Volume 104, Issue 3 p. 703-715
Standard Paper
Free Access

Bioclimatic envelope models predict a decrease in tropical forest carbon stocks with climate change in Madagascar

Ghislain Vieilledent

Corresponding Author

Ghislain Vieilledent

Cirad – UPR BSEF, F-34398 Montpellier, France

Correspondence author: E-mail: [email protected]Search for more papers by this author
Oliver Gardi

Oliver Gardi

Helvetas Swiss Intercooperation – BP 3044, 101 Antananarivo, Madagascar

Bern University of Applied Sciences – HAFL, CH-3052 Zollikofen, Switzerland

Search for more papers by this author
Clovis Grinand

Clovis Grinand

ETC Terra, F-75020 Paris, France

Search for more papers by this author
Christian Burren

Christian Burren

Wildlife Conservation Society, Soavimbahoaka, 101 Antananarivo, Madagascar

Search for more papers by this author
Mamitiana Andriamanjato

Mamitiana Andriamanjato

Ministère de l'Environnement et des Forêts – Direction Générale des Forêts, 101 Antananarivo, Madagascar

Search for more papers by this author
Christian Camara

Christian Camara

Missouri Botanical Garden, BP 3391, 101 Antananarivo, Madagascar

Search for more papers by this author
Charlie J. Gardner

Charlie J. Gardner

WWF – Madagascar and Western Indian Ocean Programme Office, BP 738, 101 Antananarivo, Madagascar

School of Anthropology and Conservation, University of Kent – Durrell Institute of Conservation and Ecology, Canterbury, Kent, UK

Search for more papers by this author
Leah Glass

Leah Glass

Blue Ventures – Blue Forests program, Ambanja, Madagascar

Search for more papers by this author
Andriambolantsoa Rasolohery

Andriambolantsoa Rasolohery

Conservation International – Africa and Madagascar Field Division, 101 Antananarivo, Madagascar

Search for more papers by this author
Harifidy Rakoto Ratsimba

Harifidy Rakoto Ratsimba

Université d'Antananarivo – Département des Eaux et Forêts, BP 175, 101 Antananarivo, Madagascar

Search for more papers by this author
Valéry Gond

Valéry Gond

Cirad – UPR BSEF, F-34398 Montpellier, France

Search for more papers by this author
Jean-Roger Rakotoarijaona

Jean-Roger Rakotoarijaona

ONE, Antaninarenina, BP 822, 101 Antananarivo, Madagascar

Search for more papers by this author
First published: 27 January 2016
Citations: 57

Summary

  1. Recent studies have underlined the importance of climatic variables in determining tree height and biomass in tropical forests. Nonetheless, the effects of climate on tropical forest carbon stocks remain uncertain. In particular, the application of process-based dynamic global vegetation models has led to contrasting conclusions regarding the potential impact of climate change on tropical forest carbon storage.
  2. Using a correlative approach based on a bioclimatic envelope model and data from 1771 forest plots inventoried during the period 1996–2013 in Madagascar over a large climatic gradient, we show that temperature seasonality, annual precipitation and mean annual temperature are key variables in determining forest above-ground carbon density.
  3. Taking into account the explicative climate variables, we obtained an accurate (R2 = 70% and RMSE = 40 Mg ha−1) forest carbon map for Madagascar at 250 m resolution for the year 2010. This national map was more accurate than previously published global carbon maps (R2 ≤ 26% and RMSE ≥ 63 Mg ha−1).
  4. Combining our model with the climatic projections for Madagascar from 7 IPCC CMIP5 global climate models following the RCP 8.5, we forecast an average forest carbon stock loss of 17% (range: 7–24%) by the year 2080. For comparison, a spatially homogeneous deforestation of 0.5% per year on the same period would lead to a loss of 30% of the forest carbon stock.
  5. Synthesis. Our study shows that climate change is likely to induce a decrease in tropical forest carbon stocks. This loss could be due to a decrease in the average tree size and to shifts in tree species distribution, with the selection of small-statured species. In Madagascar, climate-induced carbon emissions might be, at least, of the same order of magnitude as emissions associated with anthropogenic deforestation.

Introduction

Carbon sequestration by tropical forests plays an important role in the regulation of atmospheric CO2 and global climate change (Houghton 2005; Pan et al. 2011). The current carbon stock in the world's forests is estimated to be 861 ± 66 Pg C (1 Pg = 1015 g), with 55% of this carbon (471 ± 93 Pg C) being stored in tropical forests (Pan et al. 2011). Tropical anthropogenic deforestation is thus responsible for a considerable proportion (6–17%) of global carbon dioxide emissions that affect climate change (van der Werf et al. 2009). While the effect of anthropogenic tropical deforestation on carbon dioxide emissions and climate change has been intensively studied in recent years and is recognized as a global issue (Bonan 2008; van der Werf et al. 2009; Vieilledent, Grinand & Vaudry 2013b), the effect of climate change on tropical forest carbon storage, leading to a carbon-cycle feedback, remains largely uncertain (Sitch et al. 2008; IPCC, 2014).

