South 4, Site V, and Site U1) in the southern waters of the Shandong Peninsula. Based on benthic organism survey data and multi-source remote sensing environmental data from 2015 to 2024, a remote sensing-in-situ integrated machine learning prediction framework was constructed to systematically assess the spatiotemporal impact of wind farm construction and operation on soft-bottom benthic communities. The study employed the XGBoost model as the main model and the Generalized Additive Model (GAM) as the baseline model, using SHAP interpretability analysis to reveal key driving factors. The results show that the XGBoost model achieved an R² of 0.742 on the test set, significantly outperforming the GAM model (R²=0.625). The years in operation (YSI) was the most important factor affecting benthic community diversity; after a brief disturbance during the initial construction phase, the community showed a significant recovery trend after 2–4 years of operation. The artificial reef effect caused by the conversion to hard bottom near the pile foundations resulted in an approximately 13% increase in the Shannon diversity index and an approximately 40% increase in species richness compared to the control area. This study provides a reproducible methodological framework for the ecological impact assessment of offshore wind farms, and the findings can provide scientific basis for the environmentally friendly layout planning of offshore wind power in China.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-38939-0.

Introduction

Against the backdrop of the global energy transition, offshore wind power, as one of the most promising forms of renewable energy, is expanding at an unprecedented rate. According to data released by the Global Wind Energy Council (GWEC), the cumulative installed capacity of global offshore wind power exceeded 75 GW in 2023, and is expected to reach 380 GW by 2030¹. As the world’s largest offshore wind power market, China accounted for 58% of the global newly installed capacity in 2023, and its cumulative installed capacity has ranked first in the world for many consecutive years². The Shandong Peninsula, a key area for offshore wind power development in China, has built several large-scale offshore wind farms, including Nan 3, Nan 4, V site, and U1 site, forming a complete development pattern from demonstration projects to clustered development. However, while large-scale offshore wind power development promotes the optimization of the energy structure, it has also raised widespread concerns among academics and management departments regarding its potential ecological impacts.

The impact of offshore wind farm construction and operation on marine ecosystems is a complex issue involving multiple scales and processes. During the construction phase, underwater noise generated by pile driving may interfere with the behavior of marine mammals and fish, and sediment suspension and redistribution can alter the habitat of benthic organisms; during the operation phase, the hard substrate introduced by the wind turbine foundations provides new habitat for sessile organisms, cable laying changes sediment characteristics, and fishing restrictions may produce a protective effect similar to marine protected areas³. Previous studies have shown that the response of soft-bottom benthic communities to these engineering disturbances exhibits significant spatiotemporal heterogeneity: the initial construction phase is usually accompanied by a short-term decrease in biodiversity and abundance, while after several years of operation, both near-pile areas and adjacent soft-bottom zones often show increases in species diversity and biomass⁴. Long-term monitoring data from North Sea wind farms demonstrated that after 5 years of operation, species richness within wind farms could reach 1.5–2.0.5.0 times that of control areas⁵. This enhancement is attributed to two complementary mechanisms. First, the direct artificial reef effect creates novel hard-substrate habitat near foundations, supporting sessile and cryptic species absent from the original soft-bottom community. Second, indirect effects cascade to adjacent soft-bottom areas through altered food web dynamics, including increased predation by reef-associated fish, enhanced organic matter deposition from fouling communities, and modified larval supply from hard-substrate populations. Understanding this dual mechanism—direct habitat creation and indirect ecological spillover—is essential for predicting how offshore wind farms affect regional-scale benthic biodiversity, as the relative contribution of each mechanism depends on the spatial extent of hard substrate introduction and the distance over which indirect effects propagate through soft-bottom communities. Relevant researchers, based on long-term monitoring data from offshore wind farms in the North Sea, found that after 5 years of operation, the species richness within the wind farm could reach 1.5–2.0 times that of the control area, a finding that provides important theoretical support for the synergistic development of offshore wind power and marine ecology⁵.

Although a considerable amount of research on the ecological impact of offshore wind power has been accumulated internationally, existing research still has significant methodological limitations. From a research scale perspective, most studies focus on single wind farms or short-term monitoring periods (usually no more than 3 years), making it difficult to capture the long-term response trajectory and regional cumulative effects of benthic communities to engineering disturbances⁶. From an analytical method perspective, traditional studies mostly employ descriptive statistics or simple one-way ANOVA, lacking a systematic quantification of the interaction between environmental and engineering factors, and making it difficult to identify key driving mechanisms in complex nonlinear relationships. From a data integration perspective, the integration of field survey data and satellite remote sensing data is still insufficient, limiting the spatiotemporal extrapolation capabilities of research findings. In recent years, the application of machine learning methods in ecology has provided new technical pathways to overcome these bottlenecks⁷. Relevant researchers have applied gradient boosting tree algorithms to predict benthic organism distribution, achieving higher prediction accuracy than traditional statistical models; the introduction of interpretable analysis tools such as SHAP (SHapley Additive exPlanations) effectively addresses the ecological interpretation challenges posed by the black-box nature of machine learning models⁸.

Recent applications of machine learning to coastal and marine environmental assessment have demonstrated the effectiveness of XGBoost and SHAP analysis in uncovering complex nonlinear relationships between anthropogenic pressures and ecosystem responses. Notably, studies applying XGBoost-SHAP frameworks to marine ecological modeling have shown that these methods can outperform traditional statistical approaches in prediction accuracy while simultaneously providing interpretable insights into driver-response relationships^9,10. These precedents support the adoption of machine learning interpretability tools for assessing offshore wind farm impacts, where multiple interacting factors (temporal trends, spatial proximity, environmental conditions) create complex ecological patterns that are difficult to disentangle with conventional methods.

Based on the above research background and methodological gaps, this study uses four offshore wind farms in the southern sea area of the Shandong Peninsula as the research area, relying on benthic organism survey data spanning the baseline, construction, and operation periods from 2015 to 2024, and integrating multi-source satellite remote sensing environmental products such as MODIS and GOCI, to construct a remote sensing-field integrated machine learning prediction framework. The core objectives of this study include: quantifying the changes in benthic community α-diversity along temporal gradients (years of operation) and spatial gradients (distance from wind turbines, proportion of hard bottom); comparing the performance differences between XGBoost and GAM models in predicting benthic community diversity; and identifying key engineering and environmental factors affecting benthic communities through SHAP analysis and revealing their nonlinear response mechanisms^11,12. The methodological framework proposed in this study has good reproducibility and potential for cross-regional application, and the research conclusions can provide scientific reference for environmental impact assessment and ecologically friendly layout planning of offshore wind farms in China¹³.

Materials and methods

Overview of the study area

The study area is located in the southern waters of the Shandong Peninsula (36°30′N-37°10′N, 121°00′E-122°30′E), belonging to the western continental shelf of the Yellow Sea, and is an important offshore wind power development base in my country. The water depth in this area ranges from 18 to 35 m, and the seabed substrate is mainly silty clay and muddy silt, representing a typical soft-bottom sedimentary environment. Four operational offshore wind farms are located within the study area: Nan 3, Nan 4, V site, and U1 site, with a total installed capacity of approximately 2000 MW and a total of 295 wind turbines. The construction of the four wind farms shows a clear temporal gradient: Nan 3 and Nan 4, as the first batch of offshore wind power demonstration projects in Shandong Province, were connected to the grid and put into operation in December 2021. Each facility installed 58 wind turbines with a unit capacity of 5.2 MW, resulting in a total installed capacity of 301.6 MW per site; Site V was commissioned in December 2022, installing 71 turbines with unit capacities of 7.0 MW and 10.0 MW, achieving a total installed capacity of 500 MW. Site U1 was constructed in two phases. The first phase, comprising 53 turbines with a unit capacity of 8.5 MW, was commissioned in November 2023. The second phase, consisting of 53 turbines of the same model, was commissioned in October 2024, with a total installed capacity of 900 MW. This temporal gradient provides a natural space-for-time substitution research design for studying the response time of benthic communities to wind farm operation.

