Spectral diversity tracks initial restoration progress in the Eastern Amazon

Abstract

Monitoring the rehabilitation of mined lands and forest restoration is often limited by high costs and labor-intensive field surveys. This study evaluated the potential of spectral diversity to estimate the Restoration Index (RI), a metric that integrates structural and compositional vegetation attributes to indicate the degree of ecological recovery in areas undergoing rehabilitation or restoration. The analyses included plots representing different stages of recovery from four compensation and three rehabilitation areas in Pará, Eastern Amazon, Brazil. Different measures of spectral diversity (richness, Shannon index, Rao’s Q, and functional composition indices) were computed from Sentinel-2 satellite imagery using BiodivMapR package in R software, and correlations and linear regressions with the RI were tested. Among the spectral metrics tested, spectral Shannon diversity and spectral richness showed the strongest correlations with the restoration index. A regression model estimated that each 1% increase in the RI corresponds to a 0.0198 increase in spectral richness, with a predicted value of 2.29 at 70% restoration index, providing a useful benchmark for restoration assessment. Two areas were field-surveyed twice, and although spectral richness does not show significant increases between inventories, successional differences along the chronosequences were well covered in datasets from all inventories. Spectral diversity computed from images from different months correlates moderately, suggesting the robustness of spectral diversity metrics across seasons. Thus, our results demonstrate that spectral diversity offers a scalable, cost-effective, and transparent tool for reproducible restoration monitoring at landscape scales.

Introduction

To track environmental progress after restoration efforts, ecological monitoring programs are essential to measure the effectiveness of interventions and compliance with policies implemented on local biodiversity (Torresani et al., 2024). Historically, species diversity assessment has primarily relied on field inventories conducted by specialists. An emerging approach to evaluate ecological recovery is the use of composite indices that integrate structural and compositional vegetation attributes. One example is the Restoration Index (RI) that combines indicators such as species richness, basal area, canopy cover, and other ecological parameters derived from field surveys, following previous applications in tropical forest restoration monitoring (Gastauer et al., 2022). Although such indices effectively reflect ecological integrity and successional progress in rehabilitated sites, their application is costly, time-consuming, prone to bias, and often lacks transparency (McKenna et al., 2020; Rocchini et al., 2019).

In recent decades, remote sensing has emerged as a powerful and efficient tool for ecosystem monitoring, enabling the acquisition of consistent, periodic, and cost-effective data over large areas (Nascimento et al., 2020; Cavender-Bares et al., 2020; Rocchini et al., 2010). Recent studies have applied remote sensing to characterize environmental variables, e.g., estimating soil organic carbon and soil properties (Jamaoui et al., 2024; Vibhute et al., 2024), to monitor land cover and land use changes over time (Huang et al., 2024), and to track revegetation and ecological recovery in mining areas using satellite or drone imagery (Chen et al., 2024; Park and Choi, 2020). Advanced technologies, such LiDAR (Light Detection and Ranging) have proven effective for detecting biomass and measuring tree height, while multispectral and hyperspectral sensors show promise for identifying plant species (Zhong et al., 2022). Within this expanding technological context, multiple methodologies have been developed to estimate biodiversity remotely. Some approaches aim to directly map specific targets (Wang et al., 2018; Kacic and Kuenzer, 2022), while others focus in functional components of biodiversity, such as the leaf area index (Sheeren et al., 2016).

However, remote sensing methods also faces important challenges, including high data processing demands, potential classification errors, and the need for specialized expertise in image processing and data interpretation, which may limit its widespread application in restoration contexts. Furthermore, capturing successional progress in restoration areas remains difficult (McKenna et al., 2020), as these processes involve complex shifts in community composition and ecological functions, such as trophic interactions and soil biological activities, that operate across multiple spatial and temporal scales (Shi et al., 2024). Many of these functions are difficult to capture by usual tools of image analysis, as they depend on complex interactions between both above- and belowground ecological processes (Gastauer et al., 2022; James et al., 2020; Deng et al., 2020). In addition, the heterogeneity of rehabilitated or restored landscapes complicates remote monitoring, since these areas often feature varying vegetation densities, irregular topography, and a mixtures of native and non-native species, all of which can cause spectral confusion, particularly in megadiverse tropical forests.

In this scenario, an emerging biodiversity monitoring methodology arises from the Spectral Variation Hypothesis (SVH) (Torresani et al., 2024). The SVH assumes that spectral variation obtained through remote sensors reflects the functional characteristics of plant species and communities, such as canopy structure, leaf traits, and phenology (Wang and Gamon, 2019). In other words, spectral diversity may capture patterns related to the structural and functional properties of plant communities, which are influenced by species composition and functional groups (Ustin and Gamon, 2010). The central idea is that areas with higher spectral heterogeneity (i.e., higher variation in spectral data) correspond to greater ecological niche diversity, and, consequently, higher species diversity (Rocchini et al., 2019). This concept moves beyond pixel-level analysis, emphasizing the importance of spectral differences among pixels rather than their individual values (Féret and Asner, 2014). Although SVH has been tested in different ecosystems and sensors (Torresani et al., 2024), its effectiveness may vary depending on environmental conditions or ecosystem type (Gamon et al., 2020; Fassnacht et al., 2022). Therefore, the chosen analytical scale must align with the ecological questions being addressed. Rather than focusing solely on technical challenges such as spatial resolution, where low resolutions may mask spectral differences and very high resolutions can introduce redundancy and noise (Gamon et al., 2020; Wang et al., 2018), the focus should be on evaluating how well these metrics correlate with field-based restoration indices, such as RI. While field surveys remain indispensable for validating ecological conditions, the growing availability of high-resolution remote sensing data opens new avenues for integrating field and spectral information in restoration monitoring. The goal is not to replace field-based assessments, but to test whether spectral diversity can serve as a reliable proxy for key ecological indicators, thereby enhancing the scalability and efficiency of biodiversity monitoring.

