Using nonparametric modeling approaches and remote sensing imagery to estimate ecological welfare forest biomass

Chaofan Wu·Hongxiang Tao·Manyu Zhai·Yi Lin·Ke Wang·Jinsong Deng,3·Aihua Shen·Muye Gan·Jun Li·Hong Yang

Introduction

As one of the most important systems covering the land surface on the earth,forest is now recognized to have a great in fluence on the global climate,carbon storage,biodiversity,timber productions as well as ecological services to human beings(Gao and Liu 2012).While deforestation around the world has lasted for a long time,forests were once generally considered to signify the unitary timber supply for economic development worldwide(Li 2004).As a result,serious environmental problems appeared with the decrease of forest ecological ef ficacy,such stands for a broad,signi ficant range of problems across the globe,including flash flood,habitat losses and soil erosion.People began to realize the importance of reforestation and to justify ef ficacious behavior to protect the forest ecosystem for sustainable development(Yu et al.2011;Qi et al.2013).The most representative national forest protection projects in China,Natural Forest Protection Program(NFPP)and Sloping Land Conversion Program(SLCP),demonstrate China’s efforts to make a major contribution to reforestation events around the whole world(Yin and Yin 2010).

Nowadays,multi-objective forest planning is being proposed instead of exclusive timber production and more attention is being paid especially to forest ecological quality.To adapt to the development pattern of the modern forest,forest management philosophy began to change at the end of 1980s in China.The whole forest domain was divided into commercial forest and ecological forest,also named as ecological welfare forest(EWF).The former was mainly identi fied to supply fuel wood for industry construction and the latter was designed to emphasize the protection and improvement of ecological environment(Liao et al.2008).

As forest ecosystems are complex and widely distributed across the country,including tropical monsoon forest,subtropical evergreen broad-leaved forest,temperate deciduous broad-leaved forest and sub-boreal coniferous forest,forest monitoring activities,including biomass estimation,vary with multiple factors.Traditional methods like field investigation,are the most accurate but they are time consuming and prohibitively expensive,especially for complete surveys in mountainous regions.From the perspective of ef ficiency and accommodation,remote sensing offers more opportunities,especially when associated with field inventories,generating forest-related biophysical property descriptions with lower cost,faster speed,wider scope,and improved accuracy(Lu 2006;Lu et al.2014).Remote-sensing data for forestry biophysics have been ongoing since the first Landsat satellite was launched in 1972(P flugmacher et al.2012;Roy et al.2014).

A variety of remote sensing data with increasingly advanced capabilities—including hyper-spectral and high spatial resolution as well as radar and LiDAR(Light Detection and Ranging)sensors—have appeared in recent years(Koch 2010;Lu et al.2014).The performance of different sensors has varied with respect to speci fic purposes,such as which forest structural variables are being examined at multiple scales.Laser altimetry or LiDAR is an alternative way to present the three-dimensional structure of forests for more accurate biophysical measurements(Lim et al.2003).Hyperspectral data,with much more narrow spectral bands instead of corresponding few broad spectral bands,can be analyzed for properties of different forest cover types through improved discrimination of spectral features(Vaglio Laurin et al.2014).The same argument applies to increased spatial resolution imageries,such as IKONOS and QuickBird,that provide more detailed information of the surface features to avoid the mixture problem in lower or moderate resolution images and abundant attributes could be explored(Eckert 2012;Hirata et al.2014).

Biomass estimation,being one of the most important and fundamental components of forest monitoring(Roy and Ravan 1996;Lu et al.2014),is garnering considerable attention.In most cases,allometric functions are used to estimate the biomass(replacing the ancient destructive measurement method)and empirical relationships were built between field investigations based on those calculations and remote-sensing information for biomass estimation(Lu 2006;Main-Knorn et al.2011;Lati fiet al.2015).

To get a better understanding of forest ecosystems,obtaining the biomass at a local scale is still challenging.On the one hand,forest biomass monitoring especially from the perspective of ecologists,is usually limited at the quadrat scale with limited field observations,which limits the interpretation of regional biomass spatial distribution(Návar 2009;Berner et al.2015).On the other hand,the earliest coarse-resolution images,such as MODIS and VAHRR with a spatial resolution from 250 to 1000 m,contain multiple land cover types in one pixel with mixture problems(Hame et al.1997;Zhang and Kondragunta 2006;Du et al.2014).

Landsat,the most frequently used data,stands out for its suitable grain of 30 m,which is approximate to forest management size,especially at the local and regional scales.It is an important data source to enrich previous large-scale biomass estimation research(Zhu and Liu 2015).What’s more,when utilizing the high-resolution or hyper-spectral images,several factors should be taken into consideration:LiDAR data,lower availability and higher cost,as well as the time and complexity of processing highdimension information.Although the medium-resolution Landsat TM has been widely used in monitoring forest biomass(Lu 2005;Labrecque et al.2006;Du et al.2010),the free availability to the public and consecutive upgrading of Landsat archives prompted this study to explore forest biomass estimation on ecological forests in China.

