Ecological Insights from Above: Linking Habitat-Level NDVI Patterns with NDMI, LST and, Elevation in a Small Mediterranean City (Italy)

Abstract

Rapid human population growth accelerates biodiversity loss through urban habitat fragmentation, yet ecologically informed urban planning can mitigate these effects. This study evaluates whether and how vegetation characteristics, as captured by Earth observation data varies across forest habitats in a small Mediterranean city in Italy. The Normalized Difference Vegetation Index (NDVI), the Normalized Difference Moisture Index (NDMI), and Land Surface Temperature (LST) for the Functional Urban Area of Campobasso were derived from multitemporal Landsat 8 imagery (2020–2023) acquired during the growing season and combined with elevation data to account for topographic gradients. Different forest habitats were identified using the regional coeval Carta della Natura (Map of Nature) and were sampled by a random stratified strategy yielding more than 900,000 observations. A linear mixed-effects model was used to model NDVI as a function of NDMI, LST, elevation, and habitat type, while accounting for temporal and spatial dependencies. The model explained a large proportion of NDVI variability (marginal R² = 0.75; conditional R² = 0.85), with NDMI emerging as the strongest predictor, followed by weaker effects of LST and elevation. Habitat differences were also evident: oak-dominated forests (i.e., Quercus frainetto, Q. cerris, and Q. pubescens dominated habitats) exhibited the highest NDVI values, while coniferous plantations (i.e., Pinus nigra dominated habitat) had the lowest; forests dominated by Robinia pseudoacacia and riparian Salix alba showed intermediate vegetation greenness values. These results highlight the ecological importance of oak forests in Mediterranean urban landscapes and demonstrate the value of satellite-based monitoring for capturing habitat variability. The reproducible workflow applied here provides a scalable tool to support habitat conservation and planning in urban environments, also accounting for impending climate change scenarios.

1. Introduction

The rapid growth of human population is a major driver of biodiversity loss, primarily through the fragmentation and degradation of natural habitats [1], processes that are particularly prevalent in urbanized areas [2]. Although urban areas currently occupy only about 3% of the Earth’s terrestrial surface [3], their ecological footprint extends far beyond their physical boundaries, affecting climate systems, natural resources, biogeochemical cycles, and biodiversity patterns [4]. Accordingly, these impacts can reduce species richness, abundance, and ecological diversity. At the same time, cities also present important opportunities for biodiversity and ecosystem conservation [5]. Strategic urban interventions, such as habitat protection, ecological connectivity through corridors and steppingstones, and the integration of ecological principles into urban planning, can help mitigate biodiversity loss and enhance ecosystem services [6].

Forests within and around cities are more than just remnants of nature in an urban landscape: they regulate microclimates, sequester carbon, and sustain biodiversity, which directly supports human well-being [7,8,9]. In the context of rapid urbanization and accelerating climate change, monitoring the condition of urban and peri-urban vegetation is therefore essential. Reliable information on forest habitats is critical both for tailoring conservation strategies and adaptive management, and for enhancing the provision of ecosystem services in cities. Small and medium-sized Mediterranean cities are particularly interesting in this regard. They often host a mosaic of natural and semi-natural forest habitats surrounding the urban core, which can play a disproportionate role in sustaining regional biodiversity [10,11] but remain poorly studied compared to large metropolitan areas [12,13].

Satellite Remote Sensing (SRS) offers unique opportunities to monitor such habitats in a spatially explicit and repeatable way. By providing standardized, verifiable information across large areas and longtime spans, SRS allows to capture biodiversity indicators and ecological processes that would be impractical to assess solely through field surveys [14]. In urban ecology, multispectral satellite sensors have been the workhorse for decades, and indices derived from them have become indispensable tools [15]. Among these, the Normalized Difference Vegetation Index (NDVI) is by far the most widely used proxy for vegetation greenness and photosynthetic activity, with more than 17,000 studies published between 1985 and 2021 across an increasingly wide set of disciplines [16]. Complementary indices, such as the Normalized Difference Moisture Index (NDMI), provide information on canopy moisture [17], while Land Surface Temperature (LST) offers insights into vegetation thermal conditions, particularly relevant in urban heat island contexts [18,19]. Together, these metrics may enable a more nuanced characterization of vegetation functioning across different habitat types.

