Upscaling peatland mapping with drone-derived imagery: impact of spatial resolution and vegetation characteristics

ABSTRACT

Northern peatland functions are strongly associated with vegetation structure and composition. While large-scale monitoring of functions through remotely sensed mapping of vegetation patterns is therefore promising, knowledge on the interdependency between spatial resolution of acquired imagery, spatial vegetation characteristics, and resulting mapping accuracy needs to be improved. We evaluated how classification accuracy of commonly used vegetation mapping units (microforms and plant functional types) was affected by spatial resolution of acquired imagery and several spatial characteristics of the vegetation itself (size, shape, configuration, and diversity). To this end, we collected very high-resolution drone imagery (<0.05 m) from eight Irish peatlands differing in vegetation composition and pattern complexity in September 2021 and 2022. We then resampled this imagery from pixel sizes of 0.027–1 m and classified both mapping units at all unique spatial resolutions. Hereafter, we computed spatial vegetation characteristics for each of the eight classified images at 0.027 m spatial resolution to determine their role in defining minimum spatial resolution requirements for both microforms and plant functional types. We found that overall classification accuracy of microforms and plant functional types was consistently high (>90%) for all studied peatlands until average spatial resolutions were reached of 0.5 m ± 0.2 m and 0.25 m ± 0.1 m, respectively. However, variability within mapping units was considerable, with overall minimum spatial resolution ranging between 0.25 and 0.7 m for microforms and between 0.15 and 0.35 m for plant functional types. Individual classes even varied from 0.05  to 1 m. Furthermore, spatial vegetation characteristics were important drivers of minimum spatial resolution for microforms, but not for plant functional types. Particularly, peatlands with larger, compacter, and more clustered microform patches could be classified with coarser spatial resolution imagery (up to 0.7 m), while peatlands with small, complex, diverse, and more finely distributed patches required higher spatial resolutions (minimally 0.25 m). Based on these findings, we conclude that spatial vegetation characteristics strongly determine minimum required spatial resolution and thus affect upscaling of peatland vegetation mapping beyond the landscape scale by constraining the use of specific remote sensing platforms.

KEYWORDS:

• Peatlands
• vegetation patterns
• drones
• spatial resolution
• remote sensing
• upscaling

1. Introduction

Northern peatlands (hereafter, peatlands) are large, isolated, and inaccessible wetland types that store and sequester carbon in peat. These ecosystems provide key ecosystem services, like drinking water provision, water purification, recreational activities, biodiversity, and climate regulation (Bain et al. 2011; Bonn et al. 2016). Peatland services or functions are often associated with fine-scale (<10 m) vegetation structure (composition, biomass, and spatial organization; Couwenberg et al. 2011; Dieleman et al. 2015; Robroek et al. 2017). As a result, accurate assessment of current and future peatland functions at least partially depends on the spatio-temporal dynamics of peatland vegetation, which requires methods that can capture its fine-scaled heterogeneous nature, and can be expanded to the large spatial scales at which peatlands occur in the landscape.

Peatland vegetation typically exhibits spatial patterning associated with micro-topography. These micro-topographical features are called microforms and range from open water pools where the peat surface is always below the water table, to wet hollows, moist lawns, and finally to hummocks, where the peat surface reaches up to ~50 cm above the water table (Belyea and Baird 2006). Along this microtopographic gradient, the availability of water and nutrients change, creating habitats for different plant species sharing adaptations to the local conditions. As a result, microforms represent a distinct set of species and/or plant functional types that can be recognized as a coherent unit. The contrasting biogeochemical conditions also act as a filter on plant traits, subsequently affecting spectral properties of plant functional types inhabiting specific microforms (Schaepman-Strub et al. 2009). Hollows are usually dominated by aquatic peat mosses and occasional graminoids and sedges, while lawns and hummocks are occupied by more dry-adapted peat moss species, as well as graminoids, lichens, and dwarf shrubs. Both microforms and their plant functional types occur at fine spatial scales varying from 1–10 m to 0.01–1 m, respectively (Belyea and Baird 2006).

The spatial organization of microforms and plant functional types is a well-known indicator of peatland functions for multiple reasons. First, microforms affect the hydro-physical properties of a peatland (Branham and Strack 2014; Waddington et al. 2010), hereby affecting water flow and nutrient dynamics (Eppinga et al. 2008; Macrae et al. 2013). Second, the spatial organization of microforms and their associated plant functional types strongly affect carbon sequestration through differences in productivity and decomposition (Aerts, Verhoeven, and Whigham 1999; Goud et al. 2017; Johnson and Damman 1991; Laine et al. 2012; Turetsky et al. 2008). These differences are further amplified by the increasing ratio of methane to carbon-dioxide emissions through soil respiration along the water table gradient from hummock to hollow (Bubier et al. 1993; Krohn et al. 2017; Laine et al. 2007; Lunt, Fyfe, and Tappin 2019; Waddington and Roulet 1996). Third, there is evidence that retaining a diverse composition of plant communities through microform variations can increase the overall carbon sink function of a peatland (Robroek et al. 2017). Lastly, the presence of microform variations may facilitate reorganization under changing environmental conditions through lateral expansion and contraction (Belyea and Clymo 2001), hereby leading to broader ecosystem resilience.

