Changes in satellite‐derived spectral variables and their linkages with vegetation changes after peatland restoration

This is a self-archived version of an original article. This version may differ from the original in pagination and typographic details. Author(s): Title: Year: Version: Copyright: Rights: Rights url: Please cite the original version: CC BY 4.0 https://creativecommons.org/licenses/by/4.0/ Changes in satellite‐derived spectral variables and their linkages with vegetation changes after peatland restoration © 2024 The Author(s). Restoration Ecology published by Wiley Periodicals LLC on behalf of Society for Ecological Restoration Published version Räsänen, Aleksi; Jantunen, Aapo; Isoaho, Aleksi; Ikkala, Lauri; Rana, Parvez; Marttila, Hannu; Elo, Merja Räsänen, A., Jantunen, A., Isoaho, A., Ikkala, L., Rana, P., Marttila, H., & Elo, M. (2024). Changes in satellite‐derived spectral variables and their linkages with vegetation changes after peatland restoration. Restoration Ecology, Early View. https://doi.org/10.1111/rec.14338 2024 RESEARCH ARTICLE Changes in satellite-derived spectral variables and their linkages with vegetation changes after peatland restoration Aleksi Räsänen1,2,3 , Aapo Jantunen4, Aleksi Isoaho1,5, Lauri Ikkala5,6, Parvez Rana1, Hannu Marttila5, Merja Elo4,7,8 Remote sensing (RS) can be an efficient monitoring method to assess the ecological impacts of restoration. Yet, it has been used relatively little to monitor post-restoration changes in boreal forestry-drained peatlands, and particularly the linkages between changes in RS and plant species remain vague. To understand this gap, we utilize data from the Finnish peatland restoration monitoring network spanning 150 sites and a 10-year post-restoration monitoring period. We employ Bayesian joint species distribution models (Hierarchical Modeling of Species Communities) to study (1) the changes in optical Sentinel-2 and Landsat satellite spectral signatures, (2) whether the RS variables improve predictions of vascular plant and moss species and functional type occurrence and cover, and (3) what kinds of associations exist between RS variables and plant species or functional types. Our results show that peatland restoration increases the reflectance of red and near-infrared (NIR) bands in sparsely treed pine mire forests and open mires but not in densely treed spruce mire forests. Impacts on other tested RS variables consisting of moisture and greenness indices are less clear. Additionally, RS variables increase speciesor functional type-specific predictive power only modestly, and there are few clear links between the changes in RS variables and species or functional-type occurrence and cover. We suggest that red and NIR reflectance can be used as satellite-based indicators for peatland restoration success and further studies are required to develop usable methods for detecting species-specific changes with RS. Key words: bryophytes, joint species distribution models, plant functional types, remote sensing, satellite imagery, vascular plants Implications for Practice •Satellite remote sensing is suitable for monitoring postrestoration changes in ground vegetation, land cover, and wetness in peatlands with few or no trees, as trees hamper visibility to the ground. •High spatial and temporal resolution remote sensing complements field work, and it can be used to scale fieldbased knowledge to larger area extents or to other sites. •It should be further tested whether changes in reflectance can be used in operational peatland restoration monitoring and to which kind of changes the reflectance changes are attributable. •There is a need for cross-fertilization of researchers’and practitioners’knowledge to develop restoration outcome indicators that are ecologically meaningful, operationally implementable, and detectable with remote sensing. Introduction Many of the peatlands in northern latitudes have been drained to facilitate forest growth and timber production for the forestry industry (Vasander et al. 2003). However, this drainage has caused widespread and harmful environmental impacts, including loss of peatland species and habitats, greenhouse gas emissions, and deterioration of water quality in recipient water bodies (Chapman et al. 2003;Ur  ak et al. 2017; Nieminen et al. 2018). To reverse peatland degradation, ecological restoration has been conducted during the past few decades (Andersen Author contributions: AR, ME conceived and designed the research; AJ, ME, AI conducted the analyses; AR wrote the first draft of the manuscript; all authors contributed to the writing and discussed the work through all of it phases. 