The most widely used method to calculate biomass is to establish empirical models.Several distinct statistical models have been developed to explore the mathematical relationships between field data-based biomass and satellite-derived variables with environmental changes(Lu 2006).Since it is tough to measure biomass below the ground especially for remote sensing,biomass is generally referred to as above-ground biomass(AGB).AGB is estimated by establishing the relationship between field-based calculated biomass and the raw or transformed remotesensing spectral bands through the use of regression methods including linearregression,multiple linear regression,stepwise regression,and ordinary least-squares,as well as the growing use of non-linear or non-parametric modeling methods including support vector machine(SVM)regression(Mountrakis et al.2011;Avitabile et al.2012;Guo et al.2012);K-nearest neighbor(Tian et al.2014),Neural Networks(Lek 1999;Foody et al.2003;Han et al.2013);and random forest(RF)(Powell et al.2010;Dube and Mutanga 2015).These differ from traditional regression models but make no assumptions to the distribution and correlation of the input data,even with a large number of variables.

Fig.1 Study area in Zhejiang Province and the location of filed plots in Fuyang District

Many factors in fluence the distribution of forest biomass.Compared to natural conditions such as topography,temperature and precipitation,human activities and management patterns play increasingly signi ficant roles in the allocation and development of forest resources,especially at the local level(Wang et al.2001).Wang et al.(2013)investigated the direct effect of policy and management execution on spatial–temporal forest biomass change.The map of forest carbon distribution in Zhejiang Province produced by geostatistics and forest inventory indicated similar spatial distribution patterns as having corresponding in fluence of anthropogenic activities(Zhang et al.2012).Yang et al.(2007)inspected the effect of environmental factors(physiognomy,soil depth etc.)on major ecological forest biomass in Zhejiang and found that anthropogenic activities such as forest protections increased forest biomass.

The ecological forest policy was started in the middle of 1990s and have been implemented for more than 10 years(Zhang et al.2007).The main objective of this study was to(1)compare the performance of two popular nonparametric methods random forest and support vector machinemodeling approachesforestimating AGB through derived remote-sensing variables,(2)implement the preferable method for predicting AGB in the ecological forest landscape in Fuyang District,Zhejiang Province,China;and(3)to get a better understanding of the ecological forest AGB distribution linked to local management patterns for further sustainable-development decisions.

Materials and methods

Study area

The study area is located between latitudes 29°44′45′–30°11′58.5′N and longitude 119°25′–120°19.5′E in the northern part of central Zhejiang Province(see Fig.1).The Fuchun River runs across the whole county with about 70%of the land in the region covered by forest.The study area is characterized as subtropical monsoon climate with an annual average precipitation of 1477.9 mm and mean temperature of 16.7°C.The dominant forest types include coniferous forest,broadleaf forest,mixed forest,bamboo,and shrub.As the primeval forests were almost exhausted because of historic consumption,most of the current forests are secondary vegetation.The ecological forests delimited by the government are distributed in the geographic positions of high ecological signi ficance and low ecological vulnerability,including river headwaters,large reservoirs,natural reserves and scenic spots.

Field data

The field samples in our study are part of the provincial ecological forest data from Zhejiang Forestry Academy.It was obtained between 2008 and 2010 during the growing seasons by utilizing a strati fied random sampling scheme across the whole province.The properties of ecological forest subplot distributions within each county were also take into account.The allocation of hierarchical plots is based on a comprehensive consideration of geographical conditions including climate zone,tree species,age composition,and site characteristics.The size of each plot was 20 m × 20 m,and along each plot’s diagonal line,three fixed quadrats were set to 2 m×2 m to investigate the property of both shrub and grass.

Fundamental parameters including diameter at breast height(DBH)for tree,height as well as the predominant tree species were measured in each plot.AGB of individual tree was calculated using different allometric functions according to the differentspeciesand the detailed descriptions of biomass calculation can be found in(Yuan et al.2009).The final AGB is the sum of all individual trees,herbs and shrub AGB within each plot.In total,87 field-measured plots with the mean DBH ranging from 5.7 to 19.56 cm and the mean height from 3.71 to 13.08 m located in the study area were selected as candidate sample plots,and the mean aboveground biomass of plots ranged from 8.05 to 193.26 ton/ha.

Data pre-processing

The major limitation of selecting proper satellite imagery is obvious—clouds over the study area.We used the cloudfree Landsat 5 TM image Path/row 119/39(L1T product in UTM coordinates)in May 24,2010 from the United States Geological Survey(USGS)(http://earthexplorer.usgs.gov/).Six multispectral bands(except the thermal band eliminated because of its coarser spatial resolution)were preprocessed by referring the topographic maps with 1:10,000 scale to conduct the geometric correction.The total root mean square error was about 0.5 pixels.

