2.1 Study area: Lake Alaotra region

Lake Alaotra is the largest lake of Madagascar, located in the central eastern part (17–18^∘ S and 48–49^∘ E) at about 775 m above sea level (a.s.l.) (Mietton et al., 2018) (Fig. 1a). The surrounding hills rise up to 900–1300 m a.s.l. (Bakoariniaina et al., 2006). Precambrian crystalline rocks, predominantly magmatic rocks of the graphitic system of Madagascar, underlie the area. Lake Alaotra is located in a graben bounded by faults on its eastern and western sides that probably originated in the Tertiary in response to Neogene extensional tectonic events (Mietton et al., 2018). The rainfall (mean annual precipitation between 900 and 1250 mm) is characterized by a distinct wet and dry season with the occurrence of tropical cyclones. The mean annual temperature is 20.6 ^∘C, ranging from 11 ^∘C in July to 28 ^∘C in January (Ferry, 2009).

Figure 1(a) Location of the study region with indication of the sampled hillslope profiles. The dotted white line delineates the watershed of the Lake Alaotra basin. GLA and GLP are the two convex hillslopes under grassland, and F1 and F2 are the two forested convex hillslopes in the Zahamena National Park. Satellite view of the two forest transects (b) and the two grassland transects (c) (© Google Earth 2020).

Nowadays, most of the hills surrounding Lake Alaotra are covered by C₄ grasses. These areas are being used for cattle herding and are characterized by a high density of lavaka (Voarintsoa et al., 2012). Recent research has shown that lavaka in the Lake Alaotra region (Fig. S1 in the Supplement) are on average ca. 400 years old. Lavaka became far more numerous since ca. 1000 years ago, and lavaka formation rates have increased dramatically over the last 200 years. This timing and the rapid increase in lavaka erosion rates has been confirmed by floodplain sedimentation data in the same area (Brosens et al., 2022). Brosens et al. (2022) link this increase in lavaka erosion to increased environmental pressure due to growing human populations and intensified grazing based on scenario modelling and on the absence of significant climatic variations in this period.

The soils in our study area are classified as Ferralsols according to the World Reference Base for Soil Resources (2006). They are underlain by a deeply weathered regolith that is in most locations several tens of metres thick. At many locations, a hard lateritic horizon has developed in the upper part of the regolith, covering it like a “shield” (Voarintsoa et al., 2012; Wells and Andriamihaja, 1993). The lateritic soil horizon is usually between 0.5 to 2 m thick (Voarintsoa et al., 2012) and is considered to be relatively impermeable, thus favouring surface run-off, especially if there is no or little vegetation and if no cracks are present (Wells and Andriamihaja, 1993). In addition, the basement rocks on the site are metamorphic and igneous (Du Puy and Moat, 1996). Therefore, geogenic OC is not considered to be substantial, in contrast to subsoils developed from sedimentary rocks where this might be more important (Graz et al., 2010).

While most of the Lake Alaotra region is covered by C₄ grasses, a small part of the eastern side of the catchment is covered by forest which is part of the Zahamena National Park. Zahamena is the third strict nature reserve of Madagascar that was established in 1927, and since then no human occupancy has been allowed within its boundaries except for scientific research (Andriamampianina, 1984). This national park contains important remnants of the evergreen rainforest of Madagascar at low (0–800 m) and medium altitudes (800–1800 m) (Styger et al., 2007) and extends over 3710 km² (Andriamananjara et al., 2016) (Fig. S2). The rainforest of the Zahamena National Park is one of the most diverse in Madagascar with many species of flora with a high degree of endemism (Raboanarielina, 2012). This pristine forest zone is part of the eastern forest corridor of eastern Madagascar (Fig. 1), where Hackel et al. (2018) have inventoried endemic C₃ grass species in Madagascar. These endemic C₃ grasses belong to the group “forest shade clade” (Paniceae: Boivinellinea), and their diversification since the Miocene is reported to be favoured by the expansion of the Sambirano rainforest (in the north of Madagascar) (Hackel et al., 2018; Yoder and Nowak, 2006).

2.2 Sampling transects

The hillslopes surrounding Lake Alaotra are nearly all convex-sloping hills (Fig. S3), a characteristic that is commonly observed throughout the central highlands of Madagascar (Wells and Andriamihaja, 1993). These convex hillslopes typically have a rounded flat top and steepen downslope with the highest slope gradients near the bottom of the hillslope (Figs. 1b, c, S3 and S4).

