Forecasting forest areas and carbon stocks in Cambodia based on socio-economic factors/Prevision des surfaces forestieres et des stocks de carbone au Cambodge, basee sur des facteurs socio-economiques/Pronostico sobre las areas forestales y las reservas de carbono en Camboya a partir de factores socio-economicos.
Global greenhouse gas (GHG) emissions from the agriculture, forestry and other land use sector account for approximately 24% of anthropogenic emissions (IPCC 2013). REDD+, which originally consisted of reducing emissions from deforestation and forest degradation in developing countries and later had the conservation and sustainable management of forests and enhancement of carbon stocks added, has evolved as an important scheme for controlling GHG emissions. Establishing forest reference levels (RLs) or forest reference emission levels (RELs), which serve as standards or a baseline for valuating actions under the REDD+ programme, is one of the main tasks for implementing REDD+ under the UNFCCC. Forest RELs can be established based on historical data by assuming that the historical trend will continue in the future without implementing the REDD+ programme, but this has proven to be a challenging task.
Kimmins et al. (2010) note the complexity of forestry and argue that human population growth is the ultimate environmental threat to the world's forests. Vieilledent et al. (2013) forecast the intensity of deforestation to 2030 for five study sites in Madagascar as a function of human population density. Verburg et al. (2002) present a spatial dynamic land use change model, CLUE-S, to relate land use changes to socio-economic and biophysical factors, including population density, altitude, slope, distance to a town, distance to a coast, and distance to a stream and applied this model to some areas of the Philippines and Malaysia to forecast land use changes to 2017. Clark Labs (2013) provides commercial software, IDRISI LCM, that can analyse land cover changes and model future scenarios as a function of accessibility to a forest, infrastructure development and policy, among other factors. Buongiorno and Zhu (2013) forecast forest areas for various countries on the basis of the annual rate of forest area change, which is taken to be a quadratic function of income per capita through the Global Forest Products Model (GFPM). These models describe different relationships between affecting factors of deforestation and their impacts.
When combining forest area with forest carbon stocks per hectare, forest carbon stocks can be obtained; respecting the Decisions of UNFCCC, such as Decision 11/CP.19 (UNFCCC 2014), both forest area data and average forest carbon stock data are needed. Concerning Cambodia, Kiyono et al. (2010) estimated average carbon stocks in 4 carbon pools; aboveground and belowground biomass; deadwood; and litter for evergreen, deciduous and secondary forests in Cambodia using data from 12 plots. Samreth et al. (2012) estimated forest tree biomass carbon stocks for evergreen, semi-evergreen and deciduous forests using data from 100 permanent sampling plots (PSPs) set by the Forestry Administration (FA). Sasaki et al. (2013) reviewed various research results, calculated carbon stocks for every category of land uses in Cambodia, and provided forecasts of reductions and removals in Cambodia until 2043.
Deforestation in a country is mainly a result of the human activities that take place under the socio-economic circumstances specific to the country. The objective of this study was to forecast the forest areas and forest carbon stocks in Cambodia for 2011 to 2018 using panel data analysis incorporating socio-economic factors. The forecast results can be used as a reference in establishing national RELs for implementing REDD+ schemes in Cambodia and for making decisions concerning sustainable forest management. The research consists of two steps. In the first step, the relations of forest area and forecast carbon stocks to socio-economic factors are specified. In the second step, forest areas and forest carbon stocks from 2011 to 2018 in Cambodia are forecasted conditionally on the assumptions regarding the related socioeconomic factors and the estimation parameters obtained in the first step.
METHOD AND DATA
Cambodia, located in Southeast Asia, has a total area of 181 035 km2 (National Institute of Statistics (NIS) 2012) and a population of 14.68 million (NIS 2013). Cambodia witnessed its first general election in 1992 and became a peaceful country in 1998 when a civil war finally ended. Since 1999, the nation's economy has been growing rapidly. However, deforestation-the loss of forestland-occurred during the civil war and has continued with economic development since the civil war. In 1965, 73.04% of the country's territory was covered by forests (FA 2008). The first of the three most recent forest cover assessments by the FA showed that Cambodia had a forest area of 11.10 million ha in 2002, or 61.15% of the total area of the country. The forest area had decreased to 10.73 million ha, or 59.09%, by 2006 and to 10.36 million ha, or 57.07%, by 2010 (FA 2008, 2011).
National level forecasts were to be implemented; however, the three national forest area data, namely, for 2002, 2006 and 2010, are too sparse for statistically significant forecasting. Therefore, provincial-level secondary data were collected. Cambodia has 24 administrative divisions by 2010: the capital Phnom Penh and 23 provinces. Phnom Penh and the provinces of Kandal, Kep, Prey Veng, Svay Rieng and Takeo each have less than 20 000 ha of land covered by forest. Phnom Penh and these provinces were considered outliers from the perspective of data analysis and were excluded from the model specification. The remaining 18 provinces were dealt with in panel data analysis.
