Linking structural forest heterogeneity and ecological processes using Sentinel-2 and FAD-based zoning Authors: Sanjana Dutt, Jakub Wojtasik, Dimitri Justeau-Allaire, Tarmo K Remmel, Mieczysław Kunz Published: 2026 ABSTRACT AbstractStructural heterogeneity strongly influences forest ecological function, yet zone-specific spectral diagnostics remain limited. This study integrated Sentinel-2 imagery (2016, 2020, 2024) with field-observed ecological attributes across Foreground Area Density (FAD)-based structural zones in the Tuchola Forest Biosphere Reserve, Poland. The aim was to evaluate whether open Sentinel-2 vegetation indices can capture ecological variation across structurally distinct forest zones using interpretable machine-learning models. Correlation and cluster analyses of 17 vegetation indices revealed substantial multicollinearity, supporting the selection of a reduced set of spectrally distinct indices for modelling. Extra Trees (ET) and LightGBM (LGBM) produced comparable predictive performance, although ET achieved equal or lower RMSE values in most zone × year combinations and was retained for interpretation. Test-set RMSE remained below one degradation class (0.70–0.96), 2.04 moisture units, 2.35 site-type categories, and 34.4 years for stand age. Permutation importance and partial dependence analyses revealed clear zone-specific spectral-ecological relationships. Rare zones exhibited stronger stress-related spectral responses and greater variability in moisture and stand-age-related patterns, whereas Core zones displayed more stable response surfaces across years. NDRE emerged as the most consistent predictor across ecological attributes, while MCARI, NDMI, and CVI provided complementary information depending on the response variable. By combining FAD-based structural stratification, cluster-driven multicollinearity reduction, and interpretable ensemble learning, this framework provides a reproducible approach for linking spectral traits to ecological gradients across fragmentation contexts and supports open-data monitoring of fragmented forest landscapes.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-59330-z.Keywords: Forest structural heterogeneity, Ecological attributes, Sentinel-2, FAD, Interpretable machine learningSubject terms: Ecology, Ecology, Environmental sciences Introduction Understanding how forest structural continuity relates to ecological functioning remains a central challenge in remote sensing and landscape ecology. Forest fragmentation, the division of continuous forest cover into smaller and more isolated patches, disrupts processes that regulate biodiversity, hydrology, and biomass productivity1,2. Fragmentation per se (independent of habitat loss) alters patch configuration and increases edge exposure, intensifying microclimatic stress through greater solar radiation, wind, and desiccation, thereby elevating fire susceptibility3,4. These pressures are particularly pronounced in temperate conifer systems such as the Tuchola Forest Biosphere Reserve (TFBR), where even-aged Scots pine (Pinus sylvestris) stands dominate (> 90% of the area). Such structural homogeneity may increase vulnerability to moisture stress, insect outbreaks, and fire relative to more diverse forest stands5,6. The resulting changes, including canopy thinning, moisture loss, and reduced connectivity, can limit the persistence of interior-dependent species that rely on large, contiguous forest tracts7,8. Although small patches may function as stepping stones for certain taxa, extensive and well-connected interiors remain important for maintaining stable populations2,7,8. Remote sensing provides a powerful means to assess these structural and ecological patterns across scales. Sentinel-2 enables repeated monitoring of vegetation condition through indices sensitive to canopy greenness and biomass (NDVI, EVI), moisture dynamics (NDMI, GVMI), and pigment-related traits (NDRE)9–11. These vegetation indices respond to a range of disturbance regimes that frequently co-vary with fragmentation. In Polish forest systems, spectral responses have been linked to meteorological variability and climate-driven shifts in growing-season phenology12,13, while satellite-derived indices have been used to monitor forest biodiversity and ecological responses to climate change at landscape scales13. Nevertheless, relatively few studies have integrated multi-temporal Sentinel-2 data, fragmentation-derived structural zoning, and field-based ecological attributes to evaluate how local canopy continuity modulates spectral responses across structurally distinct forest zones9,10. Interpretable machine learning has become an increasingly important tool for linking spectral predictors to ecological outcomes across environmental disciplines14,15. Frameworks that combine ensemble regressors with permutation-based importance measures and partial dependence analysis offer a means of maintaining model transparency without sacrificing predictive capacity, particularly when predictors are spectrally correlated, as is often the case with vegetation indices derived from overlapping band combinations. Here, we use the Foreground Area Density (FAD) metric to stratify the landscape into Core (≥ 90%), Transitional (40–60%), and Rare (≤ 10%) zones, representing a gradient in local canopy continuity and edge exposure16. FAD quantifies neighbourhood forest density within a moving spatial window and serves as a structural proxy for fragmentation effects on canopy configuration. It captures local forest amount and canopy continuity but does not directly represent other fragmentation dimensions such as patch isolation, inter-patch distance, or shape complexity. FAD-based zones are therefore interpreted as structural context classes rather than comprehensive fragmentation characterisations, providing a consistent and spatially explicit framework for comparing ecological and spectral variation across forest density gradients2,16,17. To demonstrate this, we integrate multi-temporal Sentinel-2 composites (2016, 2020, 2024) with field-based ecological attributes like degradation, moisture content, site type, and stand age (hereafter FEAs) from the Polish Forest Data Bank (BDL). These FEAs serve as ground-observed