Vegetation varies with topography: grass vegetative cover is typically highest in the valleys and lowest at the hillslope convexities (Figs. S2 and S4).

Soil samples were collected in the lateritic horizon, along four convex hillslope transects, two of which are located in the grassland habitat, while two hillslopes under pristine forest were sampled just east of the basin boundary in the Zahamena Forest Corridor (Fig. 1). The two chosen grassland hillslopes are denoted as GLP (grassland hillslope where Lavaka is Present) and GLA (grassland hillslope where Lavaka is absent). For GLP, the chosen transect was situated outside the actual lavaka. By choosing a slope with and without lavaka, we wanted to investigate whether soils on slopes that have lavaka development may differ from slopes that do not have them. The two pristine forest transects are denoted as F1 and F2. The slope gradients (derived from the 12 m resolution TanDEM-X digital elevation model) of the four transects were comparable, with the maximum slope gradient of 25 and 30^∘ for the forest transect and 25 and 29^∘ under grassland (see Fig. S3).

Along each of the hillslope transects, six soil profiles were sampled down to 2 m deep. We named the sampling positions top (T), upper middle (UM), middle (M), lower middle (LM), bottom (B) and valley (V) (Figs. 1 and S3). Soil samples were collected using a motorized auger with 10 cm diameter for the upper 0–100 cm and 7.5 cm for the soil layer between 100 and 200 cm. The sample resolution was 5 cm for the first metre and 10 cm for the second metre. Compaction of the soil during sampling was negligible given the wide auger diameters. Soil samples were taken using Kopecky rings (diameter = 4 cm and height = 4 cm) in the open side of the auger. In total, 720 soil samples were collected, including 180 per transect and 30 per sampling location. We additionally collected leaves of the most abundant vegetation types at both the grassland and forest sampling sites.

2.3 Laboratory and data analyses

2.4 Dry bulk density and estimation of carbon stock

The dry bulk density (ρ_d) was calculated by using the dry weight of the soil samples and the known volume of the Kopecky rings. Absolute cumulative SOC stocks were calculated by integrating the carbon amount of each depth interval using the dry bulk density and the organic carbon concentration:

where q(i) is the carbon stock in soil depth interval i (units converted into Mg C ha⁻¹), %OC(i) is organic carbon content in soil depth interval i, D(i) is the thickness of horizon i (cm), and ρ_d(i) is dry bulk density of the soil depth interval i (g cm⁻³). The estimated absolute value of the cumulative SOC stock (Q(z)) is the amount of SOC stored between the soil surface and soil depth z (Cerri et al., 2007).

The relative value of the cumulative SOC stocks at depth z (R(z), %) was calculated as the ratio of the absolute value of the cumulative SOC stock at this depth (Q(z)) to the absolute value of the SOC stock for the whole soil profile (i.e. between 0 and 2 m depth, Q):

2.5 Statistical analyses

Overall comparisons between two groups were done using Student's t test if the data were normally distributed and passed an equal variance test. For non-normally distributed data, non-parametric tests (Mann–Whitney U test) were used. Differences were considered statistically significant at P<0.05. To compare more than two groups, standard one-way analysis of variance was performed, with parametric pairwise test (Tukey's honestly significant difference test) comparisons for normally distributed data. Non-parametric tests using one-way analysis of variance on ranks (Kruskal–Wallis), followed by non-parametric pairwise comparison (Dunn's test), were used for non-normally distributed data. The relationship between δ¹³C and pMC of SOC was tested using ordinary Pearson correlation and least squares linear regressions. Linear regressions were used despite not having an explicit independent variable as we sought only the presence or absence of a relationship between pMC and δ¹³C of SOC.

3.1 Soil texture for grassland and forest soils

For the transects under forest, the proportion of clays ranged between 46.3 % and 50.2 %, silt ranged between 8.5 % and 15.2 %, and the sand fraction constituted between 36.1 % and 43.6 % of the soil mass. For the transects under grassland, clays were the dominant fraction (39.5 %–52.4 %), followed by sand (28.2 %–46.7 %) and silt (3.0 %–26.9 %). There were no significant differences in texture of soils under grassland and forest (p-value = 0.7 (sand), p-value = 0.7 (silt) and p-value = 0.2 (clay)). Soil texture under both grassland and forest was thus classified as sandy clays to clays (Fig. S5).

