Geoderma 272 (2016) 20–31
Contents lists available at ScienceDirect
Geoderma journal homepage: www.elsevier.com/locate/geoderma
Digital mapping for cost-effective and accurate prediction of the depth and carbon stocks in Indonesian peatlands Rudiyanto a,b,⁎, Budiman Minasny b,⁎, Budi Indra Setiawan a,⁎, Chusnul Arif a, Satyanto Krido Saptomo a, Yudi Chadirin a a b
Department of Civil and Environmental Engineering, Bogor Agricultural University, Indonesia Faculty of Agriculture and Environment, The University of Sydney, New South Wales 2006, Australia
a r t i c l e
i n f o
Article history: Received 5 January 2016 Received in revised form 23 February 2016 Accepted 24 February 2016 Available online xxxx Keywords: Digital soil mapping Tropical peatlands Peat depth Regression tree Random Forests Artiﬁcial Neural Network
a b s t r a c t Tropical peatlands have an important role in the global carbon cycle. In order to quantify carbon stock for peatland management and conservation, the knowledge of the spatial distribution of peat and its depth is essential. This paper proposed a cost-effective and accurate methodology for mapping peat depth and carbon stocks in Indonesia. The method, based on the scorpan spatial soil prediction function framework, was tested in Ogan Komering Ilir, South Sumatra and Katingan, Central Kalimantan. A peat hydrological unit, where a peatland is bounded by at least two rivers, is deﬁned as the mapping area or extent. Peat depth is modelled as a function of topography and spatial position. Four machine learning models were evaluated to model and map peat depth: Cubist regression tree, Random Forests (RF), Quantile Regression Forests (QRF) and Artiﬁcial Neural Network (ANN). Covariates representing topography and spatial position were derived from the 1 arc-second digital elevation model (DEM) of the Shuttle Radar Topography Mission (SRTM) (resolution of 30.7 m). The spatial models were calibrated from ﬁeld observations. For model calibration and uncertainty analysis, the k-fold cross validation approach was used. Three models: Cubist, Random Forests, and Quantile Regression Forests models showed excellent accuracies of peat depth prediction for both areas where the coefﬁcient of determination values range from 0.67 to 0.92 and root mean squared error (RMSE) values range from 0.6 to 1.1 m. ANN showed inferior results. In addition, QRF and Cubist showed the best account of the uncertainty of prediction, in terms of percentage of observations that fall within the deﬁned 90% conﬁdence interval. In terms of the best predictor, elevation comes ﬁrst. Using the spatial prediction functions, peat depth maps along with their 90% conﬁdence interval were generated. The estimated mean carbon stock for Ogan Komering Ilir is 0.474 Gt and for Katingan is 0.123 Gt. Our estimate for Ogan Komering Ilir is twice larger than a previous study because we mapped the peatland hydrological unit, while the previous study only delineated peat domes. Finally, we recommend a sampling method for peat depth mapping using numerical stratiﬁcation of elevation to cover both the geographical and covariate space. We expect that the combination of an improved sampling strategy, machine learning models, and kriging will increase the accuracy of peat depth mapping. © 2016 Elsevier B.V. All rights reserved.
1. Introduction Tropical peatlands are characterized by an accumulation of partially decomposed organic matter in the water-saturated and anaerobic environment for a long period. Although the bulk density of peat is relatively low, ranging between 0.013 and 0.3 g cm−3 (Agus et al., 2011; Lähteenoja and Page, 2011; Rudiyanto et al., 2016), its carbon content is very high, ranging between 23 and 62% on mass basis (Farmer et al., 2014; Rudiyanto et al., 2016; Warren et al., 2012). In addition, tropical
⁎ Corresponding authors at: Department of Civil and Environmental Engineering, Bogor Agricultural University, Kampus IPB Darmaga, PO. BOX 220, Bogor 16002, Indonesia. E-mail addresses: [email protected]
(Rudiyanto), [email protected]
(B. Minasny), [email protected]
, [email protected]
http://dx.doi.org/10.1016/j.geoderma.2016.02.026 0016-7061/© 2016 Elsevier B.V. All rights reserved.
peats can accumulate up to a thickness of 20 m (Page et al., 2011). As a result, peatlands store the largest amount of terrestrial carbon per unit area, and it plays an important role in global carbon cycle as a carbon sink, although naturally, they also release two greenhouse gases (GHG) into the atmosphere: carbon dioxide (CO2) and methane (CH4) (Farmer et al., 2011; Sjögersten et al., 2014; Wright et al., 2011, 2013). This fact was supported by Page et al. (2011) who estimated that carbon stock in global peatlands stores between 480 and 610 Gt (Giga tonnes) or 15 to 30% of the world's carbon stock (Hugron et al., 2013), albeit global peatlands only cover between 3.8 and 4.1 million km2, with the best estimate of 3,971,895 km2 or about 2.5 to 3% of the whole lands of the earth (Page et al., 2011). Within the context of Reduce Emissions from Deforestation and Degradation (REDD+), peatlands become the main concern in the measurement, reporting and veriﬁcation (MRV) system which documents
Rudiyanto et al. / Geoderma 272 (2016) 20–31
carbon stocks and its change. Inventory of carbon in peatlands generally is calculated from the dot product of carbon content, bulk density and peat depth (Akumu and McLaughlin, 2014; Fell et al., 2016; Holden and Connolly, 2011; Parry and Charman, 2013). Range values of carbon content and bulk density have been well studied; however peat depth shows a high spatial variation. Thus the presence of accurate peat depth map is important for reliable estimate carbon stock in peatlands. The peat depth map is also required for decision policy of sustainable peatland management (Biancalani and Avagyan, 2014) and for a better understanding of peatland development (Bauer and Vitt, 2011; Esterle and Ferm, 1994) as well as its ecosystem function (Joosten and Clarke, 2002). Recently the Indonesian government has released the Regulation No 71, 2014 on the Protection of Peat Ecosystem. This regulation states that within a peatland hydrological unit if 30% of the area has a peat depth more than 3 m (considered as a peat dome) and located in the river upstream, and then it should be considered as an area of conservation. In Indonesia, peatland is estimated to cover 206, 950 km2 (Page et al., 2011), while (Ritung et al., (2011, 2012)) estimated that peatlands in the 3 main islands: Sumatra, Kalimantan, and Papua, cover 149 ,056 km2. Nevertheless, there are still much uncertainty in these ﬁgures (Hooijer and Vernimmen, 2013). Moreover, the peatlands covered a relatively large area and located at remote sites which are difﬁcult to access. Therefore, mapping peat depth remains a big challenge. Past studies on mapping peat depth commonly used kriging interpolation (Akumu and McLaughlin, 2014; Altdorff et al., 2016; Bauer et al., 2003; Jaenicke et al., 2008; Keaney et al., 2013; Proulx-McInnis et al., 2013; van Bellen et al., 2011; Weissert and Disney, 2013); however to produce a high spatial resolution map, it needs a large number of observations evenly spread throughout the area. Other works used spatial models to predict peat depth from proxy environmental information such as terrain attributes (e.g., elevation and slope) (Holden and Connolly, 2011; Parry et al., 2012). Other models include: a peat depth inference model (Holden and Connolly, 2011), a power function of the closest distance to a river (Hooijer and Vernimmen, 2013), an exponential function of elevation and slope (Parry et al., 2012), and an empirical function of elevation (Rudiyanto et al., 2015). The accuracy of these published models is usually moderate. Remote and proximal sensors such gamma radiometer, GPR, electromagnetic induction, and LiDAR have been proposed for mapping peat depth (Fyfe et al., 2014; Keaney et al., 2013; Koszinski et al., 2015; Parry et al., 2014; Rosa et al., 2009). These instruments produce highresolution data; however, they still need ground data for calibration, and may not be feasible in remote areas, furthermore, the high cost of acquiring these data does not allow a wide application. Digital soil mapping (DSM) has been successfully applied to map carbon content of mineral soils evidenced by a large number of publications in recent years (McBratney et al., 2003; Minasny and McBratney, 2015). The advances of DSM are mainly supported by the availability high-quality covariates as well as the development of machine learning algorithms such as: Random Forests (Breiman, 2001), Cubist tree model (Quinlan, 1992, 1993a,b), Artiﬁcial Neural Networks (Bishop, 1995; Günther and Fritsch, 2010), and regression kriging (Hengl et al., 2004; Odeh et al., 1994; Odeh et al., 1995). These models have been successfully applied in digital mapping of soil organic carbon (Aitkenhead and Coull, 2016; Aitkenhead et al., 2015; Grimm et al., 2008; Page et al., 2004; Song et al., 2016; Wiesmeier et al., 2011), soil physicochemical properties (e.g., bulk density, plant available water capacity, saturated hydraulic conductivity, pH, chemical concentration) (Malone et al., 2009; Motaghian and Mohammadi, 2011; Odgers et al., 2015), soil texture (i.e., percentage of sand, silt and clay) (Adhikari et al., 2013; Ballabio et al., 2016; Ließ et al., 2012), soil classes (Heung et al., 2016; Pahlavan Rad et al., 2014; Taghizadeh-Mehrjardi et al., 2015), soil parent material (Heung et al., 2014), etc. This paper seeks for a cost-effective and accurate method for mapping peat depth in Indonesia. Digital mapping techniques for peatlands
at a relatively high resolution (30 m) using widely available covariates were proposed. Four machine learning models: Cubist regression tree, Random Forests (RF), Quantile Regression Forests (QRF) and Artiﬁcial Neural Network (ANN) were evaluated and tested for peat depth modelling, mapping and uncertainty estimates. The models were tested in two peatlands in Sumatra and Kalimantan. The peat depth map combined with carbon density estimates were used to derive carbon stocks in these peatlands. In addition, based on important predictors found in the regression models, the best sampling strategy for peat depth mapping will be recommended.