Accordingly, this study aimed to evaluate the potential of spectral diversity as a remote proxy for restoration progress, represented by the RI, in post-mining rehabilitation areas—where vegetation cover and ecosystem functions are reestablished without necessarily recovering the original species composition—and restoration sites in former pastures, which aim to recover native biodiversity and structure within environmental compensation areas of the Eastern Amazon. Using freely available Sentinel-2 satellite imagery, we (i) calculated different spectral diversity indices and correlated them with the field-detected RI values from six distinct sites, (ii) correlated spectral diversity indices with additional field-surveyed indicators of community diversity, vegetation structure and ecological processes, and (iii) examined temporal trends in restoration and spectral diversity metrics. Finally, we (iv) checked the influence of seasonality on spectral diversity. Because both structural and compositional attributes influence the reflectance of the vegetation, we hypothesize that higher RI values will correspond to greater spectral diversity. This approach may contribute to advancing scalable, transparent, and cost-effective biodiversity monitoring through remote sensing.

Methods

Study sites

This study was conducted in the municipalities of Ourilândia do Norte, Canaã dos Carajás, and Parauapebas, in the state of Pará, Brazil (Figure 1). The climate in the region is Aw according to the Köppen, characterized by an average annual rainfall exceeding 2,000 mm and mean daily temperatures above 24 °C (Alvares et al., 2014). The rainy extends from October to April, and the dry season occurs from May to September, when the monthly rainfall falls below 60 mm. The landscape forms a mosaic of protected forest areas, cattle ranching lands, urban zones, and extensive mining operations (Souza-Filho et al., 2020).

FIGURE 1

Souza et al., 2020). Plots (P) measured 10 × 20 m. The compensation areas comprise pastures that regenerate naturally or are restored through seedling planting to offset the residual impacts of mining operations. Rehabilitating minelands encompass waste piles (N4–N5 iron mining complexes, Igarapé Bahia, and Águas Claras), tailing ponds (Igarapé Bahia), and mine pits (Arenitos) while forest restoration sites include Mozartinópolis 1, Mozartinópolis 2, Onça-Puma, and Sossego.

Mining in the region is primarily open-pit, resulting in the removal of original vegetation, surface soil layers, and inert rocks, thereby profoundly transforming areas (Souza-Filho et al., 2020). To mitigate the environmental impacts of mining activities, the major mining company operating in the region implements mine land rehabilitation programs that aim to restore soil quality, vegetation, and wildlife in open-pit mines, waste piles, and tailings ponds as much as possible to pre-mining conditions (Souza et al., 2016). In addition, large-scale forest restoration projects promoting natural regeneration or active seedling planting are carried out to offset the residual effects of mineral extraction (Gastauer et al., 2024). These initiatives are essential for promoting biodiversity and environmental health in impacted regions.

Analogous to the separation between primary succession, i.e., when land is exposed and colonized by living things for the first time, and secondary succession, i.e., when a disturbed area is re-colonized by plants and animals, we refer to mine land restoration as mine land rehabilitation, since mining substrates such as that from waste piless, tailings, and pit benches generally lack developed soil any pre-existing biological communities. Mineland rehabilitation focuses on stabilizing the physical environment of mining sites and reestablishing ecosystem structure and functioning as closely as possible to pre-mining conditions, though not necessarily recovering the original habitat type. Conversely, we use restoration to describe forest recovery efforts in former pastures, which are designed to compensate for the ecological impacts of mining by promoting the recovery of native vegetation and biodiversity. Thus, in this study, “rehabilitation” refers to mined sites undergoing ecological recovery, whereas “restoration” refers to reforestation projects in former pasturelands within environmental compensation areas.

Mineland rehabilitation and forest restoration areas

Overall, our study comprises three mineland rehabilitation (N4-N5, Águas Claras-Igarapé Bahia, and Arenitos) and three forest restoration sites (Mozartinópolis, Onça-Puma, and Sossego, Figure 1; Table 1), most of which include both non-intervened areas (negative references) and natural ecosystems that serve as ecological targets for restoration and rehabilitation efforts. Each site contains subsites restored or rehabilitated at different times, allowing the analysis of vegetation trajectories over time using a chronosequence approach. The age of interventions at the time of field-surveys ranged from 3 to 23 years, depending on the site. The Mozartinópolis site contains two distinct chronosequences. Furthermore, two chronosequences (N4-N5 and Mozartinópolis 1) were field-sampled twice for this study representing timeseries.

The Águas Claras – Igarapé Bahia chronosequence includes waste piles and tailing ponds derived from copper and gold mining. After mining activities ceased, rehabilitation began with the disposal of unprocessed waste rock (overburden) in piles and beneficiation tailings in waterproof ponds to prevent contamination, followed by geotechnical conformation, liming, the application of organic and mineral fertilizers, and the hydroseeding of native and exotic species (Ramos et al., 2024).

The N4-N5 chronosequence covers rehabilitation activities from three waste piles (Gastauer et al., 2022). For rehabilitation, hydroseeding was used to revegetate waste piles. For that, a standardized mixture of fertilizers, organic compost, and seeds of native and nonnative grasses and nitrogen-fixing legumes were applied to accelerate revegetation and promote ecological recovery of the affected areas (Guedes et al., 2021).

In the Arenito quarries, after sand extraction ceased, the pits were backfilled with waste from a nearby granite quarry and covered with a 30 cm topsoil layer originating from a nearby logging area associated with a manganese mine. After topsoil distribution, commercial grasses and legumes were seeded, and seedlings of native tree species, including pioneers, were planted at high densities (1.667 seedlings per hectare) (Gastauer et al., 2019).

Two distinct chronosequences where available for a forest restoration site called Mozartinópolis near the S11D Eliezer Batista iron mining complex in Canaã dos Carajás municipality. Large-scale forest restoration activities were implemented to compensate for the environmental impacts of the S11D complex (Gastauer et al., 2024). Areas with tree and shrub encroachment were fenced to trigger natural regeneration, while seedlings (1,667 plants/ha) were planted in clear-cut areas devoid of woody vegetation. Exotic grasses were controlled manually or chemically until canopy closure.