The environmental background conditions of the study area exhibit typical temperate monsoon climate characteristics. Analysis based on NOAA OISST and MODIS remote sensing products shows that the average annual sea surface temperature (SST) in the study area is 13.5–15.8 °C, exhibiting a seasonal variation pattern with high temperatures in summer (24–27 °C) and low temperatures in winter (3–5 °C); the average annual chlorophyll a (Chl-a) concentration is 1.2–3.8 mg/m³, reaching a peak of 4–6 mg/m³ during the spring algal bloom; and the total suspended matter (TSM) concentration, influenced by tides and wind waves, ranges from 5 to 25 mg/L, with higher concentrations during periods of strong wind and waves in winter¹⁴. Long-term time series analysis from 2015 to 2024 shows a slight upward trend in SST in the study area (approximately + 0.02 °C/yr), which is consistent with the overall warming trend of the Yellow Sea under the background of global warming¹⁵. To visually illustrate the geographical location of the study area, the distribution pattern of wind farms, and their construction timeline, this study has created a map showing the location of the study area and the distribution of wind farms, as shown in Fig. 1.

Fig. 1.

Location of study area and offshore wind farm distribution.

Figure 1 clearly presents the multi-scale spatial patterns and temporal evolution characteristics of the study area. Spatially, the four wind farms are distributed sequentially from west to east along the southern coastline of the Shandong Peninsula, forming a water depth gradient from nearshore shallow waters (Site V, 18–28 m) to offshore deep waters (South 3, 31–32 m). The sampling stations were strategically placed to ensure spatial representativeness, including three types: in-field stations (located within the wind turbine array), edge stations (0–2 km from the wind farm boundary), and control stations (> 5 km from the wind farm). Temporally, the construction timeline below shows that the study area experienced a complete development cycle, from a baseline period without engineering disturbance (2015–2018), through a demonstration project construction period (2019–2021), to a clustered operation period (2021–2024), providing the data foundation for the spatiotemporal comparative analysis conducted in this study.

Benthic organism sampling design

The benthic organism data in this study were obtained from systematic surveys conducted from 2015 to 2024. Surveys were conducted twice a year, once in spring (April-May) and once in autumn (September-October). A 0.1 m² grab sampler was used, and three replicate samples were collected at each station. The sampling stations were established following a stratified random sampling principle and dynamically adjusted according to the progress of the wind farm construction: during the baseline period (2015–2018), only 8–10 stations were set up in the control area outside the pre-planned site to obtain baseline data under conditions without engineering disturbance; during the construction period (2019–2020), as pile foundation construction progressed, stations within and at the edge of the site were gradually added; during the operation period (2021–2024), the sampling network was further intensified, and the stations within the site were subdivided into near-pile stations (less than 100 m from the wind turbine) and soft-bottom stations (100–500 m from the wind turbine) to capture habitat differentiation caused by hard-bottom transformation. As of 2024, a total of 204 station surveys have been completed, yielding 612 benthic organism samples (3 replicate samples per station).

To address potential bias from dynamic station adjustments, control stations maintained fixed locations throughout the study. Eight control stations established during baseline (2015–2018) were repeatedly sampled in all surveys, providing consistent temporal reference unaffected by wind farm operations. These stations were positioned 6–15 km from wind farm boundaries at comparable water depths (25–32 m) and sediment types (silty clay), ensuring observed differences reflect wind farm effects rather than environmental gradients. Within-farm and edge stations were added progressively as construction advanced. Although individual within-farm stations lack complete temporal records, the analysis does not rely on before-after comparisons at specific stations. Instead, machine learning models utilize the full spatiotemporal dataset with station identity and YSI as covariates, inferring temporal trends from control station time series and cross-sectional comparisons among wind farms. Spatial block cross-validation ensures spatially independent evaluation, mitigating pseudoreplication concerns.

The processing and identification of benthic organism samples followed the standard procedures of the “Marine Survey Specifications” (GB/T 12763.6)¹⁶. The sediment samples collected in the field were rinsed through a 0.5 mm mesh sieve, fixed and preserved with 5% formalin solution, and then brought back to the laboratory for sorting, identification, and counting. Species identification referred to the “Checklist of Marine Organisms in China” and the “Atlas of Macrobenthic Animals in the Yellow and Bohai Seas”, with identification accuracy reaching the species or genus level^17,18. A total of 156 benthic animal species were identified (after removing duplicates), belonging to major groups such as polychaetes, mollusks, crustaceans, and echinoderms. Based on the species identification results, the α-diversity indices of each sample were calculated, including the Shannon-Wiener diversity index (H′), Pielou’s evenness index (J′), and species richness (S). This study uses the Shannon index as the main response variable, and its calculation formula is:

1

In the formula  , represents the proportion of individuals of the species to the total number of individuals. To illustrate the spatiotemporal coverage of benthic organism sampling, Table 1

Table 1.

Summary of sampling stations and benthic samples (2015–2024).

Year	Inside the Plot	Edge	Control	Total number of stations	Sample size	Number of species	Stage
2015	-	-	8	8	24	45	Baseline period
2016	-	-	8	8	24	52
2017	-	-	10	10	30	58
2018	-	-	10	10	30	61
2019	4	4	8	16	48	68	Construction period
2020	6	6	8	20	60	72
2021	8	6	8	22	66	85	Initial operation
2022	12	8	10	30	90	98
2023	16	10	10	36	108	112
2024	20	12	12	44	132	125
Total	66	46	92	204	612	156*	-

* indicates the cumulative number of species identified over the years (duplicates across years have been removed); three parallel samples were collected at each station.

Table 1 shows that the sampling effort in this study increased year by year with the progress of wind farm construction. During the baseline period (2015–2018), the survey mainly focused on the control area, accumulating 108 samples and recording an increase in species from 45 to 61, reflecting a systematic assessment of the baseline biodiversity in the study area during this phase. During the construction period (2019–2020), sampling stations were established within and at the edges of the wind farm, increasing the sample size to 108 and the cumulative number of species to 72. The operational period (2021–2024) was the stage with the highest sampling density, especially in 2024 when 20 new stations were added within the wind farm, bringing the total sample size to 132 and the cumulative number of recorded species to 125. Notably, the increase in the cumulative number of species mainly originated from areas near the pile foundations, a phenomenon suggesting a gain in species diversity due to the introduction of hard substrates, which will be further analyzed in the results section.

Stations were classified into four categories based on distance from turbines and substrate type. Soft-bottom stations (> 100 m from foundations) retained natural silty clay seabed. Near-pile stations (< 100 m) had scour protection creating > 15% hard substrate cover (rock riprap, 20–50 cm diameter, 25–30 m radius). Edge stations (0–2 km from wind farm boundary) captured transition zones. Control stations (> 5 km from boundary) maintained baseline soft-bottom conditions. This classification distinguishes direct artificial reef effects from broader wind farm impacts on soft-bottom communities. Within-farm soft-bottom stations (100–500 m from turbines) provide the primary data for assessing impacts on the soft-bottom benthic community, which is the main focus of this study, while near-pile stations document the magnitude of the artificial reef effect for comparison.

Remote sensing and environmental data

This study integrated multi-source satellite remote sensing data to characterize the environmental background conditions of the study area. Sea surface temperature (SST) data were obtained from NOAA OISST (Optimum Interpolation Sea Surface Temperature) daily products (0.25° spatial resolution) and MODIS-Aqua 4 km monthly products; the two datasets were cross-validated and averaged for use¹⁴. Chlorophyll-a (Chl-a) concentration data were obtained from MODIS Ocean Color Level-3 standard products (4 km monthly values), and quality control was performed with reference to the European Space Agency OC-CCI (Ocean Colour Climate Change Initiative) products¹⁹. Total suspended matter (TSM) concentration data were mainly obtained from the South Korean GOCI (Geostationary Ocean Color Imager) satellite daily products (500 m resolution) and my country’s HY-1 series satellite products. The high temporal resolution of GOCI (one image per hour, approximately eight images per day) is particularly suitable for capturing the tidal-influenced daily variation characteristics of TSM in the study area^20,21.

