Intensive field sampling increases the known extent of carbon-rich Amazonian peatland pole forests

The study was carried out across 80 sites of the PMFB, located in the region of Loreto, in northern Peruvian Amazonia (figure 1). Three categories of field data were used (See supplementary material, table S1 (available online at stacks.iop.org/ERL/16/074048/mmedia)): (a) 571 ground reference points for remote sensing classification; (b) 102 forest plots for floristic analysis and biomass estimation; and (c) 217 measurements of carbon content and dry bulk density to quantify soil organic carbon stocks. For all ground reference points, the ecosystem type (OP: open savanna-like peatland, PF: pole forest, PS: palm swamp, SF: seasonally flooded forest, TF: terra firme, and WS: white-sand forest) was assessed following the criteria described in table S2 and peat thickness was measured following the methodology described below. The datasets included 47 sites from previous studies of [1, 2, 4, 21] and 33 new sampling sites from this study established by the MonANPeru network and the Tropical Wetlands Consortium.

Zoom In Zoom Out Reset image size

A stack of 11 remote sensing layers was used in the landcover classification model. A cloud free Sentinel 2 (S2) mosaic containing all surface reflectance bands was produced in Google Earth Engine (GEE). Band 11 (swir1) and four indices calculated from the S2 mosaic (normalized difference vegetation index [NDVI], normalized difference water index, red-Edge NDVI, and enhanced vegetation index) were retained in the final version of the model. L-band SAR annual mosaics from the ALOS-PALSAR and ALOS-PALSAR2 sensors for the years 2007–2010 and 2014–2018 (bands HH, HV and HH/HV), and SRTM 30 m digital elevation were directly imported from the GEE repository. Additionally, two satellite-derived products were used: the national map of Peruvian ecosystems [22] and a maximum inundation extent derived from the map of wetland extent at highwater [23]. Further details on the processing of the stack can be found in the supplementary information.

A supervised random forest (RF) classification (300 trees, three variables per split) of the Loreto region was undertaken in GEE. We estimated the distribution of eight landcover classes including six ecosystem types (OP, PF, PS, SF, TF, WS), open water (WA), and urban areas/river beaches (UB). For model training and validation, we created a total of 420 polygons based on the 571 ground reference points, as well as additional published polygons [2] and data [24, 25], and unpublished data of known ecosystem types (figure S1). To estimate model accuracy, 50% of the polygons were used to train the RF algorithm while the remaining half were used to validate the model predictions. In the production of the final map and the resulting total areas of each landcover class (figure 2), we used 100% of the ground data for model training. This was done due to the relatively small number of ground reference points. See supplementary information for more details.

Zoom In Zoom Out Reset image size

We estimated the area of each ecosystem type within the protected area network managed by the National Protected Areas Authority (SERNANP) by overlaying the landcover classes with the maps of national, regional and private protected areas and reserved zones [26], as well as with indigenous communities' territories [27].

We assessed the floristic composition and biomass of all individuals with diameter at breast height (dbh at 1.3 m height) ⩾10 cm in the 102 forest plots. Plot size varies from 0.1 to 1 ha (mean = 0.48 ha) with 36 small circular plots (0.1 and 0.2 ha) established on low diversity ecosystem types while rectangular (50 × 100 m) and square (100 × 100 m) plots were preferred in more diverse ecosystems (table S1). Plot establishment and tree measurement followed the standard methods described in the RAINFOR protocol (www.forestplots.net/es/in-the-field) and the data are managed at ForestPlots.net [28], except for 12 forest plots established by [21]. Arborescent palm species were identified in the field and a photograph was taken for species confirmation. Botanical samples of trees were collected for each species and deposited at the Herbarium Herrerense (HH) in Iquitos. All herbarium vouchers were standardised across all forest plots. The dataset included 30 484 individuals of which 3.7% were not collected in the field and were excluded from analyses of floristic composition.

To compare the floristic composition among plots and ecosystem types identified in the field, a floristic distance matrix was constructed using the Bray–Curtis distance for relative abundance data using the 'vegan' package [29] in R [30]. This matrix was used to create non-metric multidimensional scaling (NMDS) ordinations optimized for three axes to visualise floristic dissimilarity among ecosystem types. A non-parametric permutational multivariate analysis of variance (PERMANOVA) was conducted using the Bray–Curtis distance matrix as the response variable to test the significance of the variation in species composition among the six ecosystem types and the four river basins [31]. Dominant species were identified based on their relative abundance and frequency in each vegetation type. For this, the sum of all individuals per species and their frequency of occurrence, i.e. the number of plots that the species occurs in each vegetation type, were calculated and expressed as percentage values.