At the Sossego copper mine, environmental compensation relies primarily on assisted natural regeneration. Compensation areas were fenced, allowing propagule input from nearby forests, and enrichment plantings were carried out when necessary to accelerate biodiversity and ecosystem recovery (Araújo and Santos-Filho, 2008).

Environmental compensation of the Onça-Puma mine from Ourilândia do Norte comprises forests restoration actions in areas degraded by illegal mining, unordered spread of urban infrastructure and agricultural frontiers (Silva et al., 2021). Restoration included natural regeneration as well as the planting of heliophilous species and Brazil nut (Bertolletia excelsa Bonpl).

Restoration index

The RI was computed from nine field-surveyed environmental indicators related to community diversity (Shannon tree diversity, tree species richness, and similarity to reference areas), vegetation structure (canopy openness, tree density, and recruitment), and ecological processes (soil organic matter, functional diversity, and aboveground biomass). These attributes are key ecological metrics recommended for restoration and rehabilitation monitoring (Gamfeldt et al., 2013).

At least three permanent forest inventory plots (10 × 20 m) were established for each rehabilitation or restoration stage in every chronosequence (Table 1), totaling 288 plots. Plot number varied among chronosequences from 17 to 64. Within each plot, all trees with a diameter at breast height (dbh) of 3 cm or greater were measured and identified to the species level to determine tree density, tree species richness, and Shannon diversity of the trees. Recruitment was estimated as the number of trees with dbh values between 3 and 5 cm.

Composite soil samples (0–10 cm depth) were collected from five evenly distributed points per plot. Soil organic matter content was calculated by multiplying the soil organic carbon content (determined by the potassium dichromate method, K₂Cr₂O₇, by 1.724, assuming that organic matter contains 58% carbon.

Canopy openness was measured as the leaf area index (LAI), obtained from 10 measurements per plot with a LAI-2200C sensor (LI-COR Inc., Lincoln, NE, United States). Sky conditions were monitored continuously by a second sensor at a nearby vegetation-free site.

Aboveground biomass was estimated using wood density (Zanne et al., 2009), dbh, and tree height (measured by a digital hypsometer) following Chave et al. (2014) via the R package Biomass.

Community similarity to native forests was computed using the Bray‒Curtis index, which provides a robust measure of compositional similarity based on species abundance data.

Functional diversity was calculated from the functional traits wood density, dispersal and pollination syndrome, and ecological strategies (pioneer and nonpioneer species) via the R package FD (Laliberté and Legendre, 2010). To compute phylogenetic diversity (PD), we pruned the family phylogeny R20160415. new to the species found in this study (Gastauer and Meira Neto, 2016). This community tree was dated using age estimates from Magallón et al. (2015), and PD values were computed for each plot via the Picante package (Kembel et al., 2010).

To integrate these nine variables in a unique restoration index, a principal coordinate analysis (PCoA) was applied (Gastauer et al., 2020). The first PCA axis explained 99.7% of total variance, indicating that most of the environmental variation is captured along a single gradient. The RI was then computed as:where ΔReh-Ref is the Euclidean distance between the plot coordinates and the nearest reference plot and where ΔNR-Ref is the minimum distance from the non-revegetated plots and the reference plots.

This standardized index ranges from 0 to 100, where higher values indicate that a restoration site is more compositionally and structurally similar to native forests, reflecting greater ecological recovery. The method has been successfully applied in previous studies to describe the restoration status of tropical forests (Gastauer et al., 2021).

Remote sensing of spectral diversity

Sentinel-2 satellite images were acquired for all study sites, ensuring the best alignment with the corresponding forest inventory periods (Table 1). Because cloud cover, particularly high during the rainy season, limits the usability of imagery for the computation of spectral diversity, we selected images with less than 20% cloud cover. For inventories conducted during the rainy season, we used the first available image in this condition from the end of the rainy or start of the dry season. For inventories conducted in the dry season, the closest available image was used. All images were obtained as Level 2A (bottom-of-atmosphere reflectance) products from the Copernicus Open Access Hub.

To calculate the spectral diversity, we used the biodivMapR (version 1.14) package in R software (R Core Team, 2020) was developed to map the diversity of tropical forests on the basis of differences in reflectance in images of the forest canopy (Féret and Boissieu, 2020; Figure 2).

FIGURE 2

Preliminary filtering was performed to eliminate shaded and nonvegetated pixels from the images. In this step, the normalized difference vegetation index (NDVI) of 0.3 was used to generate masks that removed nonvegetated regions, ensuring that only areas with vegetation cover were included in the analysis. Furthermore, shaded and cloudy areas of the imagery were removed via the blue and NIR thresholds of 1,500 and 2000, respectively.

After filtering, principal component analysis (PCA) was applied to the satellite images, as described by Féret and Boissieu (2020), to reduce the dimensionality of the spectral data. On the basis of the eight principal components, the image pixels were grouped into spectral species communities via a clustering algorithm, which utilized a predefined number of 100 clusters to categorize the pixels according to their reflectance. Spectral diversity was subsequently calculated in windows of 50 × 50 m, following Gastauer et al. (2022), using the number of spectral communities and their abundance within the window as Spectral Richness (Richness), the Spectral Shannon index (H), the Spectral Functional Richness (FRic), Spectral Evenness (FEve), Spectral Functional Divergence (FDiv), Spectral Functional Distance (FDis), and Spectral Rao’s Quadratic Entropy (Mason et al., 2005). The 50 × 50 m window size was chosen because it balances ecological relevance and computational feasibility. It captures spatial heterogeneity at a meaningful ecological scale, while reducing intra-window noise. Additionally, this resolution has been shown to yield robust spectral diversity estimates in previous studies conducted in similar Amazonian environments (Gastauer et al., 2022). Based on plot coordinates, spectral diversity indices were retrieved for each plot.

Data analysis

To identify which measure of spectral diversity best represented ecological recovery, we calculated Pearson’s correlation coefficients and fitted simple linear regression models using the field-based Restoration Index (RI) as the response variable. In the first step, these analyses were performed individually for each chronosequence and forest inventory.