1 Natural Resources Institute Finland (Luke), Paavo Havaksen tie 3, 90570, Oulu, Finland 2 Geography Research Unit, University of Oulu, PO Box 8000, FI-90014, Oulu, Finland 3 Address correspondence to A. Räsänen, email aleksi.rasanen@oulu.fi 4 Department of Biological and Environmental Science, University of Jyväskylä, PO Box 35, Jyväskylä 40014, Finland 5 Water, Energy and Environmental Engineering Research Unit, University of Oulu, PO Box 4300, FI-90014, Oulu, Finland 6 Geological Survey of Finland, Teknologiakatu 7, 67101, Kokkola, Finland 7 Finnish Environment Institute (Syke), Survontie 9A, 40500, Jyväskylä, Finland 8 School of Resource Wisdom, University of Jyväskylä, PO Box 35, 40014, Jyväskylä, Finland © 2024 The Author(s). Restoration Ecology published by Wiley Periodicals LLC on behalf of Society for Ecological Restoration. This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited. doi: 10.1111/rec.14338 Supporting information at: http://onlinelibrary.wiley.com/doi/10.1111/rec.14338/suppinfo Restoration Ecology 1of15 et al. 2017). In the future, peatland restoration activities are further projected to increase globally. In the European Union (EU) alone, the restoration law targets to restore 90% of the degraded area of the ecosystems, including peatlands before 2050 (Regulation [EU] 2024/1991). In forestry-drained peatland sites, restoration includes filling or damming of ditches and removal of trees that have grown after the drainage (Haapalehto et al. 2011). The visible changes after restoration include, e.g. (1) decrease in tree cover, (2) replacement of ditches with flow-blocking structures, (3) increase in wetness and local water table level, and (4) changes in ground vegetation composition (Haapalehto et al. 2011). Of these changes, the first three occur almost immediately after restoration, while the changes in ground vegetation are slower (Haapalehto et al. 2011,2017; Menberu et al. 2016). The most recent studies have indicated that during the first 10 years after restoration, there are changes in the vegetation: particularly the more common species of pristine mires start to colonize the restored sites, but vegetation in the restored sites does not resemble that of pristine counterparts (Elo et al. 2024). Typical ecological targets of restoration are related to the return of original peatland community structure and functioning. The success can be measured, e.g. through inventories of different taxa, such as plants (e.g. Haapalehto et al. 2011,2017). Overall, monitoring restoration success is important for validating restoration methods and outcomes but also for improving our understanding of peatland ecosystem changes and processes. Nevertheless, traditional field-based monitoring methods require a considerable amount of labor and other resources, and they are restricted to a limited number of points that are not necessarily representative of the whole peatland in question. Therefore, cost-effective and spatially extensive monitoring methods are required, particularly because of the increasing amount of restoration activities. A potential solution for detecting changes over large areas cost-effectively is the utilization of satellite remote sensing (RS), as it can provide high spatial and temporal resolution observations of global land cover. The studies so far have indicated that particularly optical satellite data are usable for tracking changes in peatland wetness (Räsänen et al. 2022; Burdun et al. 2023; Isoaho et al. 2024) and land cover and vegetation such as habitat types and plant community structure (Kolari et al. 2022; Ball et al. 2023). The key strength of optical satellite imagery is its temporal availability: seamless and crosscomparable high-resolution data has been available since the 1980s (Wulder et al. 2022; Radeloff et al. 2024). In peatlands, post-restoration RS assessments have mainly focused on tracking changes in wetness (Räsänen et al. 2022;Burdun et al. 2023; Isoaho et al. 2024), while changes in spectral signatures and vegetation have gained less attention. The few studies include the work by Ball et al. (2023) analyzing whether the spectral signatures of restored sites start to resemble those of pristine peatland areas assessing the possibility of using RS for detecting changes in ground vegetation floristic gradients related to wetness and productivity. The lack of focus on vegetation changes has been evident overall in RS studies in peatlands, not just those related to restoration. This is surprising given that changes in vegetation composition and abundances of individual species are considered key indicators of peatland restoration success (Haapalehto et al. 2011; Elo et al. 2024; Kyrkjeeide et al. 2024). Even though broad-scale patterns in habitat type changes have been monitored (Kolari et al. 2022; Steenvoorden et al. 2022), more detailed