indicators of forest condition, allowing satellite-derived vegetation traits to be linked with ecological characteristics observed on the ground. Our goal is not to measure fragmentation directly but to develop a reproducible, transferable workflow that connects forest structure and function through open data and interpretable machine learning. By comparing ecological processes across FAD-derived structural zones, this study establishes a benchmark for operational, zone-based forest monitoring under increasing climatic and disturbance pressures. The specific objectives are to: Identify the vegetation indices that best capture variation in degradation, moisture content, site type, and stand age across FAD-derived zones; Identify the vegetation indices that best capture variation in degradation, moisture content, site type, and stand age across FAD-derived zones; Quantify how vegetation index–ecological relationships vary among Core, Transitional, and Rare zones across 2016, 2020, and 2024; and Quantify how vegetation index–ecological relationships vary among Core, Transitional, and Rare zones across 2016, 2020, and 2024; and Evaluate the predictive performance of interpretable ensemble models (Extra Trees and LightGBM) for field-observed ecological attributes. Evaluate the predictive performance of interpretable ensemble models (Extra Trees and LightGBM) for field-observed ecological attributes. By coupling Sentinel-2 data with FAD-based structural stratification and explainable ensemble learning, the framework provides a scalable tool for diagnosing ecological differentiation across fragmentation gradients and supports reproducible monitoring in forested landscapes. Materials and methods Study area The Tuchola Forest Biosphere Reserve (TFBR; 53° 30′ N, 17°50′ E; ~ 3,195 km2, Fig. 1) occupies nutrient-poor fluvioglacial sands on the Pomeranian outwash plain of northern Poland and contains > 900 kettle lakes and Sphagnum spp. peatlands that generate sharp hydrological and edaphic gradients18. TFBR is dominated by even-aged Pinus sylvestris plantations (> 90%), with minor Betula spp., Quercus robur, and Alnus glutinosa. Fragmentation arises primarily from silvicultural clear-cuts and salvage logging, compounded by biotic outbreaks (e.g., Panolis flammea) and extreme events (the 2012 F3 tornado and the 2017 windstorm are notable examples)18–20. As shown using Bayesian mapping, edge expansion in TFBR is strongly associated with cropland proximity, younger stands, and high wind exposure20. This configuration-driven change has progressed even where total forest area remains relatively stable, a pattern consistent with fragmentation per se effects emphasised by Fahrig and observed in other Polish landscapes8,21. These characteristics make TFBR an apt natural laboratory for examining how local forest density and neighbourhood context modulate ecological processes under fragmentation. Location and land cover of the Tucholskie Forests Biosphere Reserve (TFBR), Poland, showing land cover classes derived from CORINE Land Cover, reserve and commune boundaries, and the road network. The inset shows the location of the TFBR within Poland. The map was generated using ArcGIS 10.4 (Esri Inc., Redlands, CA, USA; https://www.esri.com). Data collection Sentinel-2 imagery We used Sentinel-2 Level-2A surface reflectance for the growing season (1 May–31 August) in 2016, 2020, and 2024. When Level-2A data were unavailable in 2016, the corresponding Level-1C scenes were converted to L2A using ESA’s Sen2Cor (atmospheric correction; also provides the Scene Classification Layer, SCL). Cloud-affected pixels were excluded using SCL classes 3 (cloud shadow), 8 (cloud-medium probability), 9 (cloud-high probability), 10 (thin cirrus), and 11 (snow/ice). Bands B2 (490 nm), B3 (560 nm), B4 (665 nm), B5 (705 nm, red-edge, 20 m native resolution), B8 (842 nm, NIR, 10 m), and B11 (1610 nm, SWIR, 20 m native resolution) were retained for their sensitivity to vegetation biochemistry and canopy structure9,11. The complete workflow, pre-processing, vegetation-index (VI) calculation, FAD-based zoning, model fitting, and interpretation, is summarised in Fig. 2. Workflow of the study. Sentinel-2 imagery and BDL ecological attributes were processed to derive 17 vegetation indices and FAD-based structural zones (Core, Transitional, Rare). After multicollinearity reduction using Pearson correlation and UPGMA clustering, representative indices were modelled using Extra Trees and LightGBM, and ecological relationships were interpreted using partial dependence analyses. Field data Forest inventory data from the BDL provided polygon-level attributes of degradation, moisture content, site type, and stand age (FEAs). Polygons correspond to operational management units delineated by the State Forests National Forest Holding. These standardised observations serve as the reference for evaluating how Sentinel-2 VIs represent on-the-ground ecological conditions (cf.16,21). Encodings for categorical FEAs are listed in Supplementary Tables S2–S4. All tabular inputs and replication code are archived in Zenodo22. Image processing and data integration All Sentinel-2 inputs were Level-2A surface reflectance (0–1). We produced seasonal median composites (May–August) in Google Earth Engine to reduce cloud and phenology noise and reprojected them to PUWG 1992 (EPSG: 2180) at 10 m. Twenty-metre bands (B5, B11) were resampled to 10 m using bilinear interpolation prior to VI calculation, while categorical layers (e.g., masks, classes) were resampled with nearest neighbour. Forested areas were restricted using a binary mask from the Dynamic World “trees” probability layer with a threshold of 0.6020,23. A total of 17 VIs (Table 1) were initially computed in Python (rasterio 1.3; NumPy 1.26) spanning four functional domains: greenness/biomass (NDVI, EVI, EVI2, GNDVI, GRNDVI, GSAVI, LAI, DVI), moisture stress (NDMI, GVMI), pigment/chlorophyll (GARI, MCARI, MTVI2, NDRE, GBNDVI), and soil/shadow correction (CVI, MSAVI)9–11,24. Mathematical formulations for all 17 indices, including the specific Sentinel-2 band combinations used, are provided in Table 1. Several indices were originally proposed for other sensors; EVI was developed for