Volume 9, Number 4 (July 2022): 3663-3676, doi:10.15243/jdmlm.2022.094.3663 ISSN: 2339-076X (p); 2502-2458 (e), www.jdmlm.ub.ac.id
Open Access 3663 Research Article
Simulating and modeling CO
2flux emitted from decomposed oil palm root cultivated at tropical peatland as affected by water content and residence time
Heru Bagus Pulunggono1*, Syva Fitriana2, Desi Nadalia1, Moh Zulfajrin2, Lina Lathifah Nurazizah3, Husni Mubarok4, Nizam Tambusai4, Syaiful Anwar1, Supiandi Sabiham1
1 Department of Soil Science and Land Resources, Faculty of Agriculture, IPB University, 16680, Indonesia
2 Graduate Program of Soil Science and Land Resources Department, Faculty of Agriculture, IPB University, 16680, Indonesia
3 Graduate Program of Agronomy and Horticulture Department, Faculty of Agriculture, IPB University, 16680, Indonesia
4 Agronomy Research, Astra Agro Lestari Tbk, Jakarta, Indonesia
*corresponding author: [email protected]
Abstract Article history:
Received 18 March 2022 Accepted 11 May 2022 Published 1 July 2022
Determining the oil palm dead roots contribution to total (Rt) and heterotrophic (Rh) respiration as a source of greenhouse gas/GHG emission in tropical peatland is urgently required, as well as predicting their magnitude to cope with difficulties of direct in-situ measurement. This study is designed to simulate the CO2 flux emitted from oil palm dead roots/Rdr in tropical peatland as affected by water content/WC and residence time/RT. The dead oil palm roots were cleaned, treated with control/15, 100, 150, 300, and 450% WC, and then incubated for three months. CO2 flux measurement, C, N, and CN ratio determination were conducted every month. This study demonstrated the importance Rdr among other CO2 emission sources, ranging from 0.05-2.3 Mg CO2 ha-1 year-1 with an average of 0.7 Mg CO2 ha-1 year-1. Rdr contribution for literature Rt and Rh were around 0.3 to 1.3 and 0.9 to 3.5%, respectively. As a product of microbial respiration, Rdr was affected by WC and RT, supported by analysis of variance, linear mixed effect model/REML, and multivariate analysis. 100-150%WC resulting in significant and highest Rdr, whereas the increase (300-450%WC) or decrease (15%WC) would generate lower emission. Rdr culminated in the first month after incubation; meanwhile, it declined in the following months. This study also emphasized non-linear relationships between CO2 flux and other root properties, which can be modeled conveniently using non-linear approach, particularly using polynomial and artificial intelligence-based models. The simulation presented in this study served as an initial attempt to separate Rdr from Rh, as well as to predict CO2 flux with reasonable accuracy and interpretable methods.
Keywords:
artificial intelligence dead root
greenhouse gas incubation time respiration
To cite this article: Pulunggono, H.B., Fitriana, S., Nadalia, D., Zulfajrin, M., Nurazizah, L.L., Mubarok, H., Tambusai, N., Anwar, S. and Sabiham, S. 2022. Simulating and modeling CO2 flux emitted from decomposed oil palm root cultivated at tropical peatland as affected by water content and residence time. Journal of Degraded and Mining Lands Management 9(4):3663-3676, doi:10.15243/jdmlm.2022.094.3663.
Introduction
Oil palm (Elaeis guineensis Jack.) is a commodity that plays an essential role in the Indonesian economy, accounting for 11.58 percent of total export in 2020
(BPS-Statistics Indonesia, 2020; BPS-Statistics Indonesia, 2022). Capturing the high demands for global palm oil, the oil palm plantation spreads around 14.66 million hectares (BPS-Statistic Indonesia, 2022) across Indonesian islands. Due to the limited
Open Access 3664 availability of permitted mineral soil areas, oil palm
plantations expanded over peatlands (1.7 million hectares) and mainly occupied Sumatran peatlands (Dohong et al., 2018; Dadap et al., 2021). Managing cultivated peatlands in this area is considered necessary, owing to its ability to sequester vast amounts of carbon over a prolonged period, thus mitigating the detrimental impact of climate change.
The development of natural peatlands into oil palm plantations significantly impacts peatland ecosystems, particularly for generating high greenhouse gas (GHG) emissions (e.g., Hooijer et al., 2012; Leifeld et al., 2019); hence, changing peatlands from CO2 sinks to sources (Miettinen et al., 2017).
Widespread studies encompass CO2 emissions from oil palm cultivated peatland, quantifying CO2 fluxes from microsites (e.g., Manning et al., 2019; Dhandapani et al., 2022) to ecosystem scales (e.g., Deshmukh et al., 2021; McCalmont et al., 2021). Especially for Rh
emission factor, many research works have been conducted involving the contribution of peat materials (Hergoualc'h et al., 2017; Pulunggono et al., 2022) as well as decomposed parts of oil palm pruned components (Wakhid and Hirano, 2021) and understory litters (Pulunggono et al., 2022).
Autotrophic respiration (Ra) from living oil palm roots was also widely reported, contributing a significant portion to total CO2 emission (Dariah et al., 2014;
Sabiham et al., 2014). Unfortunately, searching for the reports concerning their decomposed dead parts' contribution to Rt or Rh would return minimal information.
To the best of our knowledge, there are no studies investigating the potential of Rdr tropical peatland. Differently from their living one, dead roots fragment is difficult to separate from the surrounding peat materials. Notably, insufficient available information could become the confounding factor in designing an appropriate method to insulate CO2 that emanates from oil palm dead roots. However, their potential for CO2 emission sources is urged to be estimated since dead root contribution has been confirmed under different soil types and vegetations (Larionova et al., 2006; Zhou et al., 2021). Published reports in mineral soils revealed that the dry matter of living oil palm roots varied with age, ranging from 0.17±0.01 to 17±1.9 kg palm-1 year-1 (Khalid et al., 1999; Corley and Tinker, 2016; Siang et al., 2022). Oil palm roots have a higher density at the soil surface (Sabiham et al., 2014) and have undergone rapid normal dieback along with the increasing age, accounting for 82% of roots lost in the fourth year.
