Development and Evaluation of Species-Specific Biomass Models for Most Common Timber and Fuelwood Species of Bangladesh

Allometric biomass models are efficient tools to estimate biomass of trees and forest stands in a non-destructive way. Development of species-specific allometric biomass models requires extensive fieldwork and time. Our study aimed to generate species-specific allometric biomass models for the most common fuelwood and timber species of Bangladesh. We also wanted to evaluate the performances of our models relative to the performances of regional and commonly used pan-tropical biomass models. We used semi-destructive method that incorporates tree-level volume, species-specific biomass expansion factor (BEF), and wood density. We considered four base models, 1) Ln (biomass) = a + bLn (D); 2) Ln (biomass) = a + bLn (H); 3) Ln (Biomass) = a + bLn (D^2H); 4) Ln (Biomass) = a + bLn (D) + cLn (H) to develop spe-cies-specific best-fitted models for Total Above-Ground Biomass (TAGB) and stem biomass. The best-fitted model for each species was selected by the lowest value of Akaike Information Criterion (AIC), Residual Standard Error (RSE) and Root Mean Square Error (RMSE). The derived best-fitted models were then evaluated with respect to regional and pan-tropical models using a separate set of observed data. This evaluation was conducted by computing ME (Model Efficiency) and MPE (Model Prediction Error). The best-fitted allometric biomass models have shown higher model efficiency (0.85 to 0.99 at scale 1) and the lowest model prediction error (−8.94% to 5.27%) compared to the regional and pan-tropical models. All the examined regional and pan-tropical biomass models showed different magnitude of ME and MPE. Some models showed higher level (>0.90 at scale 1) of ME compared to the best-fitted specific species biomass model.


Introduction
Bangladesh has 17.48% of forestland that ecologically can be classified into three types as tropical evergreen and semi-evergreen forest, tropical moist deciduous forest and mangrove forest (FD, 2017). This country has 163.05 million of population with a density of 1116 person/km 2 that ranks 10 th position in the world (World Population Review, 2019). This large population imposes immense pressure on natural resources of this country (Reza & Sharmin, 2016) that also influences the forestry sector through deforestation and degradation of forestlands (FRA, 2000;World Bank, 2016). Recently, Bangladesh has initiated REDD+ activities to reduce emissions of greenhouse gases from deforestation and degradation, conservation and enhancement of forest-carbon stocks, and sustainable management of forests (FD, 2019). Therefore, success of REDD+ activities depends on authenticated data and information on the existing forest areas and their conditions. Forest inventory is an integral part of forest management as it provides data and information on trees and forest resources. A total of 19 forest inventories were conducted in Bangladesh by Bangladesh Forest Department since 1960. Overtime, the objectives of forest inventories have been shifted from a focus on volume for timber resources to biomass for carbon-related values to meet the demand of 21 st century (FD, 2017). Forest can act as a sink and source of carbon (Canadian Forest Service, 2007). Biomass estimation is an important tool for estimating stock and sequestration of carbon in a forested ecosystem (Golley et al., 1975;Vashum & Jayakumar, 2012;Mahmood, 2014). Biomass of trees and forest can be estimated by following destructive and non-destructive methods (Somogyi et al., 2007;Picard et al., 2012;Mahmood et al., 2015;Wakawa, 2016;Mahmood et al., 2017). Allometric biomass models are frequently used to estimate tree and forest biomass (Somogyi et al., 2007). Destructive, semi-destructive and non-destructive methods are followed to derive species-specific, regional and pan-tropical allometric biomass models (Ketterings et al., 2001;Chave et al., 2005Chave et al., , 2014Basuki et al., 2009). Destructive method of biomass model development is more accurate compared to others, but this method is usually discouraged from violating regional and/or national forest management policies (Ketterings et al., 2001).
The multi-species regional and pan-tropical biomass models are commonly used for large-scale biomass estimation (Clark & Kellner, 2012;Mahmood et Brown & Lugo (1992) was used to estimate the biomass stock in forest areas. This estimation included a common wood density value (0.57 t•m −3 ) and a fixed biomass expansion factor (6) (FD, 2007) which may result in uncertainty in biomass estimation by considering a lower wood density and a higher fixed value of biomass expansion factor (Penman et al., 2003). During the year 2009, pan-tropical model of Chave et al. (2005) was also used to estimate the biomass and carbon stock of the Sundarbans of Bangladesh (Rahman et al., 2015). The used pan-tropical model is capable to generalize poorly with its polynomial function that results in implausible relationship among biomass and diameter of trees (Sileshi, 2014). In other ways, pan-tropical biomass model of Brown et al. (1989) and Chave et al. (2005) (2012), Rahman et al. (2015). Numerous studies demonstrated that pan-tropical models generate higher bias in biomass estimation compared to locally developed models (Vieira et al., 2008;Basuki et al., 2009;Kenzo et al., 2009;Ngomanda et al., 2014;Maulana et al., 2016). Therefore, it is recommended to check the bias/ deviation in biomass estimation using the multi-species regional and pan-tropical models for a particular species and forests (Alvarez et al., 2012). Simultaneously, species-specific allometric biomass model may significantly reduce bias in biomass estimation compared to the multi-species regional and commonly used pan-tropical models because they may not able to capture the variability of tree properties (height and diameter at breast height relationship, wood density) caused by ecological and management intervention (Nam et al., 2016;Maulana et al., 2016;Mahmood et al., 2019c). Therefore, this study aimed i) to generate species-specific allometric biomass models for the most common fuelwood and timber species of Bangladesh and ii) to evaluate the performances of the derived best-fitted species-specific models in relation to the performances of multi-species regional and commonly used pan-tropical biomass models.