Building on these premises, our study aims to assess whether and how vegetation characteristics, as captured by NDVI, vary across distinct forest habitats within the Functional Urban Area (FUA) of Campobasso, a small Mediterranean city in Southern Italy. Specifically, we addressed the following questions: (i) to what extent do canopy moisture (NDMI), land surface temperature (LST), and elevation explain variation in vegetation greenness (NDVI) across a Mediterranean FUA? and (ii) do forest habitats differ in their NDVI values, and which habitat types exhibit higher levels of canopy greenness?

By integrating Earth observation data with ecological habitat mapping, we contribute to the monitoring of biodiversity in small Mediterranean cities. This approach supports evidence-based conservation and urban planning, while highlighting the ecological value of urban and peri-urban forests as strategic assets for climate adaptation and human well-being.

2. Materials and Methods

To explore whether and how vegetation, as captured by NDVI, varies across distinct forest habitats in relation to NDMI, LST, and elevation within the FUA of Campobasso, we followed a sequence of steps schematically reported in Figure 1: (a) selection of forest habitats and stratified sampling design; (b) elevation data sampling; (c) satellite imagery collection, processing, and sampling; (d) statistical analysis and NDVI modeling.

Figure 1. Workflow to assess variations in vegetation features across distinct forest habitats, as captured by the Normalized Difference Vegetation Index (NDVI), in relation to Normalized Difference Moisture Index (NDMI), Land Surface Temperature (LST), and elevation within the Functional Urban Area of Campobasso.

2.1. Study Area and Main Forest Habitat Types

The study was conducted within the Functional Urban Area (FUA) of Campobasso, a small Mediterranean city in the Molise region, Italy (Figure 2). Campobasso is a pilot city in the national framework of Italian Biodiversity Future Center (NBFC). Within this research center there is an interdisciplinary group (Spoke 5: Urban Biodiversity) aimed at improving the current knowledge on biodiversity in Italian cities and providing new insights to protect and improve nature in built-up areas (https://nbfc.it/en/spoke accessed on 24 December 2025).

Figure 2. Functional Urban Area of Campobasso, a small Mediterranean city in the Molise region, Italy, along with the analyzed forest habitat with EUNIS habitat name and abbreviation code in parentheses.

The concept of FUA, developed by the Organization for Economic Cooperation and Development (OECD) and the European Commission [20], provides a harmonized framework to delineate urban areas across Europe based on population density and commuting flows [21]. The FUA boundary of Campobasso was derived from the official dataset available on the OECD website [22].

The Campobasso FUA covers 1028 km² and lies within a hilly to mountainous physiographic context ranging between approximately 190 and 1084 m a.s.l., with the urban area situated at a mean elevation of 701 m a.s.l. The territory, located between the Apennine Range to the west and the Adriatic Sea to the east, belongs to the Apennine ecoregional province and has a temperate sub-mediterranean climate [23,24], characterized by mean annual temperature of 13.3 °C, coldest-month averages between −1 °C and 4 °C, two months exceeding 20 °C, and a mean annual precipitation of 806 mm [25,26]. Land use is defined by a relatively compact urban core embedded in an agricultural matrix with substantial forest cover [27].

According to the Carta della Natura (Map of Nature) program of ISPRA (Istituto Superiore per la Protezione e la Ricerca Ambientale) [28] and previous research [29], six main forest habitat types were considered, corresponding to EUNIS classification [30]: Southern Italic Quercus cerris forests (Qce), Southern Italic Quercus frainetto forest (Qfr), Italo-Sicilian Quercus pubescens forest (Qpu), Mediterranean riparian Populus forest (Sal), Deciduous self-sown forest of non-native trees (Rps), and Coniferous forest (Pni). These habitats are dominated, respectively, by Quercus cerris L., Q. frainetto Ten., Q. pubescens Willd., Salix alba L., Robinia pseudoacacia L., and Pinus nigra J.F. Arnold. A summary of forest habitat types, their corresponding codes, and dominant species is provided in Table 1. Their spatial distribution within the FUA is reported in Figure 2 (see Supplementary Table S1 for full descriptions of habitats after EUNIS classification and Table S2 for configuration metrics of analyzed forest habitats).

Table 1. Forest habitat types analyzed in the study, including their dominant tree species and the abbreviation codes (listed in alphabetical order).