After the radiometric calibration,atmospheric correction was implemented by converting the at-sensor spectral radiance into re flectance at the top of atmosphere,using the FLAASH(Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes)in the ENVI software.With information provided in the metadata file,input parameters for the atmospheric correction module included sensor type,elevation, flight date and time,atmospheric model,aerosol mode,initial visibility(Yuan and Niu 2008).The cosine correction method was adopted for topographic correction to alleviate the in fluence of mountainous terrain using the same resolution DEM data(Teillet et al.1982;Han et al.2013).The ecological forest map of the study area was used as reference for AGB estimation within the ecological forest.To get a better understanding of the general consciousness of AGB distribution in Fuyang District,the map of land use and cover from the local government was used as the basic imagery for forest extraction.

Table 1 Vegetation index illustration

To investigate the relationship between biomass and remote-sensing data,numerous spectral vegetation indices have been developed as they were proved to supply more information than the individual spectral bands for forest biophysical properties studying(Foody et al.2003;Zheng et al.2004;Powell et al.2010).For example,ration vegetation index(RVI)has been widely applied because it is simple but effective.Normalized difference vegetation index(NDVI)is the most widely known vegetation index for its excellent explanation of the different chlorophyll properties between the red and near-infrared spectral bands.Besides,normalized difference water index(NDWI)is in direct proportion to the water content of vegetation(Gao 1996).Other frequently used vegetation indices including soil-adjusted vegetation index(SAVI)(Baret and Guyot 1991)and visible atmospherically resistant index(VARI)(Stow et al.2005).In addition,tasseled cap(TC)transformation,a linear combination of the six multispectral bands,was calculated.The TC derived indices,viz.brightness,greenness,and wetness(Crist and Cicone 1984)as well as tasseled cap angle(TCA)(Powell et al.2010)and tasseled cap distance(TCD)(Duane et al.2010)were generated as additional potential variables.Table 1 shows the formulas of these vegetation indices.

Texture information previously explored also showed improvement for biomass estimation in addition to the original TM spectral bands(Lu 2005;Kelsey and Neff 2014).Among the different calculation approaches,Grey-Level Co-occurrence matrix(GLCM)is the most widely used method to extract texture features.Therefore,eight GLCM texture measuresincluding mean,variance,homogeneity,contrast,dissimilarity,entropy,second moment and correlation of the first band from principle component transformation were calculated using different window sizes of 3×3,5×5,and 7×7,respectively.

In summary,six spectral bands from the visible to shortwave infrared were adopted as variables in biomass estimates;other potential remote-sensing variables included:ancillary topographic variables:DEM,descended slope and aspect,tasseled cap indices,vegetation indices,the first three principal components,and texture measures.All the values were extracted through the mean values of a window size of 3×3 pixels from the corresponding plot locations in the images to establish the relationship between AGB and potential variables.The field-measured AGB was treated as the predictor variable and the remote-sensing derived features were treated as the explanatory variables.

Empirical modeling approaches

Although numbersofempiricalmodelshave been employed to explore the relationship between remotesensing variablesand field-observed measured AGB(Labrecque et al.2006),no model is recognized as the most ef ficient than the others because the accuracy of models depends on multitudinous factors,such as the processing of satellite imagery,the adjustment of parameters during the modeling procedure,or the diversity of study areas,especially in China,where the complex forest system is in fluenced by vigorous human activities and natural environmental changes in different regional development modes.

Simple linear regression is often used as the basic,most straightforward method to explore the relationship between AGB and remote-sensing variables.However,the relationship in the real word is usually complex especially when multi-dimensional features are added to the original data.Nonparametric methods are more popular than traditional parametric methods as they often perform better when the relationship between the response variables and explanatory variables is nonlinear or more complicated(Xie et al.2009;Tian et al.2014).Two prevalent nonparametric methods,RF regression and SVM regression were adopted to implement AGB estimation.Both methods were found to alleviate the over- fitting problem when the same approach was applied to a new dataset(Breiman 2001;Mountrakis et al.2011).

RF is an ensemble technique combined by multiple individual decision trees developed by Breiman(2001)both for classi fication and regression.Randomly selecting variables and samples of the dataset,producing a collection of ‘un-pruned’regression trees and estimating the errors by the remaining unselected data(also known as ‘out-ofbag’or OOB)data make the random forest result in increased precision of predictions(Dube and Mutanga 2015).There are two main important parameters:mtry,which means the number of features to split the nodes,whose default value usually is one third of the total number of variables for regression;and ntree,which means the number of trees,to be optimized in the modeling process depending on speci fic application objectives(Mutanga et al.2012).Another advantage of this approach is the built-in feature selection.The importance of each variable is calculated during the modeling process,and variables of the least signi ficance are gradually eliminated until the regression reaches the minimum error.