The spatiotemporal matching of environmental data was processed according to the following procedure: for each benthic sampling station, the average of remote sensing pixels within a 3 km × 3 km window around its coordinates was extracted to reduce the influence of single-pixel noise; temporal matching adopted a 30-day pre-sampling average strategy, i.e., the monthly average of each environmental variable within 30 days before the sampling date was extracted to reflect the cumulative environmental effects at the time of sampling. In addition, this study also obtained spatial covariates such as water depth (from the GEBCO 2023 global bathymetry dataset), distance from the shore (calculated based on Chinese coastline vector data), and sediment grain size (from historical survey reports). To intuitively display the spatiotemporal variation patterns of environmental factors in the study area over the past decade, Fig. 2 summarizes the interannual variation trends and seasonal fluctuation characteristics of SST, Chl-a, and TSM.

Fig. 2.

Spatiotemporal Variation of Environmental Factors (2015–2024).

The spatial matching protocol was implemented as follows. For each benthic sampling station coordinate (latitude, longitude), a 3 km × 3 km buffer zone was created centered on the station location. All remote sensing pixels falling within this buffer were identified, and their values were averaged using an area-weighted mean approach to account for partial pixel coverage at buffer edges. For MODIS products (4 km resolution), this typically resulted in 1–4 pixels per station. For GOCI products (500 m resolution), approximately 36 pixels were averaged per station. The 3 km window size was selected to balance two competing considerations: it is large enough to mitigate the impact of single-pixel anomalies caused by cloud contamination or sensor noise, yet small enough to preserve the spatial gradient of environmental variables across the study area.

The temporal matching protocol addressed the time lag between environmental forcing and benthic community response. For each benthic sampling event conducted on date T, environmental variables were calculated as 30-day running means covering the period from T-30 days to T. This 30-day window reflects the characteristic time scale over which water column conditions (temperature, chlorophyll, suspended matter) influence benthic metabolism and recruitment. Daily OISST data were averaged over the 30-day period. Monthly MODIS products were used directly if the sampling date fell within the product’s temporal coverage window, otherwise the nearest monthly product within the 30-day period was used. For GOCI products, which provide hourly observations, daily composites were first generated by selecting cloud-free observations, and these daily composites were then averaged over the 30-day period. Quality control included removal of pixels flagged for cloud cover, high solar zenith angle, or algorithm failure, with stations excluded from analysis if fewer than 50% of the expected observations were available within the 30-day window.

Figure 2 reveals the multi-temporal scale variation characteristics of environmental factors in the study area and their potential correlation with the construction of the wind farm. In terms of interannual variation (Fig. 2a-c), SST showed a slow upward trend, with the annual average increasing from 14.2 °C in 2015 to 15.6 °C in 2024, an annual increase of approximately + 0.02 °C, consistent with the overall warming trend in the Yellow Sea. Chl-a concentration showed significant interannual fluctuations, with peaks in 2019 (3.8 mg/m³) and 2023 (3.5 mg/m³), possibly related to the interannual variability of spring algal bloom intensity. TSM concentration showed a brief increase during the construction period (2019–2021), peaking at 25 mg/L, which is directly related to sediment disturbance during pile foundation construction; during the operation period, TSM gradually returned to the background level (8–12 mg/L), indicating that the sedimentary environment tended to stabilize. In terms of seasonal variation (Fig. 2d-e), the seasonal amplitude of SST reached over 20 °C (3–5 °C in winter, 24–27 °C in summer), and Chl-a reached its peak in spring (4–6 mg/m³), reflecting the typical temperate monsoon climate characteristics and spring algal bloom phenomenon in this sea area. The spatial heatmap in Fig. 2f shows that the SST in the nearshore shallow water area is slightly higher than that in the offshore deep water area, and this spatial gradient is consistent with the water depth distribution pattern of the wind farm²².

Engineering and spatial covariates

This study constructed three types of spatial covariates representing the intensity of engineering disturbances from wind farms. Years Since Installation (YSI) is defined as the time elapsed from the sampling time to the grid connection and operation of the wind farm where the sampling station is located (measured in years), and is used to quantify the cumulative time effect of wind farm operation; for stations that have not yet been constructed or are located in the control area, YSI is assigned a value of 0. YSI was treated as a continuous numerical variable in both XGBoost and GAM models, allowing the models to capture the gradual temporal trajectory of community response rather than imposing discrete temporal categories. The continuous YSI values in the dataset ranged from 0 (baseline and control stations) to 4.2 years (for South 3 and South 4 stations sampled in late 2024), with intermediate values reflecting the commissioning timeline of different wind farms and the sampling dates within each operational period. This continuous representation enables the identification of critical temporal thresholds (such as the 2.5-year recovery point revealed by SHAP analysis) and facilitates spatiotemporal substitution analysis across the four wind farms with staggered operational histories. Turbine Density is defined as the number of wind turbines within a circular buffer with the sampling station as the center and a radius of 2 km, divided by the area of the buffer (units: turbines/km²), used to characterize the intensity of engineering disturbance at the local scale. Hard Substrate Proportion is defined as the percentage of the seabed area covered by wind turbine foundations and scour protection structures within the buffer (units: %), based on engineering design data. The single-pile foundation diameter of each wind turbine is approximately 8 m, and the scour protection radius is approximately 25–30 m, resulting in an effective hard substrate area of approximately 2500 m²/turbine. Table 2 summarizes the main engineering attribute parameters of the four wind farms.

Table 2.

Summary of wind farm project Attributes.

Site location	Installed capacity (MW)	Number of units	Single unit (MW)	Water depth (m)	Offshore (km)	Grid connection time
South 3	301.6	58	5.2	31–32	37	2021.12
South 4	301.6	58	5.2	28–32	35	2021.12
Site V	500	71	7.0/10.0	18–28	26	2022.12
U1 Phase 1	450	53	8.5	25–35	30	2023.11
U1 Phase 2	450	53	8.5	25–35	30	2024.10
Total	~ 2000	293	-	18–35	26–37	-

Data sources: Approval documents from the National Development and Reform Commission, official announcements from State Power Investment Corporation Shandong Energy, and public reports from Qilu.com, etc.

Data sources: Approval documents from the National Development and Reform Commission, official announcements from State Power Investment Corporation Shandong Energy, and public reports from Qilu.com, etc.

Table 2 shows that the wind farms in the study area exhibit a technological evolution path from demonstration projects to large-scale development. South 3 and South 4, as the first batch of demonstration projects, utilize mature turbine models with a single-unit capacity of 5.2 MW, located in deeper waters further offshore (31–32 m). Site V introduced two large-capacity turbine models of 7.0 MW and 10.0 MW, reflecting the technological trend of larger turbines. Its location in shallower waters near the shore (18–28 m) is conducive to reducing construction difficulties and operation and maintenance costs. Site U1, the latest project, uses 8.5 MW turbines and was constructed in two phases to control investment pace and environmental risks. The commissioning time of the four wind farms spans 3 years (December 2021-October 2024), creating a continuous gradient of YSI from 1 to 4 years, providing ideal conditions for studying the temporal response of benthic communities to engineering disturbances. Figure 3 will visually present the spatial distribution patterns of YSI, turbine density, and hard bottom proportion based on spatial interpolation methods.

Fig. 3.

Spatial distribution map of engineering gradients.