2. Materials and methods 2.1. Study area Tropical peatlands in Indonesia mostly can be found along the east coast of Sumatra and in parts of Kalimantan (Page et al., 2006, 2007). This study was conducted in two peatlands in the two islands (Fig. 1a). Since the formation of peatlands depends on nutrient and oxygen availability which was controlled by ﬂooding from rivers (Anderson, 1961, 1964; Esterle and Ferm, 1994), the mapping area or extent for both peatlands was deﬁned based on the smallest unit of a peatland, namely the hydrological peatland unit, where a peatland is bounded between two rivers or sea. This is in contrast with the mapping extent deﬁned by Jaenicke et al. (2008) where they only delineated areas considered as a peat dome. The ﬁrst peatland is located in Ogan Komering Ilir (OKI) District in South Sumatra which covers latitude: S2°23′24.251″ to S3°25′16.313″ and longitude: E105°10′27.152″ to E106°5′37.075″. The total area is about 610,311 ha and bordered by Riding river in the west and Lumpur river in the south and Bangka strait in the northeast (Fig. 1b, left). The main landuse is forest plantation, and some parts are conserved. This area includes two peat domes studied by Jaenicke et al. (2008): Air Sugihan and Teluk Pulai. The second area is in Katingan District, Central Kalimantan and lays on S1°52′7.887″ to S2°6′50.785″ and E113°18′ 46.341″ to E113°51′39.727″. The total area is about 93,257 ha and bordered by Katingan river in the west and Rungan river in the east (Fig. 1b, right), mainly used as a conservation area. Note that, the study area in Katingan does not cover the whole peatland because ﬁeld observations were only limited to parts of the area. Hereafter, we refer the two peatlands as OKI and Katingan, respectively.
2.2. Collecting ﬁeld data Field data collections were carried out between years 2007 and 2009 in OKI. Peat depth data were obtained from ﬁeld surveys of drilling using the Eijkelkamp peat auger. The observations were based on transects commonly used in peat surveys with a distance between observations of about 100 to 1000 m (Fig. 1b left). Peat depth is deﬁned as the depth from the surface until the depth where mineral soil layer is found (Agus et al., 2011; Page et al., 1999). At each location of drilling, the geographical coordinates were recorded using a global positioning system (GPS). In addition, data from South Sumatra Forest Fire Management Project (SSFFMP) collected in 2005 (Prayitno and Bakri, 2005) were also included in the Ogan Komering Ilir dataset. For Katingan, the observed peat depths were obtained from published data of Boehm and Frank (2008) where the peat drillings were done in years 2006 and 2007 (Fig. 1b right) and Shimada et al. (2001). Fig. 2a and b shows the histogram of peat depth data for OKI and Katingan, respectively. In total, 840 observations were collected from OKI with peat depth ranging between 0 and 7.1 m with median = 1.9 m and 121 observations were obtained from Katingan ranging between 0 and 9 m with median = 1.39 m.
Rudiyanto et al. / Geoderma 272 (2016) 20–31
Fig. 1. (a) Study area in two peatlands located at Ogan Komering Ilir, South Sumatra (orange polygon) and Katingan, Central Kalimantan, Indonesia (pink polygon) and (b) digital elevation and locations of sampling points of peat depth surveys. (For interpretation of the references to color in this ﬁgure legend, the reader is referred to the web version of this article.)
2.3. Laboratory measurements We only have carbon content and bulk density measurement of six peat samples taken from the depth of 0–10 cm at three locations in
OKI that was analysed in Laboratorium Sumberdaya Tanah Terpadu Balai Penelitian Tanah, Bogor, MOA (Integrated Land Resource Laboratory belongs to the ministry of agriculture in Bogor, West Java). To account for bulk density at different depths, we included 14
Rudiyanto et al. / Geoderma 272 (2016) 20–31
Fig. 2. Histogram of observed peat depth at (a) Ogan Komering Ilir, South Sumatra and (b) Katingan, Central Kalimantan.
observations of bulk density from the depth of 0–100 cm in this peatland from the study of Sumawinata et al. (2014) and four observations taken from the depth of 0–40 cm depth by Brady (1997). For Katingan, 24 published data of bulk density were obtained from Boehm and Frank (2008) where samples were collected from 0 to 50 cm depth and layers above mineral soils. The bulk density was measured using the volume-gravimetric method. 2.4. Environmental covariates The freely-available 1 arc-second DEM from the Shuttle Radar Topography Mission (SRTM) was downloaded from http://earthexplorer. usgs.gov/. The DEM has a spatial resolution of 30.7 m. Since the DEM was not corrected for vegetation height, it has a substantive amount of noise in the data. Therefore, the DEM was ﬁltered using a simple window averaging ﬁlter in the SAGA GIS software. The results are shown in Fig. 1b-left for peatland in OKI and Fig. 1b-right for peatlands in Katingan. Subsequently, from the ﬁltered DEM, several terrain attributes were derived, such as slope, aspect, and Saga wetness index. Since a peat hydrological unit was bounded by rivers or sea, we included the nearest distance to the river as a covariate. Based on GPS coordinates of peat depth observations, those ﬁve covariates were extracted. These covariates will be used for training and testing spatial models as described in the next section. The ﬁve covariates' description and range of values from the observation locations are summarized in Table 1. 2.5. Peat depth model development We used the scorpan spatial soil prediction function framework in the mapping procedure (McBratney et al., 2003), where a soil property
can be mapped as an empirical function of its environmental covariates: soil, climate, organisms, relief, parent material, age, and spatial position. In this study, we only use topography and spatial position as covariates, as other factors can be assumed to be constant within a peat hydrological unit. We did not use vegetation as a covariate because of the unavailability of representative land cover. Land cover in these peatlands is quite dynamic, the conversion of secondary forest (e.g., shrub and bush) to be forest and or agricultural plantations can happen just in few years. Four machine learning models were used as spatial prediction functions: Cubist regression tree, Random Forests (RF), Quantile Regression Forests (QRF), and Artiﬁcial Neural Network (ANN) to estimate peat depth. The ﬁrst three models allow the derivation of important variables that were used in the model prediction. Cubist. The Cubist R package (Kuhn et al., 2014) was used. Cubist is extended the Quinlan's M5 model tree. Cubist produces a series of “if– then” rules, where each rule has an associated multivariate linear model. Whenever a set of covariates matches a rule's conditions, the corresponding model is used to calculate the predicted value. For the details, we refer to Quinlan (1992, 1993a,b). RF. The Random Forest R package (Liaw and Wiener, 2002, 2015) was used. Random Forest ﬁrstly was introduced by (Breiman, 2001) which is a tree-based ensemble method. RF contains not a single standard regression tree but many regression trees, like a forest. RF operates two subsets: (1) in predictor variables at each node and (2) in individual data by bootstrapping technique. The recommended number of trees is between 64 and 128 trees, which is a compromise between the accuracy and processing time (Oshiro et al., 2012), and thus, in this study we used 100 trees.