Given the importance of microforms and plant functional types in peatland functioning, as well as the spatial scales at which they occur, remote sensing with unmanned aerial vehicles (UAV) or drones has recently gained much attention for its potential role in mapping peatland vegetation. However, while peatland vegetation patterns have successfully been mapped in recent drone studies (e.g. Beyer et al. 2019; Bhatnagar et al. 2021; Palace et al. 2018; Räsänen and Virtanen 2019; Räsänen et al. 2020; Steenvoorden et al. 2022), upscaling of this approach is hampered by the largely unknown interdependency between spatial resolution, spatial vegetation characteristics, and classification accuracy in different peatlands. More precise understanding of the relationship between these factors is required not only to shed light on the potential and limitations of using drones for large-scale mapping and monitoring of peatland vegetation patterns, but also those of aerial and satellite imagery.

In this study, we investigated the effect of various spatial resolutions of drone imagery on the classification accuracy and consistency of microforms and plant functional types in eight peatlands varying in the complexity and heterogeneity of their vegetation. We hypothesize that the minimum spatial resolution required for making accurate classifications of microforms and plant functional types 1) is located between 0.1 and 1 m for all peatlands, and 2) is affected by the spatial characteristics of the vegetation itself (size, shape, configuration, and diversity), being smallest in peatlands with small, complex, finely distributed, and diverse vegetation patterns.

2. Materials and methods

2.1. Study area

In this study, we investigated eight different Irish ombrotrophic peatlands representing a broad range in the composition and complexity of their vegetation patterns (Figure 1). Peatlands in Figure 1 are sorted visually from left to right and top to bottom in order of decreasing pattern complexity from smaller, complex, finely distributed and more diverse patches to larger, compacted, more clustered and less diverse patches. Bangor Erris and Roundstone are two Atlantic blanket bogs, which occur along the west-coast of Ireland under higher precipitation rates (usually >1200 mm per year) than the other peatlands in our study. The central areas of these two peatlands exhibit a clear hummock-lawn-hollow pattern, but where Bangor Erris has relatively large and clustered pools, those of Roundstone are more interspersed with lawns and hummocks (Figure 1). Besides, the two Atlantic blanket bogs contain notably less peat moss cover than the raised bogs in our study area. Carrowbehy is a western raised bog complex of which its central areas also exhibit a clear hummock-lawn-hollow pattern with relatively frequent open water pools and high peat moss cover. Derrinea is also a western raised bog with similar vegetation patterning as Carrowbehy, but with smaller open water pools and a flush system in its central zone. Moyclare, Mongan, Ferbane, and Raheenmore are all Midland raised bogs, and exhibit less developed microtopography than the two western raised bogs, lacking permanent pools and extensive hollows. However, the vegetation patterns of the central areas of Raheenmore are even more homogeneous than the other peatlands, missing noteworthy hollows and being dominated largely by hummocks interspersed with lawns filled with graminoids and only some peat mosses.

Figure 1. Study area maps showing the characteristic vegetation patterns in the central area of each of the eight Irish ombrotrophic peatlands that were investigated in this study as seen from a drone-derived true-color orthomosaic at 120 m altitude.

ll peatlands are part of Irelands’ Special Areas of Conservation network under the EU Habitats Directive (92/43/EE; Mackin et al. 2017; National Parks and Wildlife Service 2018). Carrowbehy, Mongan, and Raheenmore have undergone hydrological restoration with drain blocking work (Fernandez et al. 2014), starting as far back as 1984 on Mongan and continuing intermittently on the four sites. However, all peatlands were subject to active EU-funded restoration measures through LIFE (L’Instrument Financier pour l’Environment program). Specifically, all six raised bogs received restoration between 2016 and 2022 through the “The Living Bog” restoration project (LIFE code: LIFE14 NAT/IE/000032). The blanket bogs included in this study have been under activate conservation since 2021 through the “Wild Atlantic Nature” restoration project (LIFE code: LIFE18 IPE/IE/000002).