Subsequently, a simple linear regression model was fitted using spectral species richness (S) as the response variable and RI as the predictor variable. Regression parameters—including slope, intercept, t-values, degrees of freedom, p-values, and coefficients of determination (R²)—were estimated to quantify the strength and significance of the relationships. Results were visualized using scatterplots and fitted regression lines, providing a practical reference framework for evaluating restoration and rehabilitation progress.

Next, within each area, we selected the spectral variable that exhibited the strongest correlation with the RI and used it to build simple linear regression models against all nine field-surveyed ecological variables. This procedure allowed us to evaluate how well the best-performing spectral indicator captured ecological attributes measured in the field.

To assess temporal trends in spectral diversity, we conducted two complementary analyses. First, paired t-tests were used to detect differences in RI and the best-performing spectral metric between monitoring periods within chronosequences that had multiple inventories. Second, we analyzed temporal changes in spectral species richness using permanent plots from the Mozartinópolis 1 chronosequence, which includes areas implemented in 2016, 2017, 2018, and under natural regeneration. Spectral diversity was extracted from Sentinel-2 images acquired in 2018, 2019, 2020, 2021, 2022, and 2023 (image IDs: S2A_MSIL2A_20180506T134211_N0500_R124_T22MET, S2A_MSIL2A_20190615T134211_N0500_R124_T22MET, S2A_MSIL2A_20200614T134221_N0500_R124_T22MET, S2A_MSIL2A_20210629T134211_N0500_R124_T22MET, S2A_MSIL2A_20220525T134221_N0400_R124_T22MET, and S2A_MSIL2A_20230629T133841_N0509_R124_T22MET).

To evaluate seasonal variation in spectral diversity, we used monthly Sentinel-2 images from the Mozartinópolis site (January to August 2023, image IDs: S2A_MSIL2A_20230120T134212_N0500_R124_T22MET, S2A_MSIL2A_20230319T134215_N0500_R124_T22MET, S2A_MSIL2A_20230425T134217_N0500_R124_T22MET, S2A_MSIL2A_20230629T133841_N0509_R124_T22MET, S2A_MSIL2A_20230729T133845_N0509_R124_T22MET, and S2A_MSIL2A_20230818T133848_N0509_R124_T22MET).

For each image, spectral richness was computed from 10,000 random points distributed across the area. Pearson correlation matrices were then constructed and visualized using the corrplot package in R to evaluate consistency in spectral richness across the year.

Results

Spectral metrics as indicators of restoration progress

The computation of spectral diversity indices was straightforward and showed significant heterogeneity within the study areas (details are shown in Supplementary Figure S1 from electronic supplementary material). However, functional diversity was sometimes overestimated in early successional stages, with sparse ground vegetation, edge effects, invasive species, and lianas. In these regions, mixtures of exposed soil and road pixels, combined with sharp contrasts among neighboring pixels, resulted in artificially high values, as the algorithm recognized such non-filtered pixels as functionally distinct.

Correlations between the RI and diversity metrics varied substantially across the evaluated sites (Figure 3). Among the evaluated metrics, spectral Shannon diversity (H) and spectral richness (S) showed moderate to strong correlations with the RI across different areas. Notably, Onça-Puma exhibited the highest correlation with spectral Shannon diversity (r = 0.86), whereas Águas Claras/Igarapé Bahia showed a stronger correlation with spectral RaoQ (r = 0.80), and Arenitos and N4-N5 areas (both surveys) were more strongly correlated with spectral richness (S). Spectral species richness also performed well in several areas, showing strong positive correlations in Águas Claras (r = 0.73), Arenitos (r = 0.67), N4-N5 (2017: r = 0.74), and moderate positive correlations in Mozartinópolis 1 (2018: r = 0.42), but weaker or negative correlations in Onça-Puma (r = −0.54) and Mozartinópolis 2 (2023: r = −0.06). Spectral RaoQ followed a similar trend, with high correlations in Águas Claras (r = 0.8), Arenitos (r = 0.47), and N4-N5 (2019: r = 0.44), but a negative value in Onça-Puma (r = −0.45). Spectral metrics of functional diversity showed more irregular patterns: Spectral FDis was positive in Mozartinópolis 1 (2018: r = 0.52) and N4-N5 (2019: r = 0.47), but strongly negative in Onça-Puma (r = −0.55). Spectral FEve and FDiv ranged from positive to negative values across sites, indicating lower reliability as indicators of ecological recovery. Overall, spectral Shannon diversity and spectral richness proved to be more sensitive indicators of ecological restoration status in the studied areas.

FIGURE 3

The observed patterns indicate a positive association between the restoration index and spectral richness. To contextualize these findings, we modeled the relationship using a linear regression across all monitored plots. The fitted model indicated that each 1% increase in the RI was associated with an average increase of 0.198 species (t = 8.61, df = 250, p < 0.001), highlighting a significant positive relationship. The model explained a substantial portion of the variation in richness (R² = 0.27, p < 0.001) (Figure 4).

FIGURE 4

At a reference RI of 70%—a commonly applied threshold in restoration assessments—the model predicted a mean richness of 14.3 species, providing a practical criterion for evaluating restoration success. Areas with richness values near or above this threshold might be considered on track to achieve ecological restoration targets.

Correlation of selected spectral variables with field-measured vegetation attributes

For each area, we first identified the spectral variable that showed the strongest correlation with the restoration index (RI). This selected spectral variable was then used to perform correlation analyses with other vegetation variables measured in the field (Figure 5). The results showed that this spectral variable also positively correlates with key ecological indicators, particularly in areas such as Onça-Puma, Águas Claras/Igarapé Bahia, and Arenitos. High correlations were observed with metrics like H, PD, and LAI, indicating that the chosen spectral variable effectively reflects both structural and functional aspects of vegetation. For example, in Onça-Puma, strong correlations were found with H (r = 0.86), PD (r = 0.75) and a moderate correlation with LAI (r = 0.60). In Águas Claras, the highest correlations were observed with PD (r = 0.67) and species richness (S = 0.63). Conversely, variables such as tree density (Des) showed weak or even negative correlations in areas like Arenitos (r = −0.63), suggesting a lower alignment with spectral diversity patterns. These results highlight the potential of carefully selected spectral variables to reflect important ecological attributes of vegetation when guided by their initial relationship with the restoration index.

