Influence of land-use change and season on soil greenhouse gas emissions from a tropical wetland: A stepwise explorative assessment

• CO2 emissions are influenced by season and not land-use change. • Conversion of wetlands to farmland promotes CH4 uptake regardless of the season. • Conversion of wetlands to farmland increases N2O emissions during the dry season. • Beyond season and land-use, analyses reveal the effects of soil parameters on emissions. • Wetland conversion to farmland jeopardizes climate change mitigation. ⁎ Corresponding author. E-mail address: ondiek.risper7@gmail.com (R.A.' Ondi 1 Equally contribution authors. https://doi.org/10.1016/j.scitotenv.2021.147701 0048-9697/© 2021 The Authors. Published by Elsevier B.V a b s t r a c t a r t i c l e i n f o


H I G H L I G H T S
• CO 2 emissions are influenced by season and not land-use change. • Conversion of wetlands to farmland promotes CH 4 uptake regardless of the season. • Conversion of wetlands to farmland increases N 2 O emissions during the dry season. • Beyond season and land-use, analyses reveal the effects of soil parameters on emissions. • Wetland conversion to farmland jeopardizes climate change mitigation.

G R A P H I C A L A B S T R A C T
a b s t r a c t a r t i c l e i n f o

Introduction
Wetlands cover around 5-8% of the earth's surface area (7-10 million km 2 ) but constitute one of the largest natural sources of greenhouse gases (GHGs) (Mitsch and Gosselink, 2007). According to Mitsch et al. (2013), even though wetlands emit CH 4 , they constitute net carbon sinks of about 830 Tg-C year −1 . For example, wetlands sequester about 1280 Tg-C year −1 of CO 2 from the atmosphere and emit 448 Tg-C year −1 of CH 4 back into the atmosphere (Mitsch et al., 2013). Therefore, wetlands are essential climate regulators, underlining the urgency of promoting their sustainable use (Batson et al., 2015;Liu et al., 2017;Zedler and Kercher, 2005).
Wetlands are among the most threatened ecosystems worldwide (Houghton et al., 2012;MEA, 2005). Their decline is mainly attributed to conversion to agriculture due to increasing food demand (MEA, 2005). Also, increased river water withdrawals due to economic development contribute to the wetland decline (Dixon and Wood, 2003;Hayes et al., 2018;Junk et al., 2013;Mitchell, 2013). During the twentieth century alone, over 50% of all wetland areas in Europe, North America, New Zealand, and Australia were converted to agriculture (MEA, 2005). For other regions, the quantities remain speculative as reliable data are missing (MEA, 2005). In Kenya, for instance, studies estimate that the losses range between 34 and 55% during the last four to five decades. In these cases, the conversion to agriculture is the primary reason for this decline Owino and Ryan, 2007). Although the conversion and drainage of wetlands contribute to crop production, soil GHG emissions from the ecosystem may also be altered, putting the climate regulating function of wetlands at stake (Nath and Lal, 2017;Zedler and Kercher, 2005).
Land-use change can alter the emissions of soil GHG such as CO 2 , N 2 O, and CH 4 . The drainage of peatlands to agricultural land, for example, can lead to enhanced emissions of N 2 O and CO 2 while simultaneously decreasing CH 4 emissions (Smith and Conen, 2004). Alteration of wetland's hydrology through drainage and cultivation interferes with the ecosystems' anaerobic conditions. Under aerobic conditions, soil organic matter decomposes much faster, increasing CO 2 emissions (Mitsch et al., 2013;Smith and Conen, 2004). In contrast, flooding inhibits the atmospheric oxygen supply to the soil, resulting in anaerobic decomposition of soil organic matter (Reddy and DeLaune, 2008), leading to increased CH 4 production. Methane is a potent GHG that, over a 100-year period, has a 25-fold global warming potential than that of CO 2 (Whiting and Chanton, 2001). Therefore, the drainage of wetlands can contribute to mitigating CH 4 emissions, as they produce 20-25% of global CH 4 emissions (ca. 115-227 Tg year −1 ) (Bergamaschi et al., 2007;Bloom et al., 2010;Whalen, 2005). Of these emissions, tropical wetlands emit a substantial share of around 138 Tg year −1 (Bergamaschi et al., 2007).
Nitrous oxide is another potent GHG. It has a global warming potential 300 times that of CO 2 over a 100-year period (IPCC, 2013). Nitrous oxide emissions from wetlands stem from coupled nitrificationdenitrification processes (Bernal and Mitsch, 2012;Hernandez and Mitsch, 2007). However, under strictly anaerobic conditions in wetlands, the major end-product of denitrification is N 2 , thus reducing N 2 O emissions Reddy and DeLaune, 2008). However, the drainage and conversion of wetlands to intensified agriculture are known to increase N 2 O emissions because of increased nitrogen input from fertilizer application and aerobic soil conditions (Reddy and DeLaune, 2008). In sub-Saharan Africa (SSA), however, smallholder farms, which encompass nearly 80% of farmland and up to 90% of farms, apply only a little or no fertilizer at all (Altieri and Koohafkan, 2008;Pelster et al., 2017). Therefore, under such nonintensive agricultural practices, it remains to be elucidated if the conversion of wetlands to agriculture enhances N 2 O emissions.
Given the importance of wetlands in biogeochemical cycles, their conversion to agriculture is expected to have considerable impacts on national and regional GHG emission budgets. The changes in GHG emissions with land-use change are also expected to be influenced by the seasonal changes related to dry and rainy periods. However, knowledge on the biogeochemistry of tropical wetlands remains limited (Bernal and Mitsch, 2013), as most relevant studies were conducted in temperate regions with entirely different climatic conditions than those in the tropics (Bridgham et al., 2006;Euliss et al., 2006;Tangen et al., 2015). Few studies assessed the effect of land-use change on wetland GHG emissions in Africa's tropical region (CO 2 : Saunders et al. (2012); N 2 O: Tshering (2012)), and, to our knowledge, no research has so far considered seasonal changes. Hence, it is crucial to quantify GHG emissions in different seasons.
Soil properties are the main drivers of soil GHGs production and consumption Davidson and Janssens, 2006). Soil properties such as moisture content control oxygen concentration in soils and subsequently microbial processes such as the degradation of organic matter, thus regulating GHG emissions Reddy and DeLaune, 2008). The production of N 2 O, for example, depends on soil conditions, including moisture, pH, temperature, oxygen concentration, carbon, ammonium (NH 4 -N), or nitrate (NO 3 -N) (Reddy and DeLaune, 2008;Zhu et al., 2013). Therefore, considering that land-use change, and season can modify soil properties, it is expected that the interrelations of these influences determine soil GHG emissions.
This study aims to assess the effects of land-use change, season, and soil properties on soil GHG emissions from wetlands in Africa's tropical region. Therefore, we conducted field measurements of the emission rates of three prevalent GHGs -CH 4 , CO 2 , and N 2 Ofrom areas converted to farmland and unconverted areas of the Anyiko wetland in Kenya. Our key research question is how season influences the variation in GHGs emissions between the converted and unconverted wetland areas. Also, we aim to identify which soil parameters influence GHGs emissions' variation. The available dataset does not meet the necessary mathematical-statistical preconditions for general linear models (e.g., multivariate distribution properties, number, structure, and selection of cases) and does not meet inferential statistical requirements. This status raises the question of which methodological approach can deal with these preconditions commonly encountered in field studies.