Feature engineering and model framework

This study used the Shannon diversity index (H′) as the response variable and integrated environmental factors, engineering factors, and spatial covariates to construct a predictive model. The initial feature set included 15 candidate variables: environmental factors (annual mean SST, seasonal amplitude of SST, annual mean Chl-a, seasonal amplitude of Chl-a, TSM concentration), engineering factors (YSI, distance to the nearest wind turbine, turbine density, proportion of hard bottom), and spatial covariates (water depth, distance from shore, sediment grain size, longitude, latitude, sampling season). Feature engineering involved two steps: all continuous variables were standardized using Z-score standardization (subtracting the mean and dividing by the standard deviation) to eliminate dimensional differences; and multicollinearity diagnosis was performed using the variance inflation factor (VIF), removing highly collinear variables with VIF > 5 (longitude and latitude were removed due to their high correlation with distance from shore), resulting in 13 variables for modeling²³.

The choice of standardization method can influence model performance, particularly for distance-based algorithms and neural networks, although tree-based methods like XGBoost are relatively insensitive to feature scaling due to their split-based decision rules. Z-score standardization (mean = 0, standard deviation = 1) was selected for this study because it preserves outlier information, which may represent ecologically meaningful extreme events (e.g., unusual temperature spikes, high suspended sediment loads during storms), and because it produces symmetric distributions suitable for both XGBoost and GAM. Alternative normalization approaches, such as min-max scaling to [0,1] range or robust scaling based on median and interquartile range, could potentially improve prediction accuracy if predictor distributions are highly skewed or contain numerous outliers. Future studies may benefit from comparing multiple normalization methods and their effects on model interpretability, as suggested by recent work demonstrating that preprocessing choices can substantially impact environmental model performance. For the present study, Z-score standardization provided adequate performance and facilitated coefficient interpretation across variables measured in different units.

Machine learning models and baseline models

This study employs the XGBoost (eXtreme Gradient Boosting) algorithm as the main model. XGBoost is an ensemble learning method based on the gradient boosting decision tree framework, and it is widely used in ecological prediction research due to its superior performance in capturing nonlinear relationships and variable interaction effects²⁴. The objective function of XGBoost consists of a loss function and a regularization term:

2

In the equation, is the loss function (mean squared error is used in this study), and is the regularization term, used to control model complexity. The specific form of the regularization term is:

3

where is the number of leaf nodes, is the weight of the leaf node, and and are penalty coefficients.

To avoid performance overestimation caused by spatial autocorrelation, this study adopted a spatial block cross-validation strategy: the study area was divided into 5 spatial blocks along the longitude direction²⁵. In each fold, one spatial block was used as the test set, and the remaining four were used as the training set, ensuring that the training and test sets were spatially independent. Hyperparameter tuning was performed using a grid search method, with the search space covering key parameters such as learning rate (0.01, 0.05, 0.1), maximum tree depth (3, 5, 7), number of iterations (100, 200, 500), and subsampling ratio (0.7, 0.8, 0.9). The average R² from 5-fold cross-validation was used as the optimization objective.

The complete set of XGBoost hyperparameters used in the final model is presented below. The learning rate (eta) controls the step size at each boosting iteration, with smaller values requiring more iterations but potentially improving generalization. The maximum depth of each tree determines model complexity, with deeper trees capable of capturing more complex interactions but at risk of overfitting. The subsampling ratio controls the fraction of training data used to build each tree, introducing stochasticity that helps prevent overfitting. Column subsampling (colsample_bytree) further reduces overfitting by randomly selecting features for each tree. The regularization parameters (alpha for L1, lambda for L2) penalize model complexity. The minimum child weight parameter prevents the creation of leaf nodes with very few samples. The gamma parameter controls the minimum loss reduction required to make a split, with higher values making the model more conservative.

These hyperparameters were selected through grid search with 5-fold spatial cross-validation (Table S1). The search space explored learning rates of 0.01, 0.05, and 0.1; maximum depths of 3, 5, and 7; numbers of trees of 100, 200, and 500; and subsampling ratios of 0.7, 0.8, and 0.9. Other hyperparameters were set to XGBoost defaults. The final configuration (learning_rate = 0.05, max_depth = 5, n_estimators = 200, subsample = 0.8) achieved the highest average cross-validation R² of 0.718.

Model performance was evaluated using four metrics: coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE). R² measures the proportion of variance explained by the model, with values closer to 1 indicating better fit. RMSE and MAE quantify prediction error in the original units, while MAPE expresses error as a percentage. The formulas for R² and RMSE are defined as follows:

4

5

To evaluate the predictive performance advantages of XGBoost, this study also constructed a Generalized Additive Model (GAM) as a baseline model¹¹. GAM captures the nonlinear relationship between the response variable and predictor variables through smoothing functions, and its mathematical form is:

6

In the equation, represents the thin-plate spline smoothing function, and the basis dimension is set to k = 10.

The complete GAM model structure is specified as follows. Each smooth term is defined as a linear combination of basis functions:

7

where are thin-plate regression spline basis functions and  are coefficients estimated by penalized maximum likelihood. The smoothing penalty λ controls the trade-off between fit and smoothness, with optimal λ values selected by generalized cross-validation (GCV). The model includes four smoothed terms (YSI, water depth, SST, hard substrate proportion) and nine linear terms (distance to nearest turbine, turbine density, Chl-a, TSM, seasonal amplitude of SST, seasonal amplitude of Chl-a, distance from shore, sediment grain size, sampling season). The basis dimension k = 10 for each smooth term provides sufficient flexibility to capture nonlinear patterns while preventing overfitting, as confirmed by basis dimension diagnostic checks (k-index > 0.9 for all terms). The model was fitted using the mgcv package in R with default settings: thin-plate regression splines, GCV for smoothing parameter selection, and restricted maximum likelihood (REML) for coefficient estimation. The full model formula is:

8

where Shannon is the Shannon diversity index ( ), and all continuous predictors were standardized prior to model fitting.

Figure 4 illustrates the four-layer architecture of the remote sensing-field integrated machine learning framework developed in this study. The data input layer integrates heterogeneous sources including remote sensing-derived environmental variables (SST, Chl-a, TSM spanning 2015–2024), benthic biodiversity observations from 612 samples across 204 stations, and engineering parameters characterizing disturbance intensity (turbine locations, commissioning dates, installed capacity). The data processing layer harmonizes multi-source inputs into a unified modeling matrix through spatial matching, standardization, and VIF-based collinearity screening. The model construction layer employs XGBoost and GAM in parallel configuration with 5-fold spatial block cross-validation to ensure robust performance evaluation. The interpretation output layer represents the framework’s core innovation by integrating SHAP analysis to quantify predictor contributions, response curves to reveal non-linear factor effects, and spatial prediction maps to achieve rasterized biodiversity mapping. The modular architecture ensures portability and facilitates extension to ecological impact assessments across diverse offshore wind installations^26,27.

Fig. 4.

Technical roadmap.

Variable importance and interpretability analysis

This study uses the SHAP (SHapley Additive exPlanations) method to perform interpretability analysis on the XGBoost model^28,29. SHAP is based on the Shapley value concept from game theory and decomposes the model’s prediction into the marginal contributions of each feature. For the j-th feature of sample x, its SHAP value is defined as:

9

In the formula, is the total number of features, is a subset of features excluding feature  , and is the model’s predicted value based on the subset  . SHAP values satisfy the additive decomposition property:

10

 is the baseline value predicted by the model.

SHAP analysis output includes two types of results: global importance ranking, obtained by calculating the average absolute SHAP values (| |) for each feature and sorting them in descending order, visualized using a swarm plot; and SHAP dependence plots, which show the scatter relationship between individual feature values and their SHAP values, with color coding indicating the influence of interaction variables, used to reveal non-linear response patterns. In addition, partial effect plots from the GAM model are used as an independent validation method, cross-referenced with the SHAP analysis results to enhance the robustness of key findings.