Code	Forest Habitat Type	Dominant Species
Pni	Coniferous forest	Pinus nigra
Qce	Southern Italic Quercus cerris forests	Quercus cerris
Qfr	Southern Italic Quercus frainetto forest	Quercus frainetto
Qpu	Italo-Sicilian Quercus pubescens forest	Quercus pubescens
Rps	Deciduous self-sown forest of non-native trees	Robinia pseudoacacia
Sal	Mediterranean riparian Populus forest	Salix alba

Ecologically, these types range from deciduous, meso- to xerothermophilous oak systems to evergreen conifer and riparian softwood forests. Qce are dominated by Quercus cerris, a deciduous, sub-Mediterranean, mesophilous oak forming mixed broadleaf stands on well-drained soils [31]. Qfr are deciduous thermophilous oak woods often found on fertile hillsides with hornbeam or other oaks, including Q. cerris, representing mesophilous oak–hornbeam assemblages ranging from the central–southern Balkans to the southern Apennines [32]. Qpu represent sub-Mediterranean, xerothermophilous oak forests prevailing on shallow calcareous or flysch substrates; Q. pubescens shows high drought tolerance via coordinated anatomical and physiological adjustments (e.g., modulation of vessel traits, stomatal conductance and leaf morpho-anatomy) that enable acclimation to amplified drought scenarios [33,34]. Sal are softwood gallery forests along lowland river corridors dominated by Populus spp. and Salix alba. The structure and function of these forests are tightly constrained by hydrogeomorphic dynamics, with tree growth, water-use efficiency and wood anatomy responding to drought severity and river flow, while S. alba typically relies on shallow sources of soil water in seasonal settings [35,36]. Rps are secondary, even-aged stands dominated by the non-native, Nitrogen-fixing Robinia pseudoacacia; they establish on disturbed, base-rich or urban–agricultural substrates, spread vigorously by root suckering, alter nutrient cycling and understory composition, and can persist as monodominant patches unless actively managed [37,38]. Finally, Pni comprise sub-Mediterranean to montane Pinus nigra plantations. The species is heliophilous, often growing on skeletal calcareous sites. Despite its suitability for restoration on degraded terrain, P. nigra growth is sensitive to warming and hot droughts, with projections indicating declines in basal-area increment across its latitudinal range under future climates and a generally high disturbance exposure in Mediterranean landscapes [39,40].

Beyond structure, these forests underpin distinct biodiversity profiles. Oak-dominated types (Qce, Qfr, Qpu) generally host species-rich and functionally diverse herb–shrub layers tightly linked to overstory composition and stand structure, highlighting how understory diversity contributes to ecosystem functioning [41]. Riparian galleries (Sal)—when in good conservation conditions—often act as local biodiversity hotspots, with higher plant richness and a greater incidence of rare species than adjacent upland woods [42]. By contrast, Rps stands of Robinia pseudoacacia tend to depress native understory and invertebrate diversity relative to native forests [37]. Pni forests—at least in non-plantation conditions—sustain variable understory assemblages modulated by canopy structure and site conditions, with richness and composition tracking overstory attributes [43]. For further climatic description of the study area and ecological characterization of these habitats, see Varricchione et al., 2024 [29].

2.2. Stratified Sampling Design and Elevation Data

To analyze and compare forest types based on SRS ecological variables and elevation data, we implemented a stratified random sampling design to ensure the representativeness of forest habitats, with each forest type defined as a distinct stratum. A 30 × 30 m grid was overlaid across the extent of the Campobasso FUA to match both the scale of the Carta della Natura (1:25,000) and the spatial resolution of the satellite imagery (30 m). The centroid of each grid cell overlapping with forest habitat polygons was computed using QGIS v. 3.34.15.

A random selection of 10% of the centroids was performed to balance computational efficiency with ecological representativeness. This procedure yielded a total of n = 20,539 sampling points, each associated with a forest type and its corresponding polygon identifier.

Elevation was included as a key predictor. In mountain systems, elevation acts as a robust proxy for complex environmental gradients, including temperature lapse rates, solar radiation exposure, and the length of the growing season, all of which are known to strongly influence vegetation productivity [44]. Elevation data were obtained from the TINITALY DTM 1.1 [45], a seamless national digital terrain model with a 10 m spatial resolution. The DTM was resampled to 30 m using nearest-neighbor interpolation to match the spatial resolution of Landsat 8 imagery. The elevation data were then sampled at the sampling point locations.