analyses of temporal vegetation changes have not been conducted. Despite the lack of assessments about temporal changes in vegetation, there have been multiple studies mapping the spatial patterns of vegetation at a specific time point. Examples of monitored vegetation characteristics include plant communities and floristic gradients (Harris et al. 2015; Räsänen et al. 2020b), plant functional types (PFTs), such as shrubs, forbs, graminoids, and mosses (Räsänen et al. 2020b;Pangetal.2024), functional traits, such as leaf-area index and plant nutrient content (Kalacska et al. 2015; Räsänen et al. 2020a), and the occurrence and cover of single species (Kalacska et al. 2013;Pang et al. 2024; Simpson et al. 2024). It can be hypothesized that the temporal changes after restoration in these characteristics can be monitored if there are systematically collected long-term monitoring data and if the scale of the changes is detectable. Furthermore, for revealing the post-restoration vegetation succession, the RS approaches should simultaneously account for several changes in land cover, including in tree cover and wetness. We utilize globally unique 10-year before-after controlimpact (c.f. Christie et al. 2020) Finnish peatland restoration monitoring initiative data spanning 150 sites that belong to six different peatland types (Elo et al. 2024). We use Bayesian joint species distribution models (Ovaskainen et al. 2017; Ovaskainen & Abrego 2020) that can be used to assess plant community change and the associations between different plant species, PFTs, and RS variables. Our objective is to study how the post-restoration land cover changes in peatlands are linked with spectral signature changes and what kinds of associations there are between spectral and vegetation changes in different peatland types and treatments (pristine, drained, and restored). Our broader objective is to contribute to the work developing RSbased ecological restoration success indicators (c.f. Skidmore et al. 2021) that can be used for automatic restoration success analysis. Our specific research questions are as follows: (1) What is the effect of peatland restoration on spectral signatures in different peatland types? (2) Do satellite imagery variables improve predictions of plant species and PFT occurrences and covers in restored, drained, and pristine sites? (3) What kinds of associations exist between RS variables and plant species and PFT occurrence/cover? Methods Study Sites and Field Data We used data from 150 sites belonging to the Finnish Metsähallitus Parks & Wildlife peatland restoration monitoring network (Fig. 1; description in Elo et al. 2024). The sites in the network Restoration Ecology2of15 Remote sensing of peatland vegetation 1526100x, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/rec.14338 by University Of Jyväskylä Library, Wiley Online Library on [10/11/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License are located throughout Finland (60–68N, 21–31E; Fig. 1; Elo et al. 2024) across elevation and climatic gradients (Fig. S1). They are divided into six different peatland types based on their vegetation: rich and poor spruce mire forests, rich and poor pine mire forests, and rich and poor open mires. For each type, there are data for 10 restored sites and 10 nearby located pristine counterparts, with the exceptions being 9 +9 sites for poor open mires and 11 +11 sites for rich open mires. Additionally, there are 30 drained control sites (4–6 sites per type). The restored sites have been drained for forestry between the 1960s and 1970s and subsequently restored between 2007 and 2014, while pristine sites have not been drained, and drained sites have been drained approximately concurrently with the restored sites but have not been restored. Spruce mire forests are densely treed by Picea abies in oligotrophic poor sites, while in meso-eutrophic rich sites, there are also some deciduous trees (esp. Betula pubescens). The ground vegetation consists of forbs, graminoids, and Sphagnum and feather mosses. Pine mire forests are sparsely treed by lowgrowth Pinus sylvestris, accompanied by B. pubescens in rich sites. Pine mire forests are in general more nutrient-poor than spruce mire forests, with poor sites being ombrotrophic and rich sites oligo-mesotrophic. Ground vegetation consists typically of various evergreen and deciduous shrubs (e.g. Rhododendron tomentosum and Vaccinium uliginosum) and Sphagnum mosses. Open mires are mostly treeless sites, with the few trees being P. sylvestris in the ombrotrophic poor sites and deciduous trees (e.g. B. pubescens) in oligo-mesotrophic rich sites. The ground vegetation in poor sites consists of Sphagnum mosses and shrubs, while in the rich sites, the cover of