A major part of the scientific studies dealing with this potential carbon-cycle feedback in tropical forest have used dynamic global vegetation models (DGVMs) coupled to global climate models (GCMs). DGVMs are process-based models which focus on the physiological response of vegetation (through photosynthesis and respiration) to increasing atmospheric CO2 concentration and changes in temperature and precipitation (Prentice et al. 2007). DGVMs generally agree that net primary productivity and carbon storage on land will increase as a result of the simultaneous enhancement of plant photosynthesis and water-use efficiency under higher atmospheric CO2 concentrations, but will decrease due to higher plant respiration rates associated with warming temperatures (Cox et al. 2013). Despite the numerous efforts made to improve DGVMs, large uncertainties remain in the predictions (Cramer et al. 2004; Quillet, Peng & Garneau 2010; Adams et al. 2013; Scheiter, Langan & Higgins 2013; IPCC, 2014). For instance, Cox et al. (2013) reported a range of 330 PgC in the projected change in the amount of carbon stored on tropical land by 2100 and Sitch et al. (2008) reported that cumulative land uptake differs by 494 PgC among DGVMs over the 21st century, which corresponds to 50 years of anthropogenic emissions at current level. In the past, DGVMs have either predicted a dieback of tropical forests (Cox et al. 2000; Sitch et al. 2008), with, for example, a decrease of 30% (from 568 PgC in 2000 to 398 PgC in 2100) of the carbon stored on land under the IPCC (Intergovernmental Panel on Climate Change) business-as-usual scenario IS92a (Cox et al. 2000) or a resilience of tropical forests to climate change (Sitch et al. 2008; Huntingford et al. 2013), with, for example, an average increase of 18% (from 165 Mg ha−1 in 1980 to 195 Mg ha−1 in 2100) of the African tropical forest carbon stock under IPCC business-as-usual scenario SRES A2 (Huntingford et al. 2013). In a study testing five DGVMs predictions of carbon fluxes using measurements of Amazon forests subjected to experimental drought, Powell et al. (2013) showed that the majority of models (four out of five) predicted negligible reductions in above-ground biomass in response to drought, which was in contrast to an observed c. 20% reduction in forest. Most of the uncertainty around DGVMs predictions is associated with the uncertain response of terrestrial vegetation to climate, that is to drought and heat stress (Sitch et al. 2008; Cox et al. 2013). DGVMs typically use a limited number of generic plant functional types (PFTs) to describe vegetation. This number is usually low to facilitate model parametrization, but consequently, PFTs might not be representative of the high diversity of tree species and functioning that can be found in the world tropical dry and moist forests (Powell et al. 2013; Schimel et al. 2015). Improving DGVMs is a difficult task limited by the complexity of the physiological and ecological processes explaining vegetation dynamics (Quillet, Peng & Garneau 2010; Scheiter, Langan & Higgins 2013) and by the limited amount of biological data available for model parametrization (Adams et al. 2013; IPCC, 2014). The question of knowing whether tropical forests are likely to compensate for anthropogenic carbon dioxide emissions through a strong positive biomass response to climate change thus remains unanswered.

In contrast to process-based approaches, simpler correlative methods could offer rapid and robust alternatives for the projection of the effects of climate change on tropical forest carbon storage. Despite known limitations (including the fact that they usually do not take into account biotic interactions, species adaptation and dispersal ability, see Pearson & Dawson (2003)), bioclimatic envelope models have been widely used to project the effect of climate change on biodiversity at both the species (Guisan & Zimmermann 2000; Thuiller et al. 2005; Vieilledent et al. 2013a) and biome level (Hannah et al. 2008; Zelazowski et al. 2011). For instance, Zelazowski et al. (2011) predicted a forest retreat in Amazonia, Central America and parts of Africa, and expansion in other regions, in particular around the Congo Basin. If one of the biological mechanisms explaining the impact of climate change on forest carbon stocks is the direct effect of climate on tree species metabolic rates and forest net primary productivity (Huxman et al. 2004; Cox et al. 2013), another mechanism that has to be considered on a larger time scale is the shift in tree species relative abundance and tree species range associated with climate change (Iverson & Prasad 1998; Feeley et al. 2011). Climate change should variably impact tree species through growth (Ouédraogo et al. 2013; Wagner et al. 2014), recruitment (Clark et al. 2011) and mortality (Allen et al. 2010; Anderegg, Kane & Anderegg 2013) with consequences on tree species’ relative abundances in the community. Tree species are characterized by functional traits (Violle et al. 2007) such as wood density (Chave et al. 2009; Vieilledent et al. 2012) and tree stature which includes tree maximal height and maximal diameter (King, Davies & Noor 2006; Poorter, Bongers & Bongers 2006; Barthelemy & Caraglio 2007). Because these traits determine tree and forest biomass (Baker et al. 2004; Vieilledent et al. 2012; Chave et al. 2014), the predicted shift in species range (Hannah et al. 2008) and species relative abundance (Feeley et al. 2011) with climate change should significantly impact forest carbon storage. Interestingly, in a study simulating potential tree species extinction scenarios in a Panamanian tropical forest, Bunker et al. (2005) underlined the fact that future carbon storage in tropical forests will be influenced strongly by future tree species composition. In their study, they simulated 18 possible extinction scenarios and showed that the loss of large-statured tree species (replaced by smaller-statured tree species) in a Panamanian forest would lead to a 29% decrease in the total forest carbon stock.

Although it is common to predict future species range using bioclimatic envelope models and climatic projections (Guisan & Zimmermann 2000; Pearson & Dawson 2003; Thuiller et al. 2005; Vieilledent et al. 2013a), it is difficult to predict changes in tree species composition and forest structure (tree height and diameter distribution), and thus forest carbon stocks, following climate change. Nonetheless, it is still possible to directly correlate forest biomass to climate. Recent empirical studies in tropical forests have shown the importance of climatic variables in determining tree stature and biomass (Feldpausch et al. 2011; Banin et al. 2012; Vieilledent et al. 2012; Chave et al. 2014) having consequences on carbon stocks at the forest stand level (Feldpausch et al. 2012). In particular, in a global study where 4004 trees from 58 tropical forest sites were harvested and measured, Chave et al. (2014) found that asymptotic tree height decreased with bioclimatic stress, that is with temperature and precipitation variability and drought intensity. Similarly, using a theoretical model based on scaling laws and energy budgets constrained by local resource limitations, Kempes et al. (2011) have demonstrated that maximal tree height increased with precipitation and decreased with temperature. Additional studies have shown that tree maximal size was highly correlated to forest biomass locally (Stegen et al. 2011; Michaletz et al. 2014). As a consequence, the climate should also have an effect at the ecosystem scale for determining forest carbon stocks. Only a few studies have investigated the link between climate and forest carbon stocks at the ecosystem scale using empirical correlative approaches (Keith, Mackey & Lindenmayer 2009; Stegen et al. 2011). Mixing data from boreal, temperate and tropical forests, the two studies by Keith, Mackey & Lindenmayer (2009) and Stegen et al. (2011) suggested that the role of climate was not consistent across forest types. As a result, they made no general predictions concerning changes in forest carbon stocks in response to changes in climate. Moreover, correlative models used to derive forest carbon maps at the global scale (Saatchi et al. 2011; Baccini et al. 2012) do not include explicative climatic variables and therefore cannot be used to predict future tropical forest carbon stocks under the effects of climate change.