Study area
The study was conducted in the Anyiko wetland in Siaya County in Kenya (Fig. 1). The wetland is in the Nzoia River Basin, the largest Kenyan sub-basin of Lake Victoria (MEMR, 2012). The Anyiko wetland currently (2018) covers an area of ca. 0.7 km 2 . Papyrus (Cyperus papyrus L.) is the dominant vegetation type in the wetland.
The Anyiko wetland receives two rainy seasons (March-May and October-December) annually. The mean annual rainfall is 1556 and ranges from 813 to 2417 mm between years. The mean annual temperature is 22°C (range: 16-30°C). The driest months are January and February with mean temperature of 23°C and rainfall of 68 mm (County Government of Siaya, 2018).
The Anyiko wetland size has been reduced by 55% (0.9 km 2 ) in the last five decades   . The main reasons for this trend are wetland drainage and conversion to farmland (−0.7 km 2 ) but also shrubland encroachment . The local community uses the converted wetland areas primarily for small-scale paddy rice production and the cultivation of cocoyam, maize, and vegetables. Paddy rice is predominantly grown at downstream areas where more water is available for irrigation; cocoyam (Colocasia esculenta), maize (Zea mays), and vegetables, e.g., collard greens (Brassica oleracea) and African nightshade (Solanum scabrum), are mostly grown at the upstream areas. Maize is flood-intolerant but is grown during the rainy seasons in relatively well-drained soils. Paddy rice is grown once a year because of inadequate water supply during the 'short' rains. Cocoyam is grown throughout the year and vegetables during the dry season. Therefore, in the downstream areas of the wetland, the converted areas are uncultivated during the dry season. At the upstream areas, the cultivation is more diverse depending on the crop.

Study design and sampling sites characteristics
We measured soil CH 4 , CO 2 , and N 2 O emissions up to four times a month from 27 December 2017 to 27 May 2018 (total number of sampling dates, n = 14). We selected four sampling sites in the wetland (two upstream and two downstream situated sites) and one site in the area converted to cocoyam farming (Fig. 1). The four sampling sites in the wetland included: Upper site 1 (US1), Upper site 2 (US2), Lower site 1 (LS1), and Lower site 2 (LS2) (Fig. 1). During the rainy season, only two wetland sitesupper site 2 (US2) and lower site 1 (LS1), were sampled.