Panel data analysis, an econometric approach, was applied in this research. Panel data analysis is widely used in clarifying drivers or affecting factors of deforestation for different countries and regions (e.g., Barbier and Burgess 2001, Michinaka and Miyamoto 2013, Michinaka et al. 2013, Nguyen Van and Azomahou 2007, etc.). Panel data analysis is suitable especially for situations where time series data are not long enough. Panel data, or longitudinal data, comprise a dataset based on observations of the same individuals over multiple periods; thus, it is wider than time-series data and longer than cross-sectional data. The panel data used in this research are composed of 18 provinces at three time points, for a total of 54 observations.
The response variables are forest area and forest carbon stock. The explanatory variables included population, agricultural gross value added (GVA), Economic Land Concession (ELC) and the interaction between agricultural GVA and ELC. Because ELC is a dummy variable with values of only one and zero, implying ELC implementation and no ELC implementation, respectively, the interaction between agricultural GVA and ELC implies agricultural GVA in the province where ELC is implemented.
The model takes the following linear functional form:
[TF.sub.it]or[CS.sub.u] = [[alpha].sub.it] + [[beta].sub.1][POP.sub.it] + [[beta].sub.2][AGVA.sub.it] + [[beta].sub.3][ELC.sub.it] + [[beta].sub.4][(AGVA:ELC).sub.it] + [[mu].sub.it] (1)
where TF stands for forest area, CS for forest carbon stock, a for intercept, POP for population, AGVA for agricultural GVA, AGVA.ELC for the interaction, [mu] for the disturbance term, i for the individuals (the provinces), and t for the time, namely the years 2002, 2006 or 2010.
Assuming parameter homogeneity, i.e., no individual or time effects, the intercept [alpha] and slope parameter [beta] remain constant for all is and ts, and the disturbance term [[mu].sub.it] is treated as random with mean 0, a pooled or pooling OLS model can be obtained (Greene 2000, Croissant and Millo 2008). Usually, the pooled OLS model is not good because of strict assumptions. The other two popular models under panel data analysis are the fixed effects (FE) model and the random effects (RE) model. The FE model assumes different intercepts, i.e., recognising the presence of individual effects [[alpha].sub.i], but the same slope [beta] across cross-sectional unit i, and the intercepts and slopes remain stable as time t changes. As for the pooled OLS model, the FE model is also estimated by an ordinary least square regression (OLS). The RE model is estimated using a generalised least square estimate (GLS). The essential difference between FE and RE models is whether there are correlations between explanatory variables and the disturbance term. When correlation exists, the FE model should be chosen; otherwise, the RE model is preferred. The Hausman test statistic and the p-values for the difference between FE and RE estimates are calculated in this research. In the presence of heteroskedasticity and serial correlation, robust estimations of the error covariance matrix are necessary.
Packages lmtest (Zeileis & Hothorn 2002) and plm (Croissant and Millo 2008) in R 3.1.0 (R Core Team 2014) were used to estimate Pooled models, FE models and RE models, and to conduct a series of tests to assess parameter heterogeneity, serial correlation, cross-sectional dependence and heteroskedasticity and to obtain robust estimates.
The three recent national forest cover maps provided by the FA were used based on manual interpretation of Landsat imagery from 2002, 2006, and 2010. These maps have eight land cover classes, i.e., evergreen forests, semi-evergreen forests, deciduous forests, bamboo thickets, other forests (e.g., rubber plantations, other plantation forests, inundated forests, mangrove forests, flooded forests, and degraded forests), evergreen and deciduous wood and shrub lands, and other land covers (e.g. agriculture, urban, water bodies, grass land and barren land); the former five classes are forests, whereas the latter three classes are non-forests (FA 2011). Because some vegetation classes were partly amended by the FA in 2006 and 2010 (e.g., forest to woodland or vice versa, reclassification between evergreen and deciduous), these changes were applied to the 2002 and 2006 maps for consistency. The National Census 2008 map was used for the provincial boundaries, which are the same as those used for the statistical data, such as the data on population and agricultural GVA. This provincial boundary map differs in some respects from the map used by the FA, particularly with respect to the boundary between Koh Kong and Preah Sihanouk. For these reasons, the provincial data on total forest areas used here differ slightly from those published by the FA.
Forest carbon stocks
Forest carbon stocks are defined here as carbon stocks stored in aboveground and belowground biomass, i.e., living biomass, in forests. Five categories are included: evergreen, semi-evergreen, deciduous forests, bamboo thickets, and other forests. The carbon stocks in the aboveground and belowground biomass for the five categories are as follows, respectively, in MgC ha-1 (tons of carbon stock per hectare): 96.2, 27.8; 98.1, 29.8; 95.1, 28.9; 87.6, 26.6; 36.4, 11.1 (Sasaki et al. 2013). The carbon stocks in evergreen and deciduous wood and shrub land are not included because they are not classified as forests, even though there is a possibility that they can be converted into forests in the future. The data of Sasaki et al. (2013) for carbon stocks per hectare were used in the analysis, but deadwood, litter and soil organic matter were not included.
Based on field surveys conducted in Cambodia and the list of the drivers or affecting factors in the Roadmap of the REDD+ program in Cambodia (Cambodia et al. 2010), Michinaka et al. (2013) drew a cause map of socio-economic factors and their relationships to forests in Cambodia. Michinaka et al. (2013) found that, in Cambodia, population, agricultural GVA, or simply called agricultural sector GDP, and ELC are the significant affecting factors of deforestation, whereas the area of cultivated rice (ha), industrial GVA (million Riels), income by household (million Riels), and house floor area by household (m2) are not significant in affecting deforestation. Based on this result, population, agricultural GVA and ELC were included in the forecasting model.