Table 1 Covariates used in the prediction of peat depth. Covariates
Ogan Komering Ilir Minimum
Elevation Slope Aspect Saga wetness index
Nearest distance to river
Filtered DEM of SRTM with spatial resolution of 30.7 m, calculated using SAGA 2.2.2 Identiﬁes the gradient or the rate of change of elevation, calculated using SAGA 2.2.2 Identiﬁes the slope direction, calculated using SAGA 2.2.2 Similar to the ‘Topographic Wetness Index’ (Moore et al., 1993) which measures the ratio between the catchment area and slope to reﬂect ﬂow accumulation, but Saga wetness index modiﬁed catchment area (Böhner et al., 2002), calculated using SAGA 2.2.2 Calculates the Euclidean distance for each cell to the closest river, calculated using ArcMap 10.0
0.0006 12.721 3.388
0.0048 13.311 6.278
Minimum 20.88 0.0002 0.013 8.434
0.0033 2.781 10.927
0.0269 6.168 12.581
Rudiyanto et al. / Geoderma 272 (2016) 20–31
QRF. The Quantile Regression Forests R Package (Meinshausen and Schiesser, 2014) was used. QRF is an extension of Random Forests (Liaw and Wiener, 2015) which allow for a non-parametric estimation of conditional quantiles of the prediction. QRF utilises the output of the Random Forests to derive the full conditional distribution of the prediction. We also used 100 trees, and the theory behind QRF is described by Meinshausen (2006). Artiﬁcial Neural Networks. We applied ANN that are consisted of three layers: input, hidden and output layers having ﬁve nodes for inputs (covariates), 10 hidden nodes and a single output for peat depth. The Sigmoid function was used as activation function in each node. Before the data were used in the calculation, the data were linearly normalized into 0 to 1 for input and 0.2 to 0.8 for output. The optimization of weights that connected between nodes in different layers was performed using the feed-forward backpropagation learning method (Hecht-Nielsen, 1989). Learning rate and momentum were set equal to 0.9 and 0.8, respectively.
The training process was terminated when iteration reached 10,000. The source code of ANN in Macro Visual Basic and Pascal is available from the ﬁrst author.
2.6. Model training, testing and evaluation We used the k-fold realizations or cross-validation approach to determine the accuracy and uncertainty of the four peat depth regression models. This approach divides randomly the observed data into k groups and then trains k models using all except one of the subsets where this left aside subset for each model was used for testing. We used 10-fold realizations, and consequently, each regression model yielded 10 models of peat depth. The performances of the models during training and testing in the k-fold cross validation procedure were evaluated using the root mean square error (RMSE) and coefﬁcient of determination, R2.
Fig. 3. Agreement between observed and predicted peat depths at (a) Ogan Komering Ilir, South Sumatra and (b) Katingan, Central Kalimantan from four models: Cubist, Random Forest (RF), QuantregForest (QRF) and Artiﬁcial Neural Network (ANN). Linear ﬁt between observed and predicted (solid lines) and concordance 1:1 line (dash lines).
Rudiyanto et al. / Geoderma 272 (2016) 20–31
Table 2 Average RMSE and R2 for prediction of peat depth from 10-fold cross validation. R2
Ogan Komering Ilir, South Sumatra Cubist 0.169 ± 0.016 RF 0.313 ± 0.008 QRF 0.359 ± 0.029 ANN 0.891 ± 0.019
0.631 ± 0.237 0.627 ± 0.105 0.662 ± 0.125 0.902 ± 0.103
0.995 ± 0.000 0.985 ± 0.000 0.980 ± 0.003 0.866 ± 0.006
0.923 ± 0.062 0.936 ± 0.024 0.928 ± 0.026 0.862 ± 0.033
0.506 0.322 0.313 0.013
0.071 0.049 0.051 0.004
Katingan, Central Kalimantan Cubist 0.442 ± 0.047 RF 0.622 ± 0.029 QRF 0.867 ± 0.044 ANN 1.288 ± 0.058
1.039 ± 0.260 1.148 ± 0.321 1.125 ± 0.437 1.339 ± 0.347
0.952 ± 0.009 0.937 ± 0.006 0.846 ± 0.019 0.594 ± 0.029
0.666 ± 0.281 0.674 ± 0.137 0.694 ± 0.148 0.570 ± 0.098
0.597 0.525 0.258 0.051
0.286 0.263 0.152 0.025
2.7. Predicting spatial peat depth and uncertainty analysis To predict spatially the peat depth (i.e., at each raster cell of 30.7 m resolution), the covariates were extracted from each cell of raster and then used as inputs to the 10 models that were generated in the k-fold realizations (k = 10, see Section 2.5). Each regression model produced 10 peat depth maps. Subsequently, assuming a normal distribution prediction from those 10 maps, the mean (μ) and standard deviation (σ) of prediction were calculated at each raster cell. After that, in order to account for the uncertainty in the model, we calculated the 90% conﬁdence interval of prediction as follows: μ ± 1.64σ. As a result, three maps of peat depth were produced for: 1) lower 5% prediction, 2) mean prediction, and 3) upper 95% prediction. When the predicted peat depth is a negative, we set the value to zero. These procedures were carried out for three models: Cubist, RF and ANN. Since QRF automatically facilitated the prediction interval, we used the prediction limits (median, 5th and 95th percentiles) set out by the algorithm. We also tested the quality of the uncertainty prediction by calculating the percentage of observations that fall within the prescribed 90% conﬁdence interval, which theoretically should be 90%. 2.8. Volume of peat and carbon stock The volume of peat for an area, Vp (m3) was calculated by summing the peat depth, T (m) at all raster cells multiplied by the cell area (Acell = ΔX × ΔY = 30.7 × 30.7 m2), which is given by:
T i¼1 i
T i¼1 i
where i is the ith cell and m is the number of cell raster in the area. Carbon stock of the peatland is obtained by multiplying the peat volume, Vp (m3) with its carbon density, Cd (Mg m−3): Cp ¼ V p Cd
where Cp is the carbon stock (in Mg or tonnes) and the carbon density, Cd (Mg m−3) is calculated as a dot product between bulk density, BD (Mg m−3) and percentage of carbon content Cc (%): C d ¼ BD C c
1 : 100
We used average BD and Cc values as we do not have measurements of BD and Cc with depth. We used an average BD from our measurement and the literature (see Section 2.3). Cc is set to 55 ± 2.2% which is an average value of carbon content for tropical peats at varying depths with BD b 0.25 g cm−3 (Rudiyanto et al., 2016).