Soil GHG sampling
We measured soil CO 2 , CH 4 , and N 2 O emissions with the closed static chamber technique (Pumpanen et al., 2004), according to the minimum requirements for GHG (Butterbach-Bahl et al., 2016;Parkin and Venterea, 2010). The chamber technique is commonly used to measure soil GHG fluxes because it is relatively inexpensive, simple to operate, and flexible (Butterbach-Bahl et al., 2011). However, the technique underestimates GHG fluxes due to an altered diffusion gradient slowing the gas diffusion from the soil after putting the chamber on the soil surface (Pumpanen et al., 2004).
We fabricated the chambers from ten-liter opaque plastic buckets comprising of two parts: a base (25 cm diameter and 15 cm height) and a lid (25 cm diameter and 25 cm height). The lid was fitted with a gas sampling port, a vent tube (50 cm long and 2.5 mm diameter), and a thermometer (Pelster et al., 2017;Tully et al., 2017). In order to minimize temperature increases inside the chambers (Butterbach-Bahl et al., 2011), we insulated them with reflective tape.
We installed the chamber bases at the sampling sites two weeks before the first GHG measurement. In the unconverted areas, we had to clear above-ground papyrus biomass to provide room for installing the chamber bases. Each sampling site consisted of three randomly selected clusters as replicates for the GHG measurements. We inserted three chamber bases 7-10 cm into the soil at each cluster, totaling nine chamber bases per site. During the rainy season, the chamber bases at US2 were removed and adjusted to 35-40 cm height due to increased water level and re-inserted into the soil. The chamber bases were also equipped with 2-4 holes to ensure water flow between the chamber and its surrounding (Butterbach-Bahl et al., 2016). The overlying water in the chambers enhances anaerobic conditions in the soil, allowing the growth of methanogens that use carbon compounds as electron donors and producing CH 4 in anaerobic respiration (Butterbach-Bahl et al., 2016). Hence, CH 4 emissions may be higher in the chamber than in the surrounding area if there are no holes in the chamber bases. The holes above the water layer were closed before sampling when the water level dropped in the chambers. Closing of the holes eliminates air diffusion from inside to outside of the chambers, thereby averting underestimation of GHG emissions. Towards the end of the dry season, farmers dug a canal along LS1 to prevent flooding into the adjacent rice paddies during the rainy season. Water flowing along LS1 was, therefore, mostly restricted to the canal. The chamber bases remained in place throughout the sampling period to minimize soil disturbance. Occasionally, some got damaged or went missing due to livestock trampling, wetland burning, or vandalization. However, these were quickly replaced (at least one week before the next gas sampling). Chamber bases were also removed during cocoyam farm weeding and re-inserted after soil management. During the rainy season, boardwalks built at each sampling site minimized CH 4 emissions due to sediment disturbance/vibrations (Butterbach-Bahl et al., 2016). We conducted gas sampling between 10:00 am and 1:00 pm and between 4:00 pm and 6:00 pm, thereby covering the diurnal cycle's average flux (Minamikawa et al., 2012;Parkin and Venterea, 2010).
At each sampling event, we cut growing shoots of papyrus in the chambers to ground level before sampling. We ensured the chamber's airtightness by using rubber seals to clamp the bases and lids together (Rochette, 2011). We closed the chambers for 30 min and collected air samples (60 ml) inside the chambers from the headspace every 10 min (0, 10, 20, and 30 min) using polypropylene syringes (60 ml capacity) fitted with Luer locks. Using the gas-pooling technique, we ensured that the soil's spatial heterogeneity in the two land-use types is appropriately depicted (Arias-Navarro et al., 2013). At each of the above-mentioned intervals, we collected 20 ml of headspace air from each of the three chambers. We used the first 40 ml of each 60 ml composite gas sample to flush the vials. We then filled the remaining 20 ml into 10 ml glass vials. This stepwise approach guarantees that gas contamination with ambient air is minimized (Rochette and Bertrand, 2003). Also, at each time interval before gas collection, we manually mixed air inside the chamber by drawing and pumping the air back and forth severally. The gas samples were transported to the Mazingira Centre, International Livestock Research Institute (ILRI) for analysis.