MODIS spectral ranges25, while MTVI2, MCARI, and GBNDVI were formulated using Landsat or airborne hyperspectral data and their application here relies on the spectral equivalence of Sentinel-2 bands at comparable central wavelengths9,11. Table 1Sentinel-2 vegetation indices, ecological domains, and references.Functional domainIndexFormula (Sentinel-2 bands)Native resolutionEcological focusRefs.Greenness/BiomassNDVI10 mLeaf area, productivity, biomass9,11EVI10 mCanopy biomass, atmospheric correction25,9EVI210 mTwo-band EVI variant, avoids blue band25,9GNDVI10 mChlorophyll content, canopy greenness9,10GRNDVI20 m*Red-edge greenness, early stress detection9,36GSAVI10 mSoil-adjusted greenness9,11LAI3.618× EVI-0.11810 mLeaf area index proxy derived from EVI25,9DVIB8-B410 mSimple reflectance difference, biomass9,11Moisture StressNDMI20 m*Canopy water content, drought stress36,10GVMI20 m*Vegetation moisture, atmospherically corrected36,10Pigment / ChlorophyllGARI10 mChlorophyll, atmospherically resistant9,36MCARI20 m*Chlorophyll absorption ratio9,36MTVI220 m*Canopy structure, LAI sensitivity9,36NDRE20 m*Red-edge reflectance, chlorophyll, stress9,36GBNDVI10 mPigment and nutrient status9,36Soil / Shadow CorrectionCVI10 mCanopy cover density proxy36,11MSAVI10 mSoil background correction9,11*Natively acquired at 20 m; resampled to 10 m using bilinear interpolation prior to VI calculation. Sentinel-2 vegetation indices, ecological domains, and references. *Natively acquired at 20 m; resampled to 10 m using bilinear interpolation prior to VI calculation. To address potential multicollinearity among the 17 VIs, Pearson correlation matrices were computed for each zone × year combination (9 matrices in total; Supplementary Fig. S10). Hierarchical cluster analysis using the unweighted pair group method with arithmetic mean (UPGMA) was applied to each correlation matrix to identify groups of functionally redundant indices. Based on the cluster structure, a reduced set of spectrally independent VIs was selected, one representative per cluster, retaining those with the greatest ecological interpretability and lowest inter-index redundancy. This reduced set formed the basis for all subsequent modelling and interpretation. FEAs were rasterized to 10 m using nearest neighbour to preserve categorical integrity and aligned to the composite grid. Where polygon boundaries did not coincide with pixel edges, attributes were assigned to the pixel containing the polygon centroid, and overlay checks flagged potential misalignments. Landscape stratification FAD was computed in GuidosToolbox16 using a moving 243 × 243 pixel window (2.43 km per side; 5.90 km2) on the 10 m grid to quantify neighbourhood forest density and highlight sparsely forested areas (Fig. 3). The Dynamic World “trees” mask (threshold 0.60; Sect. "Image processing and data integration") was applied before FAD calculation. From the six native FAD classes, we retained Core (FAD ≥ 90%), Transitional (40–60%), and Rare (≤ 10%) as structurally distinct zones representing varying canopy continuity and edge exposure20,26,27. FAD captures local forest amount and canopy continuity but does not represent all dimensions of fragmentation, including patch isolation, edge density, or shape complexity. FAD-based zones were interpreted as indicators of local structural context, consistent with previous applications at this site20. Spatial distribution of forest fragmentation zones (Core, Rare, and Transitional) within the Tuchola Forests Biosphere Reserve (TFBR) for 2016, 2020, and 2024. Forest cover was derived from the Dynamic World dataset, and fragmentation zones were delineated using GuidosToolbox (v. XX; https://forest.jrc.ec.europa.eu/en/activities/lpa/gtb/). The maps were generated using ArcGIS 10.4 (Esri Inc., Redlands, CA, USA; https://www.esri.com). FEAs (degradation, moisture content, site type, stand age) were linked to FAD classes for 2016, 2020, and 2024, then rasterized to the 10 m grid for pixel-wise modelling using nearest-neighbour interpolation to preserve categorical integrity. Degradation comprised eight ordinal classes (Table S1), moisture content thirteen ordinal classes (Table S2), site type fifteen nominal categories (Table S3), and stand age was represented as a continuous variable (years). Ordinal variables were integer-encoded to preserve rank, site type was one-hot encoded, and stand age remained numeric; this limitation is addressed in Sect. "Limitations and methodological considerations". Tree-based models required no additional feature scaling28. The combined dataset was > 99% complete; pixels with missing values (< 1%) were removed and no outliers were discarded to retain ecologically meaningful extremes. Modelling and interpretation Modelling framework Two ensemble regression algorithms, Extremely Randomized Trees (ET) and LightGBM (LGBM), were evaluated to predict FEAs within Core, Transitional, and Rare zones for 2016, 2020, and 202429–31,51. Hyperparameters were optimised independently for each FEA × zone × year combination using Bayesian optimisation with spatial k-fold cross-validation (k = 3) implemented in skopt.BayesSearchCV (random_state = 42). Spatial folds were defined using non-overlapping geographic partitions to minimise spatial leakage between training and validation data. Key ET hyperparameters, including the number of trees, maximum features per split, and minimum samples required for node splitting and leaf formation, were optimised for each model configuration (Supplementary Table S4). Model performance was assessed using Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute Error (MAE) on held-out test data. Comparative performance metrics for ET and LGBM across all FEA × zone × year combinations are reported in Supplementary Table S5. Based on predictive performance, ET was retained as the primary model for subsequent interpretation (Sect. "Spectral index selection and model performance"). Variable importance and partial dependence Variable importance was assessed using impurity-based importance and permutation importance (PI) derived from the ET models24,26,29,32. While impurity-based importance quantifies split-level variance reductions, PI provides a model-agnostic estimate of predictor