sedges, forbs, and wet brown mosses increases. Restoration aims to raise the water table and to return the canopy structure as similar as possible to the pre-drained state or an undrained reference site. Typical restoration measures in each type consist of filling in and damming the ditches as well as felling of trees at various extents, depending on the peatland type. In spruce mire forests, a relatively dense tree cover has usually been left after restoration, while in pine mire forests, only some trees have been left, and in open mires, practically all trees have been cut. In each site, vegetation has been monitored in 10 one-squaremeter squared plots. These plots are arranged in two parallel lines, with each line containing five plots spaced four meters apart from each other (see Fig. 1). The lines are located to represent typical vegetation of each site, and the minimum distance to the nearest ditch is 10 m. The exact location of the first plot has Figure 1. Finnish peatland monitoring network, with locations of the monitoring sites (A), number of monitoring sites for each peatland type and productivity (B), and sampling of vegetation inventory at each site (C). Restoration Ecology 3of15 Remote sensing of peatland vegetation 1526100x, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/rec.14338 by University Of Jyväskylä Library, Wiley Online Library on [10/11/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License been randomized within the criteria defined above. The vegetation sampling has been conducted before restoration (year 0) and 2, 5, and 10 years after restoration. In pristine and drained sites, a similar inventory interval has been utilized. During each inventory, the %-cover of each vascular plant and moss species (Table S1) has been visually estimated. For each site, we have calculated a site-level community by averaging the cover over plots for each site. While the differences between sites are larger in spruce mire forests than in pine mire forests and open mires (Elo et al. 2016), within-site variability is approximately equal between the peatland types (Fig. S1). For PFT-level analyses, we divided the plant species into the following PFTs that have been widely used in peatland research before (e.g. Räsänen et al. 2020a,2020b): deciduous shrubs, evergreen shrubs, forbs, graminoids, Equisetum,Pteridophytina,Sphagnum, and other mosses (Table S1). We further divided the shrub, forb, and graminoid PFTs by their primary habitat requirements into mire and other groups, while Sphagnum and other mosses were divided into hummock, lawn, and hollow species (Eurola et al. 1995; Finnish Biodiversity Info Facility 2024). This was done because the habitat requirements and potential restoration impact are not uniform within a PFT, but species within a single PFT can react differently to restoration. Remote Sensing Data We used five different optical RS variables: red reflectance, near-infrared (NIR) reflectance, shortwave infrared transformed reflectance (STR; Sadeghi et al. 2015), soil-adjusted vegetation index (SAVI; Huete 1988), and normalized difference moisture index (NDMI; Gao 1996). We selected variables that do not strongly correlate with each other and that have been shown to be useful in peatland studies related to land cover, vegetation, and wetness. Of the visible and NIR wavelength bands, we chose red and NIR due to their capability to track changes in peatland vegetation, habitats (Kolari et al. 2022), and wetness (Isoaho et al. 2023,2024). STR is a transformation of shortwave infrared (SWIR) reflectance, and it has been shown to function well in wetness prediction (Isoaho et al. 2024; Jussila et al. 2024). Of different vegetation greenness indices, we included SAVI due to its relatively good performance in predicting changes in productivity gradient in open and sparsely treed peatlands. We complemented the list with NIR-SWIR index NDMI that has correlated with peatland soil moisture, water table, and wet area (Meingast et al. 2014; Ludwig et al. 2019). Overall, a versatile set of variables has been recommended due to site-specific differences in the most important variables (Räsänen et al. 2022). We calculated the variables from the bottom-of-atmosphere reflectance products of 10–20 m spatial resolution European Space Agency Copernicus Sentinel-2 and 30 m spatial resolution National Aeronautics and Space Administration/United States Geological Survey Landsat 5-9 datasets that we harmonized to Landsat 8-9 reflectance (Roy et al. 2016; Zhang et al. 2018). For each variable, we calculated early summer (ES; May 1–June 15) and midsummer (MS; July 1–August 15) annual median imagery, from which we calculated median imagery for each monitoring period (1–5 years before restoration; 