Soil GHG analysis
The gas samples were analyzed for CO 2 , CH 4 , and N 2 O using gas chromatography (SRI 8610C gas chromatograph). We calculated the GHG concentrations from the sample's peak areas relative to the four regular calibration gases' peak areas. We converted the gas concentrations to mass per volume flux according to Butterbach-Bahl et al. (2011). We computed the chamber fluxes if R 2 ≥ 0.70 (Rochette and Bertrand, 2008). Also, we examined the CO 2 concentration dynamics over the 30 min period to validate the data for each chamber measurement. Chambers with a CO 2 decrease >10% between any time intervals were assumed to leak. In such a case, we discarded all gas fluxes from further analysis. The only exception was when the decrease occurred in the last measurement, as we could still calculate the flux rate with the first three measurement points (Pelster et al., 2017). However, if there was no leakage, negative CH 4 and N 2 O values were recognized as the soil's uptake of the respective GHG. The minimum detection limits for each gas were ± 0.02 mg CH 4 -C m −2 h −1 , ±3.1 mg CO 2 -C m −2 h −1 , and ±3.5 μg N 2 O-N m −2 h −1 . However, we included all the data, even those below the minimum detection limit in the analysis as recommended by Croghan and Egeghy (2003) and Parkin et al. (2012).

Measurements of soil parameters
At each gas-sampling event, we measured soil temperature at a depth of 0-20 cm next to the chamber bases with a digital thermometer (Brannan 31/162/0). Also, we collected one soil sample (depth: 0-15 cm) from each of the three randomly selected clusters per sampling site. We then determined soil moisture by oven-drying a weighed amount of the samples for 48 h at 105°C and re-weighing them after cooling. We calculated the soil moisture on a dry weight basis (Okalebo et al., 2002). The remaining soil samples were analyzed for organic carbon, total phosphorus, total nitrogen, ammonium-nitrogen, and nitrate-nitrogen. We determined total nitrogen by the Kjeldahl method, total phosphorus by the ascorbic acid method, organic carbon by the Walkley-Black method, and ammonium-nitrogen and nitrate-nitrogen by the colorimetric method (Okalebo et al., 2002). See Table 1 for an overview of all parameters.

Data analyses
The general linear model (GLM) presents itself as a suitable approach to answer our research questions. However, the dataset violated the minimum requirements or conditions for the GLM, such as multi- Table 1 Overview of parameters.

Code
Short name Description Unit Effect status dimensional normal distribution and variance homogeneity. Also, the data collection and the targeted arbitrary site selection prohibited strictly inferential statistical data analysis approaches. For such cases, however, the literature offers hypothesis-generating evaluation strategies (e.g., Tukey, 1977). Here, as an alternative to the GLM, we offer a step sequence that has decision tree analysis at its center.
Considering that the metric requirements were not given (Wall and Lienert, 1976;Lienert, 1978), we transformed each of the three continuously-scaled dependent variables (Y i ) into ordinally-scaled variables to allow the prediction of interaction ranges (Table 2). Testing these classified (trichotomized) variables with a log-linear model (LLM) for contingency supported the assumption of total independence of the three GHGs. Moreover, of the eleven independent variables, season (X 1 ′) and land use (X 2 ′) had the status of a controlling (indirect) effect (Table 1). Hence, we merged both controlling effects into one profile variable (Z 1 ). Formally, the three univariate, multi-factorial models could therefore be expressed as: For data analysis, we used the following step sequence.
In the first step, we identified relevant parameters (main and interaction effects) through decision tree analysis (classification and regression tree/CRT; Breiman et al. (1984). The decision tree method corresponds to a graphical, non-linear, scale-independent, stepwise regression/discriminant analysis. We ran a decision tree (SPSS-module: TREE/CRT) for each of the GHGs as the dependent variable (see formula 1). The variable Z 1 {season & land-use} received the highlighted status of a controlling (indirect) effect. Therefore, Z 1 was placed at the first split in all CRT models. The decision which parameters have influence was determined by the Gini-coefficient, a splitting criterion that constitutes an internal measure of homogeneity (IBM Statistics, 2016). Subsequently, we assessed the overall strength of the model fit (model accuracy) through a classification table showing the number of correctly and incorrectly classified cases for each category of the dependent variable (observed versus predicted) (IBM Statistics, 2016).
In the second step, we conducted an exploratory evaluation of the identified effects, particularly the interaction effects, by using a local cross-classification analysis (configuration frequency analysis/CFA; Von Eye, 2002;Von Eye et al., 2010). With this analysis, we examined which of the trees' end nodes contribute to model explanation (node versus categories of the respective GHG), thereby possibly indicating a statistical trend. We determined global significance through chisquare tests. Cramér's V served as an effect size measurement. Local significance was evaluated through Bonferroni-adjusted cell-residual tests to correct for alpha-inflation. Thereby, we could check which single cells affect the target variable. In this regard, it should be noted that, although inferential statistical methods are used in the second step, the reader must not interpret the results strictly. Instead, they serve as an orientation aid for the explorative interpretation (formulation of trend statements).
Although the data quality had not been checked concerning essential statistical quality criteria of a sample (i.e., validity, reliability, objectivity), this working step sequence allowed us to generate hypotheses and describe to what extent there is a possibility that the results are stable.