influence by measuring the increase in prediction error following feature randomisation. Because several VIs share spectral bands and exhibit substantial correlation, PI was preferred over attribution-based approaches such as SHAP, whose feature rankings may be sensitive to predictor collinearity33. All interpretation outputs were treated as spectral–ecological associations rather than causal mechanisms. PI was repeated with 50 random permutations per feature and the resulting spread summarized. Partial Dependence Plots (PDPs) visualize how VIs influence model predictions while averaging over the distribution of other features31. Let f(⋅) denote the trained model and split the features into a set of interest ​ and its complement The partial dependence of f with respect to is is defined as where is the marginal distribution of the complementary features. In practice, we approximate this expectation using the brute-force empirical average over the observed dataset where ​are the observed values of the complementary features for sample i, and n is the number of samples used in the average. Because fragmentation effects are expected to vary across structural contexts, we emphasised two-dimensional PDPs to evaluate interactions between key VI pairs for each FEA. The selected pairs were NDRE × MCARI (degradation), NDRE × NDMI (moisture content), NDRE × CVI (site type), and CVI × MCARI (stand age), chosen based on PI rankings and representation of distinct spectral clusters (Sect. "Image processing and data integration"). PDPs were organised in 3 × 3 grids, with rows representing FAD zones (Core, Transitional, Rare) and columns representing years (2016, 2020, 2024). All PDPs were generated using the brute-force implementation in scikit-learn34, with evaluation restricted to observed feature ranges to avoid extrapolation. PDP response values correspond to the original numerical encoding of each FEA (Sect. "Landscape stratification"; Supplementary Tables S1–S3). Model accuracy assessment Predictive performance of ET and LGBM was evaluated using Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute Error (MAE), computed separately for training and test datasets. For each FEA, predicted values ŷ_i were compared with ground-truth values y_i derived from the Polish Forest Data Bank, rasterised at 10 m resolution to align with Sentinel-2 L2A imagery (Sect. "Image processing and data integration"): Here, n represents the number of sampled pixels per fragmentation zone, drawn as a 20% stratified random sample. Metrics were computed separately for each FEA, FAD zone (Core, Transitional, Rare), and year (2016, 2020, 2024). Complete performance metrics for all model configurations are provided in Supplementary Table S5. Results Spectral index selection and model performance Pearson correlation analysis revealed substantial multicollinearity among the 17 candidate VIs across all nine zone × year combinations (Fig. 4; Supplementary Fig. S10). UPGMA clustering consistently identified a large vegetation cluster comprising EVI, LAI, EVI2, MSAVI, DVI, GSAVI, MTVI2, GBNDVI, GNDVI, and GRNDVI, with correlations typically exceeding r = 0.90. NDMI and GVMI formed a distinct moisture-related pair, whereas CVI and MCARI remained comparatively independent of the main cluster, particularly in Rare forests where correlations were consistently lower than those observed among the remaining indices. EVI and LAI were perfectly collinear (r = 1.00) across all conditions, reflecting the derivation of LAI from EVI. The contrast between Core 2016 and Rare 2024 illustrates the persistence of these relationships across structurally distinct forest conditions (Fig. 4). Pearson correlation matrices with UPGMA clustering for Core 2016 (a) and Rare 2024 (b), illustrating the tight greenness/biomass cluster (mean r > 0.95) and the consistent spectral independence of CVI and MCARI across fragmentation contexts. The matrices were generated using the clustermap() function in the seaborn package (v. 0.13.2; https://seaborn.pydata.org/) in Python v. 3.10.11. Full results for all nine zone × year combinations are in Supplementary Figure S10. Based on the clustering structure, a reduced VI set was selected for modelling. NDRE, CVI, and MCARI were retained in all nine zone × year combinations, indicating stable and complementary spectral information across fragmentation contexts. NDMI was retained in eight combinations, while GARI was retained in seven, EVI was selected as the representative greenness indicator in configurations where GARI was excluded. ET and LGBM produced comparable predictive performance across all FEAs and FAD zones (Table 2; Supplementary Table S5). ET achieved equal or lower test-set RMSE values in most zone × year combinations and was therefore retained for subsequent interpretation. Differences between algorithms were consistently small, indicating that predictive performance was not strongly dependent on the choice of ensemble method. Table 2Summary of test-set predictive accuracy of the Extra Trees (ET) model, retained for interpretation, with LightGBM (LGBM) shown for comparison. For each ecological attribute, RMSE and MAE are reported as the mean across the nine zone × year combinations (Core, Transitional, Rare × 2016, 2020, 2024), with the range in parentheses.AttributeResponse scaleET RMSE, mean (range)LGBM RMSE, mean (range)ET MAE, mean (range)Degradation8 ordinal classes0.86 (0.70–0.96)0.87 (0.73–0.97)0.72 (0.53–0.88)Moisture content13 ordinal classes1.65 (1.00–2.04)1.67 (1.01–2.05)0.73 (0.27–1.05)Site type15 nominal categories1.98 (1.29–2.35)2.00 (1.29–2.38)1.50 (0.77–1.91)Stand ageContinuous (years)27.9 (24.2–34.4)28.1 (24.5–34.5)21.1 (17.1–27.2) Summary of test-set predictive accuracy of the Extra Trees (ET) model, retained for interpretation, with LightGBM (LGBM) shown for comparison. For each ecological attribute, RMSE and MAE are reported as the mean across the nine zone × year combinations (Core, Transitional, Rare × 2016, 2020, 2024), with the range in parentheses. Predictive accuracy varied among FEAs. Degradation was predicted most accurately, with ET RMSE values averaging 0.86 (0.70–0.96) on the eight-class scale and MAE values remaining below one class across all