Furthermore, root turnover was more intense on secondaries and fine roots (31% and 57%, respectively, for each year), annually incorporating 4.5 to 11.5 Mg ha-1 of dry matter into the soil (Corley and Tinker, 2016). Although the previous information was based on different soil types, the magnitude of accumulated root dead material could be a plausible source of CO2 with respect to the lower portion of solid
fraction contained by peat soil (approximately 200 Mg ha-1 at upper 20 cm; hemic-sapric peat with bulk density 0.1 g cm-3).
Extrapolating into the field level, this study expected a higher contribution of Rdr to the Rt and Rh
than oil palm fronds (Rfr; Wakhid and Hirano, 2021).
Meanwhile, Rdr is assumed to contribute similarly or slightly lower compared to the understory litters (Ruds; Pulunggono et al., 2022) since there is a trade-off between plant, soil, and climatic processes at the soil surface. Although both (roots and fronds) components possess high lignin content, as well as share comparable amounts of C, cellulose, and hemicellulose (Pulunggono et al., 2019; Shrivastava et al., 2020; Wakhid and Hirano, 2021), the most rapid turnover was undergone on the primary and secondary roots that have much lower CN ratio compared to the fronds (Violita et al., 2016; Pulunggono et al., 2019).
Meanwhile, understory litters were contributed primarily by fern detrital materials, which also possibly possess a similar CN ratio with primary and secondary roots. Contrastingly, groundwater level (GWL) and peat water content/moisture (or water- filled pore space/WFPS) generally affect CO2
emission in tropical peatlands (Melling et al., 2005;
Ishikura et al., 2017; Manning et al., 2019). Carlson et al. (2015) find that the lower magnitudes of CO2
emission from different plantation sites are attributed to the rise of GWL near the peat surface, therefore, increasing the peat water content/WFPS. Furthermore, Pulunggono et al. (2022) mentioned the importance of latency or residence time of precipitation that may affect GWL and peat water content/WFPS and restrict the peat aeration, thus halting the respiration process.
However, under improper management or during the dry season/El-Nino, the deepening of GWL leads to decreasing peat water content/WFPS. It can significantly boost the microbial process in decomposing organic material near the peat surface, similar to microsites with the highest density of dead root fragments.
Measuring heterotrophic CO2 flux directly in tropical peatland is challenging due to time- consuming, expensive instruments, and highly variable weather. Some particular skills and expertise are needed before, within, and after the measurement process, as well as while the main detectors/sensors go unstable, flawed, or damaged. Furthermore, advanced statistical modeling development provides an alternative solution to cope with the limitation and difficulties of traditional direct measurement.
Artificial intelligence/AI-based models, for example, can "learn and recognize" patterns with reasonable accuracy, as well as less constrained from outliers, multicollinearity, heteroscedasticity, and non-linear relationships. Besides having the capability to manage either small or huge observation data, AI models also provide more flexibility than biochemical and process- based models due to their free-mechanism approach (Melling et al., 2005; Feng et al., 2019; Padarian et al.,
Open Access 3665 2020). These abilities are considered very appropriate
to overcome the interconnected and non-linear relationship problems demonstrated by peat chemical properties and CO2 flux (e.g., Melling et al., 2005;
Ishikura et al., 2017; Nurzakiah et al., 2021;
Pulunggono et al., 2022). Current reports demonstrate promising performances of AI-based models in predicting CO2 flux in subtropical peat and mineral soils at regional and global scales (Bond-Llamberty, 2018; Holl et al., 2020; Adjuik and Davis, 2022);
unfortunately, their application in tropical peatland environment are limited, particularly for root-related respiration. This study hypothesized that the AI-based models could conveniently handle Rdr prediction compared to the linear model.
Hence, this study is intended to simulate the effect GWL/WFPS and their residence times on Rdr by conducting a simulation study with varied roots' water content/WC and incubation time/IT. Furthermore, this study also develops several pedotransfer models and verifies their performance in predicting Rdr.
Materials and Methods Research site
This study was conducted from January 2020 to June 2021. The research site was a 10-15 years old oil palm plantation in PT Kimia Tirta Utama plantation, Buatan Village, Koto Gasib District, Siak Regency, Riau Province (0°44'44" N; 101°46'22" E), as shown in Figure 1. Moreover, the incubation process was held in the Soil Chemistry and Fertility Laboratory,
Department of Soil Science and Land Resources, Faculty of Agriculture, IPB University.
Research design, laboratory analysis, and CO2 flux measurement
This study was carried out using a single factor, completely randomized design consisting of five treatments. Five treatments were applied, consisting of control (without water addition, water content (WC) = 15%), 100% water content, 150%, 300%, and 450%.
Each treatment was repeated five times; hence, 25 experimental units were obtained. The incubation time lasted for four months.
The oil palm dead roots were purposively collected hemic peat, representing 0-60 cm depth. The samples were firstly air-dried. The soil particles were then removed through the flowing water (Figure 2).
Moreover, the roots were sliced into 2 cm fragments and then incubated inside the container (averaged 100.09 g per container). N total was determined using the Kjeldahl method; meanwhile, organic C was determined using the loss on ignition/LoI method. The roots were also weighted at the first and last observations (months 0 and 3, respectively).
CO2 flux measurement and analysis were carried out at the Soil Chemistry and Fertility Laboratory, Soil and Land Resources Department, Faculty of Agriculture, IPB University. The CO2 flux of decomposition roots was measured every two weeks using calibrated IRGA (Infra Red Gas Analyzer) Li- COR 830 with the closed-chamber method. IRGA was flowed by gas from the chamber, and CO2 flux concentration was read within ±2 minutes.
Figure 1. Map of the research site.
Open Access 3666 Figure 2. The oil palm root (a) after slices into 2 mm fragments and (b) their natural condition before being sliced.
The CO2 flux was measured concurrently with the temperature inside the chamber. Moreover, the height of the chamber lid to the surface of the root samples was also measured. The results are then calculated by the equation:
𝑓𝑐 =𝑃ℎ 𝑅𝑇
𝑑𝐶 𝑑𝑡
With the description, fc is the flux value (mg m-2 sec-
1), P is atmospheric pressure, h is the high cover of the surface, R is the gas constant (8.314 Pa m3 °K-1 mol-1), T is the temperature (°K), and is the linear change in CO2 concentration (Madsen et al., 2009).