In 1962, a population census was conducted, and in 1998, a General Population Census of Cambodia was conducted. The 1998 census yielded a detailed population dataset (NIS 2000). Since 1998, the Inter-Censal Population Survey of 2004, the General Population Census of Cambodia of 2008, and the Inter-Censal Population Survey of 2013 have been conducted. Because some provinces were combined in the 2004 InterCensal Population Survey, population data from the 2004 survey were not used. Although some annual population projections are published by NIS, these population projections were not used either, due to the gaps between these data and the new Census data. For consistency with the forest area data described above, population data for 2002, 2006 and 2010 were interpolated from the General Population Census data from 1998 and 2008 and the Inter-Censal Population Survey data from 2013.
The population in Cambodia increased from 11.44 million in 1998 to 13.40 million in 2008 and 14.68 million in 2013 (NIS 2000, 2009, 2013). Based on these data, NIS estimates the average annual exponential growth rate of the population of Cambodia as 1.54% from 1998 to 2008 and 1.83% from 2008 to 2013. The total population of the 18 provinces considered in this research study was 7.12 million in 1998, 8.49 million in 2008, and 9.18 million in 2013. The corresponding population growth rates were approximately 1.78% and 1.58% for the periods of 1998 to 2008 and 2008 to 2013, respectively. For comparison, the population growth rates for Phnom Penh and the other five provinces were 1.31% and 2.34%, respectively, for those two periods. Although the population growth rate in the 18 provinces decreased from the first to the second period, the average annual increase in the population was consistently approximately 137 000 persons per year.
In Cambodia, the provincial agricultural gross product is expressed in terms of the GVA in the agricultural sector. Oeur and Yim (2009) explained that, in Cambodia, the GDP (gross domestic product) is the sum of the GVA of the agricultural sector, the industrial sector and the services sector, the tax on production less subsidies and the FISIM (Financial Intermediation Services Indirectly Measured). Agriculture encompasses crops, livestock and poultry, fisheries, forestry and logging. Industrial crops, such as rubber, cassava, and sugarcane, are included in agriculture. Agricultural GVA data were sourced from the NIS of the Ministry of Planning. The constant values for agricultural GVA in Riels in 2000 were calculated based on GDP deflator data from the World Bank (2012).
In Cambodia, the productivity of agricultural production is increasing. However, agricultural GVA, or simply agricultural GDP, is growing at a slower rate than the growth rate of the GDP of the industrial sector and the GDP of the country as a whole. At 2000 constant prices, the agricultural GDP increased from 4 942 billion Riels in 1998 to 8 311 billion Riels in 2010, which corresponds to an average annual rate of 4.43%. In comparison, the GDP of the industrial sector increased from 1 936 billion Riels in 1998 to 8 088 billion Riels in 2010 (an average annual rate of 12.65%), and the total national GDP increased from 11 545 billion Riels in 1998 to 30 403 billion Riels in 2010 (an average annual increase of 8.40%). The contribution of the agricultural sector to the total GDP decreased from 42.81% in 1998 to 27.33% in 2010, whereas that of the industrial sector increased from 16.77% in 1998 to 26.60% in 2010 (NIS 2012). These trends show that, as of 2010, even though the contribution of the agriculture sector to the total GDP was decreasing, it was still larger than that of the industrial sector.
The main objectives of ELC in Cambodia are to develop intensive agricultural and industrial-agricultural activities, develop the land in an appropriate and sustainable manner, increase employment and generate state, provincial, or communal revenues (Ministry of Agriculture, Forestry and Fisheries (MAFF) 2012). However, these objectives may not be easy to accomplish (Kao et al. 2005). Based on the MAFF website profile of ELC companies, we summarised some basic facts about the implementation of ELC in each province. Implementation of ELC was reflected by a dummy variable in the model fitting. This variable was assigned a value of one if ELC was implemented in the province during that year; otherwise, a value of zero was assigned. The use of this dummy variable made it possible to take the impact of ELC into account in the analysis.
A summary statistics of the above variables were shown in Table 1.
Table 2 shows the test results for the forest area model; the results for the carbon stock model are similar and not shown here. It was found that the FE model is significantly better than pooled model, whereas the RE model is preferred to the FE model by the Hausman test. Because there are only three time points for this dataset, time effects are not analysed here. It was also clear that there was no cross-sectional dependence among provinces by the Pasaran test. The Breusch-Godfrey/ Wooldridge test was implemented to test serial correlation in disturbance terms. It was found that in the FE model, but not in the RE model, there was serial correlation. The last test is for heteroskedasticity. Breusch and Pagan (1979) argued that in the presence of heteroskedastic disturbances, the loss in efficiency may be substantial, the estimated standard errors are biased, and inference may be invalid. In this research, heteroskedasticity is tested using the Studentised Breusch-Pagan test (Breusch & Pagan 1979). Heteroskedastic issues exist for both the FE and RE models. Therefore, computing the robust variance matrix for the estimators becomes necessary to obtain unbiased standard errors.