3.2 Organic carbon content and stable carbon isotope of forest soil and vegetation

The OC content of the forest profiles ranged between 1.5 % and 4.6 % for F1 and between 2.9 % and 5.6 % for F2 for the topsoil samples (0–5 cm; Fig. 2a and b). Overall, the percent of OC (%OC) trends of the profiles were similar for the different sampling locations, where the %OC was highest in the topsoil, decreasing exponentially with depth over the first ∼ 60 cm. At 190 cm depth, %OC decreased to 0.1 %–0.3 % for both profiles. For forest transect F1, %OC content in the upper ca. 60 cm was higher at the UM and LM positions compared to the other hillslope positions. These differences were particularly apparent in the upper ca. ∼ 20 cm of the profile. This difference between the hillslope positions was less marked in F2. However, the %OC was highest for the F2-B, F2-UM and F2-LM when compared to the F2-V, F2-T and F2-M at 0–20 cm depth.

Figure 2Depth profiles of %OC for each sampling location (top, upper middle, middle, lower middle, bottom, valley) of the four transects including F1 (a), F2 (b), GLA (c) and GLP (d).

Download

At the surface, the δ¹³C values only showed minor variations, ranging between −27.1 ‰ and −25.5 ‰ (0–5 cm; Fig. 3a and b). δ¹³C values increased with depth and reached a value of $-\mathrm{24.1}\pm \mathrm{0.6}$ ‰ at a depth of 60 cm. Below this depth, δ¹³C values no longer showed a systematic variation with depth but varied within a narrow range. No distinct trends were observed for the different sampling positions, in which only the δ¹³C of F1-T was slightly higher throughout depth when compared to the other sampling positions. δ¹³C values of profiles under forest (F1 and F2) did not differ significantly in the upper 50 cm depth of the profiles (Table S1 in the Supplement). The vegetation samples from the forest showed δ¹³C values that are clearly lower than those observed in the forest soils, with values between −35.8 ‰ and −30.5 ‰ with an average δ¹³C value of $-\mathrm{33.5}\pm \mathrm{1.8}$ ‰ (n=10).

Figure 3Depth profiles of δ¹³C for each sampling location (top, upper middle, middle, lower middle, bottom, valley) of the four transects including F1 (a), F2 (b), GLA (c) and GLP (d).

Download

3.3 Organic carbon content and carbon isotope ratios of grassland soil and vegetation

The SOC content of the grassland soil in the upper layers is much lower than that in the forest soil and ranges between 0.4 % and 1.5 % for GLA and between 0.7 % and 1.7 % for GLP (0–5 cm; Fig. 2c and d). Most grassland profiles show a similar relation between SOC content and depth: the SOC content decreases from values between ca. 0.7 % and 1.3 % at the top to values of ca. 0.2 % below 1 m depth. However, there are exceptions: in the first few decimetres, the GLA-M and GLA-LM profiles have lower SOC than the other profiles. The valley profiles are also different: both have much higher SOC values than the other hillslope profiles in the top 30 cm. At the GLP profile, the SOC content of the valley profile is higher than that of other profiles down to a depth of ca. 100 cm. %OC values do not differ significantly in the upper 0–50 cm depth for GLA and GLP for all profile positions (Table S1).

On both grassland hillslopes, δ¹³C values are clearly higher for the valley profiles in comparison to the hillslope profiles, with the absolute differences being largest down to a depth of 40 cm for the GLA site and down to ca. 60 cm at the GLP site (Fig. 3c and d). The variation in δ¹³C values with depth is similar on most hillslope positions under grassland: on all sites, δ¹³C values vary within a relatively narrow range (from −24 ‰ to −20 ‰) below a depth of 40 cm. Above 40 cm, some profiles (GLP-B, GLP-UM, GLA-UM and GLA-T) show a gradual decline of δ¹³C values with depth, with maximum δ¹³C values of ca. −16 ‰ at the top. In most profiles, however, the decline of δ¹³C with depth is much smaller or even non-existent. δ¹³C values of profiles under grassland (GLA and GLP) did not differ significantly in the upper 0–50 cm of the profiles (Table S1). Soil δ¹³C values are consistently lower than those observed in the vegetation: the δ¹³C of the grassland vegetation ranges between −13.4 ‰ and −11.4 ‰ with an average value of $-\mathrm{12.8}\pm \mathrm{0.6}$ ‰ (n=7).