Results
This study aims to clarify the emission intensity of three prevalent greenhouse gases -CH 4 , CO 2 , N 2 Odepending on two land-use types, two seasons, and nine soil parameters (Table 1). The non-fulfillment of inferential statistical requirements forced us to carry out a stepwise exploratory (hypothesis-generating) analysis approach for each of the three GHGs. The analysis approach comprised of decision tree analyses to determine main and interaction effects, followed by crossclassification analyses for the exploratory evaluation of identified effects. In the next sections, we report the results for each of the three GHGs.

CH 4 emissions
The CRT method's graphical output is a tree that is split according to selected environmental variables. In addition to revealing which features explain the variation of the target variable, the tree also shows interactions between selected parameters. These interaction effects become visible when one follows a given node sequence, starting from node 0 and moving downwards at each split until an end node is reached. The number of nodes in an interaction sequence reveals a one-(one node), two-(two nodes), or more-fold effect.
The base sample's CH 4 emission tree of selected four different variables (season & land-use, nitrate-nitrogen, organic carbon, and soil moisture). The values <0.001 mg m −2 h −1 were negatives, indicating CH 4 uptake. Also, as shown by Fig. 3, five effects were essential in assessing the variation of CH 4 emissions in the study area. All five effects were characterized by two-to three-fold interactions, read as interaction paths via the tree nodes. The first interaction effect pathway (IP1: node 0 ⇒ node 1 ⇒ end node 3) indicated that regardless of the season, but only in the farmland, medium emissions were supported by NO 3 -N ≤ 0.04 g kg −1 . The second pathway (IP2: node 0 ⇒ node 1 ⇒ end node 4) shared the land-use type with the first but then signaled that NO 3 -N > 0.04 g kg −1 supported CH 4 uptake. On the contrary, the third interaction effect pathway (IP3: node 0 ⇒ node 2 ⇒ end node 6) suggested that in the wetland and irrespective of the season, high emissions were supported by organic carbon content >2.7%. The fourth interaction chain (IP4: node 0 ⇒ node 2 ⇒ node 5 ⇒ end node 7) showcased that in the wetland, CH 4 uptake was connected to organic carbon ≤2.7% and soil moisture <90%. As signaled by the fifth pathway (IP5: node 0 ⇒ node 2 ⇒ node 5 ⇒ end node 8), soil moisture >90% supported the medium emissions (Fig. 2).
Overall, the four selected variables could sufficiently illustrate the variations in CH 4 emissions, as three-quarters of all cases were correctly predicted. In detail, the high emissions group performed best. The uptake and medium emissions groups showed a slightly lower probability for correct classification (Table 3).
After identifying the interaction pathways, we assessed which effects are most probable and may be generalized from a hypothesisgenerating perspective. Therefore, we conducted cross-classification analyses with cell-wise post hoc tests. In this regard, the residuals (z) that become apparent in the analysis aid in trend interpretation if they exceed the significance threshold. The higher the residual value, the higher the probability that the results are generalizable.
As shown in Table 4, seven of 15 cells of the CH 4 tree were statistically significant. All interaction pathways except IP1 demonstrated a significant trend in one to three cells. Considering that the residual of Table 2 Class boundaries of the ordinally-scaled target variables for the exploratory analyses. the low emission group of IP5 was only slightly above the significance threshold, it must be treated with care. In contrast, the high residual values of the other pathways suggested a high hit probability, particularly those of IP2 and IP3. The second pathway revealed that at soil NO 3 -N > 0.04 g kg −1 , there was a likelihood of CH 4 uptake in farmland.
On the contrary, as shown by IP3, CH 4 emissions from the wetland were likely if soil organic carbon exceeded 2.7%. However, at lower soil organic carbon content and lower soil moisture ≤90%, there was CH 4 uptake in the wetland (IP4); if soil moisture exceeded 90%, emissions occurred (IP5).

CO 2 emissions
The CO 2 base sample tree selected two variables, season & land-use and organic carbon, and identified four effects that explained emissions rates' variation. Each of the effects constituted a two-fold interaction (Fig. 3). The first interaction effect pathway (IP1: node 0 ⇒ node 1 ⇒ end node 3) signaled that during the dry season but irrespective of land-use type, organic carbon ≤2.6% promoted medium CO 2 emissions. The second pathway (IP2: node 0 ⇒ node 1 ⇒ end node 4) shared the dry season effect with the first but then indicated that higher organic carbon content promoted the highest emissions. The third and fourth pathways were both governed by the rainy season, showing mostly low to medium emission rates, but were then split according to landuse: low emissions cases dominated in the third pathway describing the farmland (IP3: node 0 ⇒ node 2 ⇒ end node 5), whereas the fourth pathway describing the wetland (IP4: node 0 ⇒ node 2 ⇒ end node 6) was characterized by a higher share of medium emission cases and an almost entire lack of high emission cases (Fig. 3).
Overall, the four effects correctly predicted 70% of the cases. Similar to the CH 4 tree, the high emissions group performed best, followed by the medium group. Only the low group showed a lower correct classification probability; only one-third of cases were correctly predicted (Table 5).
Preceding with the cell tests, we found that more than half of the cells exceeded the threshold of significance and that each of the four interaction pathways demonstrated a significant trend in at least one of its cells. In particular, the high residual values in IP2-4 suggested a high hit probability. During the dry season, at soil organic carbon content ≤2.6% and >2.6%, CO 2 emissions were more likely to be moderate and high, respectively. However, during the rainy season, the emissions were more likely to be low in the farmland (IP3) and in wetland (IP4). Regarding the latter, emissions were also less likely to be high (Table 6).