Statistical software and tools

Data processing and statistical analysis in this study were performed using both R (version 4.3.2) and Python (version 3.10.12) platforms. In the R environment, benthic community diversity calculations and NMDS ordination analysis were performed using the vegan package (version 2.6–4.6), GAM model construction used the mgcv package (version 1.9-0.9), and spatial data processing used the sf and terra packages. In the Python environment, XGBoost model construction used the xgboost package (version 2.0.3), hyperparameter grid search was implemented through the scikit-learn package, and SHAP interpretability analysis used the shap package (version 0.44.0). Data visualization was completed using plotting toolkits such as matplotlib, seaborn, and cartopy.

Results

Spatial distribution pattern of engineering gradients

Based on wind farm engineering attribute data and spatial interpolation methods (Inverse Distance Weighting, IDW), this study generated spatial distribution maps of operational years (YSI), turbine density, and hard bottom ratio, as shown in Fig. 3, to intuitively present the spatial heterogeneity of engineering gradients in the study area. The spatial distribution of the three types of engineering factors all showed distinct site differentiation characteristics: YSI was highest at sites South 3 and South 4 (4 years), and decreased eastward to site V (3 years), U1 Phase I (2 years), and U1 Phase II (1 year); turbine density was highest at site V (1.08 turbines/km²), followed by South 3 and South 4 (approximately 1.0 turbines/km²), while the U1 site had a lower density (0.74 turbines/km²) due to its large sea area (143 km²); the spatial pattern of the hard bottom ratio was similar to that of turbine density, but showed stronger local heterogeneity—the hard bottom ratio near the pile foundations could reach 15%–20%, while the soft bottom areas within the wind farm were only 2%−5%.

The comparative analysis of Fig. 3 reveals spatial coupling and differentiation among engineering factors across wind farm installations. The YSI distribution (Fig. 3a) displays a west-to-east decreasing gradient consistent with construction chronology, whereby South 3 and South 4 as the earliest grid-connected projects (late 2021) have accumulated 4-year operational periods as of December 2024, while U1 Phase II commissioned in October 2024 exhibits only 1-year YSI, providing a spatiotemporal substitution framework for examining benthic community response trajectories. Turbine density patterns (Fig. 3b) show maximum values at the V site attributable to compact spatial design (71 turbines within 66 km²) forming orange-red high-density zones, moderate contiguous density at adjacent South 3 and South 4 installations, and minimum density at U1 reflecting large-spacing low-density configuration. Hard bottom proportion (Fig. 3c) exhibits fine-scale local heterogeneity characterized by green hotspots surrounding individual turbines due to scour protection structures, with these hotspots forming contiguous patterns at the densely-spaced V site while maintaining discrete distributions at U1. The control area (labeled Ref) demonstrates zero or near-zero values across all three indicators, establishing an effective reference baseline for spatiotemporal comparative analysis.

Spatiotemporal variation patterns of benthic communities

The α-diversity of benthic communities showed significant differentiation characteristics across both temporal stages and spatial regions. Statistical analysis based on 612 benthic samples revealed that the overall range of the Shannon diversity index (H′) was 1.15–3.28, with a mean of 2.35 ± 0.42; the species richness (S) ranged from 8 to 38 species/station, with a mean of 18.5 ± 5.8 species/station. To systematically present the variation patterns of diversity indices along spatiotemporal gradients, Fig. 5 uses grouped box plots to show the distribution characteristics of the Shannon index and species richness in different temporal stages and spatial regions³⁰.

Fig. 5.

Changes in benthic community α-diversity over time and across spatial regions.

Figure 5 reveals a temporal response pattern of benthic community diversity showing an initial decrease followed by an increase, and a spatial response pattern of differentiation within the study area. From a temporal perspective (Fig. 5a and c), the Shannon index decreased from 2.45 ± 0.32 during the baseline period to 2.08 ± 0.45 during the construction period (a decrease of 15%, p < 0.001), then recovered to 2.18 ± 0.38 in the early operation phase, and reached 2.55 ± 0.35 in the mid-operation phase (a 4% increase compared to the baseline period, p < 0.05). Species richness showed a similar U-shaped change from 18.5 ± 4.2 species/station in the baseline period to 14.3 ± 5.1 species/station during the construction period (a decrease of 23%), then recovered to 20.8 ± 4.5 species/station in the mid-operation phase (a 12% increase compared to the baseline period). This temporal pattern of initial decline followed by recovery aligns with construction disturbance-operational recovery dynamics reported in comparable international studies, demonstrating benthic community resilience to wind farm impacts. Spatial analysis (Fig. 5b, d) reveals significantly elevated diversity indices near pile foundations relative to soft-bottom and marginal zones, with Shannon index values of 2.85 ± 0.38 representing approximately 13% increases over control areas (2.52 ± 0.28, p < 0.001), while species richness reaches 26.8 ± 5.5 species/station reflecting approximately 40% enhancement compared to controls (19.2 ± 3.8 species/station, p < 0.001, Cohen’s d = 0.9 indicating large effect size). This spatial differentiation pattern strongly indicates artificial reef effects whereby pile structures introduce hard-bottom substrates that provide novel ecological niches on previously homogeneous soft sediments, facilitating colonization by sessile and cryptic taxa³¹.

The spatiotemporal differentiation of community structure was further visualized through non-metric multidimensional scaling (NMDS) analysis. Based on the Bray-Curtis dissimilarity matrix, a stable two-dimensional ordination solution was obtained after 250 iterations (Stress = 0.18, indicating a good fit). PERMANOVA analysis showed that both spatial region (R²=0.32, F = 8.5, p < 0.001) and temporal stage (R²=0.12, F = 4.2, p = 0.003) significantly explained the variation in community structure. Figure 6 shows the NMDS ordination results and their association with environmental gradients.

Fig. 6.

NMDS ordination plot showing the spatiotemporal separation of community structure.

Figure 6’s NMDS ordination space clearly shows the differentiation pattern of benthic communities along the engineering gradient. Three spatial regions occupy relatively independent positions in the ordination space: on-site samples (red ellipse) are mainly distributed in the negative region of the NMDS1 axis, consistent with the direction of the hard substrate proportion and YSI environmental vectors; control area samples (blue ellipse) are clustered in the positive region of the NMDS1 axis, close to the direction of the water depth vector; and edge area samples (yellow ellipse) are located between the two, forming a transitional zone. From a temporal perspective, baseline samples (circles) are relatively concentrated, representing the baseline community state before engineering disturbance; construction phase samples (triangles) show a significant trend of dispersion, reflecting the increased community heterogeneity caused by construction disturbance; and operation phase samples (squares) differentiate into two subgroups within the on-site area—samples near the pile foundations shift towards the negative direction of the NMDS1 axis (associated with a high proportion of hard substrate), while soft-bottom area samples tend to return to the baseline position³². The kernel density curves at the top and right further quantify this differentiation: along the NMDS1 axis, the density peaks of the three regions show a clear left-right displacement; along the NMDS2 axis, the degree of differentiation is relatively weaker, indicating that NMDS1 is the main gradient axis differentiating the communities³³. To quantitatively summarize the statistical characteristics of benthic community indicators in each region and stage, Table 3 presents the descriptive statistics of the Shannon index, species richness, abundance, and biomass³⁴.

Table 3.

Statistical summary of benthic community indicators (by region and stage).