In this study, we present the first attempt to project the effects of climate change on tropical forest carbon storage using correlative bioclimatic envelope models. To do so, we used a very large number of forest plots (1771) dispersed across Madagascar. Madagascar has a wide variety of climates and forest types and is thus particularly interesting for the study of the relationship between climate and tropical forest carbon stocks. The large climatic gradient on the island is associated with elevation and position relative to the dominant south-eastern winds (Goodman & Benstead 2003; Tadross et al. 2008) and contributes to define a large range of tropical forest types, from spiny dry forests in the subdesert Southern regions to cloud forests which are found at the top of the northern and eastern mountains. Using this large data set, we first modelled the relationship between climate and forest carbon storage and tested for a correlation between the two. We then used this correlative model to project the future national forest carbon stock under climate change and quantify the effect of climate change on the forest carbon storage. Lastly, we compared the effect of climate change to the effect of anthropogenic tropical deforestation in terms of carbon emissions and discuss the generalization of our results to other tropical regions.

Materials and methods

Compiling forest inventory data

We assembled data from nine forest inventories carried out over the period 1996–2013 [Table 1 and Vieilledent et al. (2016)]. Collaboration with different institutions (governmental institutions, conservation NGOs, research institutes) allowed us to obtain a large number of forest plots (1771) in the three forest ecoregions of Madagascar (moist, dry and spiny forest ecoregions, see Fig. 1). Ecoregions were defined on the basis of climatic and vegetation criteria using the climate classification by Cornet (1974) and the vegetation classification from the 1996 IEFN national forest inventory (Ministère de l'Environnement, 1996). For each of the forest plots, tree diameter was measured at 1·30 m (D in cm) using a metre tape and identified to the genus from vernacular names. Several trees were also measured in height (H in m) using either a clinometer (coupled to a transect tape to measure distance to tree) or an ultrasonic Vertex hypsometer. For all inventories with the exception of the IEFN inventory, all trees over 5 cm D were measured in concentric circular plots of variable radius depending on tree size. The largest radius was 30 m in moist forest (plot size of 0.28 ha) and 20 m in dry or spiny forest (plot size of 0.13 ha). The IEFN inventory (the largest inventory in the data set, providing 768 plots) used concentric circular plots for trees with 5 ≤ D < 15 cm but angle count plots (Bitterlich plots) for trees with  15 cm, with a basal area factor of 9 and 4 m2 ha−1 for moist and spiny/dry forest, respectively.

Table 1. Forest inventories used in the analysis. We combined nine forest plot inventories with above-ground carbon density data. Data were collected between year 1996 and 2013 in the three types of forest in Madagascar (moist, dry and spiny forest). Eight institutions have collaborated to provide a total of 1771 plots spread over all Madagascar
Id Name Date Ecoregion Plot nb. Institution
1 IEFN 1996 moist, dry and spiny 768 DGF
2 JariAla 2007 moist and dry 439 DGF/USAID
3 PHCF 2010 moist and spiny 92 Goodplanet/WWF
4 CAZ-COFAV 2009 moist 117 CI
5 Makira 2010 moist 130 WCS
6 CI-forêt sèche 2011 dry 126 CI
7 Ranobe-PK32 2012 dry and spiny 14 WWF/Cirad
8 Kirindy 2012 dry 15 ONE/DGF
9 Itasy 2013 moist 70 ONE/DGF
Total 1771
Details are in the caption following the image
Climatic ecoregions and location of forest plots. Left panel: Madagascar can be divided into three climatic ecoregions with three forest types: the moist forest in the East (green), the dry forest in the West (orange) and the spiny forest in the South (red). Ecoregions were defined following climatic (Cornet 1974) and vegetation (Ministère de l'Environnement, 1996) criteria. The 1771 forest plots of our study (black points) cover as much as possible the remaining tropical forest in 2010 (dark grey areas). Right panel: climate in the three forest types of Madagascar can be represented by 2D densities (coloured curves) obtained from sampling at random twenty thousand points in each forest type. The three forest types are characterized by significantly different climatic conditions and a large gradient of precipitation and temperature is observed. The 1771 forest plots of our study (dark grey points) are representative of the climatic conditions found in Madagascar forests.

Computing tree biomass with pantropical allometric models

We computed the above-ground biomass (AGBi, in Mg) of each tree i using the pantropical biomass allometric equation developed by Chave et al. (2014): AGBi = 0.0673 × (ρiDi2Hi)0.976, ρi being the tree wood density (in g cm−3). When tree height was not available, we used height–diameter allometric models based on tree height and diameter measurements from the forest inventories and one destructive study (Vieilledent et al. 2012). We used a power relationship between H and D: log(Hi) = β0 + β1 log(Di) + εi, εi ~ Normal (0,σ2), and we fitted the model independently for the three forest types. We obtained three significantly different height–diameter allometric models, representative of the climatic and soil conditions of each ecoregion (Fig. S1). For a given diameter, trees are on average higher in the moist forest than in the dry and spiny forests. When assigning wood density to each tree, we used the following rules: (i) if the tree genus was present in one of the Malagasy wood-density data bases (Rakotovao et al. 2011; Vieilledent et al. 2012), we computed the mean wood density at the genus level from these sources, (ii) if the tree genus was not present in the Malagasy data bases, we computed the mean wood density at the genus level from the global wood density data base (Chave et al. 2009), and (iii) if the tree genus was not present in the Malagasy or global wood density data bases or was undetermined (this was the case for 32% of the trees), we computed the mean wood density at the plot level using wood density values of all the trees in the plot whose genus was present in the data bases. Taking into account tree size and forest plot structure (radius for circular plots and basal area factor for Bitterlich plots), we converted each tree biomass AGBi to a value of AGB per hectare. We summed these values for all trees ≥5 cm DBH at the plot level and obtained the AGB≥5,j (in Mg ha−1) for each forest plot j. We obtained the above-ground carbon density ACD≥5,j (in Mg ha−1) for each plot j assuming a carbon ratio of 0·47: ACD≥5,j = 0·47× AGB≥5,j.