1–3 years after restoration, 4–6 years after restoration, and 9–11 years after restoration) for both seasons. We used two seasons as multitemporal analysis has been shown to boost model performance in various studies (Räsänen et al. 2020b; Pang et al. 2022; Wu et al. 2023) and as these seasons have strikingly different hydrological and phenological conditions (Sallinen et al. 2023; Isoaho et al. 2024. We calculated median imagery to filter out noise present in single images and to construct representative datasets for the selected phenological stages. During the ES season, the snow has melt, vegetation starts to emerge, and the water table is at its highest. During the MS season, vegetation peaks and the water table is typically at its lowest. We did not include imagery during late summer or autumn due to persistent cloud coverage during that season. We utilized only images with a maximum of 30% cloud cover and masked out remaining clouds, haze, snow, and shadow with Scene Landcover Classification (Sentinel-2) and Quality Assessment pixel classification (Landsat). For each variable, we calculated mean values for a 15-mradius buffer area that contained all vegetation plots in the sites. For dates with multiple Sentinel-2 or Landsat satellite image observations, we calculated the mean values over the observations. We conducted all satellite image processing in Google Earth Engine (Gorelick et al. 2017). Statistical Analysis We applied a type of Bayesian joint species distribution modeling: Hierarchical Modeling of Species Communities (HMSC; Ovaskainen & Abrego 2020; Ovaskainen et al. 2017). We conducted two different sets of HMSC analyses: (1) plant specieslevel and (2) PFT-level analyses. In both analyses, RS variables were included. In Section 3, we mostly report species-level analysis results but complement the information with PFT analysis results. HMSCs can be used to examine the species-to-species associations (here also RS variable-species and RS variable-PFT associations) when controlling for other covariates, as well as changes in plant communities in different management types. For each peatland type separately, we modeled the occupancy (presence/absence) of the species or PFT having greater than 20 occupancies by a probit model, and conditionally on the occurrence, we modeled the cover (log-transformed, normalized to zero mean and unit variance within each species) of the same species or PFT with a normal model. We included RS variables (normalized to zero mean and unit variance) as response variables in the same model to infer their associations with species. As random effects, we included site, modeled as a spatially explicit random effect and sampling year. As explanatory variables, we included treatment (a factor with three levels: restored/drained/pristine), time (a continuous variable; 0, 2, 5, and 10 since restoration or corresponding period), and its second-order polynomial to allow for unimodal responses, as well as the interaction of treatment and time squared. Restoration Ecology4of15 Remote sensing of peatland vegetation 1526100x, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/rec.14338 by University Of Jyväskylä Library, Wiley Online Library on [10/11/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License We ran the models using R package Hmsc 3.0 (Tikhonov et al. 2020). The package uses the Bayesian framework with Gibbs Markov chain Monte Carlo (MCMC) sampling. We assumed the default prior distributions, with the exception of a1 and a2 parameters for the random effect site, which were set both to 100 to increase the shrinkage and thus avoid modeling noise. We sampled the posterior distribution with four chains, each for 250 samples with thinning of 10,000, using a transient phase of 500,000 and adaptation (the number of MCMC steps at which the adaptation of the number of latent factors is conducted) of 400,000. We evaluated the chain mixing by assessing the effective size of the posterior sample with a potential scale reduction factor (Fig. S2) and assessed the explanatory power by Tjur’sr 2 (occupancy) and r 2 (cover) (Fig. S3). Based on the fitted models, we predicted the values of each RS variable in time for different treatments. From these predictions, we calculated the following three measures informing about different aspects of the effects of restoration on the RS variables. First, we calculated whether restoration affected RS variable as Resp ¼vR 10 vR 0  vD 10 vD 0  where vR 10 and vR 0are values of RS variables in restored sites in the years 10 and 0, respectively, and vD 10 and vD 0represent corresponding values in drained sites. Resp takes positive values if the change is positive in relation to change in drained sites and negative values if the