Region	Phase	H’ (Mean ± SD)	S (Mean ± SD)	N (ind/m²)	B (g/m²)
Control area	Baseline period	2.45 ± 0.32	18.5 ± 4.2	856 ± 245	15.8 ± 6.3
	Operation period	2.52 ± 0.28	19.2 ± 3.8	912 ± 268	16.5 ± 5.8
In-field soft bottom	Construction period	2.08 ± 0.45	14.3 ± 5.1	625 ± 312	11.2 ± 7.5
	Early operation	2.18 ± 0.38	15.8 ± 4.5	745 ± 285	13.5 ± 6.2
	Mid-operation	2.35 ± 0.35	17.2 ± 4.0	825 ± 256	14.8 ± 5.5
Near-pile hard bottom	Early operation	2.65 ± 0.42	22.5 ± 6.8	1250 ± 485	28.5 ± 12.3
	Mid-operation	2.85 ± 0.38	26.8 ± 5.5	1580 ± 520	35.2 ± 14.5
Edge area	Operation period	2.38 ± 0.35	17.8 ± 4.2	885 ± 275	15.2 ± 5.8

H′ = Shannon diversity index; S = species richness; N = abundance; B = biomass

The data in Table 3 further confirms the spatiotemporal differentiation patterns revealed in Figs. 5 and 6, and provides a more comprehensive comparison of community indicators. The control area showed no significant changes in any indicators between the baseline and operational periods (p > 0.05), indicating that the control area selected in this study was minimally affected by the indirect impacts of wind farm construction and can serve as a valid reference baseline. The soft-bottom area within the wind farm showed a continuous recovery trend from the construction phase to the mid-operational phase, with the Shannon index increasing from 2.08 to 2.35 and species richness increasing from 14.3 species to 17.2 species, representing recovery rates of 13% and 20%, respectively^18,35. The hard-bottom area near the pile foundations was identified as a biodiversity hotspot in this study: the Shannon index reached 2.85 during the mid-operational phase, 13% higher than the control area during the same period; species richness reached 26.8 species/station, 40% higher than the control area; abundance reached 1580 ind/m², 1.7 times that of the control area; and biomass reached 35.2 g/m², 2.1 times that of the control area. The consistency of these indicators suggests that the artificial reef effect is not only reflected in species diversity but also leads to a significant increase in community biomass and productivity. The indicators in the edge zone were intermediate between those of the wind farm and the control area, reflecting its ecological characteristics as a transitional zone^14,36.

Model performance comparison

Model performance evaluation based on 5-fold spatial block cross-validation showed that XGBoost significantly outperformed the GAM baseline model in the Shannon index prediction task. The optimal hyperparameter combination for XGBoost was: learning_rate = 0.05, max_depth = 5, n_estimators = 200, subsample = 0.8. Table 4 summarizes the performance metrics comparison of the two models across the training set, test set, and cross-validation.

Table 4.

Comparison of model performance metrics.

Model	Dataset	R²	RMSE	MAE	MAPE(%)
XGBoost	Training set	0.865	0.218	0.165	7.2
	Test set	0.742	0.285	0.215	9.5
	Cross-validation	0.718 ± 0.045	0.298 ± 0.032	0.225 ± 0.028	10.2 ± 1.8
GAM	Training set	0.712	0.315	0.242	10.5
	Test set	0.625	0.368	0.285	12.8
	Cross-validation	0.598 ± 0.058	0.385 ± 0.042	0.295 ± 0.035	13.5 ± 2.2

Cross-validation results are expressed as mean ± standard deviation; MAPE = Mean absolute percentage error

Table 4 shows that XGBoost outperformed GAM in all performance metrics. In terms of the test set R², XGBoost achieved 0.742, meaning the model explained 74.2% of the variation in the Shannon index, while GAM only achieved 0.625 (62.5%), a significant difference (ΔR² = 0.117). In terms of prediction error, the test set RMSE for XGBoost was 0.285, a 22.6% reduction compared to GAM (0.368); the MAPE was 9.5%, a 3.3% point reduction compared to GAM (12.8%). It is worth noting that the R² difference between the training and test sets for XGBoost (0.865 → 0.742, Δ = 0.123) was slightly larger than that of GAM (0.712 → 0.625, Δ = 0.087), indicating a slight tendency towards overfitting in XGBoost, but its overall generalization ability was still superior to GAM. Stability analysis of the cross-validation results showed that the standard deviation of the five-fold R² for XGBoost was 0.045, lower than GAM’s 0.058, indicating that XGBoost is less sensitive to data partitioning. In summary, XGBoost demonstrated advantages in both prediction accuracy and robustness, validating the rationality of choosing it as the main model in this study. Figure 7 visually compares the predicted and observed values of the two models using a scatter plot³⁷.

Fig. 7.

Scatter plot of model predictions vs. observed values.

To assess whether model residuals exhibit spatial dependence, we conducted Moran’s I tests on the prediction errors from both models. For the XGBoost model, the test set residuals (observed - predicted Shannon index) yielded a Moran’s I statistic of 0.08 (p = 0.12), indicating no significant spatial autocorrelation at a 0.05 significance level. Similarly, GAM residuals produced a Moran’s I of 0.11 (p = 0.09), also non-significant. These results suggest that the spatial block cross-validation strategy effectively prevented spatial dependence from inflating model performance metrics. As a supplementary diagnostic, we computed Moran’s I at multiple distance lags (0–5 km, 5–10 km, 10–20 km, > 20 km) to examine scale-dependent spatial structure in residuals. Across all distance classes, Moran’s I values remained between − 0.05 and 0.12, with no values reaching statistical significance (all p > 0.10). The absence of residual spatial autocorrelation indicates that both models successfully captured the spatial structure of benthic diversity and that the reported R² values reflect genuine predictive capacity rather than inflated correlation due to spatial dependence. This validation is particularly important in marine benthic ecology, where environmental gradients and dispersal limitation often generate strong spatial patterns that can confound statistical inference if not properly accounted for in model evaluation procedures.

Figure 7 visually demonstrates the differences in predictive performance between the two models using scatter plots. In the XGBoost scatter plot (Fig. 7A), both the training set (gray) and test set (dark) points are tightly clustered around the 1:1 reference line, with only slight deviations in the extreme regions of the Shannon index (< 1.5 and > 3.0). The linear fit line (red) highly overlaps with the 1:1 line, with a slope close to 1.0, indicating no systematic bias in the model. The embedded residual histogram shows a symmetrical normal distribution with a peak near 0, confirming the assumption of randomness in the residuals. The GAM scatter plot (Fig. 7B) shows a relatively dispersed distribution, especially in the medium-to-high Shannon index range (2.5–3.0), where there is a clear systematic underestimation—predicted values are generally lower than observed values, resulting in a fit line (purple) with a slope less than the 1:1 line. The residual distribution shows a slight left skew, suggesting that GAM has insufficient predictive power for high-diversity samples. This phenomenon may stem from the inherent limitations of GAM’s additive structure: when benthic community diversity is driven by multifactor interactions (such as the synergistic effect of YSI × hard bottom proportion), GAM struggles to capture these non-additive effects, while XGBoost’s tree structure is naturally suited for modeling variable interactions³⁸.

Identification of key driving factors

SHAP analysis revealed the key factors influencing the Shannon index of benthic communities and their relative importance ranking. Figure 8 displays the global variable importance in the form of a beeswarm plot: each point represents a sample, the x-axis represents the SHAP value (the magnitude of the impact on the model output), and the color indicates the original value of the variable (red = high value, blue = low value). Ranked in descending order of average absolute SHAP value, the top five driving factors are: Years in Service (YSI), distance to the nearest wind turbine, water depth, hard bottom proportion, and turbine density, with YSI having an average absolute SHAP value of 0.28, significantly higher than other variables.

Fig. 8.

SHAP global variable importance bar chart.