SVM is another popular machine-learning method that solves the multi-dimension prediction problem with a linear solution by projecting the multiple variables into a higher-dimensional feature space(Chang and Lin 2011).Input variables are used to form the multidimensional space,and hyperplanes are generated depending on the support vectors to separate different groups,while each group is aggregated with similar sample plots.Advanced kernel functions are applied to transform the non-linear feature space to achieve the minimum training error and maximally separate different groups(Mountrakis et al.2011).An advantage of this approach is that SVM can achieve high estimated predictions even with small numbers of training data,following the principle of structural risk minimization.The most important parameters include the kernel parameter,precision parameter ‘sigma,’and penalty ‘cost’.

Both modeling approaches were conducted through the R statistical software(http://www.r-project.org/):RF in the‘randomForest’package and SVM in the e1701 package.

Accuracy assessment

Taking the limited number of samples into consideration and to compare the accuracy of two empirical models,leave-one-out(LOO)crossvalidation wasemployed instead of the traditional method,which divides the dataset into two parts,respectively,for training and validation,it possesses higher representativeness and lower uncertainty(Du et al.2012).During the LOO procedure,training samples(all the samples with one removed)are used to optimize the modeling parameters and train the modeling method,while the removed one is used to verify the accuracy of the modeling method until all the data have been traversed through validation.Three measures were implemented to assess the accuracy of the models:root mean square error(RMSE),the coef ficient of determination(R2)between the predicted and field measured AGB values,where a value of smaller RMSE and higher R2(the ability to explain the total variation of models)demonstrated a better performance(Zhu and Liu 2015).

Results and discussion

Correlation analysis

Data pre-processing is prerequisite for developing AGB estimation models.The 3σ(standard deviation)measurement was adopted to remove the field measured AGB outliers.Plots located in relatively homogenous areas werechosen.As a result,a total of 82 plot samples were selected for the subsequent processing.Table 2 showed part of the correlation coef ficient between AGB and remote-sensing variables,among the correlations,topographic variables,and most texture features were weakly correlated with AGB,while the multi-spectral bands and vegetation indices were more closely correlated with AGB.Although many researchers have found that NDVI has a high correlation with biomass,the correlation coef ficient between NDVI and AGB in our study area was not signi ficant,the reason may due to the saturation problem(Zhu and Liu 2015).Without any assumption about particular probability density distribution for the input variables,the nonparametricbased modeling approaches were explored in the present study.

Table 2 The correlation between above ground biomass(AGB)and the input variables

Parameter optimization

Besides the variable selection,different combinations of parameters greatly in fluence the accuracy of modeling prediction(Powell et al.2010).Optimizing key parameters is another important procedure to assure more accurate biomass estimation performance of the two non-parametric methods.All the parameters were tuned to search for the most appropriate value to decrease the errors during the LOO cross validation process.

For the RF regression model,considering the randomness of feature and sample selection,the modeling process was conducted repeatedly for 30 times.In this research,the ntree parameter was set to 200,500,800,1000,and 1500 for the OOB error convergence,and the mtry was set as the default value,viz.one third of the total number of the input variables.As a result,the selected optimal number of ntree was 1000 with the mtry set to 5.In addition,we investigated variables of the most importance ranked by random forest,and the top six features were band7,band5,TC.3,NDWI,R54T_7.1,and TC.1.

For the SVM regression approach,although linear kernel function,polynomial kernel function,radial basis function(RBF)and sigmoid kernel function were alternatives for modeling,the radial basis function(RBF)kernel was selected for the SVM modeling approach.Previous studies showed that it is the most effective option for forest biophysical estimation(Mountrakis et al.2011;Guo et al.2012),and required the least parameters:cost and sigma.In the present study,cost was set in the range(0.01,0.1,1.5,10,20,and 30)and sigma was in the interval(0.001,0.01,0.1,1,10,100,and 1000).Finally,the optimal values of cost and sigma were set to 20 and 0.001 with the smallest RMSE tested by trial and error.

Fig.2 Scatter plots of measured biomass versus predicted biomass using support vector machine and random forest modeling approaches

Comparison of different regression approaches

When comparing the results of two modeling methods with LOO cross validation,both modeling methods proved to be signi ficant with P value below 0.001,implying that they were immune to the over- fitting problem.Nevertheless,the results in Fig.2 indicated that random forest achieved more satisfactory predictions with obvious higher R2(0.92)and lower RMSE (9.83 ton/ha).The highest values were underestimated and the lowest values were overestimated by the SVM approach,while random forest relieved the deviation phenomenon,which proved the same conclusion as some previous researchers,that RF obtained excellent performance when compared to other methods(Zhang et al.2014).