Above-ground biomass and below-ground biomass of roots were estimated using published allometric equations for trees and palms (table 1) and the measurements of diameter and total height taken in the forest plots. The diameter of all stems was measured using a diameter tape and the total height of palms was measured using a clinometer. In the case of trees, total height was measured to a sample of 2174 stems: ∼50 stems randomly sampled by diameter class in each 0.5 ha and 1 ha plot and the three tallest trees in each 0.1 ha plot (figure S2) and these data were used to develop an equation specific to each ecosystem type in order to estimate the height of all stems (table 1). Tree wood density was obtained from [32, 33] and the dry mass fraction of palms was obtained from [34]. Biomass values expressed as dry weight were converted into megagram per hectare (Mg ha⁻¹) and then to carbon content (Mg C ha⁻¹) using the fraction of 0.47 as the average carbon content of wood in angiosperms [35].

Peat thickness was measured in all 571 reference points using a Russian-type coring device [36] by sampling inside the 102 forest plots and every 200–500 m along 53 1–5 km-long transects (table S1): 206 measurements were previously taken by [1, 4], 107 measurements by [2], 12 by [21], and 246 new measurements were taken as part of this study (figure 1). At each sampling point, the ecosystem type was registered and the coordinates were recorded with a handheld GPS. The base of the peat was identified visually using the Troels-Smith protocol [37]. Peat was identified as partially decomposed, wet, dark brown substrate in a layer at least 30 cm thick. Peat thickness was measured from the surface to the point of the transition between peat and underlying inorganic sediment (pale grey sandy to silty clay), including only pure peat and clayey peat.

Peat cores, taken in 50 cm sections along the peat profile, were analysed for total carbon content and dry bulk density, resulting in 217 measurements (table S1) taken at 13 sites by [1, 4] and ten additional sites by this study in the Tigre river basin using the same method. Laboratory protocols and conditions varied among studies. In this study, the 56 new peat samples were analysed at the University of St Andrews; protocols are provided in the supplementary material (tables S3 and S4).

SOC (Mg C ha⁻¹) was estimated for all 571 measurement points using equations that predict SOC at different peat thickness. The equations were developed using a linear model of the 217 SOC measurements based on dry bulk density and carbon content measured along the profile of the peat cores described above (figure S3). Due to the many zeros recorded in peat thickness for non-peat forming ecosystems, SOC was estimated to a depth of 2 m, by multiplying the above-ground carbon densities of the SF, TF and WS forest plots by 0.824 [38].

Across Loreto, we estimated a total peatland pole forest area of 7540 km² and a total potential peat-forming vegetation (i.e. OP, PF, and PS) area of 55 766 km² (figure 2). The landcover classification produces an overall model accuracy (the sum of correctly classified pixels divided by the total number of pixels across all classes) of 96% and a kappa coefficient (coefficient of agreement accounting for agreement occurring by chance) of 0.95 when tested against an independent dataset. Both producer's and user's accuracy are highly variable across classes, ranging from 63% to 100% and 39% to 100% respectively (table 2). We estimated that 24% of the area of the six ecosystem types is contained in protected areas managed by SERNANP, and an additional 16% is in titled land owned by indigenous communities (table 3). Peatland pole forest is mainly located in land titled to indigenous communities of the Tigre river (e.g. Nueva York, Piura).

Overall, the floristic dataset contained 1632 species and morphospecies, of which 460 species were found in peat-forming ecosystems. The floristic composition showed significant variation among ecosystem types (PERMANOVA, p < 0.001; figure 3). Peatland pole forests were dominated, in terms of relative abundance and frequency, by tree species such as Pachira nitida (Malvaceae), Platycarpum loretense (Rubiaceae), and Hevea guianensis (Euphorbiaceae), while the palm species Mauritia flexuosa (Arecaceae) was the most dominant species in palm swamps and open peatlands (figure 4). A more diverse set of species dominated non-peatland ecosystem types.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Peatland pole forests were characterized by the high abundance of trees but low abundance of palms, with trees contributing more than palms to above-ground biomass. Trees were also abundant in seasonally flooded forest, terra firme and white-sand forest and made the greatest contribution to above-ground biomass in these three ecosystem types. Palms were more abundant and made the largest contribution to above-ground biomass in open peatlands and palm swamps compared to the other ecosystem types (figure 5). Thin stems were abundant in pole forest and white-sand forest, which both had smaller mean diameter per plot (17–18 cm) than the other ecosystem types (21–25 cm), while mean total height per plot was not different among ecosystem types (table 3).