Spatial ACD modelling from vegetation indexes, topography and climate using the Random Forests algorithm

We modelled ACD spatially using three types of explicative variables: vegetation indices, topography and climatic variables. Vegetation indices were obtained from 250 m resolution MODIS (Moderate Resolution Imaging Spectroradiometer) satellite images from the years 2000 to 2010. We also utilized the annual vegetation continuous field per cent tree cover (VCF in %) available from the global land cover facility (DiMiceli et al. 2011). The per cent tree cover describes the percentage of a pixel which is covered by tree canopy. Lastly, we computed the mean annual enhanced vegetation index (EVI, from 0 to 1) from 16-day MODIS MOD13Q1 data (NASA-LP DAAC, 2014). Compared to the standard Normalized Difference Vegetation Index (NDVI), which is chlorophyll sensitive, the EVI is more responsive to canopy structural variations, including leaf area index, canopy type, plant physiognomy and canopy architecture. Moreover, EVI has an improved sensitivity in high biomass regions. Because forest plot inventory data were collected on different dates (from 1996 to 2013), we associated the measured ACD≥5,j for plot j with the mean of the three annual values of EVI and VCF at the closest years, denoted EVIj and VCFj. For examples, for plots sampled before 2002, we computed the mean EVI and VCF of years 2000, 2001 and 2002. For plots measured after 2008, we computed the mean EVI and VCF of years 2008, 2009 and 2010. For plots measured at years y between 2002 and 2008, we computed the mean EVI and VCF of years y − 1, y and + 1. Using average values of EVI and VCF, we reduced the temporal noise associated with the year-to-year and seasonal variability of these indices. For the topography, we only used the elevation (Ej, in m) which was extracted from the 90 m resolution SRTM (Shuttle Radar Topography Mission) global elevation data set. Current (1950–2000) climatic data at 30 s (~1 km) resolution were obtained from the MadaClim website (http://madaclim.org) which provides WorldClim current climate data (Hijmans et al. 2005) specifically for Madagascar. For each plot j, we extracted the annual precipitation (Pj, in mm), the mean annual temperature (Tj, in °C) and the temperature seasonality (Sj, standard deviation of monthly temperature ×100).

We used the Random Forests algorithm to model ACDj as a function of the vegetation, the elevation and the climate: ACD≥5,j = f (VCFj,EVIj,Ej,Pj,Tj,Sj). Random Forests is an ensemble learning method based on decision trees (Breiman 2001). Because it uses a nonparametric approach, Random Forests accommodates strong nonlinear relationships between the explicative variables and the response variable. Random Forests has been successfully used in previous scientific studies that modelled forest carbon stocks (Baccini et al. 2012; Mascaro et al. 2014). For our study, we performed the statistical analysis using the randomForest package (Liaw & Wiener 2002) available with the r software (R Development Core Team, 2014).

Evaluating model performance and variable importance

To evaluate model performance, we repeated a cross-validation procedure ten times. About 70% of the data were selected at random to be used as training data to fit the model. The remaining 30% was used as test data to compute several model performance indices. First, we computed the coefficient of determination (R2, in %) which indicates the part of the observed ACD variability explained by the model. Secondly, we computed the root-mean-square error (RMSE, in Mg ha−1) which measures the average difference between values predicted by the model and observations. Thirdly, we computed the mean bias (B, in %) indicating an average over- or underestimation of ACD values by the model. We computed the mean and standard deviation of the ten values obtained for the performance indices through the repeated cross-validation procedure. We also investigated the relative importance of each variable in determining ACD by computing the percentage increase in the mean standard error when the variable was randomly permuted (%IncMSE) and observing the graphical relationship between variables and ACD using partial dependence plots (see randomForest package manual for details).

Deriving a 2010 forest carbon map

We resampled the explicative variables data at 250 m resolution and used EVI and VCF from 2010 to derive a national carbon map at 250 m resolution from our ACD model. Because our data set included values of ACD observed only in forest habitat (and not in cropland or savanna for example), we made the assumption that our carbon map was not valid in non-forested regions. We thus masked the carbon map with a map depicting Madagascar's forest in 2010. This forest map was derived from the 30 m resolution 2000 forest map by Harper et al. (2007). On Harper's map, ~200 000 ha clouds are present over the ~4.2 million ha moist forest ecoregion (4.8% of clouds). To remove these clouds, we used the 2000 cloud-free tree per cent cover map provided by Hansen et al. (2013) (also at 30 m resolution) and we chose a threshold of 75% of tree cover to decide whether to replace cloud pixels with forest or non-forest pixels (Achard et al. 2014). We thus obtained a cloud-free year 2000 forest map for Madagascar. From this map, using deforestation data from 2000 to 2010 by Hansen et al. (2013), we obtained a cloud-free forest map in 2010 at 30 m resolution. We resampled this forest map to 250 m resolution using the nearest-neighbour method. Combining the carbon map and the forest map, we obtained a forest carbon map in 2010 for Madagascar at 250 m resolution. We compared the predictions of our carbon map at 250 m resolution with the measured ACD data from the 1771 forest inventories and computed the resulting R2 and RMSE.

Comparisons with global carbon maps

We tested whether our 2010 forest carbon map for Madagascar at 250 m was more accurate than the two global carbon maps provided by Saatchi et al. (2011) and Baccini et al. (2012). Saatchi's map is at 1 km resolution and Baccini's map is at 500 m resolution. To allow a fair comparison between the predictions of the three maps at different resolutions, we resampled the three maps to 1 km resolution using a bilinear interpolation. We compared the predictions of the three maps with the measured ACD data, and we computed the resulting R2 and RMSE. At the regional scale, we also compared the mean and standard deviation of the predicted ACD values by forest type using our map at 250 m, Saatchi's map at 1 km and Baccini's map at 500 m. We also compared the values of total forest carbon stock at the national scale using the three maps.