Table 3 shows the model results with estimates of the intercepts and coefficients of explanatory variables under the RE model; t-values and p-values are calculated by robust estimations of variances. The p-values showed that all the estimates are significant at the 1% level.
The estimates indicated that a population increase of 1 000 people corresponded to an average decrease of 651.3 ha in forest area or 80 600 MgC in carbon stock. An increase in agricultural GVA of one million Riels corresponded to an average decrease of 0.091 ha in forest area or 11.5 MgC in carbon stock. In provinces where ELC was implemented, the forest area was 25 880 ha, or 3.143 TgC (1 TgC = [10.sup.6] MgC) in carbon stock, smaller than those in provinces in which ELC was not implemented. The interaction term estimate indicates that in the provinces in which ELC was implemented, an increase in agricultural GVA of one million Riels corresponded to an average increase of 0.080 ha in forest area or 9.8 MgC in carbon stock. The effect of agricultural GVA was offset by the effect of the interaction between agricultural GVA and ELC, which led to a smaller effect of agricultural sector production.
Validation of the model estimations
To validate the model fitting results, we compared (1) fitted values versus actual observations, (2) residuals (not normalized) versus fitted values, (3) residuals versus population and (4) residuals versus agricultural GVA by vision inspections. Figure 1 is the validation for forest area model estimations. Figure 1 (a) shows that the fitted values were very close to the actual values of forest areas in most cases. Figure 1 (b) shows that the residuals are scattered around the horizontal line of value zero and that no obvious trend is observable. Figures 1 (c) and (d) show some values with high deviations, but neither of these figures show any trends in the residuals. Therefore, the model can be considered to fit the data well. Validation for carbon stock model estimations were not shown here. However, obvious trend was not found either between residuals and fitted values, population and agricultural GVA.
FORECASTING FOREST AREA AND CARBON STOCK IN CAMBODIA
Forecasting using the random effects model
Forecasting is the natural step that follows model fitting. The fitted values of the response variable (here, the forest area and carbon stock) can be obtained using the model parameters and the observed values or actual values of the explanatory variables. When new values for the explanatory variables reflect future conditions in some years, the responses will also have fitted values for the corresponding years. These new fitted values can be called forecasts.
To forecast the value of [y.sup.0] associated with a vector [X.sup.0], the fitted or forecast value of the response variable can be obtained by
[y.sup.0] = [beta]'[X.sup.0] + [[epsilon].sup.0]. (2)
By applying to the Gauss-Markov theorem,
[[??].sup.0] = b'[X.sub.0] (3)
is the minimum variance linear unbiased estimator of E[[y.sup.0]] (Greene 2000).
In the forecasting, [X.sup.0] is the population, agricultural GVA, ELC, and the interaction of agricultural GVA and ELC. The NIS published the results of the 2013 Inter-Censal Population Survey in 2013. As for other years, the growth rates of the population in Cambodia by UN World Population Prospects were introduced (UN 2013). The Population Division of UN provides population estimates and forecasts for various countries and regions until 2100. In the World Population Prospects: For the 2012 Revision, with medium fertility data, the population growth rates for the years 2011 to 2018 are approximately 1.67%, 1.77%, 1.82%, 1.80%, 1.74%, 1.67%, 1.62%, 1.56%. Based on the data for the Inter-Censal Population Survey in 2013 and the above growth rates, provincial populations for 2011 to 2018 were extrapolated.
Agriculture is very important in Cambodia because 78.56% of the total population lives in rural areas, even though this rate decreased from 81.68% in 1998 to 80.49% in 2008 (NIS 2013). As of 2010, the agricultural sector was still larger than the industrial sector. The GDP growth rates in the agricultural sector were 5.5%, 5.0%, 5.7%, 5.4% and 4.0% for the years from 2006 to 2010, respectively (NIS 2012). Although the mean of these five rates is 5.12% and the median is 5.4%, to be conservative for the establishment of RLs in REDD+, the lowest rate, 4.0%, was chosen in estimating the growth in the agricultural GVA for the period from 2011 to 2018.
As for ELC, of the 18 provinces, all the provinces except Pailin had implemented it by 2011. Therefore, in the forecasting, the value of the ELC dummy variable was set as one for all the provinces except Pailin.
It was assumed that in the future, the total forest area and carbon stocks of the provinces that, along with Capital Phnom Penh, were not included in the analysis, would remain stable at the 2010 level, 58 000 ha and 6.517 TgC, respectively, summed by their forest areas and calculated using the average carbon stock data of Sasaki et al. (2013); by further calculation, the total forest carbon stocks for Cambodia in 2010 was obtained as 1.276 PgC (1 PgC = [10.sup.9] MgC). The total forest area and carbon stock forecasts for Cambodia can be obtained by adding the sum of the forecasts for the 18 provinces analysed above and the subtotal quantity of those for Phnom Penh and the other five provinces. These results are shown in Table 4. For Cambodia, the total forest area was forecasted to decrease to 9.944 million ha by 2014 and to 9.510 million ha by 2018. At this rate of deforestation, forest cover for Cambodia would decrease to 55.55% in 2014 and 53.12% in 2018. The total forest carbon stock was forecasted to decrease to 1.224 PgC by 2014 and to 1.170 PgC by 2018. The forecasts suggest that deforestation will continue until 2018, assuming the population growth rates assumed previously and a 4% annual growth rate in the agricultural GVA. By comparing the annual data, it was shown that deforestation was expected to accelerate slightly in the coming years.