Zoom In Zoom Out Reset image size

Peatland pole forests have high mean peat thickness (with a mean value ± 95% CI of 274 ± 22 cm) similar to open peatlands (282 ± 46 cm), but higher than palm swamps (161 ± 17 cm) and seasonally flooded forests (24 ± 13 cm). Peat was absent from terra firme and white-sand forests (table 3 and figure 6(a)). Sites in the Marañón and Tigre river basins had significantly greater peat thickness than sites in the Amazon and Ucayali river basins (Kruskal–Wallis, X² = 54.5, df = 3, p < 0.001; figure 6(b)).Carbon content and bulk density were negatively correlated (Pearson's r = − 0.76, t = −6.71, df = 33, p < 0.001) and showed a high level of variation within each ecosystem type (figure S3).

Zoom In Zoom Out Reset image size

Peatland pole forest stores the greatest total carbon density among the six ecosystem types, with 78.24 ± 13.17 Mg C ha⁻¹ above ground, 21.18 ± 3.08 Mg C ha⁻¹ in roots of living trees and 1033.75 ± 91.57 Mg C ha⁻¹ in the peat (table 3). Average above-ground carbon was found to vary from 41.26 ± 22.71 Mg C ha⁻¹ in open peatlands to 123.52 ± 10.91 in terra firme forest. Peat-forming ecosystem types store on average 80%–95% of the total carbon density below ground, mainly in the peat, while the other ecosystem types store 51%–56% below ground, mainly in the soil.

Our results suggest that the geographical range of pole forest is substantially greater than previously estimated [2], with a wider distribution along the Tigre river (figure S4) and new scattered areas along the Napo and Tapiche rivers. Across the PMFB, we estimate a total pole forest area of 5942 km² (table S5), which represents a 61% increase on the 3686 km² previously estimated [2]. Furthermore, our estimate of total carbon density in peatland pole forest (1133.17 ± 92.56 Mg C ha⁻¹) was the highest among the six ecosystem types of this study.

Peatland pole forest has been poorly classified and mapped in previous studies for two reasons. Firstly, the lack of field data has meant that there has been substantial confusion over the name and nature of this ecosystem type. The term peatland pole forest was introduced in English in 2014 [2, 3] and translated in Spanish as varillal hidromórfico [15, 16]. Other studies referred to this ecosystem type by various names in both English and Spanish, such as forested peatlands [1, 4], varillal de turbera or stunted peatland forest [39], or bosque húmedo de varillal [40]. At a national level, pole forest or varillal hidromórfico was ignored in the 2019 ecosystem map of Peru [22].

Secondly, mapping the distribution of peatland pole forest in satellite imagery is a significant challenge. Our model still had some difficulty in discriminating between pole forest and palm swamp (table 2), which could be caused by a variety of reasons. One likely cause is that most new data used in model training and testing were collected across transects with the aim of crossing various vegetation types. For this reason, we sampled transitional areas located in between two ecosystem types. This highlights an inherent limitation in classifying landcover types which vary continuously along a spectrum, into distinct categories. Both ecosystem types had relatively low accuracy values in our map (table 2), which is likely to be at least in part, due to our limited ground data for these classes (PF: 972 pixels; PS: 1730 pixels) compared to most other ecosystem types (>3000 pixels, except for WS forest with 152 pixels). Additionally, the small stature of the dominant tree species may make pole forests appear similar to white-sand forests in structurally sensitive data (e.g. ALOS PALSAR). However, we observed very few pixels of pole forest misclassified as white-sand forest in the model, and despite relatively few ground data and a low user's accuracy, the distribution of white-sand forests is visually very similar to the distribution observed in the field in previous studies [24, 25, 39]. The lower extent of peatland pole forest estimated by [2] can be at least partially attributed to the lack of ground reference data (17 polygons compared with 52 here). In a preliminary study using PALSAR, Landsat 7 and STRM data, more extensive areas of peatland pole forest were predicted across the PMFB (e.g. ∼16 000 km²; [41]), though it should be noted that this study was based on a more limited set of field data. Additional remote sensing products may offer further improvements in our ability to discriminate pole forests from other ecosystem types. For example, high-resolution satellite imagery such as RapidEye with a resolution of 5 m have been used to successfully discriminate the stunted structure of pole forests at small scales [42].