2.3. Satellite Imagery Collection and Processing

Satellite imagery was processed in Google Earth Engine (GEE) a cloud-computing platform [46]. For this study, we selected Landsat 8 imagery Collection 2 Level 2 (L8) [47] because it uniquely combines multispectral and thermal bands, enabling the simultaneous estimation of vegetation indices as well as LST. Although the native resolution of the TIRS thermal band is 100 m [48] (resampled to 30 m), Landsat 8 remains the highest-resolution freely available source of LST at global scale with a revisit time of 16 days. For the purposes of this study, only the red (RED), near-infrared (NIR), shortwave infrared 1 (SWIR1), and thermal infrared 1 (TIR1) bands were used. Additionally, two quality assessment bands were included: the pixel quality band for cloud cover assessment (QA_PIXEL) and the radiometric saturation quality band (QA_RADSAT).

Imagery was filtered by acquisition date, retaining only scenes captured during the vegetative season (May to October) between 2020 and 2023, and clipped to the extent of the Campobasso FUA. This temporal range was selected to align with the reference date of the Carta della Natura of the region (2021) [28], thus minimizing inconsistencies between habitat classification and SRS data. Only images with less than 70% cloud cover were retained. A complete list of the 81 images resulting from the query and used for analysis, is reported in Table S3. Cloud-contaminated pixels were masked using a bitwise algorithm based on the QA_PIXEL band, and saturated pixels were further removed using the QA_RADSAT band. Following USGS guidelines for Collection 2 Level 2 products [49], the appropriate scaling factors and additive offsets (Table 2) were applied to convert raw digital numbers into surface reflectance and land surface temperature (LST), as follows:

$$Digital Number\left(DN\right)\times Scale Factor+Offset.$$

(1)

Table 2. Scaling factors and additive offsets to convert Landsat 8 imagery pixels’ digital number into surface reflectance and land surface temperature values.

Science Product Collection 2	Scale Factor	Additive Offset	Data Type	Valid Range
Surface Reflectance	0.0000275	−0.2	Uint 16	7273–43,636
Surface Temperature	0.00341802	149	Uint 16	293–65,535

As for the LST, since it is given in degrees Kelvin, it was then converted to degrees Celsius.

From the processed imagery, two multispectral indices were computed for each L8 image: (i) the Normalized Difference Vegetation Index (NDVI), calculated as the normalized difference between the reflectance in the NIR and RED spectral bands as

$$\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}={\displaystyle \frac{\left(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\mathrm{E}\mathrm{D}\right)}{\left(\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}\mathrm{E}\mathrm{D}\right)}},$$

(2)

ranging between −1 and +1, with higher values denoting increasing photosynthetic activity and biomass, and (ii) the Normalized Difference Moisture Index (NDMI), computed as the normalized difference between the reflectance in the NIR and SWIR1 bands:

$$\mathrm{N}\mathrm{D}\mathrm{M}\mathrm{I}={\displaystyle \frac{\left(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}1\right)}{\left(\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}1\right)}}.$$

(3)

NDMI is a proxy of vegetation water content, typically ranging between −1 and +1, with higher values indicating greater canopy moisture.

Sampling was performed using the GEE function sampleRegions over the stratified point set (see Section 2.2 Stratified Sampling Design), adopting the UTM zone 33N projection (EPSG:32633). Each of the 20,539 sampling points was sampled repeatedly across dates. However, because the bit-wise cloud masking algorithm removed only cloudy pixels at each date, the number of valid samples varies by image. This resulted in a total of 947,864 valid point–date observations containing NDVI, NDMI, and LST. Moreover, for each sampled pixel we retained both sampling point attributes (i.e., sampling point id, forest habitat code, forest polygon id, geometry, elevation), and image metadata required for reproducibility (i.e., acquisition date, system time, and scene cloud cover). Records lacking any of the analysis variables (NDVI, NDMI, or LST) were discarded before export. All valid observations were retained individually. We did not apply temporal aggregation (e.g., averaging or compositing), as preserving intra-seasonal variability was important for the subsequent modeling. Temporal dependence among repeated observations was handled in the statistical analysis using acquisition date as a random intercept (see Section 2.4).