The response variable was a random variable in the model fitting. The forecasts shown as in Table 4 were point estimates of means, such as 9.510 million ha in 2018. These point estimates are the expectations of the response conditionally on the assumption of explanatory variables. To show the accuracy of the forecasts, it is necessary to calculate forecast interval.
Greene (2000) gave the calculation of forecast variance as follows:
Var[[e.sup.0]] = Var[[y.sup.0] - [[??].sup.0] ] = [[sigma].sub.2] + Var[([beta] - b)'[X.sup.0]]. (4)
This calculation showed that the forecast variance is sourced from the disturbance term, the estimation of parameters and conditional values of explanatory variables. The forecast interval (FI) can be formed as follows:
FI = [[??].sup.0] [+ or -] [t.sub.[lambda]/2]se([e.sup.0]) (5)
where se stands for the standard error, [t.sub.[lambda]/2] for the value from the t distribution that is exceeded with probability [lambda], and [lambda] for the confidence level. The value of t is 1.96 at the 95% confidence interval. In table 4, forecast intervals with lower and upper limits were calculated at a 95% probability level. The forecast interval here means that the true value of the forest area or carbon stock was expected with 95% confidence to lie within the range given by the above equation.
Forecast intervals tries to find the uncertainty based on the observations. However, uncertainty for carbon stocks also comes from the chosen average carbon stock per hectare. Sasaki et al. (2013) provide average carbon stock data for eight categories of land uses, including five categories of forest, which are convenient for calculating forest carbon stocks; however, their estimates for evergreen forests and deciduous forests are close. Kiyono et al. (2010) and Samreth et al. (2012) provide lower estimates for deciduous forests. Evergreen forests, semi-evergreen forests and deciduous forests share almost 90% of the total forests. Close average carbon stocks for these three categories might cause similar model and forecast results for carbon stocks to forest areas. Now we further fit models and implement forecasts under three new scenarios: decreasing deciduous forest carbon stocks to 90%, 80% and 70%, respectively. Table 5 reported the model estimation results and Table 6 reported forecast results under the three scenarios.
Hausman test statistic is over 0.9 for these three scenarios, therefore, RE model was chosen. It showed that all the explanatory variables are significant by robust estimations, same as those results in the previous section, but the coefficients became smaller and smaller. The quantity of carbon stock forecasts also became smaller after observations of carbon stocks in deciduous forests became smaller. The percentages of decreases for these three scenarios were 4.36-4.38%, 8.73-8.77% and 13.09-13.16% respectively compared with the results in Table 4. However, the trend of decreasing has not changed.
DISCUSSION AND CONCLUSIONS
As noted by UNFCCC in its decisions, different countries have different national circumstances and different drivers of deforestation and forest degradation. Previous studies have shown that per capita income, population or population density, accessibility to forests and road development, among other factors, can be used in forest sector forecasting (e.g., Vieilledent et al. 2013, Verburg et al. 2002, Clark Labs 2013). This research was an effort to produce forecasts of forest areas and carbon stocks for Cambodia for 2011 to 2018 using socio-economic factors by adopting panel data analysis, an econometric approach. In the first step of this research, population, agricultural gross value added and ELC implementations, with the interaction term of agricultural gross value added and ELC implementations, were confirmed to be significant drivers of deforestation for Cambodia, similar to the findings of Michinaka et al. (2013). The result that the agricultural GVA was found to be a significant driver of deforestation shows that, even the land use productivity increases, land expansion and land-grabbing for agricultural production are still important ways to develop agriculture, including large-scale development activities by economic land concessions and social land concessions and small-scale development activities by rural residents for their livelihoods. In the second step, with some assumptions of these drivers, forest areas and forest carbon stocks were forecasted using the models obtained in the first step. These forecasts can be used as a reference for establishing forest reference levels for the REDD+ programme.
Cambodia currently publishes forest cover assessments every four to five years. Since 2002, Cambodia has published data on forest cover in 2002, 2006 and 2010. The next two forest cover assessments are supposed to be conducted in 2014-2015 and 2018. The period of 2011 to 2018 was chosen to be in concord with forest inventory results. Population was used as a factor of influence, which is reasonable for a developing country such as Cambodia. Michinaka and Miyamoto (2013) find that the attitudes of people in a country toward forests change as the Human Development Index (HDI) of the country increases. As Cambodia makes progress in education, health and income, the factors that affect forests and the impacts of those factors will also change.
In this research, the forecast model was built on a national level. Accessibility to forest was not considered in the estimations. An increase in population increases the total demand for forestland. Therefore, land use conversion will also increase without appropriate policy interventions. For a specific plot of forest, the distance to a road or a town or the Capital, Phnom Penh, i.e., accessibility to forest, might be important. However, on a national scale, the land use conversion that happens in one specific place or another does not make a difference in the total deforestation.
The forecasts of forest areas for Cambodia are conditional on assumptions where we only provided one scenario of business as usual. If the population growth rate or agricultural GVA growth rate were higher than assumed, deforestation would be higher.