Compared to previous biomass estimation research,the results in this study showed relatively satisfactory performance in terms of R2and RMSE.Ji et al.(2012)estimated the AGB in the Yukon Flats ecoregion using Landsat data,explaining 73%of the variance in the observed AGB.Du et al.(2010)utilized Landsat images for bamboo AGB prediction with multiple-linear regression,achieving an R2of 0.13.Dube and Mutanga(2015)testi fied the newly launched Landsat 8 for speci fic species’AGB estimates,and found that the best result was for Eucalyptus dunii with an R2of 0.71 and RMSE of 10.66 ton/ha.Kelsey and Neff(2014)predicted the biomass map in southwest Colorado with the highestcorrelation between predicted and observed biomass values,0.86.Ultimately,RF was chosen as the final modeling approach to generate AGB estimation map for analyzing forest distribution characteristics in the study zone.

AGB prediction map analysis

Since 2000,the government in Zhejiang Province has implemented a number of measures on ecology related policies to protect the local forest ecosystem(Tang et al.2009).The ecological forest was initiated to satisfy ecological and social demand for human beings with priority given toecologicalservicesover timberproductionandtopromote sustainable development of the environment.The scope of ecological forest was delimited arti ficially,according to the location of local forest,especially the signi ficant ecologic niches with eco-environmental vulnerability.

The distribution of estimated AGB in Fuyang District within the ecological forest and the whole forest,respectively can be found in Figs.3 and 4.The average AGB value in the ecological forest and the whole forest was 86.96 and 79.76 ton/ha,with a proximal standard deviation of 26 ton/ha.This implies that the ecological forest achieved a more satisfactory performance for ecological development.At the same time,the estimated AGB map in the current ordinary forest offered potential areas for ecology-related decisions in the future.

Fig.3 The distribution of estimated AGB map in the ecological forest in Fuyang District

The area with highest AGB value was mainly located alongside the Fuchun River(see the predicted map in Fig.4),which runs across the central part in Fuyang District,and the lowest AGB was scattered mostly in the northwest part and some southeast sections on the whole.The estimated AGB distribution was in accordance to the local government’s decision that the area should be into three parts for different development patterns.In the middle part,run through by the river,the main function of forest was determined as ecological landscape.While in the northwest and southeast,the principal forest was de fined as special commercial forest with a consideration for timber supply in some way and protective forest,respectively.

Within the ecological forest,when referring to the subcompartment information(the minimal unit for forest inventory in China)about the tree species map in Fig.5,the estimated AGB map obviously showed that there were distinct biomass differences between different tree species.The main forest species in the ecological zone included broadleaf(accounted for 42%),coniferous(39%)and mixed forest(17%),white the coniferous forest achieved a higher mean value of AGB 94.42 ton/ha than the broadleaf forest’s 80.42 ton/ha and the mixed forest’s 87.53 ton/ha.

Although measures have been taken by the government to promote afforestation and reforestation,the huge historic forest consumption resulted in a low biomass density,especially for the broadleaf forest,which is signi ficantly in fluenced by the immature age structure.Despite the fact that the dominant forest species in the county is still the coniferous forest,previous researchers have pointed out that with the increase of age,the broadleaf forest will be the dominant carbon reserve in the whole province(Liu et al.2005).

Fig.4 The distribution of estimated AGB map in the whole forest in Fuyang District:a Fuchun Taoyuan Scenic Spot,b Wuchao Mountain National Forest Park and Ling Mountain Scenic Spot,c Longmen Mountain Forest Park

Fig.5 The sub-compartment measured ecological forest map in Fuyang District

For the distinct spatial variability,the distribution characteristics in the estimated AGB map were further investigated.As showed in Fig.4,we found that the biomass was signi ficantly higher when the forest were set as scenic spots and forest parks.Figure 5 represented the(a)Fuchun Taoyuan Scenic Spot,(b)Wuchao Mountain National Forest Park and Ling Mountain Scenic Spot,and(c)Longmen Mountain Forest Park.The highest AGB value was 143.67,144.03,and 140.84 ton/ha,respectively,which demonstrates that the implementation of environmental protection measures will be bene ficial to the high productivity of forest resources,re flecting that biomass is one of the most important measures of satisfactory forest quality.

Forest ecosystems,especially in the protected areas,are always greatly in fluenced by anthropomorphic activities,such as forest management.Although the policy of ecological forest conservation has been promulgated for over 10 years,the scienti fic effects of this policy for local forest conditions remains to be further investigated.The above analysis connecting the AGB distribution map with local forest management pattern shows that combining the remote-sensing imagery with field-measured data could provide an ef ficient way to inspect the countermeasures of categorized forest management implemented to promote sustainable development.