The propagation of uncertainty in the calculation of carbon density and carbon stock is calculated from the following expression (Goodman, 1960): σ 2f ¼ A2 σ 2B þ B2 σ 2A þ σ 2A σ 2B
where f = A × B, A and B are variables with std. deviations σA and σB, respectively, assuming no correlation between variable A and B. First we calculated the uncertainty of Cd, given the uncertainty in BD and Cc using Eq. (4). Assuming that there is no error in the area of a pixel, the uncertainty of peat volume is only quantiﬁed in terms of the predicted thickness, T. Thus the uncertainty of C stock is calculated from uncertainty in thickness prediction using DSM and Cd estimates. 3. Results and discussion 3.1. Peat depth modelling 3.1.1. Model accuracy The accuracy of the four regression models was assessed using the k-fold cross validation procedure (k = 10). The results are presented in Fig. 3a and b for Ogan Komering Ilir and Katingan, respectively with corresponding RMSE and R2 in Table 2. Excellent agreements were found between the observed and predicted peat depths especially for Cubist, RF and QRF for both sites. Cubist, RF and QRF models resulted in similar performances especially in the testing subset: RMSE about (≈) 0.6 m and R2 ≈ 0.98 for OKI and RMSE ≈ 1.1 m and R2 ≈ 0.67 for Katingan. The performance of ANN is inferior with RMSE = 0.9 m and R2 = 0.86 for OKI and RMSE = 1.3 m and R2 = 0.57 for Katingan. The lower model performance in Katingan is due to the limited transect observations in this area. Nevertheless, these three models showed a higher accuracy compared to other peat depth mapping studies
Table 3 Variable usage in Cubist models for peat depth prediction in Ogan Komering Ilir and Katingan. The values represent the percentage of times where each covariate was used in a condition or a linear model. Covariate
Ogan Komering Ilir, South Sumatra Elevation Saga wetness index Aspect Nearest distance to river Slope
89.9 ± 6.4 63.5 ± 13.8 55.0 ± 5.4 41.8 ± 11.5 32.3 ± 11.4
79.7 ± 6.9 67.0 ± 16.1 45.6 ± 9.8 44.5 ± 10.9 7.5 ± 6.7
Katingan, Central Kalimantan Elevation Nearest distance to river Aspect Saga wetness index Slope
72 ± 26 66 ± 30 9 3 2
65 ± 30 78 ± 20 26 27 23
Rudiyanto et al. / Geoderma 272 (2016) 20–31
Fig. 4. Variables of importance for predicting peat depth according to Random Forests (RF) for (a) Ogan Komering Ilir, and (b) Katingan.
(Holden and Connolly, 2011; Koszinski et al., 2015; Parry et al., 2012) where the reported R2 values only range between 0.5 and 0.65. Fig. 3a and b showed the plot between observed and predicted peat depths for the testing subset which are more scattered compared to
those in the training subset. The difference in model performance between training and testing were evaluated by ΔRMSE and ΔR2 and were listed in Table 2. It shows that the difference in performances of ANN from training to testing is lower (ΔRMSE = 0.01 and ΔR2 =
Fig. 5. Mean, lower limit (5% percentile), upper limit (95% percentile) prediction of peat depth maps from Cubist, Random Forest (RF), Quantreg Forest (QRF) and Artiﬁcial Neural Network (ANN) for OKI.
Rudiyanto et al. / Geoderma 272 (2016) 20–31
0.004 for OKI and ΔRMSE ≈ 0.051 and ΔR2 ≈ 0.025 for Katingan) compared to the other three models (ΔRMSE ≈ 0.3 and ΔR2 ≈ 0.05 for OKI and ΔRMSE ≈ 0.5 and ΔR2 ≈ 0.2) for Katingan. It indicates that the degree of overﬁtting for Cubist, RF and QRF is higher than ANN. 3.1.2. Variables of importance Cubist and RF showed variables of importance for peat depth prediction from the model training as shown in Table 3, and Fig. 4, respectively. Elevation comes out to be the most important predictor for those models in both peatlands. It is not surprising since generally the surface shape of peatlands is convex where peat dome is developed (Anderson, 1961, 1964; Jaenicke et al., 2008). Peat can be very thick and concentrated in the peat dome (Esterle and Ferm, 1994) generally having the highest elevation in a peatland. Rivers play an important role in peatland development (Anderson, 1961, 1964; Esterle and Ferm, 1994). Our ﬁnding shows that the nearest
distance to the river is the second most important predictor in Katingan, but it is of lower importance in Ogan Komering Ilir. This is due to the proximity of rivers in Katingan compared to the more remote Ogan Komering Ilir. Slope, aspect and wetness index did not contribute much to the models. It is due to the fact that tropical peatland surfaces are almost ﬂat and located in a lowland area (see range of its value in Table 1). 3.2. Peat depth mapping and uncertainty analysis We then applied the calibrated four models for peat depth mapping. Figs. 5 and 6 show maps of the peat depth prediction along with its upper and lower 90% conﬁdence intervals (CIs) from four models in the two peatlands. While the performances of the three models are similar, the widths of the models' conﬁdence intervals are different as shown in Table 4.
Fig. 6. Mean, lower limit (5% percentile), and upper limit (95% percentile) prediction of peat depth maps from Cubist, Random Forest (RF), Quantreg Forest (QRF) and Artiﬁcial Neural Network (ANN) for Katingan.
Rudiyanto et al. / Geoderma 272 (2016) 20–31
Table 4 The number and percentage of peat depth observations that fall inside the prescribed 90% conﬁdence interval, CI for Ogan Komering Ilir and Katingan. Number of points All
5 to 95%
5 to 95%
Cubist RF QRF ANN
Ogan Komering Ilir, South Sumatra 840 623 99 118 840 282 294 264 840 830 3 7 840 266 289 285
74 34 99 32
12 35 0 34
14 31 1 34
Cubist RF QRF ANN
Katingan, Central Kalimantan 121 94 14 121 30 56 121 117 0 121 10 111
78 25 97 8
12 46 0 92
11 29 3 0
13 35 4 0
Based on the prediction and uncertainty results, we recommend the use of Cubist models, since it is very fast in training–testing processes and available in R package that is easy to implement (Malone, 2013). From our experience, at the current stage, QRF needs a lot of computer memory for generating prediction maps. To solve this problem; we have to divide the areas into several subsets to be used in QRF prediction. Another approach that can be used to improve the prediction is the regression kriging technique, where the regression model (i.e. Cubist) is combined with kriging of the model residuals (Adhikari et al., 2013; Dai et al., 2014; Guo et al., 2015). We did not use this approach in this study, because available observations are based on transects and did not have a good spatial coverage. We note that both the regression models require further validation using random sampling to avoid overﬁtting within the training area and extrapolation of model beyond the geographical and covariate domain of observations.
3.3. Peat volume and carbon stock The width of the 90% conﬁdence interval in increasing order is as follow: ANN b RF b Cubist b QRF. We quantiﬁed the goodness of the uncertainty measurement as the percentage of observations that fall within the deﬁned conﬁdence interval. Theoretically, 90% of the observations should fall within the prescribed interval. For OKI, ANN and RF yielded the smallest uncertainty, however only about 30% of the observations fall in the deﬁned 90% CI, implying that the prescribed uncertainty is overconﬁdent. This is followed by Cubist where 74% of the observations fall in the CI and ﬁnally QRF where 99% of the observations fall in the CI. A similar pattern was observed for prediction in Katingan. While the models performed similarly, Cubist and QRF produce the most reliable CI. Nevertheless, the data used in this study is based on transect observations, and thus may not represent the whole area. In addition, this uncertainty only represents the model (or parameter) uncertainty and does not take into account the actual spatial uncertainty. Fig. 7 shows a cross section of transect a1 to a2 (between Air Sugihan to Lebung Gajah Tulung Selapan) in OKI (Fig. 1b left). It shows that the prediction values of peat depth from Cubist, RF and QRF are quite variable with elevation while ANN prediction is quite smooth following elevation. It indicates that Cubist, RF and QRF are more sensitive to the changes of covariate values than ANN. We expect that the decision trees (i.e. composed of conditions: If bN Then bN Else bN) that were used in the Cubist, RF and QRF led to this variable prediction while the continuous activation function (i.e., sigmoid function) used in ANN led to a smoother surface response. We tested the inclusion of geographical coordinates X and Y as predictors (Bonfatti et al., 2016), Cubist, RF and QRF yielded non-continuous maps where peat depth changed abruptly at some X or Y values (not shown here). Careful selection of covariates should be considered when using these tree models.