conditions. Moisture Content yielded an average RMSE of 1.65 (1.00–2.04) and MAE of 0.73 (0.27–1.05), indicating that predictions were generally within approximately one moisture class. Site Type produced intermediate errors (RMSE = 1.98, range = 1.29–2.35), although interpretation should be made cautiously because category encodings do not represent ecological distance. Stand Age exhibited the largest absolute errors, with RMSE values ranging from 24.2 to 34.4 years and MAE values ranging from 17.1 to 27.2 years. Prediction accuracy also varied across structural zones. Core forests generally exhibited lower errors than Transitional and Rare forests, particularly for Moisture Content and Site Type. This pattern likely reflects the greater spectral consistency of structurally intact forest interiors relative to edge-influenced and fragmented stands. Complete model performance metrics for all zone × year combinations are reported in Supplementary Table S5. Zone-specific ecological patterns For each FEA, a representative VI pair was selected based on permutation importance rankings within the reduced-VI framework and confirmed as representatives of distinct spectral clusters (Sect. "Image processing and data integration"; Supplementary Figs. S2, S4, S6, S8). All PDP response values correspond to the continuous numeric encoding of each FEA as described in Sect. "Landscape stratification" and Supplementary Tables S1–S3. Sensitivity comparisons using alternative VI pairs derived from the full 17-VI model are provided in Supplementary Figs. S3, S5, S7, and S9. Degradation Permutation importance (Supplementary Fig. S2) identifies NDRE as the most influential predictor of degradation across most zone × year combinations, with MCARI contributing complementary chlorophyll absorption sensitivity, particularly in Core zones. PDPs for the NDRE × MCARI pair (Fig. 5) reveal zone-differentiated spectral gradients consistent across all three years. Although the overall shape of the response surfaces remained comparable through time, the location and magnitude of peak responses varied among years, indicating temporal shifts in the spectral expression of degradation. Partial dependence plots for Degradation (NDRE × MCARI) across Core, Transitional, and Rare zones (2016, 2020, 2024). Higher response values indicate greater canopy stress. Core zones show relatively stable response surfaces concentrated at low MCARI (0.03–0.05) and intermediate NDRE (0.40–0.48), with peak values declining gradually from 2016 to 2024. Transitional zones display a more pronounced response gradient with a clear depression toward higher MCARI values, reflecting greater variability in canopy condition at edge-influenced margins. Rare zones exhibit the widest spectral spread, with peak responses shifting toward higher MCARI across all years, indicating greater variability in chlorophyll-related spectral responses under structurally open forest conditions. Moisture content Permutation importance (Supplementary Fig. S4) ranks NDMI and NDRE as the dominant predictors across most configurations, with NDRE gaining relative importance in Rare zones where canopy water signals are partially confounded by soil background reflectance. PDPs for the NDRE × NDMI pair are shown in Fig. 6. The 2016 Rare zone panel is absent because NDMI was not selected for that zone × year configuration during hierarchical clustering-based VI reduction (Sect. "Image processing and data integration"); the pair is therefore structurally incomplete for that condition. Partial dependence plots for Moisture Content (NDRE × NDMI) across Core, Transitional, and Rare zones (2016, 2020, 2024). Higher response values correspond to drier soil-canopy conditions. The 2016 Rare zone panel is absent as NDMI was not selected for that configuration (Sect. "Image processing and data integration"). Core zones show a broad region of high response values at low NDMI and intermediate NDRE, consistent with fresh to moderately moist soil conditions under dense canopy. Transitional zones display a broader, more heterogeneous response surface reflecting the moisture variability at canopy margins. Rare zones (2020 and 2024 only) show compressed response ranges with peak values concentrated at low NDMI and low NDRE, consistent with spectral signatures of dry or stressed soils under open canopy. Site type Permutation importance (Supplementary Fig. S6) identifies NDRE and CVI as the primary predictors of site-type variability, with CVI contributing canopy structural information and NDRE providing physiological status differentiation. PDPs for the NDRE × CVI pair (Fig. 7) reveal distinct habitat-level spectral separations across zones and years. Partial dependence plots for Site Type (NDRE × CVI) across Core, Transitional, and Rare zones (2016, 2020, 2024). Higher response values correspond to site types associated with more productive and mesic conditions. Core zones show peak response values consistently located at high NDRE (> 0.50) and low CVI (1.6–2.0), corresponding to moist, fertile broadleaf and mixed coniferous stands with active chlorophyll cycling. Transitional zones display intermediate response values across a wider CVI range, reflecting mixed site conditions under partial edge influence. Rare zones show response surfaces declining from high NDRE and low CVI toward low NDRE and high CVI, an inverse structural relationship relative to Core zones that reflects the canopy heterogeneity of edge-dominated fragments. Stand age Permutation importance (Supplementary Fig. S8) indicates CVI as the leading predictor of stand-age variation in Core and Transitional zones, with MCARI providing complementary pigment-based discrimination, particularly in Rare zones where structural indices alone are less reliable. PDPs for the CVI × MCARI pair (Fig. 8) show spectral relationships between canopy structure and developmental stage across the fragmentation gradient. Partial dependence plots for Stand Age (CVI × MCARI) across Core, Transitional, and Rare zones (2016, 2020, 2024). Higher response values indicate structurally mature stands. Core zones display well-organised response surfaces with peak stand age values concentrated at low CVI (1.5–1.8) and low MCARI (0.025–0.040), consistent