Conclusion

Ef ficient quanti fication of biomass estimation is the foundation to get a better understanding of local forest carbon storage and exchange.This study mainly compared two popular nonparametric modeling approaches to estimate the aboveground biomass in the forest area with an emphasis on the ecological forest.Considering estimation uncertainties in the modeling process including feature selection,parameters optimization and evaluation criteria of accuracy measures,experiments were conducted to figure out the most optimal combination for our study area.The results demonstrated that with limited field-measured sample plots,the nonparametric approach RF achieved a more satisfactory result for AGB estimation compared to SM in the same scheme.The same modeling process was recommended for studies with similar conditions.The distribution difference of predicted AGB map in the forest area,to a great extent,re flected the pattern of forest management in Fuyang District,demonstrating that associating remote sensing with field-measured biophysical properties could make for valuable instructions for sustainable forest management decisions.

Some possible solutions for further work on better performance of AGB estimation as the supplement to the present study are as follows:

The Landsat pixel and the field plots were not matched exactly because of the different resolutions(also called the‘scale issue’),which is a common problem when connecting remote sensing data with field-investigation data that could impact the estimated result(Ohmann et al.2014).

Although multi-spectral bands,vegetation indices and texture measures were derived from the Landsat TM data,the accuracy has great potential to be improved.Some researchers have pointed out that using multi-temporal imagery could alleviate saturation problem and to get better prediction performance to some extent(Powell et al.2010;Zhu and Liu 2015).In addition,an increase in the number of limited sample plots in our study area could promote separate modeling processes to be executed,according to the distinct tree species.

From this perspective,thorough,traditional field inventory is still the most important base for building reliable relationships between biomass and remote-sensing variables.

AcknowledgementsThe authors are thankful to the USGS and NASA for the open archives of Landsat imagery,and would like to give sincere thanks to the R Development Team for the open-source package for statistical analysis.The authors also thank the Editor and anonymous reviewers for their constructive comments,suggestions,and help in enhancing the manuscript.

Avitabile V,Baccini A,Friedl MA,Schmullius C(2012)Capabilities and limitations of Landsat and land cover data for aboveground woody biomass estimation of Uganda.Remote Sens Environ 117:366–380

Baret F,Guyot G(1991)Potentials and limits of vegetation indexes for lai and apar assessment. Remote Sens Environ 35(2–3):161–173

Berner LT,Alexander HD,Loranty MM,Ganzlin P,Mack MC,Davydov SP,Goetz SJ(2015)Biomass allometry for alder,dwarf birch,and willow in boreal forest and tundra ecosystems of far northeastern Siberia and north-central Alaska.For Ecol Manag 337:110–118

Breiman L(2001)Random forests.Mach Learn 45(1):5–32

Chang CC,Lin CJ(2011)LIBSVM:a library for support vector machines.ACM Trans Intell Syst Technol 2(3):1–27

Crist EP,Cicone RC(1984)A physically-based transformation of thematic mapper data—the TM tasseled cap.IEEE Trans Geosci Remote Sens 22(3):256–263

Du HQ,Cui RR,Zhou GM,Shi YJ,Xu XJ,Fan WL,LüYL(2010)The responses of Moso bamboo(Phyllostachys heterocycla var.pubescens)forest aboveground biomass to Landsat TM spectral re flectance and NDVI.Acta Ecol Sin 30(5):257–263

Du HQ,Zhou GM,Ge HL,Fan WY,Xu XJ,Fan WL,Shi YJ(2012)Satellite-based carbon stock estimation for bamboo forest with a non-linear partial least square regression technique.Int J Remote Sens 33(6):1917–1933

Du L,Zhou T,Zou ZH,Zhao X,Huang KC,Wu H(2014)Mapping forest biomass using remote sensing and national forest inventory in China.Forests 5(6):1267–1283

Duane MV,Cohen WB,Campbell JL,Hudiburg T,Turner DP,Weyermann DL(2010)Implications of alternative field-sampling designs on Landsat-based mapping of stand age and carbon stocks in Oregon forests.For Sci 56(4):405–416

Dube T,Mutanga O(2015)Evaluating the utility of the mediumspatial resolution Landsat 8 multispectral sensor in quantifying aboveground biomass in uMgeni catchment,South Africa.ISPRS J Photogramm Remote Sens 101:36–46

Eckert S(2012)Improved forest biomass and carbon estimations using texture measures from worldview-2 satellite data.Remote Sens 4(12):810–829

Foody GM,Boyd DS,Cutler MEJ(2003)Predictive relations of tropical forest biomass from Landsat TM data and their transferability between regions. Remote Sens Environ 85(4):463–474

Gao BC(1996)NDWI—a normalized difference water index for remote sensing of vegetation liquid water from space.Remote Sens Environ 58(3):257–266

Gao J,Liu YS(2012)Deforestation in Heilongjiang Province of China,1896–2000:severity,spatiotemporal patterns and causes.Appl Geogr 35(1–2):345–352