High deposit of peat was found in the middle, east and west parts of OKI (Fig. 5) where Jaenicke et al. (2008) called as Air Sugihan and Teluk Pulai peat domes. The volume of peat calculated using Eq. (1) is listed in Table 5. Jaenicke et al. (2008) estimated the carbon stock of this area by ﬁrst delineating the dome using a Landsat imagery and a 90 m SRTM DEM. The volume of the peat was then estimated by ﬁtting a surface polynomial model, and kriging was performed at a resolution of 1 km. Jaenicke et al. (2008) and Wetlands International (2003) estimated the volume for these two domes as 4.07 × 109 m3 and 1.31 × 109 m3, respectively. Our study area encompasses the whole peat hydrological unit as opposed to only the two domes. Our estimation of the peat volume in this area is larger, with a mean of 6.02 × 109 m3 and a range between 4.60 and 7.33 × 109 m3. In Katingan, the peat is mainly concentrated in the middle part, parallel between the two rivers (Fig. 6). Based on the three models: Cubist, RF and QRF; the best prediction of peat volume in this area has a mean of 1.59 × 109 m3 with a range between 1.30 and 1.83 × 109 m3. Since the ANN model yielded the worst prediction, the prediction of peat volume may be inaccurate. A large peat volume prediction from the ANN model was caused by overestimation of depths less than 4 m (see Fig. 3b in ANN testing). The dataset for bulk density data obtained from our measurement combined with the literature for OKI and Katingan are listed in the Supplementary information (Supplement 1 and 2, respectively). The average bulk density for both peatlands is quite similar: 0.144 ± 0.037 g cm−3 for OKI and 0.141 ± 0.032 g cm−3 for Katingan. Based on an average Cc of 55% for tropical peatlands (Rudiyanto et al., 2016) and Eqs. (3) and (4), the estimated carbon density and its standard
Fig. 7. Cross section a1 to a2 between Air Sugihan to Lebung Gajah Tulung Selapan at Ogan Komering Ilir (see Fig. 1b left): observed peat depth (red square symbols), peat surface (dark blue lines), predicted peat depth by Cubist (light green lines), Random Forests (violet lines), Quantreg Forest (light blue lines) and Artiﬁcial Neural Networks (orange lines). (For interpretation of the references to color in this ﬁgure legend, the reader is referred to the web version of this article.)
Rudiyanto et al. / Geoderma 272 (2016) 20–31 Table 5 Mean, lower limit (5% percentile), and upper limit (95% percentile) prediction of peat volume and carbon stock from Cubist, Random Forests (RF), Quantreg Forest (QRF) and Artiﬁcial Neural Network (ANN) models for Ogan Komering Ilir and Katingan. Prediction of peat volume (×109 m3)
Prediction of Carbon stock (×109 t = Giga tonnes)
Cubist RF QRF ANN
Ogan Komering Ilir, South Sumatra 2.632 6.327 10.422 6.104 7.328 8.551 0.763 4.603 12.983 5.131 5.816 6.610
0.126 0.299 0.062 0.243
0.498 0.577 0.363 0.458
0.915 0.856 1.532 0.711
Cubist RF QRF ANN
Katingan, Central Kalimantan 0.903 1.834 2.825 1.366 1.641 1.915 0.104 1.303 5.639 3.540 5.493 7.451
0.048 0.072 0.009 0.194
0.141 0.126 0.100 0.423
0.240 0.180 0.450 0.652
deviation for OKI and Katingan are 0.079 ± 0.021 and 0.077 ± 0.018 g cm−3, respectively. The mean carbon stock with its lower and upper limits of 90% conﬁdence interval in both peatlands calculated using Eqs. (2) and (4) are listed in Table 5. The carbon stocks in OKI range between 0.363 and 0.577 Gt with a mean of 0.474 Gt which is twice larger than previous estimates of 0.24 Gt for the two peat domes in this area (Jaenicke et al.,
2008). The difference is due to the difference in carbon density where our measurement is 79 kg m− 3, while the study by Jaenicke et al. (2008) used a mean value from the literature of 58 kg m−3. In addition, we included all areas bounded by the rivers which have accumulation of peat up to 1 m. In Katingan, the estimated carbon stock from Cubist, RF and QRF ranges between 0.100 and 0.141 Gt with a mean of 0.123 Gt. The results indicate that the difference in estimating peat volume can lead to a signiﬁcant difference in carbon stock estimation. Our approach advances the work of Jaenicke et al. (2008) in several ﬁelds: • The maps were produced at a ﬁner resolution using the 30.7 m SRTM as opposed to the 90 m SRTM and kriged to 1 km resolution. • Regression tree models were calibrated using ﬁeld data, and the tree models allow variation within elevation which leads to a more accurate modelling of peat depth. • Nearest distance to the river is also a useful covariate in addition to elevation. • Mapping the peatland as a hydrological unit (where the area is bounded by rivers) rather than just the domes, where the delineation of the boundaries can be subjective. Our results suggest that within a peat hydrological unit, areas outside of the domes contain the same amount of carbon as the domes. Nevertheless, there are few limitations in this study, i.e. the sampling locations were based on transects, and the calibrated models may not cover the whole covariate space of the mapping areas. The calculated carbon stock is based on an average value of carbon density while carbon density may change with depth and location. 3.4. Sampling strategies for peat depth mapping Conventionally, locations of sampling observations for peat survey and mapping were determined based on a grid or multiple transects across the peatlands, for example, north–south, east–west, and northeast–southwest transects with a uniform distance between observations (Agus et al., 2011). Generally, only two or three transects were implemented in the ﬁeld (Fig. 8a). Consequently, sampled points do not cover the whole area and are focussed in areas with similar elevation (Fig. 8a) which leads to the low coverage of both the geographical and covariate spaces. For digital soil mapping, we need to cover the covariate space effectively, either by conditioned Latin hypercube sampling (Minasny and McBratney, 2006) or grouping the peatland based on similarity of physiographic conditions (Aguilar et al., 2005). Since elevation is the most important predictor, we recommend future sampling for peat depth mapping as follows: • Generate some strata or clusters based on elevation using unsupervised classiﬁers such as fuzzy k-means (Bezdek, 1983; McBratney and de Gruijter, 1992), k-means (Hartigan and Wong, 1979) or isodata (Ball and Hall, 1965) methods. Note that the elevation cluster here is identical to the class of contour. • Determine the number of samples based on the available resources: time, cost, and manpower etc. More sampling points are always better. • Determine the location of the sampling points within the cluster using the following options: 1) transects across clusters: draw several transects which cover all clusters. Note that the distance between points does not need to be uniform (Fig. 8b), and 2) random samples within clusters: distributing the sampling points randomly in all clusters where the number of points in each cluster is proportional to the area of the cluster (Fig. 8c). Note that access or mobility also needs to be considered here.
Fig. 8. An illustration of the determination of sampling observations: (a) traditional approach using transects, (b) transect across clusters and (c) random point within clusters methods. Clustering was based on elevation using the fuzzy k-means approach.