The value calculated from the equation above is considered as CO2 emitted from the surface that entirely consisted of dead roots. Concerning the lack of studies which quantified the percentage of peat surface occupied by dead root, this study developed an approach to solve the issues using the percentage of incorporated root dry mass to total solid masses in peat (root dry mass+peat) per meter cubic of soil column at 60 cm depth and 1.5 m distance from the oil palm tree.
Incorporated root dry mass was calculated from root dry mass reported by Khalid et al. (1999) and the average of secondary and fine roots annualturnover (Corley and Tinker, 2016). Both root types were selected due to their high density and turnover compared to primary roots at the measured depth and distances (Sabiham et al., 2014; Corley and Tinker, 2016). Peat bulk density was set to 0.1 g cm-3. The authors assumed that the roots were uniformly distributed along the depth and distance to simplify the equation, following the generalization of root density as described by previous reports (Khalid et al., 1999;
Corley and Tinker, 2016) and the average depth of peat acrotelmic zone in the study area (Pulunggono et al., 2022).
Statistical analyses and CO2 flux modeling
Data collation was performed using Microsoft Excel.
Before the statistical analyses were conducted, the data was subjected to absolute value elimination, outlier removal, and normality inspection. Due to the
measurement error, CO2 fluxes exceeding 0.14 mg m-2 sec-1 were discarded. Then, log transformation was applied to attain normality. One-way analysis of variance/ANOVA was used to separately assess the difference of CO2 flux as affected by the given factors, followed by Tukey honestly significant difference/HSD as the post hoc test. Moreover, a linear mixed-effect model with restricted maximum likelihood/REML was applied to obtain multiple effects with interaction, incorporating organic C, total N, and CN ratio as random factors. Incorporating random factors in REML maximized the model's capacity (with imbalance observation data) in explaining variance compared to two-way ANOVA type-III. Furthermore, principal component analysis/PCA was also employed to assess underlying factors that may control CO2 emitted from decomposed oil palm roots. The entire analyses were executed in R environment using stat, multcompView (Graves et al., 2019), lme4 (Bates et al., 2021), factomineR (Husson et al., 2020), and factoExtra (Kassambara and Mundt, 2020) packages.
This study also developed several linear and non- linear-based pedotransfer models between CO2 flux with the treatments and several root properties in the R environment. The model was divided by two conditions, as represented by the following formulas:
Model1 <– CO2 Flux ~ WC + IT + organic C + total N + CN ratio
Model2 <– CO2 Flux ~ WC + IT + organic C + total N + CN ratio + Weight loss
This study used nested division, which firstly divided the randomized dataset (seed = 42) into training (70%) and validation/testing (30%) datasets. Secondly, the best predictive models were acquired using grid search with tenfold repeated cross-validation on the training datasets.
Several polynomials and AI-based models, including machine learnings/ML's and deep learnings/DL's techniques, were assembled to fit the models. This study also performed standard multiple (a) (b)
Open Access 3667 linear regression/MLR as a benchmark due to its wide
usage in soil modeling research. Multivariate adaptive regression splines/MARS (earth package;
Milborrow, 2021) and generalized additive model/GAM (mgcv package; Wood, 2021) were executed as a polynomial-based model. ML-based regression models used in this study consisted of regression tree/RT (rpart package; Therneau et al., 2022), random forest/RF (randomForest and ranger packages; Liaw and Wiener, 2018; Wright et al., 2021), support vector regression/SVR (e1071 package; Meyer et al., 2021), and gradient boosting machine/GBM (gbm package; Greenwell et al., 2020).
Furthermore, two feedforwards artificial neural network/ANN or multilayer perceptron/MLP models trained with resilient backpropagated and stochastic gradient descent (neuralnet package; Fritsch et al., 2019) were employed as DL method, based on single and double hidden layers. This study also employed a general AI algorithm aggregator: caret package (Kuhn et al., 2021) as the complement to the previously separated packages. Both the results were initially compared, and then the best model was selected based on the lowest RMSE. The full description of ML and DL algorithms can be found in Bruce and Bruce (2017) and Wiley and Wiley (2019).
Model agreements were assessed by comparing root mean square error/RMSE and R2 of predicted and measured CO2 flux. Variable importance was calculated using mean square error/MSE through a sum permutation-based approach, utilizing iml package (Molnar and Schratz, 2020).
Results and Discussion
Rdr emission factor and its contribution to Rh and Rt
As shown in Table 1, CO2 emission studies in tropical peatland cultivated by oil palm plantations had been conducted over decades without any significant progress in separating heterotrophic respiration of oil palm dead roots from their surrounding peat materials.
This study demonstrated the importance of Rdr among other CO2 emission sources, which can be interpreted as an initial approach to confirm newly compartment of Rh in oil palm cultivated tropical peatland. The measured Rdr emission was 0.05 to 2.3 Mg CO2 ha-1 year-1 with an average of 0.7 Mg CO2 ha-1 year-1. Assuming the Rt and Rh range from 13.8 to 178 Mg CO2 ha-1 year-1 and 5.2 to 66.3 Mg CO2 ha-1 year-1, respectively (Table 1); Rdr contribution was low, both for Rt and Rh. The contribution to Rt and Rh recorded around 0.3 to 1.3 and 0.9 to 3.5 percent, respectively.
The average and upper range of Rdr contribution to Rt
and Rh are still considered low, owing to their contribution that is under 5 percent.
These estimations presented above seemingly contrasted with the second hypothesis, which is lower than the emission emanates from oil palm fronds decomposition (Manning et al., 2019; Wakhid and Hirano, 2021; Table 1). Furthermore, the contribution of Rdr to Rt was outside the range of Ruds contribution, which accounted for around 14.6 to 21.9 Mg CO2 ha-1 year-1 (Wakhid and Hirano, 2021; Pulunggono et al., 2022; Table 1), which was in contrast with the previous assumption.
Table 1. The review of surface emission measurements and their contributed fractions in tropical peatland cultivated by oil palm plantations.