This research was implemented using data from 2002 to 2010, and the newly issued land allocation to farmers and social economic concession since 2011 was not considered. Due to these facts, the research result should be taken with caution. There is more uncertainty with respect to large-scale plantation development, such as ELC. Due to limitations in data availability, ELC was treated as a dummy variable, and the extent of the implementation of ELC is not well reflected in the model. As the implementation of ELC accelerates, so does deforestation. As for other land use concessions, when they are issued to individuals, their impacts can be reflected by the population factor. Otherwise, the concession issue is also uncertain.
Because the average carbon stocks estimates greatly affect the forest RELs or RLs, further evidence is still needed for estimating accurate carbon stocks. Some research results have dealt with this issue recently. Romijn et al. (2013) argue that different definitions of forest impact estimations of forest RELs. Qureshi et al. (2012) point out that the quantification of forest carbon stock involves high degrees of uncertainty and discrepancies because of the methods used. Uncertainty and the importance of improving the quality of forest carbon stock data were discussed in many literatures (e.g. Ramankutty et al. 2007, Pelletier et al. 2012, Watson et al. 2013). To reflect the uncertainty in average carbon stocks per hectare, especially those for deciduous forests, a sensitivity analysis was undertaken in the research. Even though the significant drivers have not changed but the forecasts were affected; the decrease percentage were designed from 10% to 30% in three scenarios, the forecasts of carbon stocks also decreased further.
This research finds the area of deforestation and loss of carbon stocks will increase slightly by 2014; as the UN population projection provides a smaller population growth rate for Cambodia from 2014, leading to smaller increases in population from 2015, the area of deforestation and loss of carbon stocks will be slightly smaller since then. However, the current general deforestation trend in Cambodia will continue. These forecasts of forest areas and carbon stocks can be used as a reference to establish national forest RLs for implementing the REDD+ Program in Cambodia and for related decision-making with caution. The approach used in this research, namely, applying panel data analysis to provinciallevel data to forecast national forest areas and carbon stocks, can also be used as a reference for other countries in implementing REDD+.
This collaborative research, conducted by the Forestry and Forest Product Research Institute (FFPRI) of Japan and the Forestry Administration (FA) of Cambodia, was supported by the Emergency Project to Develop the Structure for Promoting REDD+ Action, funded by the Forestry Agency of Japan. We gratefully acknowledge the comments of Mr. Chivin Leng of the FA of Cambodia on an earlier version and the comments of two anonymous reviewers. We also thank other FA officials and officials from other ministries for their cooperation in our field trip and data collection.
BARBIER, E.B. and BURGESS, J.C. 2001. The economics of tropical deforestation. Journal of Economic Surveys 15(3): 413-433.
BUONGIORNO, J., ZHU, S. 2013. Using the Global Forest Products Model (GFPM version 2013with BPMPD). STAFF PAPER SERIES #78, Department of Forest and Wildlife Ecology, University of Wisconsin-Madison, USA.
BREUSCH, T.S. and PAGAN, A.R. 1979. A Simple Test for Heteroscedasticity and Random Coefficient Variation. Econometrica 47(5): 1287-1294.
CLARK LABS. 2013. www.clarklabs.org.
CAMBODIA, FAO, UNDP, UNEP. 2010. UN collaborative programme on reducing emissions from deforestation and forest degradation in developing countries national programme document. www.unredd.net/index.php? option=com_docman&task=doc_details&gid=3487&Ite mid=53.
CROISSANT, Y. and MILLO, G. 2008. Panel Data Econometrics in R: The plm Package. Journal of Statistical Software 27(2): 1-43.
Forestry Administration [FA], Ministry of Agriculture, Forestry and Fisheries [MAFF], Cambodia, 2008. Cambodia Forest Cover. 8 p.
Forestry Administration [FA], Ministry of Agriculture, Forestry and Fisheries [MAFF], Cambodia, 2011. Cambodia Forest Cover 2010. 20 p.
GREENE, W.H. 2000. Econometric Analysis, Fourth edition, Prestice Hall International Editions. 1 004 p.
IPCC, 2013. Climate change 2013: the physical science basis. Contribution of Working Group I to the Fifth Assessment Report of Intergovernmental Panel on Climate. [Stocker, T.F., D. Qin, G.-K., Platter, M. Tignor, S.K. Allen, J. Boschung, A. Nauels, Y. Xia, V. Bex, and P.M. Midgley (eds.)] Cambridge, United Kingdom and New York, NY, USA, 1 535 p.
KAO, D., IIDA, S., INOUE, S., OHGA, S. 2005. Assessment of Cambodian forest concession management planning based on criteria and indicators of Montreal Process: a case study of CFC Company. Bulletin of the Kyushu University Forest 86: 69-84.
KIMMINS, H., BLANCO, J.A., BRAD, B., WELHAM, C., SCOULLAR, K. 2010. Forecasting Forest Futures: A hybrid Modelling Approach to the Assessment of Sustainability of Forest Ecosystems and their Values. Earthscan. 281 p.
KIYONO, Y., FURUYA, N., SUM, T., UMEMIYA, C., ITOH, E., ARAKI, M. and MATSUMOTO, M. 2010. Carbon Stock Estimation by Forest Measurement Contributing to Sustainable Forest Management in Cambodia. JARQ 44(1): 81-92.