We illustrate this with an example in Fig. 8. The transect-cluster method is proposed following the traditional transect method for peat where a cross-section of the peat dome can still be generated for pedological study. This sampling approach may enrich the covariate
Rudiyanto et al. / Geoderma 272 (2016) 20–31
information in peat depth modelling, thereby increasing the accuracy of peat depth mapping. 4. Conclusions and recommendations We proposed an accurate and cost-effective methodology for mapping peat depth in Indonesian peatlands. The method is based on the scorpan spatial prediction function framework, where ﬁeld observations selected by smart sampling techniques are combined with freely and widely available covariates and machine learning techniques. This methodology allows peat depth mapping and carbon stock prediction at a resolution of 30 m along with the uncertainty of estimates. We tested four machine learning models in two peatlands. Three models: Cubist, Random Forests, Quantile Regression Forests are highly accurate and describe the variation of the observed peat depth data higher than 0.67; while ANN showed a poor performance. All models produced comparable digital maps of peat depth, however, Cubist and QRF produced the best 90% conﬁdence interval. The models revealed that elevation is the most importance variable in peat depth modelling for both peatlands. Thus, we recommend sampling for peat depth mapping based on stratifying the area by elevation and sampling locations can be determined either by drawing transects that crossed all strata or selecting random locations within each stratum. We believe that the combination of those regression models and an improved sampling strategy will increase the accuracy of peat depth mapping. Finally, we recommend that peat depth mapping and other spatial studies in a peatland should be carried out in the whole peat hydrological unit bounded by neighbouring rivers and/or sea instead of only focussing on peat domes. Acknowledgements Rudiyanto would like to express sincere gratitude to the NSW Committee of the Crawford Fund, and the SSEAC Regional Mobility Fund of the University of Sydney, Australia for allowing him to conduct this research at the Faculty of Agriculture & Environment, The University of Sydney. The authors also thank the editor and three reviewers for their constructive comments. Appendix A. Supplementary data Supplementary data to this article can be found online at http://dx. doi.org/10.1016/j.geoderma.2016.02.026. References Adhikari, K., Kheir, R.B., Greve, M.B., Bøcher, P.K., Malone, B.P., Minasny, B., McBratney, A.B., Greve, M.H., 2013. High-resolution 3-D Mapping of soil texture in Denmark. Soil Sci. Soc. Am. J. 77, 860–876. http://dx.doi.org/10.2136/sssaj2012.0275. Aguilar, F.J., Agüera, F., Aguilar, M.a., Carvajal, F., 2005. Effects of terrain morphology, sampling density, and interpolation methods on grid DEM Accuracy. Photogramm. Eng. Remote. Sens. 71, 805–816. Agus, K., Hairiah, K., Mulyani, A., 2011. Measuring carbon stock in peat soils: practical guidelines. World Agroforestry Centre (ICRAF) Southeast Asia Regional Program. Indonesian Centre for Agricultural Land Resources Research and Development, Bogor, Indonesia. Aitkenhead, M.J., Coull, M.C., 2016. Mapping soil carbon stocks across Scotland using a neural network model. Geoderma 262, 187–198. http://dx.doi.org/10.1016/j. geoderma.2015.08.034. Aitkenhead, M.J., Donnelly, D., Sutherland, L., Miller, D.G., Coull, M.C., Black, H.I.J., 2015. Predicting Scottish topsoil organic matter content from colour and environmental factors. Eur. J. Soil Sci. 66, 112–120. http://dx.doi.org/10.1111/ejss.12199. Akumu, C.E., McLaughlin, J.W., 2014. Modeling peatland carbon stock in a delineated portion of the Nayshkootayaow river watershed in Far North, Ontario using an integrated GIS and remote sensing approach. Catena 121, 297–306. http://dx.doi.org/10. 1016/j.catena.2014.05.025. Altdorff, D., Bechtold, M., van der Kruk, J., Vereecken, H., Huisman, J.A., 2016. Mapping peat layer properties with multi-coil offset electromagnetic induction and laser scanning elevation data. Geoderma 261, 178–189. http://dx.doi.org/10.1016/j.geoderma. 2015.07.015.
Anderson, J.A.R., 1961. The Ecology and Forest Types of the Peat Swamp Forests of Sarawak and Brunei in Relation to Their Silviculture. Univ. Edinburgh. Anderson, J.A.R., 1964. The structure and development of the peat swamps of Sarawak and Brunei. J. Trop. Geogr. 18, 7–16. Ball, G.H., Hall, D.J., 1965. ISODATA, a novel method of data analysis and pattern classiﬁcation. Analysis 1–79. Ballabio, C., Panagos, P., Monatanarella, L., 2016. Mapping topsoil physical properties at European scale using the LUCAS database. Geoderma 261, 110–123. http://dx.doi. org/10.1016/j.geoderma.2015.07.006. Bauer, I.E., Vitt, D.H., 2011. Peatland dynamics in a complex landscape: development of a fen-bog complex in the sporadic discontinuous permafrost zone of northern Alberta, Canada. Boreas 40, 714–726. http://dx.doi.org/10.1111/j.1502-3885.2011.00210.x. Bauer, I.E., Gignac, L.D., Vitt, D.H., 2003. Development of a peatland complex in boreal western Canada: lateral site expansion and local variability in vegetation succession and long-term peat accumulation. Can. J. Bot. 81, 833–847. http://dx.doi.org/10. 1139/b03-076. Bezdek, J.C., 1983. Pattern Recognition With Fuzzy Objective Function Algorithms. SIAM Review. Springer http://dx.doi.org/10.1137/1025116. Biancalani, R., Avagyan, A., 2014. Towards Climate-responsible Peatlands Management. Food and Agiculture Organization of the United Nations (FAO), Rome. Bishop, C.M., 1995. Neural networks for pattern recognition. J. Am. Stat. Assoc. http://dx. doi.org/10.2307/2965437. Boehm, H.-D.V., Frank, J., 2008. Peat Dome Measurements in Tropical Peatlands of Central Kalimantan With a High-resolution Airborne Laser Scanner to Achieve Digital Elevation Models (Tullamore, Ireland). Böhner, J., Köthe, R., Conrad, O., Gross, J., Ringeler, A., Selige, T., 2002. Soil regionalisation by means of terrain analysis and process parameterisation. Eur. Soil Bur 213–222. Bonfatti, B.R., Hartemink, A.E., Giasson, E., Tornquist, C.G., Adhikari, K., 2016. Digital mapping of soil carbon in a viticultural region of Southern Brazil. Geoderma 261, 204–221. http://dx.doi.org/10.1016/j.geoderma.2015.07.016. Brady, M.A., 1997. Organic Matter Dynamics of Coastal Peat Deposits in Sumatra, Indonesia ProQuest Diss. Theses The University of British Columbia (Canada). Breiman, L., 2001. Random Forests. Mach. Learn. 45, 5–32. http://dx.doi.org/10.1023/A: 1010933404324. Dai, F., Zhou, Q., Lv, Z., Wang, X., Liu, G., 2014. Spatial prediction of soil organic matter content integrating artiﬁcial neural network and ordinary kriging in Tibetan Plateau. Ecol. Indic. 45, 184–194. http://dx.doi.org/10.1016/j.ecolind.2014.04.003. Esterle, J.S., Ferm, J.C., 1994. Spatial variability in modern tropical peat deposits from Sarawak, Malaysia and Sumatra, Indonesia: analogues for coal. Int. J. Coal Geol. 26, 1–41. http://dx.doi.org/10.1016/0166-5162(94)90030-2. Farmer, J., Matthews, R., Smith, J.U., Smith, P., Singh, B.K., 2011. Assessing existing peatland models for their applicability for modelling greenhouse gas emissions from tropical peat soils. Curr. Opin. Environ. Sustain. 3, 339–349. http://dx.doi.org/ 10.1016/j.cosust.2011.08.010. Farmer, J., Matthews, R., Smith, P., Langan, C., Hergoualc'h, K., Verchot, L., Smith, J.U., 2014. Comparison of methods for quantifying soil carbon in tropical peats. Geoderma 214-215, 177–183. http://dx.doi.org/10.1016/j.geoderma.2013.09.013. Fell, H., Roßkopf, N., Bauriegel, A., Zeitz, J., 2016. Estimating vulnerability of agriculturally used peatlands in north-east Germany to carbon loss based on multi-temporal subsidence data analysis. Catena 137, 61–69. http://dx.doi.org/10.1016/j.catena.2015.08.010. Fyfe, R.M., Coombe, R., Davies, H., Parry, L., 2014. The importance of sub-peat carbon storage as shown by data from Dartmoor, UK. Soil Use Manag. 30, 23–31. http://dx. doi.org/10.1111/sum.12091. Goodman, L.A., 1960. On the exact variance of products. J. Am. Stat. Assoc. 55, 708–713. Grimm, R., Behrens, T., Märker, M., Elsenbeer, H., 2008. Soil organic carbon concentrations and stocks on Barro Colorado Island — digital soil mapping using Random Forests analysis. Geoderma 146, 102–113. http://dx.doi.org/10.1016/j.geoderma.2008.05. 008. Günther, F., Fritsch, S., 2010. neuralnet: training of neural networks. R J. 2, 30–38. Guo, P.-T., Li, M.-F., Luo, W., Tang, Q.-F., Liu, Z.-W., Lin, Z.-M., 2015. Digital mapping of soil organic matter for rubber plantation at regional scale: an application of random forest plus residuals kriging approach. Geoderma 237-238, 49–59. http://dx.doi.org/10. 1016/j.geoderma.2014.08.009. Hartigan, J. a, Wong, M. a, 1979. Algorithm AS 136: a K-means clustering algorithm. Appl. Stat. 28, 100. http://dx.doi.org/10.2307/2346830. Hecht-Nielsen, R., 1989. Theory of the backpropagation neural network. Proc. Int. Jt. Conf. Neural Netw. 1, 593–605. http://dx.doi.org/10.1109/IJCNN.1989.118638. Hengl, T., Heuvelink, G.B.M., Stein, A., 2004. A generic framework for spatial prediction of soil variables based on regression-kriging. Geoderma 120, 75–93. http://dx.doi.org/ 10.1016/j.geoderma.2003.08.018. Heung, B., Bulmer, C.E., Schmidt, M.G., 2014. Predictive soil parent material mapping at a regional-scale: a Random Forest approach. Geoderma 214-215, 141–154. http://dx. doi.org/10.1016/j.geoderma.2013.09.016. Heung, B., Ho, H.C., Zhang, J., Knudby, A., Bulmer, C.E., Schmidt, M.G., 2016. An overview and comparison of machine-learning techniques for classiﬁcation purposes in digital soil mapping. Geoderma 265, 62–77. http://dx.doi.org/10.1016/j.geoderma.2015.11.014. Holden, N.M., Connolly, J., 2011. Estimating the carbon stock of a blanket peat region using a peat depth inference model. Catena 86, 75–85. http://dx.doi.org/10.1016/j. catena.2011.02.002. Hooijer, A., Vernimmen, R., 2013. Peatland maps for Indonesia. Including Accuracy Assessment and Recommendations for Improvement, Elevation Mapping and Evaluation of Future Flood Risk. Quick Assessment and Nationwide Screening (QANS) of Peat and Lowland Resources and Action Planning for t. Hugron, S., Bussières, J., Rochefort, L., 2013. Tree Plantations Within the Context of Ecological Restoration of Peatlands: A Practical Guide. Peatland Ecology Research Group, Université Laval, Québec.