2.2. Drone imagery capture

We retrieved drone imagery from the central and hydrologically most intact areas of all peatlands during two different field seasons (2021 and 2022) using a DJI Mavic 2 Pro drone with Hasselblad L1D-20c red–green–blue (RGB) color sensor camera, which contains a one-inch (2.54 cm) CMOS sensor that captures images of 20 MP. Drone imagery of Roundstone, Bangor Erris, Carrowbehy, and Raheenmore was collected between 11 and 21 September 2021, respectively, while drone imagery of Moyclare, Mongan, Derrinea, and Ferbane was retrieved in the same month one year later between 6 and 14 September 2022, respectively. In 2021, we used DJI Ground Station Pro to design automated flights of about 2 ha at an altitude of both 40 m (1 cm pixel size) and 120 m (2.7 cm pixel size) above ground level. We chose a flight altitude of 120 m because it is the most efficient legal flight altitude to map fine-scale vegetation patterns in Irish peatlands (Steenvoorden, Bartholomeus, and Limpens 2023). The higher spatial resolution imagery at 40 m flight altitude was used to validate vegetation classes during development of training/testing samples used in classification (see also 2.2.4). To allow for georeferencing of the stitched drone imagery, we distributed four large 30 × 40 cm ground control points around the edges of each 2 ha plot and measured their position using a Topcon HiPer HR real-time kinematic (RTK) and TopNET+ global navigation satellite system (GNSS) with 1–3 cm accuracy. For the flights in 2022, we did not use any ground control points because we designed automated flights of the whole peatlands, and used available reference maps by ESRI in ArcGIS Pro for georeferencing. We used a forward/side image overlap of 80/80% for automated flights in 2021 and 75/60% forward/side image overlap for automated flights of peatlands in 2022, where the image overlap refers to the number of shared pixels between adjacent images, allowing for more accurate stitching and 3D reconstruction. We decided upon using less image overlap for all drone flights in 2022 because tests using different image overlap (75/60% versus 80/80%) in Carrowbehy in 2021 did not show differences in the accuracy of photogrammetry products, but notably improved flight and image-processing efficiency. All drone flights were conducted within one hour of solar noon and under fully cloudy to mostly cloudy weather conditions to minimize the potential effect of shade and light variability during classification. Wind speeds during all drone flights varied between low to moderate wind speeds up to 19 km/h.

2.3. Pre-processing drone imagery

We pre-processed all drone imagery products before classifying both microforms and plant functional types within each peatland. First, we used ortho-mapping software in ArcGIS Pro to produce a stitched RGB-orthomosaic of the drone images for each flight (2.7 cm pixel size), as well as to create a Digital Terrain Model (DTM) with the same spatial resolution using the photogrammetric data contained within each individual drone image. Hereafter, we clipped a rectangular 1 ha polygon from each full orthomosaic and DTM for use in all further analyses. Second, we grouped all pixels in each stitched orthomosaic into segments using a mean-shift clustering algorithm. This algorithm groups pixels based on similarity in spectral characteristics (spectral reflectance values) and spatial characteristics (proximity between pixels with similar reflectance) within the raster image. During segmentation, we used a minimum segment size of 0.25 m because it has previously shown to be a cutoff point that yields significantly more efficient classifications than finer segmentation scales (Steenvoorden, Bartholomeus, and Limpens 2023). However, once the pixel size of the drone imagery became larger than 0.25 m (see also 2.6.), the segment size for a specific classification was automatically set to the spatial resolution of the imagery. Third, we detrended each DTM by fitting a second-order polynomial function through the elevation points in the DTM, and subsequently subtracted the DTM from the fitted trend function to correct for doming effects. Consequently, the detrended DTM represents the relative micro-topographical differences in the orthomosaic more realistically, making it functional as an additional predictor variable during classification of both vegetation patterns (Steenvoorden, Bartholomeus, and Limpens 2023). Lastly, we calculated Hue–Saturation–Value (HSV) color model values and 10 RGB-derived vegetation color indices as additional predictor variables to further emphasize spectral differences between vegetation classes of microforms and plant functional types. When using only individual RGB bands, the dark shades of different colors tended to be more spectrally similar than the same color with different brightness, reducing classification accuracy. In contrast, the hue values of pixels with the same color were more spectrally similar regardless of their brightness, diminishing potential effects of shade on spectral reflectance, increasing classification accuracy. Ultimately, we used a total of 25 predictor variables in classification of microforms and plant functional types, consisting of mean and standard deviation of RGB values, topography (minimum, maximum, and mean elevation), mean of the HSV color model values, 10 vegetation color indices combining two or more RGB bands, and three shape metrics (pixel count, rectangularity, and compactness), computed separately for all segments in each orthomosaic (see also Table S1).

2.4. Ground truth data

We divided vegetation into microforms and plant functional types, which are two commonly used mapping units for vegetation patterns in peatlands. We then subdivided each mapping unit into multiple vegetation classes based on drone-visible indicator species (or the lack thereof) and their commonly associated position along the micro-topographical gradient as seen from the newly developed orthomosaics. Microforms were subdivided into three classes: 1) hollow, 2) lawn, and 3) hummock (Table S3), while plant functional types were subdivided into five classes: 1) peat moss, 2) shrub, 3) graminoid, 4) lichen, and 5) water/bare peat (Table S3). Although water/bare peat is not officially a type of vegetation, we included it as a vegetation class in our analysis for plant functional types because it had noteworthy cover in most peatlands.