Nitrous oxide (N 2 O) emissions
The N 2 O base sample tree selected six different variables (season & land-use, total phosphorus, carbon: nitrogen ratio, nitrate-nitrogen, total nitrogen, carbon: nitrogen ratio), thereby revealing seven effects that explained the emission rate variation in the study area. These effects were characterized by up to five-fold interactions (Fig. 4). Interestingly, the N 2 O tree was the only tree of the three GHGs that exhibited an uneven split of the profile variable. As shown by the first pathway (IP1: node 0 ⇒ end node 1), the farmland during the dry season featured almost entirely high emission cases. However, this was a solitaire effect as no further interaction effects could be detected.
Overall, the N 2 O tree showed a correct classification rate of more than two-thirds. All three emission groups performed equally well in classifying the effects (Table 7).
As shown in Table 8, only five of 21 cells exhibited a high hit probability. These distinctive features were found in four of seven pathways. The high emissions group was particularly prevalent in the farmland during the dry season (IP1). In the wetland during the dry season and in both land-use types during the rainy season, emissions were likely to be moderate if total phosphorus remained below 40 ppm and carbon: nitrogen ratio below 10 (IP2), albeit this effect was only slightly above the significance threshold. In contrast, when the carbon: nitrogen ratio exceeds 10, the high emissions group becomes prevalent (IP3). Also, at a low carbon: nitrogen ratio, the emissions were distinguished by a predominance of lower cases and a lack of the medium emissions group (IP6).

CH 4 emissions
In the Anyiko wetland, there were CH 4 emissions and uptake, whereas, in the farmland, there was an uptake. Season did not influence the emission rates. The CH 4 emissions from the wetland were driven by the high soil organic carbon content (>3%) and high soil moisture (>90%). According to Reddy and DeLaune (2008), water-saturated conditions in wetland soil minimizes oxygen supply, thereby resulting in anaerobic decomposition of soil organic carbon and, consequently, the production of CH 4 . Also, soil organic carbon supply is the main driver for methanogenesis once anaerobic conditions have established (Whalen, 2005). Methanogenic bacteria use soil organic carbon as a substrate during methanogenesis, and hence its availability enhances CH 4 emissions (Segers, 1998;Torres-Alvarado et al., 2005). The CH 4 uptake in the wetland, in contrast, was attributed to the low soil moisture ≤90%. This is anticipated because soil moisture is low under waterunsaturated conditions; therefore, the aerobic surface layers of wetlands can serve as CH 4 sinks (Horz et al., 2002;Reddy and DeLaune, 2008). The drainage of wetlands increases the oxygen diffusion and reduces soil moisture content, ensuring aerobic conditions supporting methanotrophs, which leads to CH 4 uptake (Reddy and DeLaune, 2008;Smith et al., 2003). This pattern could be observed in the farmland (Fig. 2). Unlike the wetland, the farmland's CH 4 sink function was attributed to high soil NO 3 -N concentrations. This observation's likely cause is that aerobic conditions in the farmland enhance the nitrification process, thereby resulting in increased NO 3 -N concentrations in the soil, which is used to oxidize CH 4 by methanotrophs (van Grinsven et al., 2020).
Considering that the Anyiko wetland is seasonally inundated, it is expected that flooding during the rainy season would enhance CH 4 emissions, resulting in seasonal variations due to anaerobic conditions (Dalal et al., 2008;Liu et al., 2011). However, in the Anyiko wetland, flooding in LS1 was minimal due to water drainage measures. Hence, water mostly remained near the soil surface during the rainy season. Therefore, CH 4 produced in the oxygen-depleted soil zone may have been oxidized at the soil-water interface or the topsoil to CO 2 by methanotrophic bacteria, thereby reducing the CH 4 emissions (Lai, 2009). Consistent with our findings, Were et al. (2021) reported no significant seasonal variations in CH 4 emissions in a wetland in Uganda.  However, Marín-Muñiz et al. (2015) reported higher CH 4 emissions during the rainy season (992.1 ± 313 mg m −2 d −1 ) than during the dry season (160.8 ± 104 mg m −2 d −1 ) in tropical freshwater marshes on the coastal plain of Veracruz in Mexico. During the rainy season, higher emissions were attributed to the higher water depths, increasing reduced soil conditions. Marín-Muñiz et al. (2015) reported the highest CH 4 emissions when the water depth was only 10 cm deep above the soil surface. In contrast, Mitsch et al. (2010) and Nahlik and Mitsch (2011) reported the highest CH 4 emissions when the water depths were 15-30 cm and 30-50 cm above the soil surface, respectively. An increase of the water column beyond these depths creates a CH 4 oxidation zone by methanotrophs that can mitigate CH 4 diffusion to the atmosphere (Nahlik and Mitsch, 2011). In contrast, Brown et al. (2014) reported the highest CH 4 emissions when the water depth dropped to 40 to 55 cm below the soil surface. They attributed the highest emissions to the establishment of required redox potentials and available substrates for methanogenesis. The reduction in water depth can also enhance CH 4 emissions due to an increase in the CH 4 diffusivity in the unsaturated zone (Brown et al., 2014). However, Moore et al. (2011) and Yan et al. (2020) reported positive linear relationships between the water depth and CH 4 emissions. A low water table in these studies was associated with more oxidizing conditions causing the reduction in CH 4 emissions. The use of carbon isotopic signatures (δ 13 C) such as done by Marushchak et al. (2016), Münchberger et al. (2019), and Popp et al. (1999) can aid in the identification of mechanisms and pathways of CH 4 dynamics with changing water depth. Therefore, future studies should consider stable isotope analyses along with the CH 4 flux measurements.