Component CO2 Emission (Mg CO2 ha-1 year-1)
Method Source
Total Respiration (Autotrophic+
Heterotrophic) 13.8 – 178
Static chamber, trenching collars,
subsidence
Hooijer et al., 2012; Dariah et al., 2014; Hergoualc’h et al., 2017;
Manning et al., 2019; Pulunggono et al., 2022
Autotrophic Respiration
Living root 30.7 - 137.7 Static chamber,
trenching, distance differences
Meiling et al., 2013; Sabiham et al., 2014
Stem 0.5 Stem chamber Manning et al., 2019
Heterotrophic Respiration Peat materials(+OP
dead roots) 5.2 – 66.3
Trenching, trenching collars, static chamber,
distance differences
Dariah et al., 2014; Husnain et al., 2014; Marwanto and Agus, 2014;
Comeau et al., 2016; Hergoualc’h et al., 2017; Manning et al., 2019;
Pulunggono et al., 2022 Understory litters 14.6 – 21.9 Trenching collars Pulunggono et al., 2022
OP fronds 4 – 6.6 litter bag
decomposition/
gravimetric
Wakhid and Hirano, 2021
OP dead roots 0.05 – 2.3 Separation and
incubation This study
Drainage canals 1.4 – 1.6 Floating chamber Manning et al., 2019
Open Access 3668 This study treated the dead root with various WC
(Figure 3) to simulate the rise and deepening of GWL in the field. According to the figure, some of the treatments provide favorable conditions for microorganisms to thrive, resulting in maximum emissions as shown in Table 1 (2.3 Mg CO2 ha-1 year-1). This value is lower than the lower level of Rfr
and Ruds. This study’s approach calculated root incorporation for only at a single year period, which is quite less flexible compared to their continuous process in nature. That limitation thereby restricts maximum heterotrophic respiration that feeds entirely from root dead materials (45.1 Mg CO2 ha-1 year-1) to their corrected value at the field level (2.3 Mg CO2 ha-
1 year-1; Table 1). However, this study approach is still considered reasonable since there are large voids of available information, particularly in quantifying heterotrophic emission emanates from oil palm dead root fragments or basic knowledge regarding the percentage of specific oil palm root turnover near the peat surface during an extended monitoring period.
Compared to the living roots, their decomposed dead parts contribute to total CO2 emissions about 97.03 percent lower (Meiling et al., 2013; Sabiham et al., 2014; Table 1). The amount and contribution of Rdr
are likely originated from a relatively quick turnover (primarily consisting of secondaries and fine roots), low CN ratio, and high magnitude of incorporated dry matter (Corley and Tinker, 2016). However, their maximum potential contribution was technically cannot be reached at the study site due to high precipitation and shallow GWL (Busman et al., 2021;
Pulunggono et al., 2022), as well as lower and stabilized soil temperature under dense leaves shading of matured oil palm and pruned fronds coverage (Jauhainen et al., 2014; Manning et al., 2019).
Figure 3 shows the significant effect of WC in governing Rdr as assessed by one-way analysis of variance and linear mixed-effects model with restricted maximum likelihood/REML. Statistically different and highest Rdr emanated under the 100 and 150% of WC treatments. However, decreasing (control/15%) or increasing WC (300-450%) could generate lower Rdr. Furthermore, the treatment of 100 and 150%WC were also generating a relatively lower organic CN ratio and higher N compared to other treatments (Figure 4), whereas organic C seems to vary. Concerning CO2 emission as a product of the microbial process, organic material with higher N and lower CN ratios would result in a higher respiration rate, hence, increasing CO2 flux (Hutchings and Martin, 1934). These findings corroborate the general understanding that the dynamic of GWL in cultivated peatland must be maintained to sustain CO2 emission at its lowest rate (Carlson et al., 2015; Manning et al., 2019). To obtain lower Rdr, in this study case, WC must be maintained at around 300 to 400%.
Furthermore, Rdr culminated in the first month and was significantly different from the zero and last months. Rdr inclined until the last observation, with both Rdr were not significantly different from the early measurement. Rdr at the first two months seemingly represents the variation of CN ratio possessed by diverse types of dead roots rather than the RT.
Figure 3. The effects of water content (WC; 3a) and incubation time (IT; 3b) to Rdr. The letters above the subfigures 3a and 3b were derived from multiple comparisons of REML. Subfigures 3c, 3d, and 3e represent the
relationships among organic C, total N, and CN ratio to Rdr. The white asterisk and black horizontal line inside the boxplot represent mean and median values of of CO2 flux in each class/factor, respectively; grayish dotted
line represents the general averaging of CO2 flux after outlier removal.
However, the authors supposed that high Rdr in the first month represented fine roots decomposition which had a lower CN ratio (Violita et al., 2016), whereas roots with a high CN ratio would decompose slowly (Figure
3d; Figure 4c), referring to lower Rdr at the following months. With regard to the high CN ratio used in this study, possibly longer period of measurements is needed to gain a general pattern of Rdr and separation
Open Access 3669 of roots based on its diameter, order, or function as
other studies report (e.g., Moradi et al., 2013; Sabiham et al., 2014; Yang et al., 2019). Even though both fixed factors held strong effects on the dynamic Rdr, they differently relate to Rdr, as exhibited by the difference in Rdr pattern. By dropping control treatment, Rdr
seemingly possessed a linear relationship with WC (Figure 3a). However, non-linear trends showed by Rdr
along with the increase of IT (Figure 3b). The entire root properties (organic C, total N, and CN ratio;
Figure 3c, 3d, and 3e, respectively) demonstrated a
scattered pattern and possibly had non-linear relationships to Rdr, albeit their trends (showed as a red line in Figure 3) were in agreement with the general consensus of decomposition rate. Higher C content and CN ratio, as well as lower N content in organic material, would result in slower rates of microbial decomposition and vice versa (Hutchings and Martin, 1934). REML analysis also showed that incorporating organic as a random factor affects Rdr, whereas the other two covariates (N total and CN ratio) had no role in the REML model.
Figure 4. The effects of WC and IT on organic C (4a, 4b), total N (4 (4c, 4d), and CN ratio (4e, 4f) of dead oil palm root. The white asterisk and black horizontal line inside the boxplot represent mean and median values in
each class/factor, respectively; grayish dotted line represents the general averaging after outlier removal.
Descriptive multivariate analysis (PCA) used in this study revealed five principal components/PC's that accounted 100% of the total variance, as shown by screen plots in Figure 5a. This study utilized PC1, PC2 and PC3 since they had an eigenvalue of more than 1 and explained 87.5 of the total variance (Sharma, 1996; Figure 5a). CN ratio and N total significantly dominated PC1 (Figure 5b1). Meanwhile, organic C and weight loss were loaded solely by PC2 (Figure 5b2). Moreover, PC3 statistically correlated with Rdr. The figure showed that PC1 and PC2 represented dead oil palm root chemical (organic C, N total and CN ratio) and physical (weight loss) properties, whereas PC3 was designated for its emission.