Figure 8 shows SHAP distribution reveals the direction and non-linear characteristics of each factor’s influence on benthic diversity. The point distribution in the YSI row shows a right-skewed red pattern: high YSI values (red points) correspond to positive SHAP values, indicating that longer operating years contribute more positively to the Shannon index. However, in the YSI = 0–2 year range (blue-purple points), SHAP values are predominantly negative, reflecting a brief negative impact during the initial construction and early operation phases. The distance to the nearest wind turbine row shows the opposite pattern: short distances (blue points) correspond to positive SHAP values, while long distances (red points) correspond to negative or near-zero SHAP values, indicating that areas closer to wind turbines have higher diversity, consistent with the artificial reef effect hypothesis. The SHAP distribution of the water depth row shows that moderate water depths (25–30 m) have the greatest positive contribution to diversity, while shallow (< 20 m) and deep (> 32 m) areas have negative contributions, exhibiting an inverted U-shaped response. The hard bottom percentage row is similar to YSI, with high hard bottom percentages (red points) resulting in positive SHAP values, but the effect intensity tends to saturate after reaching approximately 15%. The turbine density row exhibits an interesting bimodal characteristic: moderate densities (0.3–0.5 turbines/km²) have the highest SHAP values, while extremely low and extremely high densities have smaller contributions, suggesting an optimal density range. Environmental factors (SST, Chl-a, TSM) rank relatively lower in importance, but the annual average SST still has a discernible influence: high SST (red points) tend to have positive SHAP values, possibly related to increased biological activity in warmer years.

To further understand the non-linear response mechanisms of key factors, Fig. 9 shows the SHAP dependence plots for YSI, hard bottom percentage, and turbine density, with color coding indicating the interaction effect with the second variable.

Fig. 9.

SHAP dependence plots (key factors).

The three-panel SHAP dependency analysis in Fig. 9 reveals complex non-linear responses of benthic communities to engineering factors and their interactions with environmental gradients. The Shannon index response to YSI (Fig. 9a) exhibits an S-shaped trajectory characterized by near-zero or slightly negative SHAP values (≈−0.1 to 0) during the initial 0–2 year disturbance phase following construction, followed by rapid positive increase during the 2–4 year recovery period, and eventual saturation at SHAP values around + 0.2 beyond 4 years, with color coding indicating accelerated recovery trajectories in high turbine density areas (red points) likely attributable to intensified artificial reef effects. The hard substrate proportion response (Fig. 9b) displays positive saturation dynamics with steep SHAP value increases in the 0–10% range, decelerated growth at 10–15%, and plateau formation beyond 15%, while color gradients reveal depth-mediated regulation wherein shallow waters (yellow points) exhibit enhanced substrate effects possibly linked to favorable light conditions promoting epibenthic algal growth and subsequent food web development. Turbine density response (Fig. 9c) demonstrates an inverted U-shaped pattern with peak SHAP values occurring at 0.3–0.5 turbines/km², attributed to insufficient hard substrate provision at low densities versus potential disturbance from operational noise, vibration, and electromagnetic fields repelling sensitive species at high densities, with color coding suggesting temporal dynamics whereby this optimal density threshold becomes more pronounced in mature installations (high YSI, red points)^39,40.

SHAP dependency plots reveal three key interaction effects. Figure 9a shows turbine density modulates recovery timing: high-density areas achieve positive SHAP values at YSI = 2 years versus 3 years in low-density areas, as abundant hard substrate accelerates colonization. Figure 9b demonstrates water depth modifies hard substrate effects, with stronger diversity gains in shallow waters (slope ≈ 0.03 SHAP/1% substrate) than deep waters (slope ≈ 0.01), attributable to enhanced light-driven productivity. Figure 9c shows turbine density effects shift from linear positive (new installations) to inverted U-shaped (older wind farms), indicating evolving optimal density thresholds. This temporal transition suggests habitat benefits are offset by chronic disturbances (noise, electromagnetic fields) when density exceeds ≈ 0.5 turbines/km², with optimal thresholds changing as communities mature. The smooth curve analysis of the GAM model provides independent methodological validation for the SHAP findings described above. Figure 10 shows the GAM partial effect plots for the four key factors, with each curve representing the marginal contribution of that factor to the Shannon index while controlling for other variables.

Fig. 10.

GAM smoothing curves.

The GAM smoothing curves in Fig. 10 exhibit highly consistent response patterns with the SHAP dependency plots in Fig. 9, providing robust cross-model validation of key findings. The YSI smoothing curve (Fig. 10A, EDF = 3.2, p < 0.001) displays a clear S-shaped trajectory characterized by an initial negative effect (partial effect: −10 to 0) during the 0–2 year period, followed by rapid increase crossing the zero threshold at approximately 2.5 years and reaching maximum positive effects (≈ + 10) at 4–5 years, suggesting that wind farms require an ecological adaptation period of at least 3 years during which enhanced monitoring and disturbance minimization are warranted. Water depth shows an inverted U-shaped pattern (Fig. 10B, EDF = 2.8, p = 0.003) with peak effects in the 25–30 m range, corresponding to favorable benthic conditions including moderate sedimentation rates and stable hydrodynamics. The SST response (Fig. 10C, EDF = 2.5, p = 0.012) reveals an optimal thermal window of 14.5–15.5 °C, with both colder and warmer temperatures reducing diversity and reflecting benthic community sensitivity to thermal variation. The hard bottom ratio exhibits a positive saturation pattern (Fig. 10D, EDF = 2.1, p < 0.001) consistent with SHAP results in Fig. 9b, notably displaying an inflection point at approximately 12% beyond which marginal diversity gains diminish, providing design guidance for optimizing scour protection in offshore wind installations^15,21.

Discussion

This study systematically revealed the spatiotemporal impact patterns of offshore wind farms on soft-bottom benthic communities in the Shandong Peninsula by constructing an integrated remote sensing-field-based machine learning prediction framework. The results showed that offshore wind farm expansion triggered a response trajectory of initial decline followed by recovery in benthic communities over time. The Shannon diversity index decreased from 2.45 during the baseline period to 2.08 during the construction period, and then recovered to 2.55 after 2–4 years of operation, exceeding the baseline level⁴¹.

The temporal recovery pattern identified through SHAP analysis, with a critical threshold at approximately 2.5 years post-installation, can be explained by multiple ecological mechanisms operating at different time scales. During the construction phase (0–1 years), pile driving generates underwater noise exceeding 160 dB re 1 µPa, which disrupts benthic foraging behavior and induces temporary emigration of mobile fauna. Concurrently, sediment resuspension increases total suspended matter concentrations by 50–200% above baseline levels, reducing light penetration and primary productivity, which cascades to lower food availability for deposit feeders and suspension feeders. These acute disturbances account for the initial diversity decline observed in the SHAP dependency plot (negative SHAP values for YSI = 0–2 years).

The recovery phase (2–4 years) reflects three successional processes. First, sediment stabilization occurs as hydrodynamic patterns adjust to foundations, with scour protection inducing fine sediment deposition that creates habitat for pioneering polychaetes. Second, hard substrate provides attachment sites for sessile invertebrates (barnacles, mussels, hydroids), which increase structural complexity and create secondary habitat for cryptic crustaceans and juvenile fish, enhancing species richness via the artificial reef effect. Third, fishing restrictions create a de facto marine protected area, reducing trawling disturbance and enabling recovery of slow-growing species. These processes explain why diversity recovery steepens after 2 years (positive SHAP values for YSI > 2.5 years) and plateaus after 4–5 years (saturation for YSI > 4 years) as communities approach equilibrium.

Spatially, the artificial reef effect caused by the conversion to hard bottom near the pile foundations increased the diversity index by approximately 13% and species richness by approximately 40% compared to the control area. SHAP analysis identified the years of operation (YSI) as the most important driving factor affecting benthic diversity, with an average absolute SHAP value of 0.28, significantly higher than environmental and engineering variables such as water depth and hard bottom proportion, indicating that the cumulative effect over time plays a dominant role in community response. The XGBoost model test set R² reached 0.742, an 18.7% improvement compared to the GAM baseline model, verifying the superior performance of the gradient boosting tree algorithm in capturing nonlinear relationships and variable interaction effects, providing a methodological reference for subsequent ecological prediction studies⁴².