3.4 Soil organic carbon stocks for grassland and forest soil profiles

For all hillslope positions, the absolute values of the cumulative SOC stocks were consistently higher in forest profiles, irrespective of the lower depth considered to integrate the SOC stock (Fig. 4). However, in the valley profile, the absolute cumulative SOC stock under grassland is similar to that under forest for every soil layer. Both for the forested (F1 and F2) and grassland transects (GLA and GLP), no major differences in SOC stocks were observed between both profiles (Table S1). SOC stocks at the valley position were significantly different from all other hillslope positions under grassland (Table S2).

Figure 4Absolute values of cumulative SOC stocks of hillslopes F1, F2, GLA and GLP plotted together for each sampling location: (a) Top, (b) Upper Middle, (c) Middle, (d) Lower Middle, (e) Bottom and (f) Valley.

Download

3.5 ¹⁴C analyses of bulk SOC

We measured the ¹⁴C activity of bulk SOC at different depths in both grassland and forest profiles (5–10, 20–25, 50–60 and 100–110 cm). The ¹⁴C activity of the bulk SOC is plotted against the δ¹³C of SOC in Fig. 5. Along the forest transect, all surface samples at the top and valley positions showed a pMC value greater than 100 %, indicating a significant contribution from bomb-derived carbon (post-1950). The pMC values of the forest soil samples (F2-T and F2-V) decrease to 42 %–52 % at a depth of 50 cm. This decrease in pMC with depth is associated with a modest increase in δ¹³C.

Figure 5Comparison of the average value of ¹⁴C pMC of bulk SOC plotted against δ¹³C of SOC of grassland and forested soil. Each point is labelled by the transect position and the top depth interval. Continuous and dotted black lines represent the regression lines for grassland soil data and forest soil data, respectively.

Download

For the grassland transect, samples from the soil surface layer (5–10 cm) on the hillslopes show clearly lower pMC values in comparison to the forest sites, ranging between 78 % and 85 %. At the valley position, topsoil samples from both grassland transects have pMC values greater than 100 % (106 % and 103 % respectively), slightly higher than the pMC values observed for topsoils in the valley positions under forests. The pMC values of bulk SOC decrease with increasing depth down to 60 % at 100 cm depth (GLP-T; Fig. S6). The pMC values under grassland at depth are generally higher than the pMC values observed under forest at similar positions. Under grasslands, the decrease in pMC with depth is associated with a decrease in δ¹³C (Fig. 5).

4.1 Difference in carbon sources between grassland and forest soils

δ¹³C depth profiles of forest soils (Fig. 3a and b) are similar to those observed in the primary vegetated forest of Didy, located ca. 60 km south of our forest sampling locations (Winowiecki et al., 2017). They reported a median value of −27 ‰ in the forest topsoil and also found that δ¹³C increased with depth. One partial explanation for the ¹³C enrichment with depth under forest is the Suess effect which leads to a gradual decrease in the δ¹³C values of atmospheric CO₂ over time (Keeling et al., 2017). In our forest profiles, an increase in δ¹³C values of 2 ‰–3 ‰ over the upper 60 cm is observed (Fig. 3a and b). Given the magnitude of the Suess effect (2 ‰ over the past 250 years; Keeling et al., 2017), the variations in δ¹³C that we observe in the forest soil profiles may, to a large extent, be ascribed to the Suess effect. An additional explanation for ¹³C enrichment with depth commonly observed in soil profiles is that the soil microbial biomass is generally more ¹³C-enriched than its substrate, and soil organic matter therefore typically becomes more enriched in ¹³C due to the incorporation of microbial biomass or microbially derived C (Ehleringer et al., 2000). It has been shown that SOC in deeper soil layers is in a more advanced stage of decomposition and is increasingly composed of stable microbially derived SOC (Domeignoz-Horta et al., 2021; Kallenbach et al., 2016; Schlesinger, 1977).