Forecasting the effects of climate change on forest carbon storage

We used the correlative bioclimatic envelop model we fitted previously (see Section ‘2.3’) to predict the potential future forest carbon stocks at 250 m for Madagascar in 2050 and 2080 considering climate change. Because changes due to climatic variables were our key focus, we assumed no land-use change (no deforestation) between 2010 and 2080. Also, we assumed no change in the values of the vegetation indices in the future. As a consequence, our predictions only indicate the marginal effect of the climatic change on forest carbon stocks through changes in the values of the climatic variables.

For future climatic data in 2050 and 2080, we used the projections of seven global climate models (GCMs), following the representative concentration pathways (RCPs) 4.5 and 8.5. GCMs were obtained from the Coupled Model Intercomparison Project Phase 5 (CMIP5) of the Intergovernmental Panel on Climate Change (IPCC). We used the following seven GCMs: ACCESS 1.0, CCSM4, GISS-E2-R, HadGEM2-ES, IPSL-CM5ALR, MIROC5 and NorESM1-M. Climatic data were obtained from the MadaClim website (http://madaclim.org) which provides the CCAFS GCM future climate data (http://www.ccafs-climate.org/data/) specifically for Madagascar. We then compared these changes in forest carbon stock induced by climate change to carbon emissions that would be associated with a spatially homogeneous deforestation rate of 0.5% y−1. This value corresponds to the last estimate (period 2000–2010) of the national deforestation rate for Madagascar (Hansen et al. 2013; ONE et al., 2013).

Results

Spatial distribution of forest carbon stocks in Madagascar

We obtained very coherent spatial patterns of forest carbon stocks at the national scale for Madagascar. In particular, we were able to clearly differentiate the carbon stocks for the spiny, dry and moist forests (Fig. 2) for which we obtained mean carbon stocks of 17 (±6), 52 (±25) and 150 (±33) Mg ha−1, respectively (Table 2). The highest carbon stocks for Madagascar (> 200 Mg ha−1) were found in the Makira and Masoala peninsula forests around the Antongila bay in north-east Madagascar (Fig. 2). These hilly forests are known to form the largest tracts of intact moist forest remaining in Madagascar. For the 250 m map (with spatial grid cells of 6.25 ha), the maximal predicted value of the forest carbon stock was 261 Mg ha−1, which is realistic for Madagascar forests at this resolution. In comparison, the maximal carbon stock measured on much smaller forest plots (< 0.3 ha) was of 348 Mg ha−1. Moreover, the effect of climatic and altitudinal gradients on the forest carbon stocks was clearly visible on the forest carbon map (Fig. 2). In particular, we were able to observe a rapid drop of the forest carbon stock in the extreme south of Madagascar which corresponds to the rapid transition from the eastern moist forest receiving high precipitation to the southern xeric spiny forest in the mountains’ rain shadow (Fig. 2).

Details are in the caption following the image
Relationships between explicative variables and ACD. Graphics shows the marginal effect of the variable on ACD. Range of predicted ACD is reduced compared to the observed range of ACD in forest plots because variables other than the target variable are set to their mean values. Hash marks at the bottom of the plot indicate the deciles of the explicative variable. Percentage in the top-left or top-right corner of each panel is the percentage of increase in mean square error when the variable was randomly permuted, which indicates the variable relative importance in determining ACD.
Table 2. Comparing observed and predicted mean ACD (Mg ha−1) by forest type with global carbon maps. Forest areas in 2010 are in thousands of ha and mean carbon stocks are in Mg ha−1. Numbers in parentheses represent the standard deviations. Column ‘Mean obs.’ gives observed mean carbon stocks from forest plot inventories and column ‘Mean 250 m’ gives predicted mean carbon stocks from the national forest carbon map at 250 m. While global carbon maps gave relatively good estimates of the mean carbon stocks for the moist forest in Madagascar, they failed at providing good estimates and differentiating clearly the carbon stocks in the dry and spiny forests
Ecoregion Forest area Plot nb. Mean obs. Mean 250 m Saatchi 1 km Baccini 500 m
Moist 4625 1190 136 (67) 150 (33) 128 (53) 113 (36)
Dry 2725 367 51 (33) 52 (25) 41 (21) 33 (20)
Spiny 1741 214 18 (11) 17 (6) 34 (19) 22 (9)

Also, we were able to find much higher carbon stocks in the Amber Mountain (850 to 1475 m of altitude) at the extreme north of Madagascar, where forests benefit of a microclimate with high precipitation in the middle of the dry ecoregion.

Model performance and variable importance

The cross-validation results indicated that the model had good predictive abilities with a relatively high coefficient of determination (R2 = 64(±2)%), and relatively low root-mean-square error and bias [RMSE = 44(±2) Mg ha−1 and B = +31(±4)%]. In terms of explicative importance, variables were classified in the following decreasing order: temperature seasonality (S, 68% of increase in the mean square error when the variable was randomly permuted), precipitation (P, 49%), mean annual temperature (T, 48%), elevation (45%), EVI (40%) and VCF (38%). All variables were highly explicative of the amount of forest carbon stock, but climatic variables (temperature seasonality, annual precipitations and mean annual temperature) were found to be the most important, relative to elevation and vegetation indices. Analysing the relationship between explicative variables and ACD, we observed biologically coherent patterns (Fig. 3). ACD was higher at intermediate elevation (500–1200 m) and increased almost linearly with EVI and VCF. Regarding climatic variables, ACD decreased dramatically with increasingly stressful climatic conditions. We identified the climatic tipping points beyond which the ACD dropped: for T > 21°C, S > 2100 and P < 1100 mm y−1 (Fig. 3).