The benthic community response patterns discovered in this study can be mechanistically explained from the perspective of habitat reconstruction caused by the conversion from soft bottom to hard bottom. The introduction of wind turbine foundations and scour protection structures provided hard bottom attachment substrates for the originally single soft-bottom sedimentary environment, promoting the colonization of sessile organisms such as barnacles, mussels, and ascidians, which in turn led to an increase in benthic crustaceans and fish resources through food web transfer effects. The sediment redistribution process also affected the spatial pattern of benthic functional groups: the scour pits around the pile foundations provided fine-grained sediment enrichment areas for burrowing polychaetes, while the surface of the scour protection structures created a suitable hydrodynamic environment for filter-feeding bivalves. Comparing the results of this study with similar international studies reveals that the community recovery time of wind farms on the Shandong Peninsula (approximately 2.5 years) is slightly shorter than the reported values for North Sea wind farms (3–5 years)⁴³. This difference may be due to the higher primary productivity and warmer water temperatures in the study area, which accelerated the biological colonization process. The intensity of the artificial reef effect observed in this study (a 40% increase in species richness) is of the same magnitude as the monitoring results from the Horns Rev wind farm in Denmark and the North Hoyle wind farm in the UK, indicating that the ecological effects of wind farms in temperate waters follow certain general patterns. These findings align with the broader literature on offshore structure colonization dynamics, which demonstrates that the time scale of ecological responses to marine infrastructure is influenced by regional environmental factors such as temperature and productivity⁴¹.

While this study focuses on soft-bottom communities, artificial reef effects near foundations may indirectly influence adjacent soft-bottom areas through three pathways. First, enhanced fish abundance increases predation pressure on soft-bottom infauna within 200–500 m, favoring deeper-burrowing over surface-dwelling species. Second, organic matter flux from fouling communities provides supplementary food for deposit feeders, potentially explaining elevated within-farm biomass (14.8 g/m²) versus control areas (12.5 g/m²). Third, altered hydrodynamics affect larval dispersal, facilitating hard-bottom specialist recruitment into soft-bottom areas. These indirect effects likely decay within 200–300 m based on predation and particle settlement patterns in similar structures. Our soft-bottom classification (> 100 m from foundations) may capture residual artificial reef influence, explaining why within-farm diversity (H’ = 2.35) only marginally exceeds controls (H’ = 2.52). Future studies should employ finer spatial sampling (50 m intervals) within 100–500 m to distinguish indirect reef effects from direct operational impacts.

The core methodological contribution of this study lies in proposing a reproducible and transferable integrated remote sensing-field-based machine learning prediction framework. The key innovation of this framework is not in the results of a single case study, but in constructing a complete technical process from multi-source data integration, feature engineering, and model training to interpretability analysis. At the data integration level, the framework achieves spatiotemporal matching of satellite remote sensing time series data from MODIS and GOCI with field benthic survey data, overcoming the bottleneck of difficulty in obtaining environmental covariates in traditional studies. At the model interpretation level, the introduction of SHAP analysis successfully addresses the black box problem of machine learning, allowing prediction results to be directly translated into ecological insights. It should be noted that SHAP analysis reveals the strength of statistical associations between variables and responses, not strict causal relationships; further research needs to combine controlled experiments or structural equation models to further verify causal mechanisms. This framework has good cross-domain transfer potential: for other offshore wind power bases in China (such as Rudong in Jiangsu and Yangjiang in Guangdong), regional prediction models can be quickly established by simply replacing local remote sensing products and benthic survey data; for other types of marine engineering projects (such as offshore oil and gas platforms, marine ranches, and artificial islands), the framework can be adapted by adjusting the definition of engineering factors⁴⁴.

The findings of this study have direct management implications for environmental impact assessment and ecologically friendly layout planning of offshore wind farms. During the site selection and layout density decision-making stages, the prediction model established in this study can be used to predict benthic sensitive areas in candidate sea areas, prioritizing avoidance of high-diversity hotspots or core habitats of important functional groups⁴⁵. During the engineering design phase, negative impacts on key benthic functional groups can be mitigated by controlling foundation types (single pile vs. jacket structure), scour protection scope, and construction timing. This study reveals a saturation effect of hard substrate proportion (threshold approximately 12%−15%), indicating that larger scour protection areas are not necessarily better; moderate introduction of hard substrates can maximize ecological benefits. This study also provides scientific support for exploring a “wind power + marine ranching” synergistic layout model: the areas near the pile foundations have already shown significant artificial reef functions. Reasonable placement of artificial reefs and seaweed cultivation facilities within or around the wind farm could potentially achieve multiple benefits, including clean energy development, marine ecological restoration, and fisheries enhancement, based on the artificial reef community²⁶.

This study has several limitations that warrant attention. The 4-year monitoring period remains insufficient to verify the persistence and successional trajectory of artificial reef effects over decadal timescales, necessitating long-term fixed-point observations to capture complete community dynamics. Turbine-related disturbances including noise, vibration, and electromagnetic fields were indirectly represented through turbine density rather than direct field measurements, constraining mechanistic interpretations of observed patterns. The biological assessment focused exclusively on macrobenthos while omitting microbial communities and higher trophic levels such as benthic fish and cephalopods, thereby limiting evaluation of food web responses. Spatiotemporal comparisons relied on control areas and space-for-time substitution without true pre-construction baseline data from within wind farms, which could be addressed through baseline surveys during the planning phase. Although spatial block cross-validation was employed to mitigate spatial autocorrelation, systematic testing of residual spatial structure (e.g., Moran’s I) was not conducted and should be incorporated to validate model assumptions. Future research directions include applying environmental DNA metabarcoding to enhance species detection coverage and temporal resolution, constructing multi-trophic ecosystem models to quantify wind farm impacts on food web structure and energy flow, and embedding this predictive framework into marine spatial planning decision support systems to enable real-time environmental assessment for offshore wind development^11,12.

Conclusion

This study focused on four offshore wind farms in the southern waters of the Shandong Peninsula. Based on benthic organism survey data from 2015 to 2024 and multi-source remote sensing environmental products, a remote sensing-in situ integrated machine learning prediction framework was constructed to systematically assess the spatiotemporal impact of wind farm construction and operation on soft-bottom benthic communities. The XGBoost model test set achieved an R² of 0.742, significantly outperforming the GAM baseline model (R²=0.625), validating the effectiveness of machine learning methods in ecological prediction tasks. SHAP interpretability analysis identified the years in operation (YSI) as the core driving factor affecting benthic diversity. The Shannon index of the benthic community showed a recovery trajectory of initial decline followed by an increase, with approximately 2.5 years being the critical threshold for the community’s shift from negative to positive effects. The artificial reef effect caused by hard-bottom conversion in the near-pile foundation area resulted in an approximately 13% increase in the Shannon diversity index and an approximately 40% increase in species richness compared to the control area, indicating that offshore wind turbine foundations have functionally evolved into artificial reefs. The inverted U-shaped response of turbine density suggests an optimal density range of approximately 0.3–0.5 turbines/km², providing quantitative reference for wind farm layout optimization. The predictive framework proposed in this study has a good modular design and portability, and can be extended to other offshore wind farm ecological impact assessments. The research conclusions provide a scientific basis for the environmentally friendly development of offshore wind power in China.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

L.W. and Y.Z. conceived and designed the study. L.W., L.Z., and H.Z. performed the experiments. L.W., K.S., and X.G. analyzed and interpreted the data. L.W., X.G. and C.Z. drafted the manuscript. Y.Z. supervised the project, provided critical revisions, and acquired funding. All authors reviewed and approved the final manuscript and agreed to be accountable for all aspects of the work.

Funding

This research was supported by the Open Fund of Shandong Provincial Key Laboratory of Marine Ecological Environment and Disaster Prevention and Mitigation (202309).

Data availability

The processed datasets supporting the findings of this study are available from the corresponding author upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.