The complete, reproducible GEE workflow in JavaScript, including filters, masking, scaling, index computation, sampling, and export, is publicly available at the following repository: https://code.earthengine.google.com/?accept_repo=users/chabottaro/shared_unimol (accessed on 24 December 2025).

2.4. Statistical Modeling

A linear mixed-effects model (LMM) was used to model NDVI as a function of forest habitat type, NDMI, LST, and elevation. To account for repeated observations across space and time, polygon ID (index_r) and image acquisition date (year–month) were included as random intercepts. The model was fitted using the lmer function of the lme4 R package (v 1.1-36) [50]. Prior to modeling, NDVI, NDMI, LST, and elevation values were standardized to zero mean and unit variance using z-score standardization. This enabled direct comparison of regression coefficients across predictors.

Multivariate outliers were identified using the Mahalanobis distance, computed as

$$D^2={\left(x-\mu \right)}^T{{\mathrm{S}}^{-}}^1\left(x-\mu \right),$$

(4)

where x is the vector of standardized NDVI, NDMI, and LST values for each observation, μ is the vector of mean values, and S⁻¹ is the inverse of the covariance matrix. Observations exceeding the critical threshold of the chi-square distribution (p < 0.001, df = 3) were considered outliers and excluded from further analysis. This procedure resulted in the removal of 33,795 cases from the original dataset, yielding a final sample size of n = 910,809 observations.

Multicollinearity among standardized continuous predictors (NDMI, LST, Elevation) was evaluated using Variance Inflation Factors (VIF), via vif function in car R-package [51]. All VIF values were below 1.2, indicating no multicollinearity concerns.

Residual spatial dependence was evaluated to verify that inference was not driven by unmodelled spatial structure. Pearson residuals were extracted from the final LMM and aggregated at the polygon level (index_r; one mean residual per forest polygon) to avoid pseudo-replication arising from repeated temporal observations at the same spatial location. Distances were computed in a projected metric coordinate reference system (UTM zone 33N; EPSG:32633). Residual spatial dependence was assessed using (i) a Monte-Carlo Moran’s I test based on k-nearest-neighbor weights (k = 10; 999 permutations) and (ii) a Moran’s I correlogram computed using uniformly spaced 0.5 km distance classes up to 40 km and 1000 permutations, reporting both nominal and Benjamini–Hochberg–adjusted p-values. These analyses were used as diagnostics; the main model retained the original random-effects structure to preserve consistency and computational feasibility on the full dataset.

The explanatory power of the fitted model was quantified with marginal R² (variance explained by fixed effects) and conditional R² (variance explained by both fixed and random effects). To further evaluate predictive performance, we conducted five-fold cross-validation (K = 5) with the cross_validate function of the cvms R package [52]. To preserve spatial independence between training and testing sets, folds were assigned at the polygon level (index_r). Predictive accuracy was measured in terms of root mean square error (RMSE) and mean absolute error (MAE).

To assess differences in NDVI responses between forest habitats, post hoc comparisons of estimated marginal means with 95% confidence intervals (Sidak-adjusted) were computed by using the emmeans R package [53], and compact letter displays were generated with the cld function of the multcomp package [54]. All analysis were conducted in R (v 4.4.3) [55].

3. Results

3.1. Model Performance, Residual Spatial Diagnostics and Predictor Effects

The linear mixed-effects model explained a substantial proportion of the variation in NDVI. The marginal R², representing the variance explained by fixed effects, was 0.75, whereas the conditional R², which includes both fixed and random effects, was 0.85. The inclusion of random intercepts for polygon identifier (index_r) and image acquisition date (year–month) accounted for additional spatial and temporal structure in the model. The remaining unexplained variability in NDVI (σ²) represents residual variance not captured by either the fixed or random effects (Table S4).

Residual spatial diagnostics indicated weak but detectable spatial structure. A Monte-Carlo Moran’s I test on polygon-aggregated Pearson residuals showed positive residual autocorrelation (Moran’s I = 0.124; p = 0.001). A residual correlogram revealed that autocorrelation is primarily short-ranged, with positive correlations at small distances that decay toward zero at intermediate distances (overall mean across distance classes: mean I = −0.017). Although a substantial proportion of distance classes was statistically significant, effect sizes were consistently small in magnitude, suggesting limited residual spatial dependence that is unlikely to affect the substantive interpretation of fixed effects.