The δ¹³C values of forest vegetation ($-\mathrm{32.5}\pm \mathrm{1.71}$ ‰) are consistent with a C₃ vegetation signature. However, they are ca. 3 ‰ lower than the value estimated from the empirical relationship of Kohn (2010) that predicts the δ¹³C values of C₃ vegetation based on annual precipitation, latitude and altitude (−29 ‰). As the vegetation samples were collected by hand and thus taken not far above the soil surface, the lower measured values could be due to the “understorey effect”. The understorey effect or canopy effect found in tropical forest is mainly related to the gradient in δ¹³C of ambient CO₂ along the vertical gradient: δ¹³C-CO₂ is low close to the ground due to elevated CO₂ concentrations via the contribution of soil respiration. Higher up in the canopy, the δ¹³C-CO₂ values are closer to the average atmospheric CO₂ composition. Kohn (2010) recommends −31.5 ‰ as a cut-off value for identifying the understorey effect in a closed-canopy forest such as the Zahamena National Park.

In contrast to the forest profiles, SOC profiles on the grassland transects show a decline of δ¹³C with depth (Fig. 3c and d). Some of our profiles, such as GLP-UM and GLP-B, show a very clear decrease in δ¹³C in the upper 40 cm of the soil profile (from ca. −16 ‰ to ca. −22 ‰; Fig. 3d). In most profiles, however, δ¹³C values of −20 ‰ down to −23 ‰ are observed in the top layer, and values decrease very gradually over the whole depth of the profile with minimum values of −23 ‰ to −24 ‰ at 2 m depth.

The δ¹³C soil value of the SOC depends on the δ¹³C value of the aboveground vegetation and the corresponding litter that gets incorporated into the soil profile over the period which is represented by the remaining SOC (Burney, 2004; Bird et al., 2003). The δ¹³C average value of the vegetation in the grassland transects ($-\mathrm{12.8}\pm \mathrm{0.6}$ ‰) is a typical value for grassland plant species (C₄ plants) and is consistent with the value found in a C₄-species-dominated basin in Madagascar ($-\mathrm{12.2}\pm \mathrm{0.7}$ ‰; Marwick et al., 2014). If C₄ vegetation would have been dominant for a long time on the hillslopes, we would therefore expect topsoil δ¹³C values typically between −15 ‰ to −13 ‰, with little variation with depth as observed in the few studies documenting soil δ¹³C profiles under long-term C₄ vegetation (Conrad et al., 2017). In contrast, while a few of our grassland soil δ¹³C profiles have topsoil δ¹³C values that are consistent with the current C₄ vegetation (GLP-UM, GLP-B and GLA-T), most surface soil samples show a δ¹³C value that is lower than would be expected of soils under C₄ vegetation. This suggests that, even in the topsoil, an important fraction of the SOC is derived from C₃ vegetation. Using the average δ¹³C of the forest soils as the C₃ endmember (δ¹³C = $-\mathrm{25.3}\pm \mathrm{0.9}$ ‰) and using the values found by Conrad et al. (2017) as the C₄ endmember (values range from −15 ‰ and −13 ‰ with mean value of −13.8 ‰), the contribution of C₃ plant material to the SOC present in the upper 0–55 cm of these grassland soil profiles is estimated at ca. 70 %, with the exception of the valley position (Fig. 6).

Figure 6Variation with depth of the fraction of remaining forest SOC for the grassland profiles. The black line represents the variation in %C₃ by taking −13.8 ‰ as the reference value of δ¹³C SOC under permanent C₄ vegetation. The grey area represents the range of %C₃ by using a reference range of −15 ‰ to −13 ‰ as the δ¹³C SOC under permanent C₄ vegetation.

Download

In contrast to the other grassland hillslope profiles, surface δ¹³C values of the valley positions are closer to the expected δ¹³C value of grassland vegetation, with an average value of $-\mathrm{15.3}\pm \mathrm{0.8}$ ‰. For GLP-V, the δ¹³C values increase between the surface and 50 cm. This suggests a much higher incorporation of OC from C₄ vegetation in the valleys in comparison to the hillslopes (Fig. 6).