FIGURE 5

Temporal trends in restoration and spectral diversity metrics

In the analyzed areas, the RI did not show a statistically significant difference between the years 2016/17 and 2019 for the N4–N5 area (t = 1.06, df = 28, p = 0.3). However, spectral richness (S) increased marginally in N4–N5 during this period (t = 2.06, df = 28, p = 0.05). For Mozartinópolis 1, both the RI (t = 4.12, df = 30, p < 0.001) and spectral richness (t = 5.23, df = 30, p < 0.001) increased substantially from 2018 indicating marked improvements in environmental quality over time (Table 2).

We evaluated the temporal dynamics of spectral richness in the Mozartinópolis 1 area, considering different years of implementation (2016, 2017, 2018, and natural regeneration) (Figure 6). In 2018, S showed the greatest dispersion, with wide variation among plots and overlap between implementation groups, indicating high heterogeneity in the early stages of vegetation development. In the following years, a stabilization trend was observed, especially between 2020 and 2023, with more stable values and reduced differences among groups. This pattern suggests a possible convergence in community composition over time, potentially driven by successional advancement and increasing structural complexity of the vegetation.

FIGURE 6

Areas implemented in 2016 and 2017 generally exhibited higher and more stable richness values throughout the monitoring period, indicating a more consolidated recovery trajectory. In contrast, areas under natural regeneration showed more variable behavior, with occasional high values but no clear pattern of increase. Areas implemented in 2018 showed a gradual increase in richness until around 2021, followed by stabilization.

Seasonality in the Carajás region does not affect spectral richness

Monthly comparisons of spectral richness showed correlations ranging from 0.32 to 0.64 (Figure 7). Higher correlations were found between temporally closer months, such as July 27th and August 18th (r = 0.60), while lower correlations occurred between more distant months, such as March 19th and August 18th (r = 0.32). These results indicate variation in species richness correspondence throughout the year, with no consistent temporal pattern. The moderate range of correlation coefficients suggests that the method may exhibit some sensitivity to seasonal variation, although not to a predominant extent.

FIGURE 7

Discussion

Spectral diversity derived from Sentinel-2 imagery showed a clear positive association with the Restoration Index (RI) across rehabilitation and forest restoration areas surrounding the Carajás National Forest. This result corroborates the findings of Gastauer et al. (2022) for waste rock piles in the same region. The use of freely available satellite imagery thus represents a cost-effective approach for monitoring restoration progress. However, reliable estimation of ecological indicators requires technical expertise in remote sensing and data processing, a factor that should be considered when upscaling such methods. Overall, the observed patterns reinforce the potential of spectral diversity as a scalable, repeatable, and policy-relevant tool for restoration monitoring—an approach consistent with the goals of the UN Decade on Ecosystem Restoration (United Nations, 2019).

Among the analyzed indices, spectral Shannon diversity emerged as the most reliable metric, showing correlation coefficients above 0.50 in six of nine study areas. This finding aligns with previous studies (Wang and Gamon, 2019) and underscores its sensitivity to ecological gradients. Spectral richness also performed well, especially in Águas Claras/Igarapé Bahia, Arenitos, and N4–N5 (2017), although its performance varied in Onça-Puma and Mozartinópolis 2. Such variability suggests that the reliability of spectral richness may depend on local factors such as floristic composition or the type of restoration intervention (Cavender-Bares et al., 2020).

When the spectral variable most strongly correlated with the RI was compared to other ecological metrics, positive associations were found with Shannon diversity (H), phylogenetic diversity (PD), and leaf area index (LAI)—particularly in Onça-Puma, Águas Claras/Igarapé Bahia, and Arenitos. This reinforces that appropriately selected spectral variables can capture both structural and functional aspects of vegetation, providing meaningful proxies for ecological attributes relevant to restoration monitoring.

In contrast, spectral functional diversity metrics (FDis, FDiv, FEve, FRic) displayed inconsistent and sometimes negative correlations. These irregular patterns likely stem from edge effects and spectral mixing in early successional areas, where exposed soil, invasive species, or lianas introduce high spectral variability unrelated to true functional diversity (Conti et al., 2021). This artificial inflation of spectral variance reduces the reliability of functional indices. Applying stricter masking or filtering methods could help mitigate such effects, though excessive filtering risks excluding ecologically relevant early-stage signals.

Seasonal analyses revealed moderate correlations of spectral richness among months (r = 0.32–0.64), indicating some temporal variability but no pronounced seasonal effect. While spectral richness may not remain entirely stable between the wet and dry periods, the observed variation is moderate and does not undermine its utility for long-term monitoring. These findings emphasize the importance of image acquisition timing when comparing data from different months or years but confirm that the metric captures enduring structural patterns rather than transient seasonal changes.

The positive correlation between RI and spectral richness indicates that sites showing greater restoration progress also exhibit higher species richness. This pattern validates the use of spectral metrics as quantitative indicators of restoration success. Nevertheless, because our study was conducted within a single region, extrapolation to other tropical areas should be done cautiously, as local environmental factors may modulate these relationships. In chronosequences with multiple inventories (e.g., Mozartinópolis 1 and N4–N5), spectral richness increased over time even when RI remained statistically unchanged, suggesting that spectral metrics can detect subtle vegetation changes overlooked by traditional indices. Similar findings were reported by Schmidt et al. (2017). The observed stabilization of spectral values in later years likely reflects successional convergence and increasing structural complexity of plant communities.

Previous studies (Gholizadeh et al., 2022; Hauser et al., 2021) have shown that plot size and sampling scale strongly influence correlations between spectral heterogeneity and biodiversity. Although not directly tested here, this insight highlights the importance of calibrating spectral indices with field-based measurements, especially in tropical ecosystems characterized by high spatial heterogeneity.