PCA results supported this study's previous information generated from one-way ANOVA and REML, particularly in revealing underlying factors
that affect dead oil palm root characteristics and their relationships with Rdr. According to color grouping on the entire observation plot in Figures 5c1.1, 5c2.1, 5c3.1, the treatment (combined WC and IT) used in this study affected dead oil palm root characteristics, either its chemical and physical properties or emission, as shown by treatment group (blue points) separated from control (red points). However, changing oil palm dead root WC into several classes could not fully separate the dead root characteristics, as exhibited by several overlapping groups (Figures 5c1.2, 5c2.2, 5c3.2). Furthermore, visual grouping generally resulted in four groups, merging 100 and 150%WC.
The gradual change was observed based on IT at all observation plots (Figures 5c1.3, 5c2.3, 5c3.3), indicating the entire dead oil palm root characteristics were affected by RT.
Open Access 3670 Figure 5. Statistic description of PCA analysis (a) scree plot percentage of explained variances of each PC; (b) contribution (expressed as percentages) of each variable in (b1) PC1, (b2) PC2, (b3) PC3; and biplot consisting of loading and observation plot of (c1) PC1:PC2, (c2) PC1:PC3, and PC2:PC3 (c3), grouped by (.1) status, (.2)
water content, and (.3) incubation time. Note that 15%WC represented control treatment.
5a 5b1 5b2 5b3
5c1.1 5c1.2 5c1.3
5c2.1 5c2.2 5c2.3
5c3.1 5c3.2 5c3.3
Open Access 3671 Modeling Rdr based on treatments and several root
properties
This study highlighted the potential capability of polynomial and AI-based models as alternative pedotransfer models that can predict Rdr from cultivated tropical peatlands, as affected by WC and IT, with an acceptable agreement. Concerning the importance and significance of GHG emission in tropical peatland ecosystems, this study's results can be used as starting point to develop more complex models of CO2 emission representing Rh, Ra, Rt, and net ecosystem exchange/NEE in oil palm cultivated peatland.
In model 1, DL and ML-based models (particularly for ensemble learning: RF and GBM) outperformed linear and polynomial models. In general, both performed comparable RMSE, whereas SVR and GBM were considered more powerful at predicting CO2 flux with the inclusion of weight loss parameters. The ensemble learning models were seemingly stable and superior in R2 term (Figure 7), as shown by the close distance between predicted and observed values, as well as a converged pattern of their predicted value close to the red line. MARS model, surprisingly, possessed lower RMSE with relatively similar R2 to more complex AI-based models performances (i.e., ML-based models: TR, SVR, and GBM) in Model1. The high agreement achieved by several models in this study is considered reasonable since a nested separation strategy (single splitting and ten-fold cross-validation) was applied to avoid model overfitting.
With respect to higher R2 acquired from of some the best models, the modeling approach in this study is considered successful in capturing the non-linear relationship between Rdr and its predictors (Figures 3, 6, and 7) compared to other researchers who applied the classical linear approach (e.g., Marwanto and Agus, 2013; Batubara et al., 2019; Pulunggono et al., 2022). However, the remarkable accuracy that advanced ML-based models can achieve often expenses the most functional basic possessed by the model itself: "the interpretability." Many researchers, soil scientists, and practitioners relied mostly on the model interpretability to link the new information provided by the model to their current and established knowledge and then further combine this "new insight"
into the decision-making process or explain complex soil process/relations (Rossiter, 2018; Padarian et al., 2020). The most interpretable ML-based model, TR,
generally obtained reasonable accuracy in this study, especially in Model1 (Figure 7), higher than Melling et al. (2005), who modeled CO2 flux from soil temperature, bulk density, relative humidity, and WFPS. Based on TR (Figures 8 and 9), the incubation time was the most influential factor in predicting Rdr, followed by organic C. Lower Rdr (averaged: 0.03 mg m-2 sec-1) was divided from the higher Rdr (averaged:
0.06 mg m-2 sec-1) by incubation time more than three months. Furthermore, organic-C more than 55%
separated the lower Rdr (averaged: 0.044 mg m-2 sec-1) from the higher Rdr (0.09 mg m-2 sec-1). The recursive partitioning algorithm of TR can be easily extracted and interpretable, constructing novel information with regard to Rdr prediction.
Interestingly, more powerful techniques -DL- based model (ANN1 and ANN2)-exhibited lower performance compared to more simplified models, for instance, MARS (polynomial-based models), and MLR (linear model). ANN also demonstrated worst agreement in Model2, particularly for R2, returning a high discrepancy between their predicted against actual/observed values. These results were contrasted with Holl et al. (2020) and Dou et al. (2018), who denoted that ANN performed better than MLR-based models in predicting CO2 flux measured from eddy covariance sensors. However, the worse performance of ANN had been well documented in Feng et al.
(2019). These indicate that traditional shallow ANN (single or double hidden layers) that were used in this study could not conveniently handle small datasets (compared to both previous reports). To overcome this limitation, Feng et al. (2019) proposed combination techniques, exploiting the deep architecture of neural networks, as well as introducing pre-training and fine- tuning analyses.
Based on the model comparison, this study supposed that eliminating the weight loss variable significantly improved the model accuracy. According to Figures 6 and 7, Model1 generally exhibited good agreement at the entire metrics, particularly at ML and polynomial-based models. Meanwhile, Model2 had a lower agreement, both in RMSE and R2, compared to Model1. Moreover, several best performance models (i.e., GAM, TR, RF, GBM) in Figure 9 also indicated some variables that gained less importance. Based on that, this study suggested that increasing variable quantities, even if supported by established knowledge (in this study case: theoretically related to CO2 flux and microbial respiration), does not always relate to the increasing model accuracy
Figure 6. The comparison of RMSE (expressed as percentages) among linear (MLR), polynomial (MARS and GAM), ML (TR, RF, RF, SVR, and GBM), and DL (NN1 and NN2)-based models on prediction Rdr.