To classify microforms, we first created 500 randomly placed points within each orthomosaic and used the orthomosaic at 40 m altitude as a reference dataset to validate the created training/testing samples for use in classification. This led to approximate area-proportional allocation of our training/testing samples. For classification of plant functional types, we used a targeted sampling approach instead with equal allocation, where we placed 100 points per plant functional type per orthomosaic at 120 m flight altitude to prevent underrepresentation of uncommon plant functional types within each peatland. We used a total of 500 sampling points per mapping unit because a previous study on mapping the same peatland vegetation types showed by means of a sensitivity analysis that consistent classifications (>90%) are still retrieved even when only about 200–250 points of the total training/testing sample remained. By keeping the number of sampling points to 500 in this study, we made certain to capture all variation within each orthomosaic. To ensure all plant functional types were sampled across the whole orthomosaic, we created a 10 × 10 fishnet grid over each orthomosaic and took one sample per plant functional type from each sub-grid. However, some peatlands contained less than 100 individual samples for specific vegetation classes. In such cases, we used all available samples as training/testing samples for those vegetation classes. Because relative proportions of microforms differed among peatlands, and not all microforms and plant functional types were present or widely distributed within a peatland, the final allocation of training/testing samples differed per peatland for some vegetation classes. In the end, the number of training/testing samples ranged between 0 and 132, 62 and 171, and 197 and 375 for hollow, lawn, and hummock, respectively, and between 63 and 100, 25 and 100, and 0 and 100 for peat moss, lichen, and water/bare peat, respectively (Table S2). The vegetation classes shrub and graminoid contained 100 individual training/testing samples for all peatlands.

2.5. Vegetation classification

We classified the vegetation of each mapping unit for each orthomosaic using the ensemble classifier Random Forest (RF; Breiman 2001) in Python’s Scikit-learn module (Pedregosa et al. 2011). Here, we employed the 25 predictor variables at each spatial resolution as input datasets (see also 2.6.), using selected segments from the training/testing dataset. Hereafter, we split the training/testing samples of each orthomosaic using five-fold stratified cross-validation (analogous to a ratio of 80/20 for training/testing). We chose this cross-validation approach because it minimized variance of model performance in our study as compared to repeated K-Fold cross-validation or a larger number of folds (Jiang and Wang 2017). Classification performance per vegetation class and final classification performance of each orthomosaic were computed by averaging precision, recall, and F1-score over all folds in the RF model using the testing samples. We averaged classification performance over all folds to reduce potential variability related to the specific samples chosen for training/testing in each fold, hereby providing a more consistent and generalizable evaluation of the overall performance of the RF model. Besides, we used F1-score as preferred classification error metric because it is more robust to imbalanced training/testing datasets, as was often the case in our study (Table S2). An average F1-score of 1 indicates a perfect prediction, while any value in between 0 and 1 (or 0–100%) indicates the probability that a testing sample is correctly classified. We classified all segments within each orthomosaic by taking the most frequently classified vegetation class for each segment over all folds.

2.6. Determining minimum spatial resolution

To assess the minimum spatial resolution required for mapping peatland vegetation, we first resampled the orthomosaic and DTM of each peatland at 120 m flight altitude in 20 increments from 2.7 cm to 1 m using the bilinear interpolation. This algorithm determines the new value of a cell based on a weighted distance average of the four nearest input cell centers. After resampling, we performed a sensitivity analysis by iteratively classifying both microforms and plant functional types for all peatlands using each combination of resampled drone imagery products. We then determined the minimum spatial resolution required to achieve consistent classifications for each mapping unit and for each peatland by averaging overall classification accuracy over all folds in each RF model. Minimum spatial resolution was reached once averaged overall classification accuracy of the next (coarser) spatial resolution became lower than 90%. To validate the use of the 90% overall classification accuracy limit in determining minimum spatial resolution, we used stratified estimation (Olofsson et al. 2013) to doublecheck to what extent 90% classification accuracy corresponded with expected 10% uncertainty in mapped area over all vegetation classes as derived from the confusion matrix. Stratified estimation uses both the confusion matrix and mapped class areas (m²) from a classification to compute the effect of misclassifications on estimated areas of each class. This represents uncertainty in mapped areas due to misclassifications more realistically, because solely using sample counts from the confusion matrix tends to underestimate or overestimate the true estimated area of vegetation classes, especially when their areas in the field are variable (Olofsson et al. 2014). While calculation of uncertainty through stratified estimation showed that studied peatlands had slightly higher uncertainty in estimated areas than 10% for both mapping units (12.4% ± 3.0% and 12.4% ± 5.2% for microforms and plant functional types, respectively; Figure S5, S6), both average uncertainty and their confidence intervals suggest that using a classification accuracy of 90% leads to relatively uniform uncertainty estimates over all studied peatlands, validating its use as a cutoff value.

2.7. Spatial vegetation characteristics

To evaluate the role of the vegetation patterning of microforms and plant functional types on their minimum required spatial resolution, we created ordinary least squares univariate linear regression models using the Python package Statsmodels version 0.14.0 (Seabold and Perktold 2010). We used the computed overall minimum spatial resolution for each mapping unit from all peatlands as response variable and used five spatial patch metrics as predictor variables: 1) “mean area” (patch size), 2) “standard deviation of area” (patch size), 3) “landscape shape index” (patch shape; Patton 1975), 4) “conditional entropy” (patch configuration; Nowosad and Stepinski 2019), and 5) “Shannon diversity index” (patch diversity; Figure 2; Shannon 1948). These metrics were selected from a total of 22 main metrics available in the Python package PyLandStats version 2.4.2 (Bosch and Rocchini 2019; excluding the six distribution-statistics metrics available per landscape metric) based on their recognized use in describing and quantifying the spatial organization of patterns on a landscape scale. To compute each spatial patch characteristic, we used the classified images of microforms and plant functional types with the highest spatial resolution of 2.7 cm. We employed the highest spatial resolution maps because these most realistically reflected the real patchiness of the vegetation, and because computation of spatial vegetation characteristics using coarser resolution imagery leads to notable loss of spatial patch complexity (Räsänen and Virtanen 2019; Virtanen and Ek 2014).