Details are in the caption following the image
Forest carbon map in 2010 for Madagascar. We derived a national forest carbon map in 2010 for Madagascar at 250 m resolution. We fitted our model using above-ground carbon density for 1771 forest plots measured between 1996 and 2013. Our model included six explicative variables: two vegetation indexes (VCF and EVI from 2000-2010 MODIS satellite images at 250 m), one topographic variable (elevation from SRTM at 90 m) and three climatic variables [mean annual temperature, mean annual precipitation and temperature seasonality from WorldClim at 30 arc-seconds (~1 km)]. Our predictions are limited to the extent of the forest in 2010. Clear differences appear for the forest carbon stocks between the three ecoregions including moist, dry and spiny forest (see black lines for delimitations).

Comparison with global carbon maps

At 1 km resolution, our map provided much more accurate predictions of ACD values (R2 = 0.64, RMSE = 44 Mg ha−1) than Saatchi's or Baccini's maps (R2 = 0.26, RMSE = 64 Mg ha−1 and R2 = 0.17, RMSE = 63 Mg ha−1, respectively). The best predictions were obtained using our model at 250 m resolution (R2 = 0.70, RMSE = 40 Mg ha−1) (Fig. S2). At the forest type level (Table 2), the global carbon maps gave relatively good estimates of the mean carbon stocks for the moist forest in Madagascar, but conversely, they failed to provide good estimates and differentiate clearly between the dry and spiny forests in terms of carbon stocks. Using our map at 250 m (Fig. 2), we estimated Madagascar's total forest carbon stock to be 873086 Gg (1Gg = 109 g) in 2010. Computing the total forest carbon stock using mean ACD estimates by forest type from Saatchi's and Baccini's maps (Table 2) led to lower estimations of the total carbon stock for Madagascar c. 2010 (764168 (−12%) and 652896 (−25%) Gg, respectively).

Potential impact of climate change on carbon emissions

Comparing the current climate in Madagascar with the projected climate in 2080 following RCP 8.5, temperature seasonality and mean annual temperature are predicted to increase (of +138 SD × 100 and +3.7°C on average, respectively), while precipitation is predicted to decrease (of −107 mm y−1 on average) over almost the entire forest area (Fig. S3). Using climatic projections following the RCP 8.5 and our forest carbon stock model, we forecasted the changes in forest carbon stocks by the year 2080 (Fig. 4). Most of Madagascar forests are likely to experience a decrease of their carbon stock, up to −150 Mg ha−1 on the period 2010–2080 for the Madagascar moist forest. In some areas, in particular in the spiny forest, a limited increase of the carbon stock is expected (< 50 Mg ha−1). The increase of the forest carbon stock is associated with the predicted future increase of the precipitation in these areas (Fig. S3). At the national scale, these changes should lead to an average forest carbon stock loss of 17% (range: 7–24%) by the year 2080 (Fig. 4 and Table S1). Following RCP 4·5, the forest carbon stock loss ranged from 2 to 13% (Table S1). Following RCP 8·5, the total forest carbon stock would drop progressively from 873086 Gg in 2010, to 799 097 Gg (−8%) in 2050 and 720944 Gg (−17%) in 2080 (Fig. 4). For a comparison, a spatially homogeneous deforestation of 0·5% y−1 from 2010 to 2080 over Madagascar would lead to a residual carbon stock of 614714 Gg (−30%).

Details are in the caption following the image
Forest carbon stock evolution with climate change in Madagascar. (a–c): Maps of Madagascar forest carbon stocks for the years 2010, 2050 and 2080. Using climate projections in 2050 and 2080 (from seven IPCC CMIP5 global climate models following the RCP 8.5) and our carbon stock model, we estimated the average forest carbon stock in 2050 and 2080 for Madagascar. At the national scale, climate change would result in a progressive decrease of the total forest carbon stock (see C stocks, in Gg = 109 g) of −9% and −17% for years 2050 and 2080, respectively. (d): Distributions of the 250 m resolution forest pixels as a function of the carbon stock change for the periods 2010–2050 and 2010–2080. (e–f): Carbon stock change on the periods 2010–2050 and 2010–2080. Moist forest in the East should experience the greatest changes with a drastic decrease of the forest carbon stock.

Discussion

An accurate forest carbon map for Madagascar

Using a large data set including 1771 forest plots spread across the entire country and representative of the three tropical forest types existing in Madagascar, we were able to fit a good predictive model (R2 = 64%, RMSE = 44 Mg ha−1 and B = +31%). We obtained an accurate (R2 = 0.70, RMSE = 40 Mg ha−1) national forest carbon map at 250 m resolution for Madagascar for year 2010. In a study mapping forest carbon on 16 million hectares in the Western Amazon and using a more sophisticated model with spatial autocorrelation, Mascaro et al. (2014) obtained a map with a lower R2 (R2 = 59%) but a smaller RMSE (RMSE = 26 Mg ha−1). In their case, the initial variability of observed ACD was limited as they used 1 km LiDAR-based estimates (and not estimates at a point), thus potentially reducing the residual values between predicted and observed ACD. Regarding forest carbon maps that can be used at the national scale for Madagascar, the two global maps by Saatchi et al. (2011) and Baccini et al. (2012) have a coarser resolution (500 m and 1 km resolution). We showed that these two global maps were less accurate than the map we obtained at 250 m resolution (R2 ≤ 26 and RMSE ≥ 63 Mg ha−1) and that they failed to differentiate clearly between the carbon stocks in the spiny and dry forests of Madagascar. When used at the national level, the two global maps underestimated (by more than 12%) the total carbon stock of Madagascar, in comparison with the map produced herein. The lower accuracy of the global carbon maps can be explained by two factors. First, only Saatchi's map (which is the most accurate of the two maps according to our study) utilized forest plot data from Madagascar (202 plots) to calibrate the statistical models used to predict ACD. Regarding the lack of distinction between spiny and dry forest carbon stocks, the plots used to calibrate Saatchi's model were located in the moist forest and were thus not representative of Madagascar's other forest types. The fact that a substantial number of carefully established field plots are necessary to derive accurate ACD maps has been previously underlined by Mitchard et al. (2014). Secondly, neither Baccini's nor Saatchi's maps included climatic variables as explicative factors of ACD, while we have shown in our study that they were more important explicative variables than the altitude and the vegetation indices.