CO 2 emissions
The drainage of wetlands creates aerobic conditions that accelerate soil organic carbon decomposition, consequently increasing CO 2 emissions (Hirano et al., 2012;Mitsch et al., 2013;Syvitski et al., 2009). In our study, CO 2 emissions from the wetland and the farmland were not significantly different. Similar findings were reported by Inubushi et al. (2003). Medium to high CO 2 emissions (>50 mg m −2 d −1 ) during the dry season and low emissions (<50 mg m −2 d −1 ) during the rainy season may be associated with more oxidized conditions during the dry season. According to Smith et al. (2003), oxic conditions enhance aerobic decomposition of soil organic carbon, thus increasing CO 2 emissions. Furthermore, high and moderate CO 2 emissions were attributed to soil organic carbon content. Organic matter available for aerobic decomposition increases with decreasing soil water table during the dry season (Jauhiainen et al., 2008). We did not measure the water table in our study sites. However, from our observation during the rainy season, the water depths were well above the soil surface   (approximately 15 cm) in US2 and near the soil surface in LS1 and some parts of the farmland. Therefore, this water-logged situation in the rainy season could have contributed to the observed low CO 2 emissions from both land-use types. During the dry season, the study sites' water levels were below the soil surface and therefore expected to enhance oxidized conditions, reduce CH 4 emissions, and increase CO 2 emissions. Therefore, an increase in CO 2 emissions during the dry season can partly be ascribed to reduced methanogenesis and the considerable potential for microbial CH 4 oxidation in the oxygenated soil layer (Preuss et al., 2013;Segers, 1998).
Our findings are consistent with Jun-Qin et al. (2009), who reported an increase in CO 2 emissions with declining water saturation in a laboratory incubation experiment of Zoige Alpine wetland soils in China. Were et al. (2021) though reported higher CO 2 emissions than our findings, they also had higher emissions during the dry than rainy seasons. Similarly, Marín-Muñiz et al. (2015) also reported findings with higher mean CO 2 emissions during the dry season (11.7 ± 1.1 g m −2 d −1 ) than during the rainy season (4.2 ± 0.52 g m −2 d −1 ).