The same processes that cause an increase in δ¹³C with depth under C₃ vegetation (Suess effect and microbial processing) would be expected to also lead to an increase in δ¹³C with depth for grassland profiles if the latter would have fully developed under C₄ vegetation. However, we observed an opposite pattern: δ¹³C decreases with depth under grassland vegetation (Fig. 3c and d). Desjardins et al. (2013) observed a similar depletion in δ¹³C with depth in soils under tropical savannah (C₄) vegetation in central Cameroon. At the surface, the δ¹³C values ranged from −16.4 ‰ to −13.9 ‰ and decreased towards −21.9 ‰ and −22.3 ‰ at between 2 and 4 m depth. They interpreted this decrease in δ¹³C with depth as the result of a C₃ to C₄ vegetation shift. While the relatively young SOC at the top of the profiles was primarily derived from C₄ plants (resulting in a high δ¹³C), the older SOC in the lower part of the profile was primarily a relic from the time when the landscape was covered by forest (with a low δ¹³C). Sanaiotti et al. (2002) reported similar results from a savannah landscape in the Amazon region where the δ¹³C signatures of soil organic matter under savannah areas converged to the signature of a nearby forest at greater depths. Therefore, it was assumed that the savannah areas had replaced forested areas during the Holocene period (Sanaiotti et al., 2002). Importantly, a decline of δ¹³C with depth towards the values reported in these studies and similar to those we observed is not always observed under C₄ vegetation. For instance, Martin et al. (1990) describe profiles under C₄ savannah where δ¹³C values are always above −17 ‰ down to depths of 120 cm and showed little systematic variation with depth, except for somewhat lower values for the deepest horizon that was sampled. This is to be expected in systems which have been under C₄ vegetation for very long time spans: indeed, both the Suess effect and fractionation during microbial decomposition of SOC are expected to lead to an increase in δ¹³C under C₄ vegetation as well.

Similar to Desjardins et al. (2013) and Sanaiotti et al. (2002), we observed a clear convergence of the grassland soil profiles towards the forest signature of −24.0 ‰ to −23.0 ‰ with depth, except for the valley sites (Fig. S7). At the surface, the average difference is ∼ 6.6 ‰, declining to ∼ 1.5 ‰ at 1 m depth. The average difference never becomes zero, except for the UM slope position, suggesting that there may be some contribution from C₄ plants to the SOC pool at all depths.

This pattern, where the δ¹³C value of grassland converges towards the δ¹³C value of forested soil with depth, is best explained by assuming that the vegetation of the grassland hillslopes we sampled has changed rather dramatically from a vegetation that was most likely dominated by C₃ plants (thus most likely a forest or wooded savannah) to an open vegetation completely dominated by C₄ vegetation. The higher δ¹³C values we observed in the upper part of some of the profiles (GLA-T, GLP-UM and GLP-B) suggest that deforestation may have occurred in patches and that the locations of these profiles were deforested first, resulting in topsoil SOC pools with a higher δ¹³C value in comparison to more recently deforested locations. The fact that the δ¹³C profiles measured at GLA-T, GLP-UM and GLP-B also converge to the forest profiles at depth suggests that also these locations originally had a tree-rich vegetation cover but that this tree cover was removed earlier than in the other locations.

The results for the valley locations are different: here the SOC has a δ¹³C value reflecting a C₄ vegetation down to 50 cm, and also below this depth, the forest and grassland profiles remain significantly different, with higher δ¹³C values under grassland (p-value < 0.001). The difference at depth compared to other locations suggests that the vegetation in the valleys has had an important fraction of C₄ plants for a very long time allowing the slow-cycling SOC pools stored at greater depth to acquire a δ¹³C signature significantly affected by C₄ vegetation (ca. 43 %). In addition, we observed more dense grass vegetation in the valley positions compared to other profile positions (Figs. S2 and S4). The higher δ¹³C values observed in the upper part of the valley profiles (0–40 cm for GLP-V) suggest that also in these landscape positions the relative importance of C₄ vegetation has recently increased, but lateral transfers may also play a role via erosion of recent C₄ inputs from the upper slopes and subsequent deposition in the valley.