Effects of climate on forest carbon stocks

The tree height is a key variable determining tree biomass (Vieilledent et al. 2012; Chave et al. 2014). Recent empirical and theoretical studies have emphasized the importance of climatic variables in determining the tree height–diameter relationship (Banin et al. 2012; Chave et al. 2014; Feldpausch et al. 2011) and the tree maximal height (Kempes et al. 2011). In a pantropical study, Chave et al. (2014) have demonstrated that tree height decreased with increasing temperature seasonality, precipitation seasonality and climatic water deficit. Chave et al. (2014) have shown that temperature seasonality (a proxy for the length of the growing season) was by far the most important climatic variable in explaining the tree height–diameter relationship and consequently tree biomass. In our study, we also identified temperature seasonality as being the most important variable determining forest biomass at the plot level (with a decrease of ACD with temperature seasonality), thus confirming the results obtained by Chave et al. (2014). In a theoretical study, based on scaling laws and energy budgets constrained by local resource limitations, Kempes et al. (2011) have demonstrated that maximal tree height increased with precipitation and decreased with temperature, thus corroborating the results of our study.

At the forest ecosystem level, Fischer et al. (2014) found a pattern similar to ours regarding the relationship between forest carbon stock and precipitation in the moist forests of Madagascar. The authors identified a tipping point around 2000 mm yr−1 for precipitation under which the forest carbon stock was rapidly decreasing. While they argued that a temporary 20% reduced rainfall would have a moderate impact on the forest carbon stock (due to the soil water content that would support tree growth), they found that a decrease of 50% of rainfall would lower the carbon stocks by about 20%. In a large-scale study using data from 271 plots from temperate and tropical forests in North, Central and South America, Stegen et al. (2011) found weak relationships between forest biomass and climate. No significant relationships was found between forest biomass and mean annual temperature across forest types and a weak, albeit significant, positive relationship was found between forest biomass and annual precipitation. In another study using biomass data from 136 plots from the world's most carbon-dense primary forests, Keith, Mackey & Lindenmayer (2009) concluded that the highest biomass carbon density occurs in cool, moderately wet climates in temperate moist forest biomes. However, Fig. 3 in Keith, Mackey & Lindenmayer (2009) and Fig. 4 in Stegen et al. (2011) show that for sites with a tropical climate (excluding boreal and temperate sites with mean annual temperature < 15C), forests under warmer and drier climates (with higher mean annual temperature and lower annual precipitation) appear to have lower carbon stocks. Results obtained by Keith, Mackey & Lindenmayer (2009) and Stegen et al. (2011) are then in line with the results obtained by Chave et al. (2014) regarding the relationship between tree biomass and climate in tropical forests and are also comparable with the results of our study. More work is still required in order to understand more precisely the effect of climate on forest carbon storage across other forest types such as boreal and temperate forests (Keith, Mackey & Lindenmayer 2009; Stegen et al. 2011; Koven 2013). One difficulty is that temperate regions have a diversity of forest types that support a wide range of mature carbon stocks or have a long land-use history with reduced carbon stocks. As a consequence, past studies comparing carbon stocks in tropical and temperate forests have led to contrasting conclusions. For example, Houghton (2005) and Saatchi et al. (2011) have shown that tropical forests have a much higher mean biomass (c. 130 Mg ha−1) than temperate forests (c. 50 Mg ha−1), which contradicts the results reported by Keith, Mackey & Lindenmayer (2009).

The results of our study at the forest ecosystem scale, together with the previous results obtained by Chave et al. (2014) at the tree level, are coherent with the known physiological and ecological mechanisms explaining the relationship between forest biomass and climate. Several studies have shown that climate (mainly through growing season length and water availability) directly impacts tree species metabolic rates such as respiration and photosynthesis, having direct (∼ daily or yearly) repercussions on tree species growth and forest net primary productivity (Huxman et al. 2004; Cox et al. 2013). On a larger time scale (∼ decades), climate drives tree species demographic rates such as mortality (Allen et al. 2010; Anderegg, Kane & Anderegg 2013) and fecundity (Clark et al. 2011), thus determining tree species composition and forest structure (Feeley et al. 2011) which in turn define forest carbon stocks (Bunker et al. 2005). On this point, it is interesting to see that the estimate of 29% decrease in the forest carbon stock reported by Bunker et al. (2005), which is associated with a replacement of large-statured tree species with smaller-statured tree species, is close to our estimate of 7–24%. At larger geographical scales and on an even larger time scales (∼ centuries), tree species composition is assumed to be the result of a selection pressure with an adaptation to local climates (Iverson & Prasad 1998; Clark & McLachnan 2003; Michaletz et al. 2014). Metabolic scaling theory (Michaletz et al. 2014) and recent empirical data (Stegen et al. 2011) led to the conclusion that forest biomass is strongly correlated to the size of the largest individual, which is in itself strongly dictated by tree species architectural characteristics (Poorter, Bongers & Bongers 2006; Barthelemy & Caraglio 2007) and the local climate (Feldpausch et al. 2011; Kempes et al. 2011; Banin et al. 2012; Vieilledent et al. 2012; Chave et al. 2014). In Madagascar, tree species composition and climatic conditions in the moist, dry and spiny forest are substantially different (Goodman & Benstead 2003; Harper et al. 2007) resulting in significantly different sizes for the largest individual between forest types (Fig. S1 and Vieilledent et al. (2012)) and in significantly different forest carbon stocks (Fig. 2). Significant changes in tree species distribution are expected in Madagascar (Hannah et al. 2008; Vieilledent, Grinand & Vaudry 2013b). Tree species adapted to drier conditions and to shorter growing seasons are likely to increase their relative abundance in forest communities locally (Feeley et al. 2011) or to expand their geographical range (Hannah et al. 2008). It has been demonstrated that these tree species are usually characterized by lower asymptotic heights (Vieilledent et al. 2012; Chave et al. 2014) and lower biomass. Climate change would thus indirectly induce a decrease of the forest carbon stock through the selection of small-statured tree species.