Guo Y,Li ZY,Zhang X,Chen EX,Bai LN,Tian X,He QS,Feng Q,Li WM(2012)Optimal support vector machines for forest above-ground biomass estimation from multisource remote sensing data.In:Geoscience and remote sensing symposium(IGARSS),2012 IEEE International,pp 6388–6391

Hame T,Salli A,Andersson K,Lohi A(1997)A new methodology for the estimation of biomass of coniferdominated boreal forest using NOAA AVHRR data. Int J Remote Sens 18(15):3211–3243

Han N,Du HQ,Zhou GM,Xu XJ,Cui RR,Gu CY(2013)Spatiotemporal heterogeneity of Moso bamboo aboveground carbon storage with Landsat Thematic Mapper images:a case study from AnjiCounty,China.IntJ Remote Sens 34(14):4917–4932

Hirata Y,Tabuchi R,Patanaponpaiboon P,Poungparn S,Yoneda R,Fujioka Y (2014)Estimation of aboveground biomass in mangrove forests using high-resolution satellite data.J For Res 19(1):34–41

Ji L,Wylie BK,Nossov DR,Peterson B,Waldrop MP,Mcfarland JW,Rover J,Hollingsworth TN(2012)Estimating aboveground biomass in interior Alaska with Landsat data and field measurements.Int J Appl Earth Obs Geoinf 18:451–461

KelseyK,NeffJ(2014)Estimatesofabovegroundbiomassfromtexture analysis of Landsat imagery.Remote Sens 6(7):6407–6422

Koch B(2010)Status and future of laser scanning,synthetic aperture radar and hyperspectral remote sensing data for forest biomass assessment.ISPRS J Photogramm Remote Sens 65(6):581–590

Labrecque S,Fournier RA,Luther JE,Piercey D(2006)A comparison of four methods to map biomass from Landsat-TM and inventory data in western Newfoundland.For Ecol Manag 226(1–3):129–144

Lati fiH,Fassnacht FE,Hartig F,Berger C,Hernández J,Corvalán P,Koch B(2015)Strati fied aboveground forest biomass estimation by remote sensing data.Int J Appl Earth Obs Geoinf 38:229–241

Lek S,Guégan JF(1999)Arti ficial neural networks as a tool in ecological modelling,an introduction.Ecol Model 120(2–3):65–73

Li WH(2004)Degradation and restoration of forest ecosystems in China.For Ecol Manag 201(1):33–41

Liao YH,Chen HY,Zheng W,Chen CG,Liu LW(2008)Study on community of Fengshui woods and the value of application in the construction of ecological public welfare forest.J Subtrop Resour Environ 3(2):42–48

Lim K,Treitz P,Wulder M,St-Onge B,Flood M(2003)LiDAR remote sensing of forest structure. Prog Phys Geogr 27(1):88–106

Liu QX,Chang J,Jiang B,Yuan WG,Qi L,Zhu JZ,Ge Y(2005)The biomass of the evergreen broad-leaved ecological public-welfare forests in Zhejiang,East China.Acta Ecol Sin 25(9):2139–2144

Lu DS(2005)Aboveground biomass estimation using Landsat TM data in the Brazilian Amazon.Int J Remote Sens 26(12):2509–2525

Lu DS(2006)The potential and challenge of remote sensing-based biomass estimation.Int J Remote Sens 27(7):1297–1328

Lu DS,Chen Q,Wang GX,Liu LJ,Li GY,Moran E(2014)A survey of remote sensing-based aboveground biomass estimation methods in forest ecosystems.Int J Digit Earth 9(1):63–105

Main-Knorn M,Moisen GG,Healey SP,Keeton WS,Freeman EA,Hostert P(2011)Evaluating the remote sensing and inventorybased estimation of biomass in the western Carpathians.Remote Sens 3(12):1427–1446

Mountrakis G,Im J,Ogole C(2011)Support vector machines in remote sensing:a review.ISPRS J Photogramm Remote Sens 66(3):247–259

Mutanga O,Adam E,Cho MA(2012)High density biomass estimation for wetland vegetation using WorldView-2 imagery and random forest regression algorithm.Int J Appl Earth Obs Geoinf 18:399–406

Návar J(2009)Allometric equations for tree species and carbon stocks for forests of northwestern Mexico.For Ecol Manag 257(2):427–434

Ohmann JL,Gregory MJ,Roberts HM(2014)Scale considerations for integrating forest inventory plot data and satellite image data for regional forest mapping.Remote Sens Environ 151:3–15

P flugmacher D,Cohen WB,Kennedy RE(2012)Using Landsatderived disturbance history(1972–2010)to predict current forest structure.Remote Sens Environ 122:146–165

Powell SL,Cohen WB,Healey SP,Kennedy RE,Moisen GG,Pierce KB,Ohmann JL(2010)Quanti fication of live aboveground forest biomass dynamics with Landsat time-series and field inventory data:a comparison of empirical modeling approaches.Remote Sens Environ 114(5):1053–1068