0.22 0.25
0.14 0.17
0.19 0.32
0.16 0.17
0.13 0.16
0.22 0.15
0.16 0.15
0.13 0.25
0.11 0.36
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
RMSE1 RMSE2
MLR MARS GAM TR RF SVR GBM NN1 NN2
Open Access 3672 Figure 7. Model agreement as represented by the coefficient of determination (R2) of predicted against observed
Rdr.
Figure 8. Model1 regression tree of Rdr modeled by its most important predictors (incubation time and organic C). The value inside and under the boxes are Rdr means and observation number, respectively.
Open Access 3673 Figure 9. Feature importance computed using a permutation-based approach, utilizing mean square error/MSE as
a loss function.
Open Access 3674 Conclusion
This study emphasized separating Rdr from Rh in oil palm cultivated peatland by conducting a simulation study as an effort to mitigate greenhouse gas emissions. Rdr emissions were 0.05 to 2.3 Mg CO2 ha-1 year-1 with an average of 0.7 Mg CO2 ha-1 year-1. Rdr to Rh and Rt contribution were around 0.3 to 1.3%
and 0.9 to 3.5%, respectively; based on the wide reports of Rt and Rh in oil palm cultivated peatlands (13.8 to 178 Mg CO2 ha-1 year-1 and 5.2 to 66.3 Mg CO2 ha-1 year-1, respectively). According to evidence supported by various statistical analyses (i.e., one-way ANOVA, REML, and PCA), WC and IC chiefly affected Rdr, which favorable condition for heterotrophic respiration in dead root was achieved at 100-150%WC and the first month after incubation.
Advanced pedotransfer models based on polynomial (MARS) and AI (RF and GBM) employed in this study provide new insight into predicting Rdr as a new compartment in Rh research, as well as can be used as starting point to develop more complex models of CO2
emission representing Rh, Ra, Rt, and net ecosystem exchange/NEE in oil palm cultivated peatland.
Acknowledgements
The authors acknowledged The Indonesian Oil Palm Plantation Fund Management, which funded this research.
The authors also thanked the management and field staff of Astra Agro Lestari Tbk. for their support and assistance during the fieldwork.
References
Adjuik, T.A. and Davis, S.C. 2022. Machine learning approach to simulate soil CO2 fluxes under cropping
systems. Agronomy 12: 197,
doi:10.3390/agronomy12010197.
Bates, D., Maechler, M., Bolker, B., Walker, S., Christensen, R.H.B., Singmann, H., Dai, B., Scheipl, F., Grothendieck, G., Green, P., Fox, J., Bauer, A. and Krivitsky, P.A. 2021. Package' lme4.' Linear Mixed- Effects Models using 'Eigen' and S4. Retrieved from https://cran.r-project.org/web/packages/
lme4/index.html.
Batubara, S.F., Agus, F., Rauf, A., and Elfiati, D. 2019.
Impact of soil collar insertion depth on microbial respiration measurements from tropical peat under an oil palm plantation. Peat and Mires 24(6):1-11, doi:10.19189/MaP.2018.DW.373.
Bond-Lamberty, B. 2018. New techniques and data for understanding the global soil respiration flux. Earth's Future 6: 1176–1180, doi:10.1029/2018EF000866.
BPS-Statistic Indonesia. 2020. Indonesian Oil Palm Statistics 2020. Jakarta (ID): BPS-Statistic Indonesia.
BPS-Statistics Indonesia. 2022. Statistical Yearbook of Indonesia 2022. Jakarta (ID): BPS-Statistic Indonesia Bruce, P. and Bruce, A. 2017. Practical Statistics for Data
Scientists. California (US): O'Reilly Media, Inc.
Busman, N.A., Maie, N., Ishak, C.F., Sulaiman, M.F. and Melling, L. 2021. Effect of compaction on soil CO2 and CH4 fluxes from tropical peatland in Sarawak, Malaysia.
Environment, Development and Sustainability 23:11646-11659, doi:10.1007/s10668-020-01132-y.
Carlson, K.M., Goodman, L.K. and May-Tobin, C.C. 2015.
Modeling relationships between water table depth and peat soil carbon loss in Southeast Asian plantations.
Environmental Research Letters 10(7):074006, doi:10.1088/1748-9326/10/7/074006.
Comeau, L.-P., Hergoualc’h, K., Hartill, J., Smith, J., Verchot, L.V., Peak, D. and Salim, A.M. 2016. How do the heterotrophic and the total soil respiration of an oil palm plantation on peat respond to nitrogen fertilizer application?. Geoderma 268:41-51, doi:10.1016/j.geoderma.2016.01.016.
Corley, R.H.V. and Tinker, P.B. 2016. The Oil Palm, Fifth Edition. West Sussex (UK): Blackwell Science Ltd., doi:10.1002/9781118953297.
Dadap, N.C., Hoyt, A.M., Cobb, A.R., Oner, D., Kozinski, M., Fua, P.V., Rao, K., Harvey, C.F. and Connings, A.G.
2021. Drainage canals in Southeast Asian peatlands increase carbon emissions. AGU Advances 2 e2020AV000321, doi:10.1029/2020AV000321.
Dariah, A., Marwanto, S. and Agus, F. 2014. Root- and peat- based CO2 emissions from oil palm plantations.
Mitigation and Adaptation Strategies for Global Change 19(6):831-843, doi:10.1007/s11027-013-9515-6.
Deshmukh, C.S., Julius, D., Desai, A.R., Asyhari, A., Page, S.E., Nardi, N., Susanto, A.P., Nurholis, N., Hendrizal M., Kuniano, S., Suardiwerianto, Y., Salam, Y.W., Agus, P., Astiani, D., Sabiham, S., Gauci, V. and Evans, C.D. 2021. Conservation slows down emission increase from a tropical peatland in Indonesia. Nature Geoscience 14(7):484-490, doi:10.1038/s41561-021-00785-2.
Dhandapani, S., Girkin, N.T. and Evers, S. 2022. Spatial variability of surface peat properties and carbon emissions in a tropical peatland oil palm monoculture during a dry season. Soil Use and Management 38(1):381-395, doi:10.1111/sum.12741.
Dohong, A., Abdul Aziz, A. and Dargusch, P. 2018. A review of techniques for effective tropical peatland restoration. Wetlands 38(2):275-292, doi:10.1007/s13157-018-1017-6.
Dou, X., Yang, Y. and Luo, J. 2018. Estimating forest carbon fluxes using machine learning techniques based on eddy covariance measurements. Sustainability 10(1):203, doi:10.3390/su10010203.