Figure 2. Visualization of the spatial patch metrics used in ordinary least square regression models to evaluate the role of spatial vegetation characteristics on the minimum image resolution requirements for mapping microforms and plant functional types. Presented are a hypothetical “low” and “high” values for mean area (patch size) standard deviation (SD) of mean area (patch size), landscape shape index (patch shape), conditional entropy (patch configuration), and Shannon diversity (patch diversity).

3. Results

3.1. Minimum spatial resolution requirements

3.1.1. Microform classifications

Microform classifications showed consistently high accuracies (>90%) for all peatlands until a spatial resolution of 0.5 m ± 0.2 m on average. However, variation in the minimum required spatial resolution between peatlands ranged from 0.25 m in Roundstone to 0.7 m in Mongan (Figure 3; Figure S1). Minimum required spatial resolution indeed also strongly differed between different microform classes (Figure 5). For example, hummock had a minimum required spatial resolution of 0.65 m ± 0.2 m on average. In the case of Mongan, hummock was even consistently classified up until the coarsest spatial resolution of 1.0 m. Conversely, lawn and hollow had, respectively, ranged in minimum required spatial resolution between 0.20 and 0.55 m and 0.25 and 0.75 m. This was expected, because lawn and hollow in Irish peatlands are generally smaller and more elongated than hummocks (Figure S7). The variation in minimum required spatial resolution between microform classes partly coincided with their frequency of occurrence, hummock being the dominant microform in all peatlands (Figure S3; Figure S7).

Figure 3. Bar graphs visualizing the minimum required spatial resolution for consistent mapping of microform vegetation classes for each of the eight studied peatlands. “Overall” represents the average over all vegetation classes. Raheenmore is missing a bar for hollow because it was not present in the peatland.

3.1.2. Plant functional type classifications

Plant functional type classifications showed consistently high accuracies for all peatlands until a spatial resolution of 0.25 m ± 0.1 m on average (Figure 5), which is almost twice as detailed as those for microforms. Overall minimum spatial resolution for plant functional types varied notably less between peatlands than for microforms, ranging from 0.15 m in Moyclare to 0.35 m in Carrowbehy (Figure 4; Table S4). Like microforms, minimum required spatial resolutions for plant functional types varied strongly between vegetation classes. For instance, while peat moss was well classified up until 0.45 m ± 0.2 m spatial resolution (Figure 5) and ranged from 0.15 to 0.65 m, other plant functional types like shrub, graminoid, and lichen had notably higher minimum spatial resolution requirements between 0.15 and 0.30 m (Table S4). Water/bare peat had the coarsest required spatial resolutions of all plant functional types (Table S4), which can likely be attributed to the relatively large open water pools that were present in many of the studied peatlands (Figure S8). Nevertheless, the variation in minimum required spatial resolution between vegetation classes was only partly aligned with their occurrence frequency as shrub and graminoid were by far the most occurring plant functional types in all peatlands, followed by peat moss, lichen, and water/bare peat (Figure S4; Figure S8).

Figure 4. Bar graphs visualizing the minimum required spatial resolution for consistent mapping of plant functional type vegetation classes for each of the eight studied peatlands. “Overall” represents the average over all vegetation classes. Ferbane, Moyclare, and Raheenmore are missing a bar for water/bare peat because it was not present in their images.

Figure 5. Bar graphs visualizing the minimum required spatial resolutions for consistent classification of the studied mapping units: microform (left) and plant functional type (right). “Overall” represents the average over all vegetation classes within a mapping unit. Error bars represent standard deviation of minimum required spatial resolution over all eight peatlands.

3.2. The role of spatial vegetation characteristics

3.2.1. Microform characteristics

For microform classifications, the minimum spatial resolution per peatland was strongly related to the spatial organization of the microforms (Figure 6). We found strongly significant linear relationships (R² > 0.7, p < 0.05) between minimum spatial resolution and all spatial patch metrics, except for the Shannon diversity index, which showed a weak relationship. Specifically, minimum required spatial resolution decreased with an increase in mean and standard deviation of patch area (Figure 6; Table S5), suggesting that that coarser spatial resolution imagery can be used for mapping peatlands with larger microform patches and lower variability in microform patch area than those with smaller and more variable microform patches. Also, minimum required spatial resolution showed a strong negative relationship with the landscape shape index and conditional entropy (Figure 6; Table S5), indicating that peatlands with more compact, round, and clustered microform configuration such as Mongan, Raheenmore, and Derrinea can be classified consistently using imagery with coarser spatial resolutions than peatlands with less organized patch forms. Lastly, the Shannon diversity index showed only a moderate relationship with the minimum required spatial resolution of microforms (Table S5). This suggests that the minimum required spatial resolution for mapping microforms is less sensitive to variation in landscape-scale diversity and evenness than their size and shape metrics.