MICHINAKA, T. and MIYAMOTO, M. 2013. Forests and Human Development: An Analysis of the Socio-Economic Factors Affecting Global Forest Area Changes. Journal of Forest Planning 18(2): 141-150.
MICHINAKA, T., MIYAMOTO, M., YOKOTA, Y., SOKH, H., LAO, S. and MA, V. 2013. Factors Affecting Forest Area Changes in Cambodia: An Econometric Approach. Journal of Sustainable Development 6(5): 12-25.
Ministry of Agricultrue, Forestry and Fisheries [MAFF], Cambodia. 2012. Economic Land Concession. www.elc. maff.gov.kh/en/objective.html.
NGUYEN, V.P. and AZOMAHOU, T. 2007. Nonlinearities and heterogeneity in environmental quality: An empirical analysis of deforestation. Journal of Development Economics 84(1): 291-309.
NIS, Ministry of Planning, Cambodia. 2000. General Population Census of Cambodia 1998. Phnom Penh, Cambodia. 200 p.
NIS, Ministry of Planning, Cambodia. 2009. General Population Census of Cambodia 2008. Phnom Penh, Cambodia. 308 p.
NIS, Ministry of Planning, Cambodia. 2012. Statistical Yearbook of Cambodia 2011. 351 p.
NIS, Ministry of Planning, Cambodia. 2013. Cambodia Inter-Censal Population Survey 2013, Provisional Report. 28 p.
OEUR, S., YIM, S. 2009. Current Status of National Accounts in Cambodia, International Workshop, Beijing, China, 8-10 June 2009 [PowerPoint slides]. http://unstats.un.org.
RAMANKUTTY, N., GIBBS, H.K., ACHARD, F., DEFRIES, R., FOLEY, J.A. and HOUGHTON, R.A. 2007. Challenges to estimating carbon emissions from tropical deforestation. Global Change Biology 13: 51-66.
PELLETIER, J., KIRBY, K.R., POTYIN, C. 2012. Significance of carbon stock uncertainties on emission reductions from deforestation and forest degradation in developing countries. Forest Policy and Economics 24(2012): 3-11.
QURESHI, A., PARIVA, BADOLA, R., HUSSAIN, S.A. 2012. A review of protocols used for assessment of carbon stock in forested landscapes. Environmental Science & Policy 16: 81-89.
R CORE TEAM. 2014. R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. www.R-project.org.
ROMIJN, E., AINEMBABAZI, J.H., WIJAYA, A., HEROLD, M. , ANGELSEN, A., VERCHOT, L., MURDIYARSO, D. 2013. Exploring different forest definitions and their impact on developing REDD+ reference emission levels: A case study for Indonesia. Environmental Science & Policy 33: 246-259.
SAMRETH, V., CHHENG, K., MONDA, Y., KIYONO, Y., TORIYAMA, J., SAITO, S., SAITO, H. and ITO, E. 2012. Tree biomass carbon stock estimation using permanent sampling plot data in different types of seasonal forests in Cambodia. JARQ 46(2): 187-192.
SASAKI, N., CHHENG, K., SOKH, H., KHUN, V., MIZOUE, N. 2013. Carbon emission reductions and removals in Cambodia under the REDD+ scheme. Asia Forest Workshop 2013: Interdisciplinary Discussion on Multiple Impacts of Forestry and Land use Change in Tropical Asia. 3-4 December 2013. Phnom Penh, Cambodia.
United Nations (UN), Department of Economic and Social Affairs, Population Division. 2013. World Population Prospects: The 2012 Revision, DVD Edition. www.esa. un.org/wpp.
United Nations Framework Convention Climate Change (UNFCCC), 2014. www.unfccc.int/.
VERBURG, P.H., SOEPBOER, W., VELDKAMP, A., LIMPIADA, R., ESPALDON, V., MASTURA, S.S. 2002. Modeling the spatial dynamics of regional land use: the CLUE-S model. Environmental Management 30(3): 391-405.
VIEILLEDENT, G., GRINAND, C., VAUDRY, R. 2013. Forecasting deforestation and carbon emissions in tropical developing countries facing demographic expansion: a case study in Madagascar. Ecology and Evolution 3(6): 1702-1716.
WATSON, C., MOURATO, S. and MILNER-GULLAND, E.J. 2013. Uncertain emission reductions from forest conservation: REDD in the Bale Mountains, Ethiopia. Ecology and Society 18(3): 6.
World Bank. (2012). GDP deflator for Cambodia. http://data. worldbank.org/country/cambodia.
ZEILEIS, A. and HOTHORN, T. 2002. Diagnostic Checking in Regression Relationships. R News 2(3): 7-10.