Cross-validation confirmed the stability of the model, with a mean RMSE of 0.032 ± 0.001 and a mean MAE of 0.025 ± 0.001 across folds. Importantly, cross-validated R² values (R²m = 0.75, R²c = 0.85) closely matched those obtained from the fitted model, indicating strong generalizability. Among the continuous predictors, NDMI exhibited the strongest positive effect on NDVI, while LST had a smaller positive effect and elevation negative effect (all p < 0.001). Forest habitat type also contributed significantly to NDVI variation (p < 0.001). A complete summary of parameter estimates, and confidence intervals is provided in Supplementary Table S4.

3.2. Difference in NDVI Across Forest Habitats

Estimated marginal means revealed clear differences in NDVI among forest habitat types (Figure 3). The lowest values were observed in coniferous forests dominated by Pinus nigra (Pni, mean = 0.765, 95% CI [0.756–0.774]), which formed a distinct group (p < 0.001). Deciduous self-sown forests dominated by Robinia pseudoacacia (Rps, mean = 0.793, 95% CI [0.785–0.800]) and riparian forests dominated by Salix alba (Sal, mean = 0.798, 95% CI [0.788–0.807]) exhibited significantly higher NDVI values than coniferous forests but did not differ significantly from each other.

Figure 3. Estimated marginal means of NDVI for forest habitat types, adjusted for NDMI, LST, and elevation. Error bars represent 95% confidence intervals. Letters denote statistically similar groups (Sidak-adjusted pairwise comparisons).

Oak-dominated habitats showed the highest NDVI values. Quercus pubescens forests (Qpu, mean = 0.808, 95% CI [0.801–0.815]) had significantly higher values than riparian habitats and deciduous self-sown forests, while Q. cerris forests (Qce, mean = 0.817, 95% CI [0.809–0.824]) exhibited even higher values (group “d”). Southern Italic Q. frainetto forests (Qfr, mean = 0.830, 95% CI [0.807–0.852]) reached the overall highest NDVI, but their confidence intervals overlapped with those of Q. cerris, resulting in partial group overlap. Taken together, these results indicate that oak-dominated habitats consistently exhibit higher NDVI values compared with coniferous plantations, deciduous self-sown forests of non-native trees, and riparian forests.

4. Discussion

4.1. Drivers of NDVI Variability

In both Mediterranean natural forests and the urban forests of the Functional Urban Area of Campobasso, vegetation greenness (NDVI) is mainly driven by canopy moisture content (NDMI), with smaller contributions from land surface temperature (LST) and elevation.

The strong and positive effect of NDMI highlights the central role of canopy water status in determining photosynthetic activity, confirming findings from previous studies that link vegetation moisture to NDVI [56,57]. In Mediterranean forests, higher canopy water content supports photosynthetic capacity by maintaining leaf turgor and enabling sustained stomatal opening under water-limited conditions. Well-hydrated foliage allows greater CO₂ diffusion into mesophyll tissues, enhancing carbon assimilation and primary productivity, especially during seasonal drought. Conversely, when water availability declines, stomatal closure and reduced mesophyll conductance inhibit photosynthesis, constraining growth. Empirical work in young Mediterranean mixed stands (e.g., Pinus pinea, Quercus ilex, Juniperus oxycedrus) has shown a strong sensitivity of photosynthetic parameters to leaf water potential and soil moisture [58]. Likewise, measurements in dry and mesic Mediterranean forests confirm that photosynthesis and stomatal conductance decline sharply under reduced water availability [59]. It is important to note that both NDVI and NDMI are indirect proxies of vegetation function, and their relationship partially reflects mathematical dependence due to the shared use of the NIR band. Nonetheless, empirical studies have demonstrated strong correlations between NDVI and canopy structure and photosynthetic potential [60] or leaf area index in Mediterranean forests [61], as well as between NDMI and canopy water content and physiological stress responses [62]. Therefore, while caution is needed when attributing direct physiological meaning to spectral indices, the patterns observed here are consistent with established links between canopy greenness, water status, and productivity in Mediterranean ecosystems.