with spectral signatures associated with mature and structurally uniform forest stands, where canopy closure reduces within-stand variability. Transitional zones show markedly lower and more variable peak responses, with a notable increase in 2024 potentially reflecting the maturation of previously disturbed stands. Rare zones present the most complex surfaces, with compressed MCARI ranges and peak values that are inconsistent across years. The overlap between regenerating edge canopy signatures and those of mature interior stands in Rare zones means the model may overestimate stand maturity under structurally open conditions; this spectral ambiguity is addressed in Sect. "Stand age and structural maturity". Discussion Ecological drivers and degradation trends across FAD zones Degradation predictions were driven primarily by NDRE and MCARI across most FAD class-year combinations (Supplementary Fig. S2), highlighting the importance of red-edge and chlorophyll sensitive spectral responses for characterising variation in forest condition. NDRE consistently ranked among the most influential predictors, while MCARI frequently emerged as a complementary indicator, particularly in Transitional and Rare zones. Their recurring importance suggests that degradation-related variation is strongly associated with changes in canopy pigment status and photosynthetic functioning, consistent with previous studies linking red-edge reflectance and chlorophyll-sensitive indices to vegetation stress and canopy condition9,10,35–37. The partial dependence surfaces further revealed distinct response patterns among fragmentation classes (Fig. 5). Core zones generally exhibited smoother and more consistent NDRE and MCARI response gradients across years, indicating relatively stable spectral–ecological relationships within structurally continuous forest interiors. In contrast, Transitional and Rare zones displayed more heterogeneous response surfaces, with stronger localised gradients and greater variation among years. These patterns suggest that degradation-related spectral signals become increasingly variable as forest structure becomes more fragmented and spatially heterogeneous. Temporal differences were also evident. Although NDRE remained an important predictor throughout the study period, the relative contribution of other indices varied among years and zones. For example, NDMI contributed strongly in several Core-zone models, whereas GARI and MCARI assumed greater importance in some Transitional and Rare configurations (Supplementary Fig. S2). Such shifts indicate that multiple spectral dimensions associated with canopy condition contribute to degradation predictions, and that their relative importance depends on both fragmentation context and year-specific environmental conditions. The observed relationships should be interpreted as model derived spectral–ecological associations rather than direct evidence of causal ecological processes. Nevertheless, the consistent importance of NDRE and MCARI across structurally contrasting forest conditions demonstrates their utility as sensitive indicators of degradation-related variation within fragmented temperate forest landscapes. Moisture dynamics and spectral predictability Moisture-content predictions were driven primarily by NDRE and NDMI, which consistently ranked among the most important predictors across years and FAD classes (Supplementary Fig. S4). Their dominant contribution is consistent with the sensitivity of red-edge and moisture-related spectral responses to variations in canopy condition and water status reported in previous studies9,36. Although the relative importance of individual indices varied among years and fragmentation classes, NDRE remained the most stable predictor throughout the analysis. The NDMI–NDRE partial dependence surfaces revealed clear differences among fragmentation classes (Fig. 6). Core zones generally exhibited smoother response gradients, whereas Transitional and Rare zones displayed more heterogeneous response surfaces, indicating greater variability in the relationship between moisture-content predictions and spectral characteristics. These differences suggest that fragmentation context influences how moisture-related spectral information is expressed across the landscape. Inter-annual variation was also evident in both the permutation-importance rankings (Supplementary Fig. S4) and the partial dependence surfaces (Fig. 6). Because the analysis was based on FAD-stratified spectral observations, these temporal differences likely reflect a combination of changing forest structure and year-specific environmental conditions. The current framework does not explicitly separate these influences, and therefore the observed year-to-year variability should be interpreted as the combined effect of fragmentation context and broader environmental variation. The absence of the 2016 Rare-zone panel in Fig. 6 reflects the clustering-based variable-selection procedure rather than missing data. In that particular zone-year combination, NDMI was not retained as the representative moisture-related index because its correlation with GVMI was substantially weaker than in the remaining configurations (Supplementary Fig. S10), resulting in a different cluster structure and the exclusion of the NDMI–NDRE variable pair from the reduced-index framework. Site type as a landscape filter Site-type variation was primarily associated with NDRE and CVI, indicating sensitivity to ecological gradients commonly linked to vegetation condition, site productivity, and moisture availability9,10,38. The selection of CVI over NDVI in the reduced VI framework reflects the clustering results, where NDVI consistently grouped within the greenness cluster and contributed limited independent information, whereas CVI occupied a distinct spectral space that captured additional canopy variability9,10. Consistent with the permutation-importance results (Supplementary Fig. S6), NDRE remained the dominant predictor across all years and FAD zones, while CVI provided complementary explanatory power. PDP surfaces showed relatively smooth response gradients in Core zones but greater variability in Transitional and Rare zones (Fig. 7), suggesting that site-type