Finally, our data demonstrate that areas with higher RI values tend to approach the spectral richness of reference forests, although this relationship is not strictly linear. For instance, Mozartinópolis 2 (RI = 78.68; S = 12.88) and Mozartinópolis 1 (2021: RI = 71.86; S = 15.03) illustrate this asymptotic trend. Unlike Gastauer et al. (2022), who examined a single site through time, our study integrated multiple areas, years, and restoration strategies. This broader perspective enhances the generalizability of results and reinforces the applicability of spectral diversity as a robust tool for monitoring ecological recovery in both compensation and mine-rehabilitation projects.

Conclusion

Our findings demonstrate that spectral diversity derived from Sentinel-2 imagery provides a sensitive and cost-effective means of assessing ecological restoration progress in mining landscapes of the Eastern Amazon. Spectral Shannon diversity exhibited the strongest correlation with the Restoration Index, followed by spectral richness. The observed increases in spectral diversity over time in chronosequences such as Mozartinópolis 1 and N4–N5 confirm its capacity to capture successional advances. Moreover, its relative stability across seasons indicates that spectral metrics primarily reflect structural vegetation attributes resilient to short-term climatic variation.

By demonstrating the utility of spectral diversity in both forest restoration and mine rehabilitation contexts, this study provides empirical evidence that remote-sensing metrics can complement and enhance traditional field monitoring. Future research should focus on integrating spectral diversity with additional ecological and environmental predictors—such as soil properties, topography, and landscape configuration—to improve the interpretation of restoration outcomes. Testing these approaches across other tropical biomes and over extended timeframes will further refine restoration targets and strengthen predictive models for ecosystem recovery.

References

• 1

  AlvaresC. A.StapeJ. L.SentelhasP. C.GonçalvesJ. L. M.SparovekG. (2014). Köppen's climate classification map for Brazil. Meteorol. Z22 (6), 711–728. 10.1127/0941-2948/2013/0507

  • CrossRef
  • Google Scholar
  • View reference in article

• 2

  AraújoF. E. S.Santos-FilhoM. P. (2008). Monografia Especialização em Gestão de Recursos Hídricos e Gestão ambiental– Universidade Federal do Pará, Canãa dos Carajás, 2008.

  • Google Scholar
  • View reference in article

• 3

  Cavender-BaresJ.GamonJ. A.TownsendP. A. (2020). Remote sensing of plant biodiversity. Springer Nature.

  • Google Scholar
  • View reference in article

• 4

  ChaveJ.Réjou-MéchainM.BúrquezA.ChidumayoE.ColganM. S.DelittiW. B. C.et al (2014). Improved allometric models to estimate the aboveground biomass of tropical trees. Glob. Chang. Biol.20 (10), 3177–3190. 10.1111/gcb.12629

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 5

  ChenZ.HouH.ZhangS.CampbellT.YangY.TuM.et al (2024). Using unmanned aerial vehicle multispectral data for monitoring the outcomes of ecological restoration in mining areas. Land Degrad. Dev.35 (4), 1599–1613. 10.1002/ldr.5010

  • CrossRef
  • Google Scholar
  • View reference in article

• 6

  ContiL.MalavasiM.GallandT.KomárekJ.LagnerO.CarmonaC. P.et al (2021). The relationship between species and spectral diversity in grassland communities is mediated by their vertical complexity. Appl. Veg. Sci.24 (3), avsc.12600. 10.1111/avsc.12600

  • CrossRef
  • Google Scholar
  • View reference in article

• 7

  DengJ.BaiX.ZhouY.ZhuW.YinY. (2020). Variations of soil microbial communities accompanied by different vegetation restoration in an open-cut iron mining area. Sci. Total Environ.704, 135243. 10.1016/j.scitotenv.2019.135243

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 8

  FassnachtF. E.MüllerováJ.ContiL.MalavasiM.SchmidtleinS. (2022). About the link between biodiversity and spectral variation. Appl. Veg. Sci.25 (1), e12643. 10.1111/avsc.12643

  • CrossRef
  • Google Scholar
  • View reference in article

• 9

  FéretJ.-B.AsnerG. P. (2014). Mapping tropical forest canopy diversity using high-fidelity imaging spectroscopy. Ecol. Appl.24 (6), 1289–1296. 10.1890/13-1824.1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 10

  FéretJ.-B.BoissieuF. (2020). biodivMapR: an R package for α- and β-diversity mapping using remotely sensed images. Kylie Scales Methods Ecol. Evolution/British Ecol. Soc.11 (1), 64–70. 10.1111/2041-210x.13310

  • CrossRef
  • Google Scholar
  • View reference in article

• 11

  GamfeldtL.SnallT.BagchiR.JonssonM.GustafssonL.KjenllanderP.et al (2013). Higher levels of multiple ecosystem services are found in forests with more tree species. Nat. Commun.4, 1340. 10.1038/ncomms2328

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 12

  GamonJ. A.WangR.GholizadehH.ZuttaB.TownsendP. A.Cavender-BaresJ. (2020). “Consideration of scale in remote sensing of biodiversity,” in Remote sensing of plant biodiversity. Editors Cavender-BaresJ.GamonJ. A.TownsendP. A. (Springer Nature), 425–447.