Our models identified NDMI as the most significant driver of NDVI variability, while LST showed a much weaker, secondary effect. This is statistically coherent, as water availability (NDMI) is widely recognized as the primary limiting factor for photosynthesis in Mediterranean ecosystems, particularly during the growing season [58,59]. The subordinate role of LST can be further interpreted from an ecological perspective. All the analyzed forest habitats, regardless of their specific composition (e.g., Quercus, Pinus, or mixed stands), provide a significant thermal buffering service [63]. This generalized cooling effect relative to the much warmer non-forested urban matrix compresses the range of LST values within our dataset. Although studies have shown that different forest types (e.g., broadleaf vs. coniferous) can produce distinct thermal regimes, these differences are often subtle or primarily dictated by structural attributes like LAI and canopy density rather than species alone [64,65]. This limited thermal variance across our sites makes LST a weaker predictor of NDVI changes compared to the overriding influence of water availability. Thus, our findings suggest that while the thermal buffering role is a comparable ecosystem service across these habitats, their differences in greenness (NDVI) are more closely linked to biodiversity and structural attributes that regulate moisture retention.

The analysis also revealed a negative, albeit weak, relationship between NDVI and elevation. This trend is ecologically consistent with the general principle that higher altitudes experience lower temperatures and shorter growing seasons, which typically constrain vegetation productivity [44].

However, the limited explanatory power of elevation in our models is almost certainly attributable to the modest range within our specific study area. This compressed gradient limits the variable’s statistical leverage. In studies encompassing more pronounced topographic variability, elevation often emerges as a primary driver of vegetation patterns, capable of explaining significant portions of NDVI variance due to strong climatic controls [66]. Our findings therefore suggest that while elevation remains a relevant ecological factor, its statistical contribution becomes secondary in landscapes where other factors, such as water availability (NDMI), exhibit much stronger variability.

4.2. Habitat-Level Differences

Habitat type significantly contributed to NDVI variability, with oak-dominated forests showing consistently higher values than riparian, alien-dominated and coniferous habitats. In particular, Quercus frainetto and Q. cerris forests displayed the highest NDVI, suggesting increased photosynthetic potential under local climatic conditions. Oak forests are characterized by high species richness, hosting several sclerophyllous taxa, which tolerate prolonged aridity and can capture atmospheric moisture even in the absence of rainfall [29,67], as well as, semi-deciduous or late-deciduous species that extend photosynthetic activity into mild autumns through delayed leaf fall [68]. These eco-morphological adaptations promote denser and persistent canopies, enhancing near-infrared reflectance and sustaining higher NDVI values. By contrast, coniferous plantations exhibited the lowest NDVI, a pattern that may reflect both their artificial origin and reduced structural and floristic complexity [69]. Riparian and R. pseudoacacia forests, while greener than coniferous stands, were less productive than oak forests, likely due to their dependence on hydrological dynamics and potential sensitivity to seasonal water availability [70]. The species richness of these habitats is also lower than oaks forest, largely because of the restrictive environmental conditions of riparian strips, where only a limited number of species are well-adapted to withstand frequent flooding [71,72], but most likely also due to the allelopathic effects exerted by the alien species R. pseudoacacia [73]. Moreover, these forests are characterized by a lighter canopy structure and by foliage that is highly vulnerable to seasonal drought stress [74]. In Mediterranean conditions, these species tend to exhibit earlier leaf senescence or a pronounced physiological decline during summer, thereby shortening the period of maximum photosynthetic activity. As a result, they consistently display lower NDVI values.

These findings support the ecological importance of oak forests in maintaining high canopy greenness and primary productivity in Mediterranean landscapes, and the consequent great interest that has received Q. cerris lately [75,76,77]. From a management perspective, canopy greenness (NDVI) is a widely used proxy for stand productivity, physiological performance, and drought resilience [78,79]. Lower NDVI often reflects reduced carbon uptake capacity, simplified stand structure, or higher vulnerability to hydrological stress [80], all of which are key attributes informing forest condition assessments in Mediterranean landscapes.

4.3. Model Performance and Robustness

The LMM framework proved effective in disentangling the effects of environmental gradients and habitat types while accounting for spatial and temporal dependence. The high marginal R² (0.75) indicates that fixed effects captured the majority of NDVI variation, while random effects associated with polygon identity and acquisition date accounted for additional variability.

Residual spatial diagnostics indicated weak but detectable spatial autocorrelation. However, its magnitude was small and primarily short-ranged. Together with the hierarchical random-effects structure, this suggests that remaining spatial dependence is limited and unlikely to affect the substantive interpretation of fixed-effect estimates.