N 2 O emissions
High N 2 O emissions from the farmland during the dry season may be attributed to the farmland's enhanced nitrification processes. The drainage of wetlands coupled with low soil moisture during the dry season enhances the soil's oxic conditions, increasing mineralization of organic nitrogen and subsequent nitrification processes (Jauhiainen et al., 2012;Venterink et al., 2002). Under low oxygen conditions, nitrifiers reduce nitrite (NO 2 -N) produced during the nitrification process to N 2 O (Reddy and DeLaune, 2008;Signor and Cerri, 2013). According to van Lent et al. (2015), higher N 2 O emissions from the farmland would occur in the first ten years after conversion and decline if the farmland remains unfertilized. Also, Owino et al. (2020) reported higher N 2 O emissions from fertilized (4.4 ± 3.2 μg m −2 h −1 ) than unfertilized (−3.6 ± 2.6 μg m −2 h −1 ) paddy rice fields converted for the last four decades from the Anyiko wetland in Kenya. The conversion decreases soil organic carbon and nitrogen, which are substrates for the nitrification and denitrification processes (Cameron et al., 2013). In our study, the farmland was established 1-2 years before the GHG sampling. Also, usage of other sources of nitrogen such as livestock manure in the farmland is unlikely as each farmer owns around three cattle (range: 1-8 cattle), and livestock manure is mainly used in the infertile upland farms (Ondiek et al., unpublished data). Therefore, in SSA, higher N 2 O emissions from farmland, particularly during the dry season, are possibly linked to the number of years the converted areas have been cultivated, even though there are non-intensive agricultural practices. However, this phenomenon warrants further studies. These studies should also extend GHG sampling for a minimum of one year to increase the accuracy of cumulative seasonal GHG emissions (see Butterbach-Bahl et al., 2016).
Furthermore, our results showed that, regardless of the season, N 2 O emissions from the wetland and the farmland during the rainy season were quite diverse. If total phosphorous was below 40 ppm and carbon: nitrogen ratio > 10 in sediments, higher emissions were observed, whereas a lower carbon: nitrogen (≤10) ratio supported moderate emissions. Carbon: nitrogen ratios less than 30 support net nitrogen mineralization and associated nitrification, increasing the availability of NH 4 -N and NO 3 -N, which are substrates for coupled nitrificationdenitrification processes . Therefore, in this study, net nitrogen mineralization and subsequent substrates were enhanced when C: N ratios were >10 and reduced when the ratios were ≤10. The low N 2 O emissions were possibly linked to reduced nitrification process since soil NO 3 -N content were low. Under more anaerobic conditions than aerobic, nitrification process is reduced (Jiang et al., 2009;Reddy and DeLaune, 2008). N 2 O emissions through the denitrification process may decline due to low concentrations of NO 3 -N (Page and Dalal, 2011). Also, under strict anaerobic conditions, the major end-product of denitrification is N 2 , thus explaining the very low N 2 O emissions from the wetland in both seasons and from the farmland during the rainy season. Our study did not investigate seasonal changes in the emissions of N 2 O: N 2 ratios and therefore needs further research to quantify this aspect.

Bootstrapping validation
Considering that the dataset did not fulfill an experiment's requirements from a statistical point of view, we had to interpret the results presented herein from a hypothesis-generating perspective. Nevertheless, it would be possible to further evaluate the stability of the detected effects to validate the results, thereby strengthening the notion that results can be generalized. To achieve this, the classical bootstrapping approach may be used to re-sample the base sample and re-run the decision trees. However, to also detect additional effects and longer interaction chains, a modified variant of the bootstrapping method, cumulative/pooled bootstrapping , provides a more viable procedure.
Re-running the tree analyses with an inflated dataset through cumulative bootstrapping shows that the splits related to land-use and season remain stable, confirming the influence of this controlling effect. In the case of all three GHGs, the decision trees of the cumulative Note: Global test results: CHI 2 = 80.49, df = 12, p = 0.000; Cramér's V = 0.592. Shown here are observed counts (n), expected counts (e) for the hypothesis of total independence, and adjusted residuals (z) to determine typical/over frequented (T) and atypical/ under frequented (AT) cells. Bonferroni-adjusted level of significance: z {* = α/frequency of cells = 0.05/(7 × 3) = 0.00238} = 3.04. The bold values are statistically significant adjusted residuals.
bootstrapping dataset split the tree into broader and deeper levels, allowing the identification of more interaction effect pathways than the base sample versions. While the base sample trees of CH 4 , CO 2 , and N 2 O selected four, two, and five parameters, respectively, the trees of the cumulative bootstrap samples chose five, six, and five parameters, respectively. Hence, the bootstrapped trees extended the list of parameters already detected in the base sample trees. Overall, the cumulative bootstrapping simulation exhibits an increased performance capability compared to the standard version. Hence, the modified variant deserves a more detailed demonstration and a more prominent and broader methodological discussion. In our opinion, this can and should be realized in a follow-up paper.

Conclusion
The conversion of the Anyiko wetland areas to farmland resulted in larger CH 4 uptakes irrespective of season, and high N 2 O emissions during the dry season. Although CO 2 emissions were not influenced by land-use, season was a primary influence, as higher emissions were observed in the dry than in the rainy season. The CH 4 uptake and higher CO 2 emissions were related to soil NO 3 -N and organic carbon content, respectively. The high N 2 O outputs in the farmland during the dry season was a solitaire effect as no further interaction effects could be detected. In light of hypothesis-exploration approaches, our findings should be subject to follow-up studies. In this regard, it is recommended to conduct sampling for at least one year. Such an approach would increase the accuracy of cumulative seasonal and annual GHG emissions as well as provide a larger sample size, therefore possibly indicating which soil properties potentially drive the emissions.

Funding
We are thankful to the Austrian Development Corporation (ADC) and the Austrian Partnership Programme in Higher Education and Research for Development (APPEAR), a programme of the Austrian Development Cooperation (ADC) and implemented by Austria's Agency for Education and Internationalisation (OeAD) for funding this study.

Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.