change is negative in relation to change in drained sites. Second, values for RS in drained and in restored sites may differ as they were not randomly selected. To assess the reliability of inferences of how RS variables respond to restoration, we calculated whether they differed at the beginning of the experiment between the drained and restored sites: Diff1 ¼vR 0vD 0 Third, we calculated whether the difference between restored and pristine control sites grew smaller (or larger) during the study period: Diff2 ¼abs vR 10 vP 10  abs vR 0vP 0  For all three measures, we calculated the median as well as the posterior probability for the median being larger than zero. We considered the measure to have high support for the median being positive/negative if the posterior probability is greater than 95% and moderate support if the posterior probability is greater than 80%. We calculated the same measures for abundance of each species, or PFT (probability of occurrence cover given occurrence). As the species-specific responses to restoration merely reinforce the previous findings (Elo et al. 2024), we present them only in Fig. S4 together with the PFT-level information. To answer whether RS variables improve predictive power of the speciesor PFT-specific models, we first calculated twofold cross-validation. Then, we performed conditional crossvalidation, where we used data from each RS variable, one at a time, and it’s estimated associations with the species or PFTs to calculate the predictions. Finally, we compared whether including information on the RS variable yielded an improvement in predictive power by subtracting the cross-validated predictive power from the conditionally cross-validated predictive power (CCV). We did the cross-validations with parameter values based on thin =10 due to the high computational demand of the calculations and because predictive powers tend to converge with a relatively low number of thinning. Furthermore, variance partitioning of the explanatory variables remained similar when thinning of 10 or 10,000 was used. Finally, we calculated association matrices, which represent the residual associations of RS variables and species, or PFTs, after controlling for the treatment, time, and their interaction. Results Effect of Restoration on Spectral Signatures Almost all RS variables were affected by restoration (Fig. 2). Especially, both ES and MS red and NIR reflectance increased after restoration in most peatland types. Moreover, SAVI MS increased, whereas for SAVI ES, the response had low statistical support (posterior probability <95%). STR and NDMI decreased or showed no highly supported response to restoration, with STR showing highly supported response in more peatland types than NDMI. The only peatland type where no high support was seen in any of the RS variables was rich spruce mire forests, whereas the clearest effects were seen in pine mire forests and open mires. In pine mire forests and open mires, the red and NIR reflectance of restored sites had similar values than drained sites before restoration and approached those of pristine sites after restoration (Figs. 3&4). For other variables and peatland types, the temporal trends in restored, pristine, and drained sites were less clear, and the spectral signatures in restored sites did not clearly move closer to the signatures in pristine sites (Figs. 3–5). Improvement in Predictive Power From RS Variables For most species or PFTs, at least one of the RS variables improved the predictive power and resulted in a predictive power higher than 0 (Figs. 6&S5). There were differences between species and peatland types, which RS variables improved the predictive power, and none of the variables was clearly better than the others (Fig. 7). The resulting predictive powers were generally relatively modest both for species and PFTs. The mean was typically circa 0.1–0.2, but for some species, CCV was very high (up to 0.65; Table S2). The same applied to the improvement in predictive power when including the best RS variable: typically, the improvement was small (<0.1), but for some species, it was very high (up to 0.52; Table S2). Restoration Ecology 5of15 Remote sensing of peatland vegetation 1526100x, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/rec.14338 by University Of Jyväskylä Library, Wiley Online Library on [10/11/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Associations Between Species or PFTs and Remote Sensing Variables When concentrating on those RS-species or RS-PFT linkages in which (1) restoration had an effect on both the RS variable and species or PFT and (2) incorporating RS variable increased predictive power (Figs. 8&S6), the improvement in predictive power was typically small (<0.05 in improvement