Rudiyanto et al. / Geoderma 272 (2016) 20–31 International, Wetlands, 2003. Map of Peatland Distribution Area and Carbon Content in Sumatera 1990–2002 (Bogor). Jaenicke, J., Rieley, J.O., Mott, C., Kimman, P., Siegert, F., 2008. Determination of the amount of carbon stored in Indonesian peatlands. Geoderma 147, 151–158. http:// dx.doi.org/10.1016/j.geoderma.2008.08.008. Joosten, H., Clarke, D., 2002. Wise Use of Mires and Peatlands — Background and Principles Including a Framework for Decision-making. International Mire Conservation Group / International Peat Society. Keaney, A., McKinley, J., Graham, C., Robinson, M., Ruffell, A., 2013. Spatial statistics to estimate peat thickness using airborne radiometric data. Spat. Stat. 5, 3–24. http://dx. doi.org/10.1016/j.spasta.2013.05.003. Koszinski, S., Miller, B.A., Hierold, W., Haelbich, H., Sommer, M., 2015. Spatial modeling of organic carbon in degraded peatland soils of northeast Germany. Soil Sci. Soc. Am. J. 79, 1496–1508. http://dx.doi.org/10.2136/sssaj2015.01.0019. Kuhn, M., Weston, S., Keefer, C., Coulter, N., Quinlan, R., 2014. Cubist: rule- and instancebased regression modeling. (R package version 0.0.18). Lähteenoja, O., Page, S., 2011. High diversity of tropical peatland ecosystem types in the Pastaza-Marañón basin, Peruvian Amazonia. J. Geophys. Res. Biogeosci. 116, G02025. http://dx.doi.org/10.1029/2010JG001508. Liaw, a, Wiener, M., 2002. Classiﬁcation and Regression by randomForest. R news 2 pp. 18–22. Liaw, A., Wiener, M., 2015. Package “randomForest”. Breiman and Cutler's Random Forests for Classiﬁcation and Regression. CRAN Ref. Man. Ließ, M., Glaser, B., Huwe, B., 2012. Uncertainty in the spatial prediction of soil texture. Geoderma 170, 70–79. http://dx.doi.org/10.1016/j.geoderma.2011.10.010. Malone, B., 2013. Use R for Digital Soil Mapping. Soil Security Lab, The University of Sydney, Sydney. Malone, B.P., McBratney, A.B., Minasny, B., Laslett, G.M., 2009. Mapping continuous depth functions of soil carbon storage and available water capacity. Geoderma 154, 138–152. http://dx.doi.org/10.1016/j.geoderma.2009.10.007. McBratney, A.B., de Gruijter, J.J., 1992. A continuum approach to soil classiﬁcation by modiﬁed fuzzy k-means with extragrades. J. Soil Sci. 43, 159–175. http://dx.doi.org/10. 1111/j.1365-2389.1992.tb00127.x. McBratney, A., Mendonça Santos, M., Minasny, B., 2003. On digital soil mapping. Geoderma 117, 3–52. http://dx.doi.org/10.1016/S0016-7061(03)00223-4. Meinshausen, N., 2006. Quantile Regression Forests. J. Mach. Learn. Res. 7, 983–999. http://dx.doi.org/10.1111/j.1541-0420.2010.01521.x. Meinshausen, N., Schiesser, L., 2014. quantregForest: Quantile Regression Forests. Minasny, B., McBratney, A.B., 2006. A conditioned Latin hypercube method for sampling in the presence of ancillary information. Comput. Geosci. 32, 1378–1388. http://dx. doi.org/10.1016/j.cageo.2005.12.009. Minasny, B., McBratney, A.B., 2015. Digital soil mapping: a brief history and some lessons. Geoderma 264, 301–311. http://dx.doi.org/10.1016/j.geoderma.2015.07.017. Moore, I.D., Gessler, G.A., Petersen, G.A., 1993. Terrain attributes: estimation methods and scale effects. In: Jakeman, A.J., Beck, M.B., McAleer, M. (Eds.), In Modeling Change in Environmental Systems. Wiley, London, pp. 189–214. Motaghian, H.R., Mohammadi, J., 2011. Spatial estimation of saturated hydraulic conductivity from terrain attributes using regression, kriging, and artiﬁcial neural networks. Pedosphere 21, 170–177. http://dx.doi.org/10.1016/S1002-0160(11)60115-X. Odeh, I.O.A., McBratney, A.B., Chittleborough, D.J., 1995. Further results on prediction of soil properties from terrain attributes: heterotopic cokriging and regression-kriging. Geoderma 67, 215–226. http://dx.doi.org/10.1016/0016-7061(95)00007-B. Odeh, I.O.A., McBratney, A.B., Chittleborough, D.J., 1994. Spatial prediction of soil properties from landform attributes derived from a digital elevation model. Geoderma 63, 197–214. http://dx.doi.org/10.1016/0016-7061(94)90063-9. Odgers, N.P., McBratney, A.B., Minasny, B., 2015. Digital soil property mapping and uncertainty estimation using soil class probability rasters. Geoderma 237-238, 190–198. http://dx.doi.org/10.1016/j.geoderma.2014.09.009. Oshiro, T., Perez, P., Baranauskas, J., 2012. How many trees in a Random Forest? In: Perner, P. (Ed.), Machine Learning and Data Mining in Pattern Recognition SE-13, Lecture Notes in Computer Science. Springer, Berlin Heidelberg, pp. 154–168 http://dx.doi. org/10.1007/978-3-642-31537-4_13 Page, S.E., Rieley, J.O., Shotyk, W., Weiss, D., 1999. Interdependence of peat and vegetation in a tropical peat swamp forest. Philos. Trans. R. Soc. Lond. Ser. B Biol. Sci. 354, 1885–1897. http://dx.doi.org/10.1098/rstb.1999.0529. Page, S.E., Wűst, R.A.J., Weiss, D., Rieley, J.O., Shotyk, W., Limin, S.H., 2004. A record of Late Pleistocene and Holocene carbon accumulation and climate change from an equatorial peat bog (Kalimantan, Indonesia): implications for past, present and future carbon dynamics. J. Quat. Sci. 19, 625–635. http://dx.doi.org/10.1002/jqs.884. Page, S.E., Rieley, J.O., Wüst, R., 2006. Tropical lowland peatlands of Southeast Asia. In: Martini, I.P., Martinez Cortizas, A., Chesworth, W. (Eds.), Peatlands: Evolution and Records of Environmental and Climate Changes. Elsevier B.V., pp. 145–171 http://dx.doi. org/10.1016/S0928-2025(06)09007–9. Page, S.E., Banks, C.J., Rieley, J.O., 2007. Tropical Peatlands: Distribution, Extent and Carbon Storage — Uncertainties and Knowledge Gaps. Peatlands Int. 2/2007, pp. 26–27. Page, S.E., Rieley, J.O., Banks, C.J., 2011. Global and regional importance of the tropical peatland carbon pool. Glob. Chang. Biol. 17, 798–818. http://dx.doi.org/10.1111/j. 1365-2486.2010.02279.x. Pahlavan Rad, M.R., Toomanian, N., Khormali, F., Brungard, C.W., Komaki, C.B., Bogaert, P., 2014. Updating soil survey maps using random forest and conditioned Latin