  • Google Scholar
  • View reference in article

• 13

  GastauerM.Meira NetoJ. A. A. (2016). An enhanced calibration of a recently released megatree for the analysis of phylogenetic diversity. Braz. J. Biol.76, 619–628. 10.1111/j.0030-1299.2005.13886.x

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 14

  GastauerM.Souza FilhoP. W. M.RamosS. J.CaldeiraC. F.SilvaJ. R.SiqueiraJ. O.et al (2019). Mine land rehabilitation in Brazil: goals and techniques in the context of legal requirements. Ambio48 (1), 74–88. 10.1007/s13280-018-1053-8

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 15

  GastauerM.CaldeiraC. F.RamosS. J.TrevelinL. C.JafféR.OliveiraG.et al (2020). Integrating environmental variables by multivariate ordination enables the reliable estimation of mineland rehabilitation status. J. Environ. Manag.256, 109894. 10.1016/j.jenvman.2019.109894

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 16

  GastauerM.SarmentoP. S. M.CaldeiraC. F.CastroA. F.RamosS. J.TrevelinL. C.et al (2021). Shannon tree diversity is a surrogate for mineland rehabilitation status. Ecol. Indic.130, 108100. 10.1016/j.ecolind.2021.108100

  • CrossRef
  • Google Scholar

• 17

  GastauerM.Nascimento JúniorW. R.CaldeiraC. F.RamosS. J.Souza FilhoP. W. M.FéretJ.-B. (2022). Spectral diversity allows remote detection of the rehabilitation status in an Amazonian iron mining complex. Int. J. Appl. Earth Observation Geoinformation106, 102653. 10.1016/j.jag.2021.102653

  • CrossRef
  • Google Scholar
  • View reference in article

• 18

  GastauerM.PinheiroT.CaldeiraC. F.RamosS. J.CoelhoR. R.FonsecaD. S.et al (2024). Large-scale forest restoration generates comprehensive biodiversity gains in an Amazonian mining site. J. Clean. Prod.1, 140959.

  • Google Scholar
  • View reference in article

• 19

  GholizadehH.DixonA. P.PanK. H.McMillanN. A.HamiltonR. G.FuhlendorfS. D.et al (2022). Using airborne and DESIS imaging spectroscopy to map plant diversity across the largest contiguous tract of tallgrass prairie on earth. Remote Sens. Environ.281, 113254. 10.1016/j.rse.2022.113254

  • CrossRef
  • Google Scholar
  • View reference in article

• 20

  GuedesR. S.RamosS. J.GastauerM.Caldeira JúniorC. F.MartinsG. C.Nascimento JúniorW. R.et al (2021). Challenges and potential approaches for soil recovery in iron open pit mines and waste piles. Environ. Earth Sci.80 (18), 640. 10.1007/s12665-021-09926-7

  • CrossRef
  • Google Scholar
  • View reference in article

• 21

  HauserL. T.FéretJ.-B.BinhN. A.van Der WindtN.SilÂ. F.TimmermansJ.et al (2021). Towards scalable estimation of plant functional diversity from Sentinel-2: in-situ validation in a heterogeneous (semi-) natural landscape. Remote Sens. Environ.262, 112505. 10.1016/j.rse.2021.112505

  • CrossRef
  • Google Scholar
  • View reference in article

• 22

  HuangY.LiX.LiuD.DuanB.HuangX.ChenS. (2024). Evaluation of vegetation restoration effectiveness along the Yangtze River shoreline and its response to land use changes. Sci. Rep.14, 7611. 10.1038/s41598-024-58188-3

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 23

  JamaouiE. I.SánchezM.SirventC.ManaA.LópezS. (2024). Machine learning-driven modeling for soil organic carbon estimation from multispectral drone imaging: a case study in Corvera, Murcia (Spain). Model. Earth Syst. Environ.10, 1–22. 10.1007/s40808-024-01963-y

  • CrossRef
  • Google Scholar
  • View reference in article

• 24

  JamesW. R.LesserJ. S.LitvinS. Y.NelsonJ. A. (2020). Assessment of food web recovery following restoration using resource niche metrics. Sci. Total Environ.711, 134801. 10.1016/j.scitotenv.2019.134801

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 25

  KacicP.KuenzerC. (2022). Forest biodiversity monitoring based on remotely sensed spectral diversity—a review. Remote Sens.14 (21), 5363. 10.3390/rs14215363

  • CrossRef
  • Google Scholar
  • View reference in article

• 26

  KembelS. W.CowanP. D.HelmusM. R.CornwellW. K.MorlonH.AckerlyD. D.et al (2010). Picante: r tools for integrating phylogenies and ecology. Bioinformatics26 (11), 1463–1464. 10.1093/bioinformatics/btq166

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 27

  LalibertéE.LegendreP. (2010). A distance-based framework for measuring functional diversity from multiple traits. Ecology91 (1), 299–305. 10.1890/08-2244.1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 28

  MagallónS.Gómez-AcevedoS.Sánchez-ReyesL. L.Hernández-HernándezT. (2015). A metacalibrated time-tree documents the early rise of flowering plant phylogenetic diversity. New Phytol.207 (2), 437–453. 10.1111/nph.13264

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 29

  MasonN. W. H.MouillotD.LeeW. G.WilsonJ. B. (2005). Functional richness, functional evenness and functional divergence: the primary components of functional diversity. Oikos111, 112–118. 10.1111/j.0030-1299.2005.13886.x

  • CrossRef
  • Google Scholar
  • View reference in article

• 30

  McKennaP. B.LechnerA. M.PhinnS.ErskineP. D. (2020). Remote sensing of mine site rehabilitation for ecological outcomes: a global systematic review. Remote Sens.12 (21), 3535. 10.3390/rs12213535

  • CrossRef
  • Google Scholar
  • View reference in article

• 31

  NascimentoF. S.GastauerM.Souza-FilhoP. W. M.NascimentoW. R.Jr.SantosD. C.CostaM. F. (2020). Land cover changes in open-cast mining complexes based on high-resolution remote sensing data. Remote Sens.12, 611. 10.3390/rs12040611

  • CrossRef
  • Google Scholar
  • View reference in article

• 32

  ParkS.ChoiY. (2020). Applications of unmanned aerial vehicles in mining from exploration to reclamation: a review. Minerals10 (8), 663. 10.3390/min10080663

  • CrossRef
  • Google Scholar
  • View reference in article

• 33

  R Core Team (2020). R: a language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing. Available online at: https://www.r-project.org.