Feng, S., Zhou, H. and Dong, H. 2019. Using deep neural network with small dataset to predict material defects.
Materials and Design 62:300-310, doi:10.1016/j.matdes.2018.11.060.
Fritsch, S., Guenther, F., Wright, M.N., Suling, M. and Mueller, S.M. 2019. Package 'neuralnet'. Training of Neural Networks. Retrieved from https://cran.r- project.org/web/packages/neuralnet/index.html.
Graves, S., Piepho, H.-P., Selzer L. and Dorai-Raj. S. 2019.
Package' multcompView.' Retrieved from https://cran.r- project.org/web/packages/multcompView/index.html.
Greenwell, B., Boehmke, B. and Cunningham, J. 2020.
Package 'gbm'. Generalized Boosted Regression Models.
Retrieved from https://cran.rproject.org/web/
packages/gbm/index.html.
Hergoualc’h, K., Hendry, D.T., Murdiyarso, D. and Verchot, L.V. 2017. Total and heterotrophic soil respiration in a swamp forest and oil palm plantations on peat in Central Kalimantan, Indonesia. Biogeochemistry 135(3):203–
220, doi:10.1007/s10533-017-0363-4.
Open Access 3675 Holl, D., Pfeiffer, E.-M. and Kutzbach, L. 2020. Comparison
of eddy covariance CO2 and CH4 fluxes from mined and recently rewetted sections in a northwestern German cutover bog. Biogeosciences 17(10):2853-2874, doi:10.5194/bg-17-2853-2020.
Hooijer, A., Page, S., Jauhiainen, J., Lee, W.A., Lu, X.X., Idris, A., and Anshari, G. 2012. Subsidence and carbon loss in drained tropical peatlands. Biogeosciences 9:
1053-1071, doi:10.5194/bg-9-1053-2012.
Husnain, H., Wigena, I.G.P., Dariah, A., Marwanto, S., Setyanto, P. and Agus, F. 2014. CO2 emissions from tropical drained peat in Sumatra, Indonesia. Mitigation and Adaptation Strategies for Global Change 19(6):845- 862, doi:10.1007/s11027-014-9550-y.
Husson, F., Josse, J., Le, S., and Mazet, J. 2020. Package
‘FactoMineR’. Multivariate Exploratory Data Analysis and Data Mining. Retrieved from https://cran.r- project.org/web/packages/FactoMineR/index.html.
Hutchings, I. J. and Martin, T. L. 1934. Influence of the carbon-nitrogen ratios of organic matter on rate of decomposition in the soil. Agronomy Journal 26(4):333, doi:10.2134/agronj1934.00021962002600040012x.
Ishikura, K., Yamada, H., Toma, Y., Takakai, F., Morishita, T., Darung, U., Limin, A., Limin, S.H. and Hatano, R.
2017. Effect of groundwater level fluctuation on soil respiration rate of tropical peatland in Central Kalimantan, Indonesia. Soil Science and Plant Nutrition 63(1):1-13, doi:10.1080/00380768.2016.1244652.
Jauhiainen, J., Kerojoki, O., Silvennoinen, H., Limin, S. and Vasander, H. 2014. Heterotrophic respiration in drained tropical peat is greatly affected by temperature - a passive ecosystem cooling experiment. Environmental Research Letters 9(10):105013, doi:10.1088/1748- 9326/9/10/10501.
Kassambara, A. and Mundt, F. 2021. Package: 'factoextra'.
Extract and Visualize the Results of Multivariate Data Analyses. Retrieved from https://cran.r- project.org/web/packages/factoextra/index.html.
Khalid, H., Zin, Z.Z. and Anderson, J.M. 1999.
Quantification of oil palm biomass and nutrient value in mature plantation. II Below-ground biomass. Journal of Oil Palm Research 11(2):63-71.
Kuhn, M., Wing, J., Weston, S., Williams, A., Keefer, C., Engelhardt, A., Cooper T., Mayer Z., Kenkel, B., R Core Team, Benesty, M., Lescarbeau, R., Ziem A., Scrucca, L., Tang, Y., Candan, C., and Hunt, T. 2021. Package
‘caret’. Classification and Regression Training.
Retrieved from https://cran.r-
project.org/web/packages/caret/.
Larionova, A.A., Sapronov, D.V., Lopez de Gerenyu, V.O., Kuznetsova, L.G. and Kudeyarov, V.N. 2006.
Contribution of plant root respiration to the CO2
emission from soil. Eurasian Soil Science 39(10):1127- 1135, doi:10.1134/s1064229306100103.
Leifeld, J., Wüst-Galley, C. and Page, S. 2019. Intact and managed peatland soils as a source and sink of GHGs from 1850 to 2100. Nature Climate Change 9:945-947, doi:10.1038/s41558-019-0615-5.
Liaw, A. and Wiener, M. 2018. Package 'randomForest'.
Breiman and Cutler's Random Forests for Classification and Regression. Retrieved from https://cran.r- project.org/web/packages/rpart/index.html.
Madsen, R., Xu, L., Claassen, B. and McDermitt, D. 2009.
Surface monitoring method for carbon capture and storage projects. Energy Procedia 1(1):2161-2168, doi:10.1016/j.egypro.2009.01.281.
Manning, F.C., Kho, L.K., Hill, T.C., Cornulier, T. and Teh, Y.A. 2019. Carbon emissions from oil palm plantations on peat soil. Frontiers in Forests and Global Change 2:
37, doi:10.3389/ffgc.2019.00037.
Marwanto, S. and Agus, F. 2013. Is CO2 flux from oil palm plantations on peatland controlled by soil moisture and/or soil and air temperatures? Mitigation and Adaptation Strategies for Global Change 19(6):809-819, doi:10.1007/s11027-013-9518-3.
McCalmont, J., Kho, L.K., Teh, Y.A., Lewis, K., Chocholek, M., Rumpang, E. and Hill, T. 2021. Short‐ and long‐term carbon emissions from oil palm plantations converted from logged tropical peat swamp forest. Global Change Biology 27(11):2361-2376, doi:10.1111/gcb.15544.
Melling, L., Cindy Soo, Y.T., Kah, J.G. and Hatano, R. 2013.
Soil microbial and root respirations from three ecosystems in tropical peatland of Sarawak, Malaysia.