relationships became increasingly heterogeneous across structurally open forest conditions. Although site type was modelled as a continuous response to facilitate comparison among zones and years, this representation may smooth transitions among ecologically distinct categories. Integrating spectral indices with LiDAR-derived structural metrics or soil information could further improve discrimination of site conditions in heterogeneous forest mosaics39. Overall, the consistent importance of NDRE and CVI indicates that red-edge and canopy-related spectral information provide useful indicators of site-type variation at landscape scale. Stand age and structural maturity Stand-age patterns linked CVI and MCARI with canopy developmental stage, capturing the combined structural and pigment signals associated with forest succession. Core and Transitional zones produced relatively organised PDP surfaces characteristic of mature and semi-mature stands, indicating that optical indices can capture stand-age gradients under structurally stable conditions40. In contrast, Rare zones exhibited greater overlap among PDP response surfaces, suggesting reduced separation between developmental stages. Similar CVI and MCARI combinations were associated with a broad range of stand-age responses in Rare forests, indicating increased spectral ambiguity relative to Core and Transitional zones. This pattern is consistent with observations from post-disturbance and fragmented landscapes, where optical indices may not fully distinguish canopy developmental stage from canopy condition41. These findings highlight a broader limitation of optical indicators for assessing forest maturity in structurally heterogeneous landscapes. Integrating LiDAR- or GEDI-derived canopy height metrics with spectral indices would provide complementary three-dimensional structural information and may improve discrimination among developmental stages in fragmented forest mosaics39,42. From process detection to management application Zone-specific spectral analysis provides a structural framework for tailoring management interventions to fragmentation context. Core zones exhibited the most stable spectral responses across FEAs and years, consistent with the ecological stability generally associated with structurally intact forest interiors7,43. Transitional zones showed intermediate and more variable response patterns, suggesting that they may represent priority areas for management actions aimed at maintaining connectivity and supporting ecological resilience8. Rare zones displayed the greatest spectral variability and the strongest stress-related responses, indicating that restoration efforts focused on improving canopy continuity and reducing structural isolation may be particularly relevant in these areas4,20. Although the present analyses do not directly evaluate management outcomes, the observed spectral patterns are directionally consistent with broader landscape-ecological evidence showing that improved connectivity and structural continuity can enhance ecosystem resilience and ecological functioning within fragmented forests20,44. It is important to note that the present study does not derive quantitative ecological thresholds from the PDP surfaces directly; the management recommendations above reflect the qualitative direction of spectral-ecological associations rather than operationally validated intervention criteria. Future work combining PDP-based diagnostics with field-validated threshold analysis would strengthen the operational specificity of zone-based forest management. Because the workflow relies on freely available Sentinel-2 imagery, open-source analytical tools, and reproducible code22, comparable diagnostics may be transferable to other temperate and boreal forest systems, subject to local ecological validation and the prerequisites outlined in Sect. "Limitations and methodological considerations". Limitations and methodological considerations Several considerations qualify interpretation of these findings. Treating ordered categorical attributes as quasi-continuous variables was necessary to accommodate uneven category distributions across zone × year combinations but may smooth ecological transitions between adjacent classes, particularly in Rare zones. High correlations within the greenness cluster (mean r > 0.95; Supplementary Fig. S10) indicate that permutation importance reflects the contribution of a spectral functional group rather than individual vegetation indices45. Residual correlations among retained indices further suggest that PI values should be interpreted as descriptive spectral associations rather than independent contributions33. A second limitation concerns the scale mismatch between Sentinel-2's 10 m resolution and the 2.43 km FAD window (~ 5.9 km2). A pixel classified as Rare may still occur within a locally intact forest patch, meaning that neighbourhood-based FAD classes may not fully capture fine-scale structural variability. Integrating FAD with pixel-level structural information from LiDAR or GEDI would help address this uncertainty42. The absence of meteorological covariates means that inter-annual variation in vegetation-index responses likely reflects both structural and climatic influences12,46,47. Differences in spectral sensitivity among Sentinel-2, Landsat, and MODIS sensors further limit direct cross-platform comparisons of vegetation-index values13,46. Because the analyses are based on observational relationships, PDP surfaces should be interpreted as model-derived associations rather than evidence of direct causal mechanisms. Transferability beyond TFBR requires a dominant vegetation community with consistent spectral responses to environmental stressors, meaningful structural stratification using neighbourhood-based density metrics, and field inventory data of comparable quality. Under these conditions, the framework may provide a transferable template for zone-based ecological monitoring, consistent with other interpretable ensemble-learning applications in environmental systems14,15,48,49. Despite these limitations, the framework revealed consistent relationships between structural heterogeneity and ecological gradients across three years and three