Figure 6. Ordinary least squares univariate regression lines for models assessing the relationship between spatial patch metrics and minimum required spatial resolution for microforms. a) Mean area (β = 0.1245, R² = 0.92, p < 0.0001), b) standard deviation of area (β = 0.0029, R² = 0.86, p < 0.0001), c) Shannon diversity index (β = −0.5857, R² = 0.57, p < 0.0020), d) landscape shape index (β = −0.0060, R² = 0.83, p < 0.0145), and e) conditional entropy (β = −1.7808, R² = 0.81, p < 0.0107). Legend with peatland data points is sorted by minimum required spatial resolution in descending order.

3.2.2. Plant functional type characteristics

In contrast with microforms, minimum spatial resolution requirements for plant functional types were unaffected by their spatial vegetation characteristics. Where microforms had strong linear relationships with four out of five of the computed spatial vegetation characteristics, minimum spatial resolution of plant functional types had only a weak negative relationship with mean patch area (Table S5). However, this negative relationship indicates that mapping plant functional types with a higher mean patch area requires imagery with lower pixel size, and plant functional types with a small mean patch area require imagery with a high pixel size.

4. Discussion

4.1. The relationship between spatial resolution requirements and vegetation patterning

4.1.1. Differences between vegetation mapping units

We found that microforms can be classified and mapped with coarser spatial resolution imagery than plant functional types (0.5 m versus 0.25 m on average, respectively), likely because of their larger and more variable size, shape, and configuration than the individual plant functional types of which they are composed. Indeed, shrub, lichen, and graminoid had among the same spatial resolution requirements with relatively low variability (Figure 5; Table S4). Nevertheless, some specific plant functional types such as peat moss and water/bare peat were consistently classified well until spatial resolutions were reached that were similar to the overall minimum required spatial resolution for microforms. The overall high minimum required spatial resolutions for peat moss (0.45 m) compared to other plant functional types is likely explained by its carpet-like growth form in both lawns and hummocks (van Breemen 1995; Waddington et al. 2015). Besides, peat mosses are often well distinguished because their spectral signature generally differs markedly from other plant functional types (Bubier, Rock, and Crill 1997; Harris and Bryant 2009b; McPartland et al. 2019; Schaepman-Strub et al. 2009). The blanket bogs in our study required higher spatial resolutions to classify peat mosses than the raised bogs (Figure 4), which likely resulted from the fact that blanket bogs generally have lower and more dispersed peat moss cover than raised bogs. Lastly, open water pools were also consistently classified well with coarser spatial resolutions than other plant functional types (0.5 m), which we attribute to the relatively large open water pools in some peatlands such as Bangor Erris and Carrowbehy. However, spatial resolution requirements for individual microforms were probably affected by the sampling strategy of our study. This is because area-proportional allocation of training/testing samples in combination with the structural dominance of hummocks in all studied peatlands but Roundstone led to increased sensitivity for misclassifications of the less-occurring lawn and hollow (Table S2; Figure S3, S4).

4.1.2. The role of spatial vegetation characteristics

We show that spatial resolution requirements of microforms were not only determined by their corresponding vegetation types but also by their spatial characteristics: peatlands that exhibit larger, compacter, more clustered, and less diverse microforms could be mapped at coarser spatial resolution imagery (up to 0.7 m) than peatlands with smaller and less organized microforms. This result has two main implications for mapping peatland functions through vegetation patterns: 1) a quick visual inspection of the size and organization of vegetation patterns may help to plan flight campaigns and subsequent analyses more efficiently, as reductions in resolution lead to notably quicker image capture and processing (Steenvoorden, Bartholomeus, and Limpens 2023), and 2) as microform configurations may change with peatland condition (Steenvoorden et al. 2022), mapping efforts targeting changes in peatland functioning for example as a result of degradation through time or recovery since restoration, should be adapted to the condition demanding the highest spatial resolution.

In contrast to microforms, the spatial resolution requirements of plant functional types showed no significant relationships with spatial patch metrics for plant functional types, but only a weak negative relationship with mean patch area. However, the direction of this relationship suggests that more spatial detail is needed if patches become larger, which is in contrast with our second hypothesis and with our microform results. We attribute the lack of strength for these relationships predominantly to the lack of range in spatial patch metrics of plant functional types among peatlands in our study, which on average were approximately three times smaller than those of microforms (see also Figure 7). While we argue that the absolute range in spatial vegetation characteristics of plant functional types is much lower than those of microforms, we cannot exclude that the limited range in spatial characteristics in our study was an artifact of our purposeful sampling approach. We took care to minimize artifacts by systematically sampling a variety of patches per plant functional type throughout the whole peatland (see also 2.4). We therefore believe the targeted sampling approach is not the main cause of the limited variation in spatial vegetation characteristics for plant functional types observed in our study. Based on these findings, we argue that spatial vegetation characteristics play a notable role in determining spatial resolution requirements for microforms, but not for plant functional types.