Tropical forests as a carbon sink: comparing DGVMs and bioclimatic envelope model outputs

Our model estimates that, in Madagascar, changes in climatic conditions predicted to occur between 2010 and 2080 will cause the loss of 17% (range: 7–24%) of the forest carbon stock and would result in the emissions of 152142 Gg of carbon into the atmosphere. Climate change induced emissions would be of the same order of magnitude as the emissions associated with projected anthropogenic deforestation on the same period of time, with a loss of 30% of the forest carbon stock and emissions of 258372 Gg of carbon into the atmosphere. At present, most DGVMs also simulate a decrease of the vegetation carbon stock in response to reduction in precipitation and increase in temperature (Sitch et al. 2008; Cox et al. 2013; Huntingford et al. 2013). For example, Huntingford et al. (2013) estimated a decrease from 5 to 30 Mg ha−1 of the vegetation carbon stock in tropical Africa between 1860 and 2100 under the effect of increasing temperature (Fig. SI2 in the cited paper). But the strong decrease of the vegetation carbon stock due to climate change would be largely compensated by the fertilization effect of CO2 modelled in DGVMs. Accounting for both climate change and CO2 fertilization, Huntingford et al. (2013) estimated an average increase of 30 Mg ha−1 (+18%) of the African tropical forest carbon stock between 1980 and 2100 under IPCC business-as-usual SRES A2 scenario. Thus, most actual DGVMs predict a resilience of tropical forests to climate change due to an increase of tree growth (Sitch et al. 2008; Cox et al. 2013; Huntingford et al. 2013). Compared to DGVMs, our model does not include the direct effect of CO2 fertilization on forest growth and resulting forest carbon stock. Indeed, higher CO2 concentration is expected to raise plant photosynthetic rates and enhance water-use efficiency (Bonan 2008). However, there is no empirical evidence that these physiological responses do increase carbon sequestration in natural tropical forests (van der Sleen et al. 2015). For example, CO2 fertilization might lead to an increase of tree growth but also to an increase of tree mortality due to faster growth (Brienen et al. 2015). Moreover, some studies showed that productivity may eventually become constrained by nutrient limitation (Norby et al. 2010) or climate stress (Perry et al. 2013). Finally, CO2 enrichment experiments showing the CO2 fertilization effect have been conducted in temperate forests (Norby et al. 2010) but not yet in tropical forests (but see Tollefson (2013)). As a consequence, DGVMs might overestimate the sink capacity of tropical forests (van der Sleen et al. 2015). Although not taking into account the potential effect of CO2 fertilization, our forest carbon model provides more insight about the vegetation response to changes in temperature and precipitation, which constitutes the largest source of variability in DGVM outputs (Sitch et al. 2008; Cox et al. 2013; Huntingford et al. 2013).

Our approach shares the same hypothesis as when bioclimatic envelope models are used to predict the effects of climate change on species distribution (Pearson & Dawson 2003). In particular, for the forecast of future forest carbon stocks, we assumed that tree species will be able to track climate change (full dispersal hypothesis). Also, we assumed no change in the values of the vegetation indices in the future whereas climate change should impact them. These two hypotheses are conservative regarding the main result of our study which predicts a decrease of the forest carbon stock in Madagascar with climate change. Given the velocity of climate change (Loarie et al. 2009) and the difficulties experienced by tree species to migrate and track climate change (Zhu, Woodall & Clark 2012), shifts in species distribution at large geographical scale are uncertain. More likely, tree species will experience range contraction (Zhu, Woodall & Clark 2012; Vieilledent, Grinand & Vaudry 2013b) and climate change could lead to widespread tree-mortality events (Anderegg, Kane & Anderegg 2013; Brienen et al. 2015). Also, it is unlikely that vegetation indices will increase with climate change. Constant or lower values of vegetation indices due to an increase in the abundance of small-statured tree individuals and species are expected. Considering these two hypotheses, it is likely that our model underestimates the loss of forest carbon stock that would be associated with climate change.

At the global scale, climate observations and model predictions suggest continued increases in temperature (Hulme & Viner 1998) and decreases in precipitation over much of the humid tropics (Malhi & Wright 2004) thus suggesting a potential decrease in tropical tree growth and height in the future. Regarding seasonality, changes in rainfall seasonality are expected throughout the tropics (Feng, Porporato & Rodriguez-Iturbe 2013), but more research is still needed to be able to interpret these seasonal changes in terms of potential impacts on tree species size and forest carbon storage. Considering our results and those from previous studies (Bunker et al. 2005; Chave et al. 2014; van der Sleen et al. 2015), a substantial part of the tropical forest carbon stock is likely to be released into the atmosphere under the effect of climate change. On this basis, the recent simulated resilience of tropical forests to climate change with DGVMs, which relies mostly on the strong CO2 fertilization effect (Sitch et al. 2008; Cox et al. 2013; Huntingford et al. 2013) may be questioned. The decrease of tropical forest carbon stocks with climate change, leading to a carbon-cycle feedback, should not be neglected and might be of the same order of magnitude as tropical anthropogenic deforestation regarding CO2 emissions in the atmosphere.

Acknowledgements

Financial support was provided by Cirad, AFD (Agence Française de Développement) through the AT-REDD+ Madagascar project and FRB–FFEM (Fondation pour la Recherche sur la Biodiversité – Fond Français pour l'Environnement Mondial) through the BioSceneMada project (project agreement AAP-SCEN-2013 I). We would also like to thank all the people who contributed to the forest inventories in the field. Lastly, this work would not have been possible without the willingness of governmental officials, conservation NGOs and research institutes in Madagascar to share their forest inventory data for the establishment of a national forest carbon map. We thank them for their support.

    Data accessibility

    The following data are accessible on the Dryad digital repository associated with this study (see Vieilledent et al. (2016) and http://dx.doi.org/10.5061/dryad.9ph68):
    • Above-ground carbon density (ACD) data.
    • Present and future climatic data for Madagascar.
    • Additional explicative variables (elevation, EVI and VCF) used to derive the forest carbon map.
    • R script that can be used to fully reproduce the results of the study.
    • Random Forests model used to derive the forest carbon maps.
    • Madagascar 2010 forest carbon map at 250 m resolution.