hypercube sampling in the loess derived soils of northern Iran. Geoderma 232-234, 97–106. http://dx.doi.org/10.1016/j.geoderma.2014.04.036. Parry, L.E., Charman, D.J., 2013. Modelling soil organic carbon distribution in blanket peatlands at a landscape scale. Geoderma 211-212, 75–84. http://dx.doi.org/10. 1016/j.geoderma.2013.07.006. Parry, L.E., Charman, D.J., Noades, J.P.W., 2012. A method for modelling peat depth in blanket peatlands. Soil Use Manag. http://dx.doi.org/10.1111/j.1475-2743.2012.00447.x (n/a–n/a). Parry, L.E., West, L.J., Holden, J., Chapman, P.J., 2014. Evaluating approaches for estimating peat depth. J. Geophys. Res. Biogeosci. 119. http://dx.doi.org/10.1002/2013JG002411 (2013JG002411). Prayitno, M., Bakri, 2005. Kajian Karakteristik Gabu di Wilayah HPH Hutan Kayu Agung Kabupaten Ogan Komering Ilir (Palembang). Proulx-McInnis, S., St-Hilaire, A., Rousseau, A.N., Jutras, S., 2013. A review of groundpenetrating radar studiesrelated to peatland stratigraphy with a case study on the determination of peat thickness in a northern boreal fen in Quebec, Canada. Prog. Phys. Geogr. Quinlan, R., 1992. Learning with continuous classes. Proceedings of the 5th Australian Joint Conference on Artiﬁcial Intelligence, pp. 343–348. Quinlan, R., 1993a. C4.5: Programs for Machine Learning. Morgan Kaufmann Publishers. Quinlan, R., 1993b. Combining instance-based and model-based learning. Proceedings of the Tenth International Conference on Machine Learning, pp. 236–243. Ritung, S., Wahyunto, Nugroho, K., Sukarman, Hikmatullah, Suparto, Tafakresnanto, C., 2011. Peta Lahan Gambut Indonesia Skala 1:250.000 (Indonesian peatland map at the scale 1:250,000) Bogor, Indonesia. Ritung, S., Wahyunto, Nugroho, K., 2012. Karakteristik dan Sebaran Lahan Gambut di Sumatera, Kalimantan dan Papua. In: Husen, E., Anda, M., Noor, M., Mamat, H.S., Maswar, Fahmi, A., Sulaiman, Y. (Eds.), Pengelolaan Lahan Gambut Berkelanjutan, pp. 47–61 (Bogor, Indonesia). Rosa, E., Larocque, M., Pellerin, S., Gagné, S., Fournier, B., 2009. Determining the number of manual measurements required to improve peat thickness estimations by ground penetrating radar. Earth Surf. Process. Landf. 34, 377–383. http://dx.doi.org/10. 1002/esp.1741. Rudiyanto, Setiawan, B.I., Arief, C., Saptomo, S.K., Gunawan, A., Kuswarman, Sungkono, Indriyanto, H., 2015. Estimating distribution of carbon stock in tropical peatland using a combination of an empirical peat depth model and GIS. Procedia Environ. Sci. 24, 152–157. http://dx.doi.org/10.1016/j.proenv.2015.03.020. Rudiyanto, Minasny, B., Setiawan, B.I., 2016. Further results on comparison of methods for quantifying soil carbon in tropical peats. Geoderma 269, 108–111. http://dx.doi.org/ 10.1016/j.geoderma.2016.01.038. Shimada, S., Takahashi, H., Haraguchi, A., Kaneko, M., 2001. The carbon content characteristics of tropical peats in Central Kalimantan, Indonesia: estimating their spatial variability in density. Biogeochemistry 53, 249–267. http://dx.doi.org/10.1023/A: 1010618807469. Sjögersten, S., Black, C.R., Evers, S., Hoyos-Santillan, J., Wright, E.L., Turner, B.L., 2014. Tropical wetlands: a missing link in the global carbon cycle? Glob. Biogeochem. Cycles 28. http://dx.doi.org/10.1002/2014GB004844 (2014GB004844). Song, X.-D., Brus, D.J., Liu, F., Li, D.-C., Zhao, Y.-G., Yang, J.-L., Zhang, G.-L., 2016. Mapping soil organic carbon content by geographically weighted regression: a case study in the Heihe River Basin, China. Geoderma 261, 11–22. http://dx.doi.org/10.1016/j. geoderma.2015.06.024. Sumawinata, B., Djajakirana, G., Suwardi, Darmawan, 2014. Carbon Dynamics in Tropical Peatland Planted Forests: One-year Research Findings in Sumatra, Indonesia. PT Penerbit IPB Press, Bogor, Indonesia. Taghizadeh-Mehrjardi, R., Nabiollahi, K., Minasny, B., Triantaﬁlis, J., 2015. Comparing data mining classiﬁers to predict spatial distribution of USDA-family soil groups in Baneh region, Iran. Geoderma 253-254, 67–77. http://dx.doi.org/10.1016/j.geoderma.2015. 04.008. van Bellen, S., Dallaire, P.-L., Garneau, M., Bergeron, Y., 2011. Quantifying spatial and temporal Holocene carbon accumulation in ombrotrophic peatlands of the Eastmain region, Quebec, Canada. Glob. Biogeochem. Cycles 25, GB2016. http://dx.doi.org/10. 1029/2010GB003877. Warren, M.W., Kauffman, J.B., Murdiyarso, D., Anshari, G., Hergoualc'h, K., Kurnianto, S., Purbopuspito, J., Gusmayanti, E., Aﬁfudin, M., Rahajoe, J., Alhamd, L., Limin, S., Iswandi, A., 2012. A cost-efﬁcient method to assess carbon stocks in tropical peat soil. Biogeosciences 9, 4477–4485. http://dx.doi.org/10.5194/bg-9-4477-2012. Weissert, L.F., Disney, M., 2013. Carbon storage in peatlands: a case study on the Isle of Man. Geoderma 204-205, 111–119. http://dx.doi.org/10.1016/j.geoderma.2013.04. 016. Wiesmeier, M., Barthold, F., Blank, B., Kögel-Knabner, I., 2011. Digital mapping of soil organic matter stocks using Random Forest modeling in a semi-arid steppe ecosystem. Plant Soil 340, 7–24. http://dx.doi.org/10.1007/s11104-010-0425-z. Wright, E.L., Black, C.R., Cheesman, A.W., Drage, T., Large, D., Turner, B.L., Sjögersten, S., 2011. Contribution of subsurface peat to CO2 and CH4 ﬂuxes in a neotropical peatland. Glob. Chang. Biol. 17, 2867–2881. http://dx.doi.org/10.1111/j.1365-2486. 2011.02448.x. Wright, E.L., Black, C.R., Turner, B.L., Sjögersten, S., 2013. Environmental controls of temporal and spatial variability in CO2 and CH4 ﬂuxes in a neotropical peatland. Glob. Chang. Biol. 19, 3775–3789. http://dx.doi.org/10.1111/gcb.12330.