Journal of Oil Palm Research 25:44-57.
Melling, L., Hatano, R. and Goh, K.J. 2005. Soil CO2 flux from three ecosystems in tropical peatland of Sarawak, Malaysia. Tellus B: Chemical and Physical Meteorology 57(1):1-11, doi:10.3402/tellusb.v57i1.16772.
Meyer, D., Dimitriadou, E., Hornik, K., Weingessel, A., Leisch, F., Chang, C.C. and Lin, C.C. 2021. Package' e1071.' Misc Functions of the Department of Statistics, Probability Theory Group (Formerly: E1071), TU Wien.
Retrieved from https://cran.r-
project.org/web/packages/e1071/index.html.
Miettinen, J., Hooijer, A., Vernimmen, R., Liew, S.C. and Page, S.E. 2017. From carbon sink to carbon source:
extensive peat oxidation in insular Southeast Asia since 1990. Environmental Research Letters 12:024014, doi:10.1088/1748-9326/aa5b6f.
Milborrow, S. 2021. Package 'earth.' Multivariate Adaptive Regression Splines. Retrieved from https://cran.r- project.org/web/packages/earth/index.html.
Molnar, C. and Schratz, P. 2020. Package' iml.' Interpretable Machine Learning. Retrieved from https://cran.r- project.org/web/packages/iml/index.html.
Moradi, A., Teh, C.B.S., Goh, K.J., Husni, M.H.A. and Ishak, C.F. 2013. Decomposition and nutrient release temporal pattern of oil palm residues. Annals of Applied Biology 164(2):208-219, doi:10.1111/aab.12094.
Nurzakiah, S., Sutandi, S., Djajakirana, G., Sudadi, U. and Sabiham, S. 2021. The contribution of organic acid on heterotrophic CO2 flux from tropical peat: a trenching study. Journal of Degraded and Mining Lands
Management 9(1):3035-3044,
doi:10.15243/jdmlm.2021.091.3035.
Padarian, J., Minasny, B., and McBratney, A.B. 2020.
Machine learning and soil sciences: a review aided by machine learning tools. Soil 6(1):35-52, doi:10.5194/soil-6-35-2020.
Pulunggono, H.B., Anwar, S., Mulyanto, B. and Sabiham, S.
2019. Decomposition of oil palm frond and leaflet residues. AGRIVITA: Journal of Agricultural Science 41(3):524-536, doi:10.17503/agrivita.v41i3.2062.
Pulunggono, H.B., Siswanto, Mubarok, H., Widiastuti, H., Tambusai, N., Zulfajrin, M., Anwar, S., Taniwiryono, D., Sumawinata. B. and Sabiham. S. 2022. Seasonal litter contribution to total peat respiration from drained tropical peat under mature oil palm plantation. Journal of Degraded and Mining Lands Management 9(2):3247- 3263, doi:10.15243/jdmlm.2022.092.3247.
Open Access 3676 Rossiter, D.G. 2018. Past, present & future of information
technology in pedometrics. Geoderma 324:131-137, doi:10.1016/j.geoderma.2018.03.009.
Sabiham, S., Marwanto, S., Watanabe, T., Funakawa, S., Sudadi, U. and Agus, F. 2014. Estimating the relative contributions of root respiration and peat decomposition to the total CO2 flux from peat soil at an oil palm plantation in Sumatra, Indonesia. Tropical Agriculture and Development 58(3): 87-93, doi:10.11248/jsta.58.87.
Sharma, S. 1996. Applied Multivariate Techniques. New York (US): Wiley.
Shrivastava, P., Khongphakdi, P., Palamanit, A., Kumar, A.
and Tekasakul, P. 2020. Investigation of physicochemical properties of oil palm biomass for evaluating potential of biofuels production via pyrolysis processes. Biomass Conversion and Biorefinery 11(5):1987-2001, doi:10.1007/s13399-019-00596-x.
Siang, C.S., Wahid, S.A.A., and Sung, C.T.B. 2022.
Standing biomass, dry-matter production, and nutrient demand of tenera oil palm. Agronomy 12:426, doi:10.3390/agronomy12020426.
Therneau, T., Atkinson, B. and Ripley, B. 2022. Package 'rpart'. Retrieved from https://cran.r- project.org/web/packages/rpart/index.html.
Violita, V., Triadiati, T., Anas, I. and Miftahudin, M. 2016.
Fine root production and decomposition in lowland rainforest and oil palm plantations in Sumatra, Indonesia. HAYATI Journal of Biosciences 23(1):7-12, doi:10.1016/j.hjb.2015.10.008.
Wakhid, N. and Hirano, T. 2021. Contribution of CO2
emission from litter decomposition in an oil palm plantation on tropical peatland. IOP Conference Series:
Earth and Environmental Science 648:012133, doi:10.1088/1755-1315/648/1/012133.
Wiley, M. and Wiley, J.F. 2019. Advanced R Statistical Programming and Data Models: Analysis, Machine Learning, and Visualization. Berkeley (US): Apress, doi:10.1007/978-1-4842-2872-2.
Wood, 2021. Package' mgcv.' Mixed GAM Computation Vehicle with Automatic Smoothness Estimation.
Retrieved from https://cran.r-
project.org/web/packages/mgcv/index.html.
Wright, M.N., Wager, S. and Probst, P. 2021. Package 'ranger'. A Fast Implementation of Random Forests.
Retrieved from https://cran.r-
project.org/web/packages/ranger/index.html.
Yang, S., Cheng, R., Xiao, W., Shen, Y., Wang, L., Guo, Y.
and Sun, P. 2019. Heterogeneity in decomposition rates and nutrient release in fine-root architecture of Pinus massoniana in the three gorges reservoir area. Forests 11(1):14, doi:10.3390/f11010014.
Zhou, W., Zhu, J., Ji, H., Grace, J., Sha, L., Song, Q., Liu, Y., Bai, X., Lin, Y., Gao, J., Fei, X., Zhou, R., Tang, J., Deng, X., Yu, G., Zhang, J., Zheng, X., Zhao, J. and Zhang, Y. 2021. Drivers of difference in CO2 and CH4
emissions between rubber plantation and tropical rainforest soils. Agricultural and Forest Meteorology 304-305:108391, doi:10.1016/j.agrformet.2021.108391.