Description of the Study Area
Sampled trees of this study were collected from the natural patches and plantations of tropical wet evergreen and semi-evergreen forest, tropical moist deciduous forest of Bangladesh during 2018. Bangladesh lies between 20˚34' and 26˚38' north latitude and 88˚01' and 92˚41' east longitude ( Figure 1). The rainfall ranges from 1500 mm in the northwest to 5000 mm in the northeast. Mean monthly maximum temperature is 24˚C to 37˚C, while mean monthly relative humidity found to vary from 63% to 83%. Soil texture is silty loam to clay loam and pH range is 5.5 to 8.3 (Banglapedia, 2014

Sampling of Trees for Biomass Expansion Factor (BEF)
Twenty individuals of each studied species, which yielded 280 sample trees, were felled from the natural patches and plantations of the study areas. The species

Field Measurement and Laboratory Analysis
Total height and DBH of the sampled tree were measured and felled at ground level. The felled trees were separated into leaves, small branches (diameter < 7 cm), bigger branches (diameter > 7 cm) and stem. Species wise fresh weight of these components of individual sampled tree were measured and recorded in the field (Picard et al., 2012;Mahmood et al., 2019aMahmood et al., , 2019bMahmood et al., , 2019c. Ten sub samples (0.25 kg) of leaf, smaller branch, and ten disk of disk of bigger branches and stem) of individual species were taken randomly from the felled trees. These sub-samples were oven-dried at 105˚C until a constant weight to estimate the fresh to oven-dry weight conversion factor. The respective conversion factors were used to estimate the oven-dry weight of individual sampled trees (Mahmood et al., 2019a(Mahmood et al., , 2019b(Mahmood et al., , 2019c. Finally, biomass expansion factor (BEF) of individual sampled trees was calculated from the ratio of Total Above-ground Biomass (TAGB) and oven-dry stem biomass, and species-specific average BEF was derived for further use (Taeroe et al., 2015).

Data Collection and Compilation
This study used stem volume data of 2490 individuals of 14 most common timber and fuelwood species of Bangladesh that were collected from the natural patches and plantations of tropical wet evergreen and semi-evergreen forest and tropical moist deciduous forest. The mean value with ranges of DBH, H and W of the sampled tree species are presented in Table 1. Bangladesh Forest Research Institute collected the stem volume data from the natural patches and plantations of tropical evergreen and semi-evergreen forest, tropical moist deciduous forest of Bangladesh. Stem biomass (kg) of individual sampled tree was estimated from their stem volume (m 3 ) and wood density (W) (kg•m −3 ) value of the respective tree species as derived by Sattar et al. (1999). TAGB of individual trees was estimated from the stem biomass and species-specific mean BEF.

Allometric Model Development and Evaluation
The independent variables (D and H) and dependent variables (Stem biomass and TAGB) were transformed to Ln (natural logarithm) to improve the linearity and homoscedasticity. Tree volume data was collected in two different occasions.
A total of 2490 sample trees (data set A) were selected to derive species-specific allometric biomass model. While, data set B contained 614 individual which was used to validate the derived best-fitted model and comparison with regional and pan-tropical biomass models. We considered four Ln base models, 1) Ln (biomass) = a + bLn (D); 2) Ln (biomass) = a + bLn (H); 3) Ln (Biomass) = a + bLn (D^2H); 4) Ln (Biomass) = a + bLn (D) + cLn (H) to develop species-specific allometric biomass models for TAGB and stem biomass according to (Picard et al., 2012). The best-fitted models were selected based on the lowest Akaike Information Criterion (AIC), Residual Standard Error (RSE) and Root Mean Square Error (RMSE); and highest coefficient of determination (Adjusted R 2 ) values (Sileshi, 2014;Mahmood et al., 2019aMahmood et al., , 2019bMahmood et al., , 2019c. Data were analyzed using R (3.2.3) statistical software. A correction factor (CF) was calculated for each equation to minimize the systematic bias during the back transformation to biomass value (Sprugel, 1983). The derived best-fitted TAGB models were compared and evaluated with the multi-species regional and common pan-tropical models ( Table 2) in terms of Model Efficiency (ME) and Model Prediction Error (MEP) (Mayer & Butler, 1993).