Our floristic analysis shows that pole forest sites are floristically distinct from palm swamps and all other ecosystem types of Loreto, consistent with the findings of [3]. Pole forests were usually dominated by thin-stemmed trees of Pachira nitida (Malvaceae), Platycarpum loretense (Rubiaceae), and Hevea guianensis (Euphorbiaceae) in the four river basins. While these three most dominant species occurred in most pole forest plots (figure 4), we observed that other tree species could also be the most abundant species at individual pole forest sites, such as Ilex laureola (Aquifoliaceae) in CUH-12 (22% of stems) and CUH-01 (15%), Iryanthera paradoxa (Myristicaceae) in FLR-11 (19%), Virola duckei (Myristicaceae) in RB-7 (23%) and MLV-11 (18%), Qualea acuminata (Vochysiaceae) in MOV-12 (45%), Sagotia racemosa (Euphorbiaceae) in NYO-02 (18%), Macrolobium GFL5604 (Fabaceae) in OLL-01 (19%). Pole forest plots are found mainly in the Tigre basin and variation in composition may reflect the existence of a diversity of successional pathways that are strongly influenced by priority effects [43]: species that happen to disperse to and establish in the ombrotrophic conditions created by the accumulation of peat during the natural dynamics of these ecosystems may achieve high local abundance. Further studies in the Tigre basin may enable us to understand whether, and to what extent this variation in floristic composition corresponds to changes from/to minerotrophic (river, run-off and ground water fed) to/from ombrotrophic (rain fed) conditions (e.g. [44]).

Pole forest plots had low species diversity as previously reported by [2, 3] and this ecosystem type includes tree species that remain undescribed. For example, a three-foliolate species of the genus Macrolobium (Leguminosae; GFL5604) was surprisingly abundant in the plots and determined as likely new to science. We also corrected the identification of the most abundant species, Pachira nitida, which was previously wrongly identified as Pachira brevipes [3]. Pachira brevipes was only present in white-sand forest plots in this study. Interestingly, palms did not contribute substantially to the abundance of individuals or biomass of pole forests, except for four plots in the Marañón basin (OLL-03, OLL-04, ROQ-05, and WEM-01) where the slender, cespitose palm Mauritiella armata was the most abundant species (figure 4). Mauritiella armata was highly abundant also in one of five swamps studied in Ecuador [45] and common in our plots in palm swamp sites. Due to the absence or rare occurrence of Mauritia flexuosa in these plots, the vegetation classification of these four forest plots remains uncertain.

Previous studies have suggested that some of the pole forest sites in the Tigre basin may be degraded palm swamps, rather than distinctive natural ecosystems [21, 46]. Here, we found no evidence that these peatland pole forests are degraded based on the similar gender ratio of populations of Mauritia flexuosa between peatland pole forests and palm swamps. The gender ratio of M. flexuosa populations is a direct measure of degradation of these ecosystems as cutting female stems of this dioecious palm for fruit harvesting leads to strongly male-skewed populations [8, 10]. The dataset from the Tigre basin (table 4) shows that the mean male to female ratio of M. flexuosa adult individuals in pole forests (mean ± 95% CI: 1.9 ± 0.5; n = 13 sites) is not significantly higher than palm swamps (4.4 ± 3.5; n = 6 sites). Indeed, the highest levels of degradation (indicated by a male:female sex ratio >10) caused by the cutting of female palms during fruit harvesting were recorded in palm swamps in Monteverde and Nueva York in the Tigre basin, rather than the pole forest sites (table 4).Visible hunting trails inside the pole forest plots, rather than cut stems, suggest that hunting is the principal use of these ecosystems. This finding is consistent with a palaeoecological study of the vegetation dynamics in peatlands where the floristic composition of the pole forest represents a natural, late successional stage of a swamp [44]. In this example, pole forest developed around 200 years ago after a 1700 year history as a palm swamp with no indicators of human activity throughout this period. In summary, our results present evidence that peatland pole forests have not been subject to degradation through fruit harvesting. Therefore, studies that label these intact ecosystems as degraded forests are erroneously downplaying their conservation value.

The current low level of protection of peatland pole forests—only 15% of pole forest area is within a protected area—represents an opportunity for conservation. In particular, our results demonstrate that the Tigre river basin is a priority for protection due to the presence of extensive pole forests, with low levels of degradation, that store large amounts of carbon. The need to develop effective conservation strategies in the Tigre basin has recently been highlighted by the threats to this region from planned road paving projects to connect Iquitos to the national road network [15–17, 47]. Implementing conservation and sustainable management projects (e.g. M. flexuosa fruit harvesting with the use of climbing equipment to prevent cutting of female palms [48, 49], handicrafts, ecotourism, sustainable fisheries) with the Kichwa indigenous and riverine communities in the Tigre basin is one route for conservation. Such projects should also reinforce the cultural values, knowledge, and traditions of indigenous communities [6, 11]. Promoting sustainable management of resources and avoiding extensive agriculture or ill-planned infrastructure in peatlands is also important for preventing the emission of greenhouse gases from forest degradation and deforestation [16, 20, 50]. All these efforts would contribute to reducing national carbon emissions and could be proposed as activities that will fulfil nationally determined contributions and international commitments to mitigate climate change (e.g. [50]).