in crossvalidated predictive power [CCV-CV]). The exceptions were mainly the negative associations of graminoids with NDMI and red reflectance (Carex chordorrhiza in poor pine mire forests and rich open mires, and also C. lasiocarpa cover in rich open mires; Fig. 8) and graminoid PFTs in rich open mires (Fig. S6). Additionally, a somewhat clear improvement (>0.05 in CCV-CV) was seen for Vaccinium uliginosum cover associated negatively with NDMI ES in rich pine mire forests. There were also other associations, both negative and positive, but in these cases, RS variables increased the predictive power little (<0.05 in CCV-CV). Discussion Our results show that (1) peatland restoration affects satellitederived spectral signatures, (2) satellite image variables increase modestly speciesor PFT-specific predictive power in joint species distribution models, and (3) there are few clear links between the changes in RS variables and the changes in post-restoration species or PFT occurrence and cover. Restoration Effects on Spectral Signatures Over time, the spectral signatures of restored sites moved closer to those of pristine sites. As pristine-like ecosystem structure and functioning is the goal for restoration, the result suggests that restoration can be successfully monitored with RS data. The trend toward pristine was particularly evident in red and NIR reflectance for sparsely treed pine mire forests and open mires, whereas for other tested variables and especially for densely treed spruce mire forests, the trends were not as clear. These findings align with Ball et al. (2023), who observed the convergence between restored and pristine sites with optical Sentinel-2 and synthetic aperture radar Sentinel-1 data. In their analysis, the similarity increased relatively strongly during the first 10–15 years after which the signatures between restored and pristine sites were close to each other. We could not verify this finding due to our 10-year post-restoration monitoring period but instead showed that during the first 10 years, the harmonization in variable values between restored and pristine sites was evident only for certain variables and peatland types. Ball et al. (2023) did not analyze the trends in different bands and indices but focused on overall spectral similarity using Mahalanobis distance and limited analysis to 1-year sampling of peatlands restored during different years. Therefore, our analysis complements the work by Ball et al. (2023) by showing (1) that there are differences between peatland types and RS variables and (2) what kind of trend is seen after restoration. Our results indicate that reflectance of the red, NIR, and SWIR (STR is transformed SWIR reflectance and negatively correlated with it) increases after restoration, particularly in pine Figure 2. The response to restoration of remote sensing variables in different peatland types. Note that the values are based on the original values of each remote sensing variable; therefore, the range for shortwave infrared transformed reflectance (STR) is much larger than for non-transformed bands (STR is circa 50 and 1 for shortwave infrared reflectance of 1 and 26%, respectively). In the figure, NDMI refers to normalized difference moisture index, SAVI to soil-adjusted vegetation index, ES to early summer, and MS to midsummer. The statistical support is “Positive 95%”if the posterior probability of the median being larger than zero is greater than 95%; “Negative 95%”if the posterior probability of the median being smaller than zero is greater than 95%; and “Weak”if the posterior probability of the median being larger/smaller than zero is less than 95%. Restoration Ecology6of15 Remote sensing of peatland vegetation 1526100x, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/rec.14338 by University Of Jyväskylä Library, Wiley Online Library on [10/11/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License mire forests and open mires. This is probably largely attributed to the felling of trees and increased openness in the landscape in these peatland types, as especially red reflectance and RSmeasured albedo is negatively associated with woody canopy cover (Yang & Prince 1997; Kuusinen et al. 2016). Felling of trees is conducted during restoration to make room for Figure 3. Changes in remote sensing variables over time in drained, pristine, and restored open mires. The plots are drawn only for those changes in remote sensing variables that responded either positively or negatively to restoration with a high support (a posterior probability of the median being larger/smaller than zero greater than 95%; Fig. 2). In the figure, NDMI refers to normalized difference moisture index, SAVI to soil-adjusted vegetation index, STR to