fragmentation contexts using entirely open data22. The combination of cluster-based vegetation-index selection, multi-model comparison, and zone-stratified PDP analysis strengthens the interpretability of spectral-ecological diagnostics in fragmented forest landscapes33,45. Conclusion This study examined how structural forest heterogeneity influences degradation, moisture content, site type, and stand age across Core, Transitional, and Rare zones in the Tuchola Forest Biosphere Reserve using Sentinel-2 vegetation indices and interpretable machine learning. Cluster-based analysis of the 17-index feature space revealed substantial multicollinearity, motivating the selection of a reduced set of spectrally distinct indices that improved the interpretability of variable-importance rankings and partial dependence analyses33,45. Across all fragmentation contexts, NDRE emerged as the most consistent predictor, while MCARI, NDMI, and CVI provided complementary information depending on the ecological attribute being modelled. Zone-stratified PDPs revealed distinct spectral-ecological relationships among Core, Transitional, and Rare forest zones, indicating that fragmentation context influences how canopy condition, moisture-related responses, site characteristics, and stand-age patterns are expressed spectrally. These relationships should be interpreted as model-derived associations rather than direct causal evidence of ecological processes33,50. ET and LGBM produced comparable predictive performance, although ET achieved equal or lower RMSE values in most zone × year combinations and was therefore retained for interpretation51. Test-set RMSE remained below one degradation class, 2.04 moisture units, 2.35 site-type categories, and 34.4 years for stand age across all conditions, demonstrating that open Sentinel-2 indices contain ecologically meaningful information across structurally distinct forest zones. By combining FAD-based structural stratification with interpretable ensemble learning, the proposed workflow provides a reproducible framework for linking spectral traits to ecological gradients across fragmentation contexts30,31,50. The explicit treatment of multicollinearity, zone-specific interpretation through PDPs, and use of entirely open data and reproducible code strengthen the transparency and transferability of the approach, where comparable field inventory data are available22. Future integration of LiDAR, GEDI-derived structural metrics, and meteorological covariates may further improve the separation of structural and climatic influences on ecological responses39,42. More broadly, the study demonstrates how structural zoning can move forest remote sensing beyond descriptive canopy mapping toward interpretable assessment of ecological variation across fragmented landscapes. Such approaches provide a foundation for scalable, open-data monitoring of forest condition and resilience under ongoing environmental change. Supplementary Information Supplementary Information 1. Acknowledgements The authors thank Dr. Melaine Aubry Kientz and Prof. Grégoire Vincent for valuable advice and guidance during the internship at the AMAP Laboratory, Montpellier. We also thank the Statistical Analysis Center at Nicolaus Copernicus University for access to their computational cluster and expert statistical feedback. We acknowledge the Bank Danych o Lasach – Lasy Państwowe (Forest Data Bank – State Forests) for providing high-quality forest inventory data collected by professional foresters. Author contributions S.D. conceived the study, curated data, performed analyses, developed methods and software, validated results, prepared visualizations, and wrote the manuscript. J.W. contributed software, statistical analysis, visualization, validation, and manuscript review and editing. D.J-A. provided supervision and manuscript review and editing. T.K.R. contributed to manuscript review, interpretation of results, and critical revisions. M.K. provided resources, supervision, map visualizations, and manuscript review and editing. Funding This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Article processing charges were funded by the YUFE CONNECTS, European Platform for Academic Cooperation of the Nicolaus Copernicus University project. Data availability Derived datasets and code to reproduce all results are archived at Zenodo (10.5281/zenodo.17241364). The archive includes: (i) tabularized vegetation and ground indices for 2016, 2020, and 2024 across FAD-classified zones (Core, Transitional, Rare); (ii) Python code for preprocessing, FAD zoning, and model replication; and (iii) documentation (README, metadata, data dictionary) and example panels. Sentinel-2 Level-2A imagery and Dynamic World “trees” probabilities are publicly available from the Copernicus Open Access Hub and Google Earth Engine. Ground attributes were obtained from the Polish Forest Data Bank (BDL) under open-data provisions. A formal data citation is included in the reference list. Declarations Competing interests The authors declare no competing interests. Ethics statement This research did not involve any experiments on humans, human data, or live vertebrates/higher invertebrates. All analyses were based on open-access remote sensing datasets and publicly available forest inventory data from the Polish State Forests. Therefore, no ethical approval or informed consent was required. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Citations Supplementary Materials Supplementary Information 1. Data Availability Statement Derived datasets and code to reproduce all results are archived at Zenodo (10.5281/zenodo.17241364). The archive includes: (i) tabularized vegetation and ground indices for 2016, 2020, and 2024 across FAD-classified zones (Core, Transitional, Rare); (ii) Python code for preprocessing, FAD zoning, and model replication; and (iii) documentation (README, metadata, data dictionary) and example panels. Sentinel-2 Level-2A imagery and Dynamic World “trees” probabilities are publicly available from the Copernicus Open Access Hub and Google Earth Engine. Ground attributes were obtained from the Polish Forest Data Bank (BDL) under open-data provisions. A formal data citation is included in the reference list.