Within our profiles, we only observed pMC values > 100 % in some upper layers (<10 cm) of forest sites and at the valley position of the grassland soil transects (Fig. 5). In contrast, the SOC at 5–10 cm depth in the grassland hillslope sites always showed pMC values lower than 85 %. At the same time, the δ¹³C data indicate that a significant fraction (ca. ∼ 65 %) of the SOC at these sites is still C₃-derived in most grassland profiles (Fig. 6) despite the current dominance of C₄ grasses. Thus, the ¹⁴C data corroborate the δ¹³C results and suggest that a significant fraction of the grassland soil SOC is relatively old and was formed when these locations were still under forest. This, in combination with a low degree of SOC replacement after the vegetation transition, results in the presence of a mixed δ¹³C signal in combination with a relatively low pMC. The low replenishment/replacement of SOC from the grassland vegetation may be partly explained by (over-)grazing and fire activity, both of which should reduce the organic inputs from C₄ vegetation after deforestation (Abril et al., 2005; Bond and Keeley, 2005). A second mechanism possibly contributing to the lower δ¹³C and pMC values of surface soils along the grassland transect is that the soil might have been subjected to more intense surface erosion after the vegetation shift, thereby constantly removing young SOC (with a higher δ¹³C value) towards the valley position. This is confirmed by soil erosion rates derived from in situ ¹⁰Be analysis of the topsoil samples (5–15 cm) which indicates that on a convex hillslope, both under grassland and forest, erosion rates increase from the top towards the valley position, where the erosion rates are consistently higher under grassland when compared to forest (Liesa Brosens et al., unpublished data).

For the grassland transects, significant differences were observed in %OC, δ¹³C and SOC stocks between the valley and other hillslope positions (T, UM, M, LM and B; Table S2). SOC stocks and surface δ¹³C values on the upper slope were lower than at the valley positions. δ¹³C values on the upper slope positions (T, UM, M, LM and B) were slightly higher than those of forest soils at corresponding depths (Fig. S7). Erosion rates in the upslope positions are expected to have increased after deforestation. Restrepo et al. (2015) have reported a 33 % increase in erosion rates 30 years after deforestation in the Colombian Andes. This erosion leads to losses of SOC (Don et al., 2011; Rabetokotany-Rarivoson et al., 2015). Even though surface erosion itself is not associated with isotope fractionation, it indirectly influences SOC δ¹³C values by removing the surface layer and exposing the deeper soil layer, where SOC is a mixture of SOC from the subsoil and fresh OC from the vegetation. Thus, erosion can lead to different δ¹³C values (Häring et al., 2013), and this process could contribute to the higher SOC stocks and δ¹³C values observed in the valley positions for the grassland transect. However, under forest (C₃ vegetation), no substantial differences in the δ¹³C are observed along the transects (Table S2). This suggests that erosion might not play a major role in the variation in δ¹³C in a forest-dominated landscape as significant erosion would have led to SOC accumulation in the valley position.

The high pMC of SOC in the grassland valley position (GLP-V) is consistent with the idea that this site acts as a depositional area for recently eroded C. It can be noticed that also in the valley sites, pMC decreases with depth (Fig. S6), which is consistent with a significant local production of SOC. On the other hand, pMC values in the valley sites are clearly higher than those observed on the forest and grassland hillslopes, at least for depths <55 cm: this suggests a significant addition of young eroded SOC. Surface erosion is expected to be variable across topographic positions along the hillslope transect, with minimal impact at the top and maximum impact on the steeper sections towards the bottom of the hillslope transect. However, no consistent variation in topsoil pMC values was observed along the grassland transect (Fig. S6). Within the grassland profiles, there is a clear positive relationship between δ¹³C values and ¹⁴C activity (R² = 0.7, P-value = 0.006; Fig. 5), whereby older SOC corresponds to a higher contribution of C₃-derived (forest) C. In contrast, for sites under current forest vegetation, the upper soil layer contains mostly young OC (pMC > 100 %), and these profiles are dominated by C₃-derived C throughout the range of ¹⁴C activities they represent. In addition, pMC values of grassland soil profiles converge to the pMC values observed in the forest soil profiles at greater depths (Fig. 5).

4.2 Response of SOC stocks to vegetation transition

The average SOC stock we measured in the upper layer (0–30 cm) of our grassland transect of 33 ± 9 Mg C ha⁻¹ and 30±11 Mg C ha⁻¹ (Table S3) is clearly lower than the average stock found in Ferralsols under grassland in Madagascar, with an average of 77±33 (n=8) Mg C ha⁻¹, and is in the same range as Luvisols and Arenosols under grassland with an average of 49±33 Mg C ha⁻¹ (n=4) (Chevallier et al., 2020).