Consistently, cross-validation confirmed the predictive robustness of the model, as out-of-sample metrics closely matched in-sample performance, indicating absence of overfitting despite the large dataset.

4.4. Implications and Perspectives

From a management perspective, our results emphasize the functional importance of oak-dominated forests as key habitats sustaining vegetation greenness. These native broadleaf systems are critical as they typically host a high biodiversity well-adapted to fluctuating Mediterranean conditions, which is foundational to their resilience despite the ever-growing evidence of oak-decline in the Mediterranean [81]. Similarly, riparian forests emerged as hotspots of productivity, a finding consistent with their high NDMI values. These habitats act as essential green corridors and microclimatic buffers in fragmented peri-urban landscapes, and their preservation is paramount for maintaining connectivity and biodiversity-related ecosystem services [82].

Conversely, the coniferous plantations in our study displayed significantly lower NDVI. This suggests a reduced capacity for primary production and related services. This finding aligns with evidence from other Mediterranean areas, where monospecific coniferous stands, often resulting from past afforestation policies, can exhibit lower functional diversity and reduced resilience to disturbances such as extreme drought compared to native mixed or broadleaf forests [83,84].

The methodological framework applied here, integrating satellite-derived vegetation metrics, open-access cloud platforms, and mixed-effects modeling, is replicable in other urban and peri-urban areas, providing a scalable tool for habitat monitoring and prioritizing conservation efforts under global change scenarios [85,86].

4.5. Limitations and Future Directions

Despite the robustness of our approach, some limitations should be acknowledged. First, NDVI and NDMI are indirect proxies of several vegetation functions and properties rather than direct physiological measurements [87]. In addition, these indices may saturate in dense canopies, potentially underestimating differences among habitats, especially during the peak growing season when greenness and moisture levels approach their upper limits [88].

While recent field campaigns in the region have produced valuable forest structural and compositional data [29], stand-level measures of leaf area index for Quercus cerris [76] and carbon stocks for oak stands [89], these datasets were not designed for the calibration or physiological validation of spectral indices. A promising direction for future research in the region would be to integrate targeted field measurements with satellite data to calibrate models that translate spectral metrics into more explicit physiological information. In this respect, Landsat 8 provides reliable and well-calibrated spectral and thermal data [48], and despite the relatively coarse spatial detail of its TIR sensor, it remains one of the most robust sources for land surface temperature estimation at landscape scales.

From a modeling perspective, although weak and short-ranged residual spatial dependence was detected, its influence on inference is mitigated by the hierarchical structure of the model and the use of spatially blocked validation. Future work could explore alternative sampling densities, incorporate additional predictors (e.g., soil type, stand age, stand density or disturbance history) to refine habitat-level assessments, or explicitly model spatial correlation structures where computationally feasible.

Finally, the modeling framework is spatially structured and could be extended to generate spatially explicit maps of predicted NDVI. Such outputs could be integrated with spatial prioritization or decision-support tools to identify areas where habitat functioning is comparatively low (e.g., coniferous plantations), or where conservation actions should be targeted, (e.g., riparian corridors). In addition, future work could couple these spatially explicit functional indicators with landscape connectivity or fragmentation analyses, enabling a joint assessment of habitat functioning and structural connectivity in urban and peri-urban forest networks. This would further enhance the operational value of the analysis for land managers and urban planners.

5. Conclusions

This study combined satellite-derived vegetation indices with mixed-effects modeling to assess the drivers of NDVI variability across forest habitats within the Campobasso Functional Urban Area. Results highlighted canopy moisture (NDMI) as the dominant predictor of NDVI, with smaller contributions from LST and elevation. Habitat type also played a key role, with oak-dominated forests sustaining the highest levels of greenness compared to riparian stands and coniferous plantations. The modeling framework demonstrated strong explanatory power and predictive accuracy, underscoring the reliability of integrating satellite multispectral data with statistical modeling in habitat monitoring.

By evidencing the functional importance of oak forests and the comparatively lower greenness of coniferous plantations, our findings provide valuable insights for prioritizing restoration efforts, guiding the management of coniferous plantations, and supporting the conservation of native oak and riparian forests in Mediterranean peri-urban landscapes. The approach is broadly transferable and offers a replicable pathway for monitoring vegetation function in other urban and peri-urban contexts under changing environmental conditions.