  • Google Scholar
  • View reference in article

• 34

  RamosS. J.OliveiraS. S. H. S.RibeiroP. G.MartinsG. C.Silva JuniorE. C.GastauerM.et al (2024). Combining approaches for environmental assessment of rehabilitated gold-mining areas in the Eastern amazon. Land Degrad. and Dev.35 (6), 2313–2325. 10.1002/ldr.5062

  • CrossRef
  • Google Scholar
  • View reference in article

• 35

  RocchiniD.BalkenholN.CarterG. A.FoodyG. M.GillespieT. W.HeK. S.et al (2010). Remotely sensed spectral heterogeneity as a proxy of species diversity: recent advances and open challenges. Ecol. Inf.5 (5), 318–329. 10.1016/j.ecoinf.2010.06.001

  • CrossRef
  • Google Scholar
  • View reference in article

• 36

  RocchiniD.MarcantonioM.Da ReD.ChiriciG.GalluzziM.LenoirJ.et al (2019). Time-lapsing biodiversity: an open source method for measuring diversity changes by remote sensing. Remote Sens. Environ.231, 111192. 10.1016/j.rse.2019.05.011

  • CrossRef
  • Google Scholar

• 37

  SchmidtJ.FassnachtF. E.LauschA.SchmidtleinS. (2017). Assessing the functional signature of heathland landscapes via hyperspectral remote sensing. Ecol. Indic.73, 505–512. 10.1016/j.ecolind.2016.10.017

  • CrossRef
  • Google Scholar
  • View reference in article

• 38

  SheerenD.FauvelM.JosipovićV.LopesM.PlanqueC.WillmJ.et al (2016). Tree species classification in temperate forests using Formosat-2 satellite image time series. Remote Sens.8. 734. 10.3390/rs8090734

  • CrossRef
  • Google Scholar
  • View reference in article

• 39

  ShiX.EisenhauerN.PeñuelasJ.FuY.WangJ.ChenY.et al (2024). Trophic interactions in soil micro-food webs drive ecosystem multifunctionality along tree species richness. Glob. Chang. Biol.30 (3), e17234. 10.1111/gcb.17234

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 40

  SilvaE. F. da.MeirelesR. R.HianesA. da C.PereiraF. M.SilvaD. F. da.CostaG. M. M.et al (2021). Análise dos impactos socioeconômicos e socioambientais da mineração no município de Ourilândia do Norte (PA). Rev. Ibero-Americana Ciências Ambient.12 (10) 572–581. 10.6008/CBPC2179-6858.2021.010.0045

  • CrossRef
  • Google Scholar
  • View reference in article

• 41

  SouzaF. M.SilvaJ. M. C.RodriguesR. R. (2016). Restoration of mining areas in the amazon: a study of vegetation dynamics in Carajás, Pará, Brazil. J. Environ. Manag.182, 682–689. 10.1016/j.jenvman.2016.08.016

  • CrossRef
  • Google Scholar
  • View reference in article

• 42

  SouzaC. M.Jr.ShimboJ.RosaM. R.ParenteL. L.A. AlencarA.RudorffB. F. T.et al (2020). Reconstructing three decades of land use and land cover changes in Brazilian biomes with landsat archive and Earth engine. Remote Sens.12, 2735. 10.3390/rs12172735

  • CrossRef
  • Google Scholar
  • View reference in article

• 43

  Souza-FilhoP. W. M.CavalcanteR. B. L.NascimentoW. R.NunesS.GastauerM.SantosD. C.et al (2020). The sustainability index of the physical mining environment in protected areas, Eastern Amazon. Environ. Sustain. Indic.8, 100074. 10.1016/j.indic.2020.100074

  • CrossRef
  • Google Scholar
  • View reference in article

• 44

  TorresaniM.RossiC.PerroneM.HauserL. T.FéretJ.-B.MoudrýV.et al (2024). Reviewing the spectral variation hypothesis: twenty years in the tumultuous sea of biodiversity estimation by remote sensing. Ecol. Inf.82, 102702. 10.1016/j.ecoinf.2024.102702

  • CrossRef
  • Google Scholar

• 45

  United Nations (2019). Strategy: UN decade on ecosystem restoration 2021–2030. Food and agriculture organization of the united nations; Nairobi (Rome: United Nations Environment Programme). Available online at: https://openknowledge.fao.org/server/api/core/bitstreams/1f6a3012-9580-4ec4-9a34-5694076491be/content.

  • Google Scholar
  • View reference in article

• 46

  UstinS. L.GamonJ. A. (2010). Remote sensing of plant functional types. New Phytol.186 (4), 795–816. 10.1111/j.1469-8137.2010.03284.x

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 47

  VibhuteA.KaleK.GaikwadS. (2024). Machine learning-enabled soil classification for precision agriculture: a study on spectral analysis and soil property determination. Appl. Geomatics16, 181–190. 10.1007/s12518-023-00546-3

  • CrossRef
  • Google Scholar
  • View reference in article

• 48

  WangR.GamonJ. A. (2019). Remote sensing of terrestrial plant biodiversity. Remote Sens. Environ.231, 111218. 10.1016/j.rse.2019.111218

  • CrossRef
  • Google Scholar
  • View reference in article

• 49

  WangR.GamonJ. A.Cavender‐BaresJ.TownsendP. A.ZygielbaumA. I. (2018). The spatial sensitivity of the spectral diversity-biodiversity relationship: an experimental test in a prairie grassland. Ecol. Appl.28 (2), 541–556. 10.1002/eap.1669

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 50

  ZanneA. E.Lopez-GonzalezG.CoomesD. A.IlicJ.JansenS.LewisS. L.et al (2009). Data from: toward a worldwide wood economics spectrum. Dryad. 10.5061/dryad.234

  • CrossRef
  • Google Scholar
  • View reference in article

• 51

  ZhongH.LinW.LiuH.MaN.LiuK.CaoR.et al (2022). Identification of tree species based on the fusion of UAV hyperspectral image and LiDAR data in a coniferous and broad-leaved mixed forest in Northeast China. Front. Plant Sci.13, 964769. 10.3389/fpls.2022.964769

  • Pubmed Abstract
  • CrossRef
  • Google Scholar