shortwave infrared transformed reflectance, ES to early summer, and MS to midsummer. Restoration Ecology 7of15 Remote sensing of peatland vegetation 1526100x, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/rec.14338 by University Of Jyväskylä Library, Wiley Online Library on [10/11/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License excavators along the ditches, bring back pristine-like canopy structure, and decrease evapotranspiration by trees. Although the effects of restoration occur short-term and are evident in 2 years after restoration data, our modeling approach, where time is used as a continuous variable, tends to extend this effect. However, the felling of trees during the time of restoration does Figure 4. Changes in remote sensing variables over time in drained, pristine, and restored pine mire forests. The plots are drawn only for those changes in remote sensing variables that responded either positively or negatively to restoration with a high support (a posterior probability of the median being larger/smaller than zero greater than 95%; Fig. 2). In the figure, NDMI refers to normalized difference moisture index, SAVI to soil-adjusted vegetation index, STR to shortwave infrared transformed reflectance, ES to early summer, and MS to midsummer. Restoration Ecology8of15 Remote sensing of peatland vegetation 1526100x, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/rec.14338 by University Of Jyväskylä Library, Wiley Online Library on [10/11/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Tikhonov G, Opedal ØH, Abrego N, Lehikoinen A, de Jonge MM, Oksanen J, Ovaskainen O (2020) Joint species distribution modelling with the R-package Hmsc. Methods in Ecology and Evolution 11:442–447. https://doi.org/10.1111/2041-210X.13345 Ur ak I, Hartel T, Gallé R, Balog A (2017) Worldwide peatland degradations and the related carbon dioxide emissions: the importance of policy regulations. Environmental Science & Policy 69:57–64. https://doi.org/10.1016/j. envsci.2016.12.012 Vasander H, Tuittila E-S, Lode E, Lundin L, Ilomets M, Sallantaus T, Heikkilä R, Pitkänen M-L, Laine J (2003) Status and restoration of peatlands in northern Europe. Wetlands Ecology and Management 11:51–63. https://doi.org/ 10.1023/A:1022061622602 Wu S, Tetzlaff D, Daempfling H, Soulsby C (2023) Improved understanding of vegetation dynamics and wetland ecohydrology via monthly UAV-based classification. Hydrological Processes 37:e14988. https://doi.org/10. 1002/hyp.14988 Wulder MA, Roy DP, Radeloff VC, Loveland TR, Anderson MC, Johnson DM, et al. (2022) Fifty years of Landsat science and impacts. Remote Sensing of Environment 280:113195. https://doi.org/10.1016/j. rse.2022.113195 Yang J, Prince SD (1997) A theoretical assessment of the relation between woody canopy cover and red reflectance. Remote Sensing of Environment 59:428– 439. https://doi.org/10.1016/S0034-4257(96)00111-3 Zhang HK, Roy DP, Yan L, Li Z, Huang H, Vermote E, Skakun S, Roger JC (2018) Characterization of sentinel-2A and Landsat-8 top of atmosphere, surface, and nadir BRDF adjusted reflectance and NDVI differences. Remote Sensing of Environment 215:482–494. https://doi.org/10.1016/j. rse.2018.04.031 Supporting Information The following information may be found in the online version of this article: Figure S1. Box and whiskers plots of elevation (m a.s.l.), mean temperature (C), annual precipitation (mm), number of plant species, and plant species beta diversity gradients for different treatments and peatland types. Figure S2. Convergence of beta and omega parameters for rich open mires, with thinning of 10,000. Figure S3. Variance partitioning of explanatory variables for rich open mires, with thinning of 10,000. Figure S4. Species and plant functional type median response to restoration, separately for each peatland type. Figure S5. Analysis whether the remote sensing variables increased the conditionally cross-validated predictive power for different plant fuctional types. Figure S6. The associations between the plant functional types and the remote sensing variables. Table S1. Species abbreviations, full scientific names and plant functional types for each species. Table S2. Conditionally cross-validated predictive power and the improvement in cross-validated predictive power when including the best performing remote sensing variable. Coordinating Editor: Raja Hussain Received: 23 August, 2024; First decision: 3 October, 2024; Revised: 25 October, 2024; Accepted: 28 October, 2024 Restoration Ecology 15 of 15 Remote sensing of peatland vegetation 1526100x, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/rec.14338 by University Of Jyväskylä Library, Wiley Online Library on [10/11/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License