Figure 7. Ordinary least squares univariate regression lines for models assessing the relationship between spatial patch metrics and minimum required spatial resolution for plant functional types. a) Mean area (β = −0.1760, R² = 0.47, p < 0.0013), b) standard deviation of area (β = −0.0013, R² = 0.10, p < 0.0006), c) Shannon diversity index (β = 0.1666, R² = 0.19, p < 0.0001), d) landscape shape index (β = −0.0003, R² = 0.00, p < 0.0001), and e) conditional entropy (β = 0.2001, R² = 0.02, p < 0.0001). Legend with peatland data points is sorted by minimum required spatial resolution in ascending order. Please note that the x-axis of all subplots (a–e) in this graph has not been normalized and does not include a 0 value for each spatial patch characteristic. This was done because the range in values for each spatial patch characteristic was so small that the graph would lose visual clarity.

4.2. Implications for mapping, monitoring, and upscaling

4.2.1. The added value of drones for mapping peatland vegetation

On the one hand, our results show that drone imagery is probably not necessary to map all mapping units in heterogeneous landscapes like peatlands, because consistent classifications of specific vegetation classes were obtained in most peatlands using spatial resolutions that parallel those of the highest spatial resolution aerial imagery and pan-sharpened imagery of some commercial satellites, which is about 0.3 m. On the other hand, our study also suggests that the use of drones for mapping of fine-scale vegetation patterns and functions has several advantages over the use of aerial and satellite remote sensing.

First, simply not all microforms and/or plant functional types occur at a spatial scale that is currently visible on aerial or even the highest spatial resolution pan-sharpened satellite imagery. Consequently, mapping vegetation patterns on a landscape-scale using aerial and satellite imagery may miss spatial heterogeneity in some peatlands, whereas they can be captured using drone imagery.

Second, segmented patch boundaries at the highest spatial resolution imagery followed vegetation patterns in the field with remarkably higher accuracy than resampled imagery with spatial resolutions mimicking aerial and satellite imagery. Räsänen and Virtanen (2019) also found that spatial vegetation characteristics in peatlands were most complex when classifications were based on drone imagery segmentation. This increasing loss of spatial complexity at lower spatial resolutions is logical, because small patches of vegetation become aggregated into a rectangular pixel if spatial resolutions become coarser than the size of the vegetation class under investigation. This loss of spatial complexity may not be problematic if vegetation classes show adequate spectral differences with coarser spatial resolution imagery and/or if the spatial characteristics of vegetation are not the object of study. However, accurate information on spatial vegetation characteristics may become important if studied vegetation classes are more spectrally similar or if different vegetation classes show notable spatial variability in the landscape. For instance, hummocks are generally large, round, and more compact, while lawns and hollows are often smaller and more elongated. Our results show that differences in these spatial patterns can currently only be captured using segmentation of drone imagery.

Third, collection and processing of very high spatial and spectral resolution aerial imagery and lidar data are costly (Anderson and Gaston 2013; Lovitt, Rahman, and McDermid 2017). This often leads to areas with missing spatial data or areas containing spatial data with low temporal resolution. In these cases, and given the added value of topographical data in classification of fine-scale peatland vegetation patterns (Räsänen et al. 2020; Steenvoorden et al. 2022), drones offer a far cheaper and flexible solution to consistently collect both multispectral and topographical data in many areas with relatively limited extent. Besides, while we obtained consistent classifications of both microforms and plant functional types using our DJI Mavic 2 Pro with RGB camera sensor, several studies have shown the added value of a near-infrared (NIR) sensor on drones for discriminating different peatland vegetation types (Knoth et al. 2013; Räsänen et al. 2020). NIR imagery is also well-suited to separate vegetation from open water in peatlands because water tends to absorb NIR more than light in the visible spectrum. This can be especially valuable if open water pools are already partially filled in with vegetation, like in some of our studied peatlands. Consequently, the inclusion of NIR sensors in classification of microforms and plant functional types could reduce spatial resolution requirements of some vegetation classes that are sensitive to changes in NIR, like hollow, lawn and peat moss. Because NIR sensors are becoming increasingly available and affordable on consumer-grade drones, we recommend their inclusion for future studies investigating vegetation patterns in peatlands where possible.

Fourth, plant phenology is important in peatlands both because it indirectly affects the identification and differentiation of specific vegetation patterns (Cole, McMorrow, and Evans 2014) but also because it directly affects peatland carbon fluxes (Antala et al. 2022; Koebsch et al. 2020; Kross et al. 2014). While multispectral satellite imagery has shown potential for broadly capturing peatland phenology across the growing season (Arroyo-Mora et al. 2018; Linkosalmi et al. 2022), this type of data may not be available during time-periods of the year where phenological stages of plants are most optimal for their characterization. Synthetic Aperture Radar (SAR) could be valuable here as it can reach spatial resolutions of 1 m, can penetrate vegetation canopies, is sensitive to water and vegetation structure, and has the capacity to retrieve imagery in all weather conditions. However, the potential of SAR in discriminating fine-scale vegetation phenology has to our knowledge not yet been studied in peatlands. Besides, our results suggest that the ability to reliably determine the different phenological stages of microforms and plant functional types throughout a peatland may require imagery on finer spatial and/or temporal scales than both multispectral and radar satellite imagery can currently provide. The flexibility and spatial resolution of drone imagery thus proves an invaluable tool for mapping at ecosystem to landscape scales. Nevertheless, it should be noted that drones can also be limited by extreme weather conditions occurring in peatland areas during some time-periods of the year, including high wind, fog, rain, heat, and cold.