T. MICHINAKA (1), M. MATSUMOTO (1), M. MIYAMOTO (1), Y. YOKOTA (1), H. SOKH (2), S. LAO (2), N. TSUKADA (1), T. MATSUURA (1) and V. MA (2)
(1) Forestry and Forest Products Research Institute (FFPRI), 1 Matsunosato, Tsukuba, Ibaraki, 305-8687, Japan
(2) Forestry Administration (FA), Ministry of Agriculture, Forestry and Fisheries, 40 Preah Norodom Blvd., Phnom Penh, Kingdom of Cambodia
Email: firstname.lastname@example.org, email@example.com, firstname.lastname@example.org, email@example.com, sokhhengpiny@ yahoo.com, firstname.lastname@example.org, email@example.com, firstname.lastname@example.org, email@example.com
TABLE 1 Summary statistics of forest area, carbon stock and other variables Statistics Forest Carbon Population ([10.sup.3] ha) (TgC) ([10.sup.3] pers.) Range 45-1 336 6-166 36-1 710 Median 514 63 352 Mean 596 73 455 Standard deviation 400 50 410 Statistics AGVA ELC ([10.sup.6] Riels) Range 14-1 388 0-1 Median 275 1 Mean 332 0.76 Standard deviation 282 0.43 TABLE 2 Test results for forest area model Test items Methods Results (1) Comparing F test for F = 1 907, models: FE vs individual effects df1 = 17, pooled (alternative df2 = 32, hypothesis: p-value = 0.000 significant effects) (2) Comparing Hausman test chisq = 0.792, models: FE vs RE df = 4, p-value = 0.939 (3) Testing for Pasaran CD test FE: z = 1.055, cross-sectional (alternative p-value = 0.291 dependence or hypothesis: cross- RE: Z = 1 082 contemporaneous sectional p-value = 0.279 correlation dependence) (4) Testing for Breusch-Godfrey/ FE: chisq = 12.351, serial correlation Wooldridge test df = 3, p-value (alternative = 0.006 hypothesis: serial correlation in RE: chisq = 3.933, disturbance terms) df = 3, p-value = 0.268 (5) Testing for studentised Breusch- BP = 43.318, df = heteroskedasticity Pagan test 21, p-value = 0.002 Test items Remarks (1) Comparing FE is better than models: FE vs pooled model pooled (2) Comparing RE is better than FE models: FE vs RE (3) Testing for No cross-sectional cross-sectional dependence dependence or contemporaneous correlation (4) Testing for Serial correlation serial correlation in disturbance terms: Yes in FE, but not in RE (5) Testing for Heteroskedastic heteroskedasticity TABLE 3 Estimation results for the forest area and carbon stock models Variables Estimates t-values p-values Forest area (in ha) model: Intercept 919 170 8.408 0.000 Population (persons) -0.6513 -5.909 0.000 Agricultural GVA (million -0.0912 -4.867 0.000 Riels) ELC -25 880 -2.845 0.006 Interaction of AGVA and ELC 0.0801 4.532 0.000 Forest carbon stock (in 1,000 MgC) model: Intercept 113 570 8.388 0.000 Population (persons) -0.0806 -5.823 0.000 Agricultural GVA (million -0.0115 -4.736 0.000 Riels) ELC -3 143 -2.775 0.008 Interaction of AGVA and ELC 0.0098 4.344 0.000 Note: T-values and p-values were calculated by using robust standard errors. TABLE 4 Forecasts and forecast intervals Year Forest areas (million ha) Lower limits Forecasts Upper limits 2011 6.737 10.276 13.815 2012 6.623 10.169 13.715 2013 6.503 10.057 13.611 2014 6.382 9.944 13.507 2015 6.262 9.833 13.404 2016 6.144 9.723 13.302 2017 6.029 9.616 13.203 2018 5.915 9.510 13.105 Year Carbon stocks (PgC) Lower limits Forecasts Upper limits 2011 0.828 1.266 1.704 2012 0.813 1.252 1.691 2013 0.798 1.238 1.678 2014 0.783 1.224 1.665 2015 0.768 1.210 1.652 2016 0.754 1.197 1.640 2017 0.739 1.183 1.627 2018 0.725 1.170 1.615 Note: Forecast intervals were calculated at a 95% level. TABLE 5 Model estimation results under different deciduous forest carbon stocks Scenario 1 Scenario 2 Scenario 3 Intercept 108 145 102 712 97 272 Population (persons) -0.0760 -0.0714 -0.0668 Agricultural GVA -0.0109 -0.0104 -0.0099 (million Riels) ELC -2917 -2692 -2467 Interaction of AGVA -0.0089 -0.0081 -0.0073 and ELC Note: Robust estimates for ELC are significant at 5% level, while all the others are significant at 1% level. TABLE 6 Forecasts of forest carbon stocks Year Carbon stocks (PgC) Scenario 1 Scenario 2 Scenario 3 2011 1.210 1.155 1.099 2012 1.197 1.143 1.088 2013 1.184 1.130 1.076 2014 1.171 1.117 1.064 2015 1.157 1.104 1.052 2016 1.144 1.092 1.040 2017 1.132 1.080 1.028 2018 1.119 1.068 1.017
|Printer friendly Cite/link Email Feedback|
|Author:||Michinaka, T.; Matsumoto, M.; Miyamoto, M.; Yokota, Y.; Sokh, H.; Lao, S.; Tsukada, N.; Matsuura, T.|
|Publication:||International Forestry Review|
|Date:||Mar 1, 2015|
|Previous Article:||The role of women in early REDD+ implementation: lessons for future engagement/Le role des femmes au debut de la mise en oeuvre de la REDD+:...|
|Next Article:||Policy options for effective REDD+ implementation in Indonesia: the significance of forest tenure reform/Options de politiques pour une mise en...|