The SOC stocks in our forest transect of 70–74 Mg C ha⁻¹ (0–30 cm) and 137–139 Mg C ha⁻¹ (0–100 cm) are similar to those of Grinand et al. (2017), who conducted a field study in the southeast humid forest of Madagascar and measured SOC stocks of 70–131 Mg C ha⁻¹ in the upper 30 cm and 139–296 Mg C ha⁻¹ in the upper 100 cm. In addition to this, they found that the SOC stock in the upper 30 cm accounts for half (49 %) of the SOC stored in the upper 100 cm. Rabetokotany-Rarivoson et al. (2015) observed a wide range of SOC stocks between 39 and 233 Mg C ha⁻¹ for the upper 30 cm in the humid eastern forest. The average values they reported for both soil layers were 110 Mg C ha⁻¹ (0–30 cm) and 226 Mg C ha⁻¹ (0–100 cm), showing that 49 % of the SOC of the 0–100 cm soil layer is stored in the upper 30 cm. Andriamananjara et al. (2017) measured SOC stocks sampled in the eastern humid forest of Madagascar and reported an average cumulative SOC stock of 90 Mg C ha⁻¹ (0–30 cm) and 137 Mg C ha⁻¹ (0–100 cm), which are also similar to the reported values in our study.

By comparing the absolute value of cumulative SOC stocks under forest and grassland, it can be seen that total stocks are clearly higher under forest, except in the valley (Fig. 4).

The first-order control on the SOC stock of the soil is the balance between inputs and outputs of carbon in soils (Davidson and Janssens, 2006). These inputs include vegetation such as litter and roots from biomass, while the outputs include CO₂ emissions, erosion and leaching of dissolved organic matter into the groundwater. Deforestation has a significant impact on SOC stocks, primarily because carbon inputs are reduced. Grinand et al. (2017) estimated an average change in SOC stocks of −10.7 % and −5.2 % in the 0–30 and 0–100 cm soil layers when forest is converted to grassland in the humid ecoregion of Madagascar over a 20-year time span. In addition, a study of tropical soils by Don et al. (2011) reports an estimated −12 % change in SOC stock when the primary forest is converted to grassland over a 25-year time span. However, when comparing the grassland SOC stocks with the forest SOC stocks for our profiles, an average loss of −55.7 % is observed in the upper 30 cm soil layer, which is much higher than in the aforementioned studies. Several factors may explain this difference. The studies of Don et al. (2011) and Grinand et al. (2017) report changes over a period of 20–25 years after deforestation. The fact that we find a much larger reduction in SOC stock may be due to the fact that deforestation has occurred much longer ago at the sites we sampled. It is often assumed that the major change in SOC occurs during the first 5 years after deforestation, and a new SOC equilibrium is reached between 20 and 40 years (Cerri et al., 2007). Our data suggest that this may not be the case in the environment we studied: this may be due to the large difference in SOC production between grasslands and forests. Indeed, our data suggest that even today, a large fraction of the SOC that is present under grassland was produced when the grassland sites were under C₃ vegetation. Similar observations were reported by several authors in landscapes where forest vegetation had been transformed into grassland more than a 1000 years ago, showing evidence that C₃-derived carbon can remain present for millennia after the replacement of the original vegetation (Desjardins et al., 2013; Guillet et al., 2001; Schwartz et al., 1986). Rabetokotany-Rarivoson et al. (2015) and Razafindrakoto et al. (2018) found a strong difference in SOC stocks between the initial forest vegetation and the final stage of deforestation (i.e. a landscape dominated by grasses). This suggests that due to the low SOC productivity of savannah systems, the time needed to reach a new equilibrium SOC stock is indeed in the order of centuries to millennia rather than decades (Wutzler and Reichstein, 2007).

As opposed to what we found on the other hillslope positions, the SOC stocks under forest and grassland are relatively similar in the valley positions (Fig. 4). There are two mutually, non-exclusive mechanisms that can explain this observation: first, the valleys may have been permanently covered by C₄ vegetation that is much more productive than the vegetation on the hillslopes. This is indeed highly likely: the valleys are clearly much wetter than the hillslopes and are locations where nutrients eroded from the hillslopes may accumulate. Second, part of the SOC found at the valley locations may not have been produced locally but may result from lateral transport of SOC through erosion from the hillslopes.