The maximum spatial scale at which drone imagery analyses could still be considered the most efficient approach over aerial or satellite imagery is obviously unclear and dependent on many factors, including the study area, the chosen drone, and flight and image-processing parameters. However, using our chosen flight parameters of 120 m flight altitude and 75/60 forward/side overlap, it most certainly reaches extents of up to 1000 hectares because our DJI Mavic 2 Pro could fly areas of about 265 hectares per hour. This means an area of about 1000 hectares can theoretically be flown within 1.5 hours of solar noon on the same day using most consumer-grade drones.

4.2.2. Transitioning to aerial and satellite imagery

Ultimately, the decision or necessity of transitioning from drone to aerial and/or satellite imagery when mapping and monitoring peatland vegetation patterns is affected by the scale of the study area, spatial characteristics and spectral separability of the peatland vegetation pattern under investigation, and the research goal. Spatial resolution requirements are likely lower in cases where the relative proportions of vegetation patterns within peatlands only need to be categorized broadly and there is no need to maximize classifier performance for all vegetation classes, or if peatlands are dominated by only few vegetation classes that are easy to detect. If the goal is to evaluate the role of vegetation as ecological indicators for broader ecosystem functioning, one could use remote sensing platforms with a spatial resolution equal or higher than the overall minimum required spatial resolution for that specific vegetation class. When using satellite imagery, one could exploit the fact that most Earth observation satellite missions with freely available imagery like Landsat and Sentinel-2 carry multispectral sensors with NIR and Shortwave Infrared (SWIR), which are of particular added value in discriminating important peatland vegetation classes like peat mosses, as well as open water (Bhatnagar et al. 2020, 2021; Harris and Bryant 2009a; Krankina et al. 2008). In contrast, when vegetation maps of individual peatlands are linked to their biogeochemical cycles such as carbon fluxes, the relative proportions of all existing vegetation classes need to be calculated with high accuracy because different microforms and their associated plant functional types often have markedly contrasting rates of carbon accumulation and decomposition (Loisel and Yu 2013; Lunt, Fyfe, and Tappin 2019; Riutta, Laine, and Tuittila 2007; Strack et al. 2006; Waddington and Roulet 1996). When fluxes are related to maps of plant functional types or plant phenology, one would likely need drone imagery, but when fluxes are coupled to microforms, aerial imagery and very-high resolution satellite imagery (~0.3 m) can be sufficient.

Given the large, isolated, and inaccessible nature of many peatlands, upscaling of remote sensing analyses in these ecosystems will require methods that combine the strengths of consumer-grade drones (e.g. potential for very high spatio-temporal resolution, flexibility) with those of aerial and satellite remote sensing (e.g. regional and global coverage, multiple sensors). Drones could be used here to train, test, validate, and calibrate coarser spatial resolution but larger-scale commercial aerial and satellite imagery (Bhatnagar et al. 2021; Riihimäki, Luoto, and Heiskanen 2019). While some recent studies show the added value of so-called nested drone-aerial or drone-satellite mapping approaches in peatlands (Bhatnagar et al. 2021; Carless et al. 2019; Connolly, Holden, and Ward 2007; Räsänen and Virtanen 2019), consistent development of these methodologies is only just emerging. Besides, although some of these nested drone-aerial or drone-satellite mapping approaches may require some radiometric calibration, we justify the use of resampled drone imagery as a proxy for aerial and satellite imagery because RGB reflectance values and vegetation indices using NIR between both remote sensing platforms often have very high correlation (Assmann et al. 2020; Fawcett et al. 2020; Jiang et al. 2022; Padró et al. 2018).

5. Conclusion

Our study shows that minimum required spatial resolutions for mapping peatland microforms and plant functional types range from 0.1 to 1 m and are variable across peatlands as well as between mapping units and their individual vegetation classes. We also found that spatial vegetation characteristics were important drivers of minimum spatial resolution for microforms, but not for plant functional types, which required imagery of at least 0.25 m spatial resolution in all peatlands on average. Based on these findings, we conclude that spatial vegetation characteristics strongly affect upscaling of peatland vegetation mapping beyond the landscape scale by constraining the use of specific remote sensing platforms.

Author contributions

JL and JS conceived ideas for this research. JS and JL designed methodology and JS collected drone imagery in September 2022. Data were analyzed by JS and interpreted together with JL. Writing of the manuscript was led by JS. All authors contributed to the draft and gave final approval for publication.

Acknowledgments

We thank the Irish National Parks and Wildlife Service (NPWS) for providing access to the studied peatlands, and Harm Bartholomeus, Daniel Kooij, Jochem van der Zaag, and Rúna Magnússon for insightful discussions regarding the methodology, interpretation of results, and visualization of relevant data.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

All data used in this study will be uploaded to open-access database DANS-EASY upon publication.

Supplementary material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/15481603.2023.2267851