Qi XK,Wang KL,Zhang CH(2013)Effectiveness of ecological restoration projects in a karst region of southwest China assessed using vegetation succession mapping.Ecol Eng 54:245–253

Roy PS,Ravan SA(1996)Biomass estimation using satellite remote sensing data—an investigation on possible approaches for natural forest.J Biosci 21(4):535–561

Roy DP,Wulder MA,Loveland TR,Allen RG,Anderson MC,Helder D,Irons JR,Johnson DM,Kennedy R,Scambos TA(2014)Landsat-8:science and product vision for terrestrial global change research.Remote Sens Environ 145:154–172

Stow D,Niphadkar M,Kaiser J(2005)MODIS-derived visible atmospherically resistant index for monitoring chaparral moisture content.Int J Remote Sens 26(17):3867–3873

Tang LN,Shao GF,Dai LM(2009)Roles of digital technology in China’s sustainable forestry development.Int J Sust Dev World Ecol 16(2):94–101

Teillet PM,Guindon B,Goodenough DG(1982)On the slope-aspect correction of multispectral scanner data.Can J Remote Sens 8(2):84–106

Tian X,Li ZY,Su ZB,Chen EX,van der Tol C,Li X,Guo Y,Li LH,Ling FL (2014)Estimating montane forest above-ground biomass in the upper reaches of the Heihe River Basin using Landsat-TM data.Int J Remote Sens 35(21):7339–7362

Vaglio Laurin G,Chen Q,Lindsell JA,Coomes DA,Frate FD,Guerriero L,Pirotti F,Valentini R(2014)Above ground biomass estimation in an African tropical forest with lidar and hyperspectral data.ISPRS J Photogramm Remote Sens 89(26):49–58

Wang XK,Feng ZW,Ouyang ZY(2001)The impact of human disturbance on vegetative carbon storage in forest ecosystems in China.For Ecol Manag 148(1–3):117–123

Wang XC,Shao GF,Chen H,Lewis BJ,Qi G,Yu DP,Zhou L,Dai LM(2013)An Application of remote sensing data in mapping landscape-level forest biomass for monitoring the effectiveness of forest policies in Northeastern China.Environ Manag 52(3):612–620

Xie YC,Sha ZY,Yu M,Bai YF,Zhang L(2009)A comparison of two models with Landsat data for estimating above ground grassland biomass in Inner Mongolia,China.Ecol Model 220(15):1810–1818

Yang J,Yuan WG,Jiang B,Zhu JR,Wen LN(2007)Effect of environment on biomass of major ecological forest in Zhejiang.J Zhejiang For Sci Technol 27(02):20–23

Yin RS,Yin GP(2010)China’s primary programs of terrestrial ecosystem restoration:initiation,implementation,and challenges.Environ Manag 45(3):429–441

Yu DY,Shi PJ,Han GY,Zhu WQ,Du SQ,Xun B(2011)Forest ecosystem restoration due to a national conservation plan in China.Ecol Eng 37(9):1387–1397

Yuan WG,Jiang B,Ge YJ,Zhu JR,Shen AH(2009)Study on biomass model of key ecological forest in Zhejiang Province.J Zhejiang For Sci Technol 29(2):1–5

Yuan JG,Niu Z(2008)Evaluation of atmospheric correction using FLAASH.In:2008 International workshop on earth observation and remote sensing applications,EORSA,pp 1–6

Zhang F,Du Q,Ge HL,Liu AX,Fu WJ,Ji BY(2012)Spatial distribution of forest carbon in Zhejiang Province with geostatistics based on CFI sample plots.Acta EcolSin 32(16):5275–5286

Zhang J,Ge Y,Chang J,Jiang B,Jiang H,Peng CH,Zhu JR,Yuan WG,Qi LZ,Yu SQ(2007)Carbon storage by ecological service forests in Zhejiang Province,subtropical China.For Ecol Manag 245(1–3):64–75

Zhang J,Huang S,Hogg EH,Lieffers V,Qin Y,He F(2014)Estimating spatial variation in Alberta forest biomass from a combination of forest inventory and remote sensing data.Biogeosciences 11(9):2793–2808

Zhang XY,Kondragunta S(2006)Estimating forest biomass in the USA using generalized allometric models and MODIS land products.Geophys Res Lett 33(9):370–386

Zheng DL,Rademacher J,Chen JQ,Crow T,Bresee M,Le Moine J,Ryu SR(2004)Estimating aboveground biomass using Landsat 7ETM+dataacrossamanagedlandscapeinnorthern Wisconsin,USA.Remote Sens Environ 93(3):402–411

Zhu XL,Liu DS(2015)Improving forest aboveground biomass estimation using seasonal Landsat NDVI time-series.ISPRS J Photogramm Remote Sens 102:222–231