Model Evaluation and Comparison
The model efficiency and model prediction error values of the best-fitted TAGB models of the studies species found to vary from 0.85 to 0.99 (at scale 1) and −8.94% to 5.27% respectively. Lower model efficiency and higher prediction error were observed for S. saman, S. macrophylla and A procera, while the highest model efficiency and lower prediction error were observed for A. mangium, E. camaldulensis and T. grandis (Table 5).
The best-fitted TAGB model of all the studied species except A. procera, L. speciosa and S. saman showed higher performance in biomass estimation compared to the regional and commonly used pan-tropical allometric models in relation to model efficiency and model prediction error values. TAGB model of Chave et al. (2014) and Chave et al. (2005) have appeared as more efficient in biomass estimation of A. procera and L. speciosa respectively. While, TAGB models of Mahmood et al. (2019a) and Chave et al. (2014) can efficiently estimate the biomass of S. saman compared to the derived model (Table 5).

Discussion
Alometric biomass models are important tools to estimate biomass of standing trees and stands (Golley et al., 1975;Basuki et al., 2009) and the accuracy in the estimation depends on model efficiency (Sileshi, 2014). Method of model development, involvement of independent variables and model selection criteria influence the efficiency of allometric biomass model (Sileshi, 2014;Picard et al., 2012). Generally, wood density (W), Diameter at Breast Height (DBH) and total height (H) are considered as independent variables of allometric biomass models (Picard et al., 2012). Wood density of a species varies among the ecoregion (Zanne et al., 2009). This study considered DBH and H as independent variables. But, W was not included because the sample trees were collected from the same ecoregion that likely to have similar wood density for a particular species     (Kusmana et al., 2018). Moreover, W is not recommended as independent variable for species-specific allometric models development due to its lower performance and robustness in use (Njana et al., 2016). Inclusion of more than one independent variable likely to increase the efficiency of allometric models to capture more variabialities (Ketterings et al., 2001;Chave et al., 2005). Our study showed that models with H and DBH have higher efficiency for all the studied species and similar findings were also reported by Rutishauser et al. (2013); Kusmana et al. (2018) and Khushi et al. (2019). Allometric models with single independent variable (DBH) are robust in the field measurement and biomass estimation (Ketterings et al., 2001;Chave et al., 2014;Istrefi et al., 2019). But, DBH as single independent variable has shown lower efficiency in model selection parameters of this study.
Model validation is an important stage to precribe best-fitted allometric biomass model for a group of species or single species (Sileshi, 2014). Different predictive performance (goodness-of-fit) statistics like ME, MPE, Roor Mean Squared Relative Prediction Error, graphical presentation of 1:1 line etc. are followed to evaluate performance of best-fitted models (Makungwa et al., 2013;Sileshi, 2014;Huy et al., 2016;Mahmood et al., 2019aMahmood et al., , 2019bMahmood et al., , 2019c. Best-fitted model of the studied species (except A. procera, S. saman and L. speciosa) showed higher predictive performance in biomass estimation compared to the Open Journal of Forestry regional and pan-tropical models. The variation in estimated biomass may be due to the differences in tree species, climatic conditions, site conditions, forest types with its composition and management practices which ultimately influence the architecture of tree and biomass partitioning (Poorter et al., 2006;Iida et al., 2011;Mugasha et al., 2016;Nam et al., 2016). Development of species-specific allometric models is quite laborious and time-consuming efforts (Picard et al., 2012). Therefore, regional and pan-tropical allometric biomass models are frequently used to estimate biomass of tree species those do not have species-specific model (Chave et al., 2014). Such application of regional or pantropical models may produce higher variation in biomass estimation compared to species specific allometric biomass models (Ketterings et al., 2001;Ngomanda et al., 2014).
However, the species-specific best-fitted allometric models of A. procera, L. speciosa and S. saman in this study showed lower efficiency compared to some regional and pan-tropical models. Therefore, the regional and pan-tropical biomass models can be used to estimate the species-specific biomass prior checking the range of variation generated by using those (Alvarez et al., 2012).

Conclusion
Allometric biomass models with DBH and H showed higher efficiency in model selection parameters for all the studied species. Most of the best-fitted biomass models showed higher model efficiency and lower model prediction error compared to the regional and pan-tropical models. Our study suggests using species-specific allometric models for biomass estimation for higher accuracy. In absence of species-specific models, ME and MPE need to be checked for the regional-and pan-tropical models to reduce uncertainties in large scale biomass estimation.