Stratified aboveground forest biomass estimation by remote sensing data

https://doi.org/10.1016/j.jag.2015.01.016Get rights and content

Highlights

  • We focus on stratification in remote sensing-assisted biomass models.

  • We used dataset based on hyperspectral and LiDAR predictors.

  • Benefits from stratification were assessed in a factorial design with other model choices.

  • The stratification of measurement units was marginally advantageous.

  • Input data type and statistical prediction showed to be most influential on model performances.

Abstract

Remote sensing-assisted estimates of aboveground forest biomass are essential for modeling carbon budgets. It has been suggested that estimates can be improved by building species- or strata-specific biomass models. However, few studies have attempted a systematic analysis of the benefits of such stratification, especially in combination with other factors such as sensor type, statistical prediction method and sampling design of the reference inventory data. We addressed this topic by analyzing the impact of stratifying forest data into three classes (broadleaved, coniferous and mixed forest). We compare predictive accuracy (a) between the strata (b) to a case without stratification for a set of pre-selected predictors from airborne LiDAR and hyperspectral data obtained in a managed mixed forest site in southwestern Germany. We used 5 commonly applied algorithms for biomass predictions on bootstrapped subsamples of the data to obtain cross validated RMSE and r2 diagnostics. Those values were analyzed in a factorial design by an analysis of variance (ANOVA) to rank the relative importance of each factor. Selected models were used for wall-to-wall mapping of biomass estimates and their associated uncertainty. The results revealed marginal advantages for the strata-specific prediction models over the unstratified ones, which were more obvious on the wall-to-wall mapped area-based predictions. Yet further tests are necessary to establish the generality of these results. Input data type and statistical prediction method are concluded to remain the two most crucial factors for the quality of remote sensing-assisted biomass models.

Introduction

The estimation of aboveground forest biomass from remotely-sensed data is currently of great interest, due to important applications ranging from forest management to environmental and climate policy. Forest biomass is directly linked to carbon stocks, which are crucial for establishing future mitigation scenarios under climate change. The importance of forest biomass in the context of such mitigation strategies is demonstrated by international initiatives such as reducing emissions from deforestation and forest degradation (REDD and REDD+) (e.g., Hill et al., 2013). Furthermore, biomass estimates can support surveys assessing the bioenergy potential of certain landscapes and help to monitor the sustainability of forest resources (e.g., Rosillo Calle et al., 2008).

Metrics from light detection and ranging (LiDAR) data have been frequently reported to provide good estimates of aboveground biomass across different geographical units (e.g., Hall et al., 2005, Næsset and Gobakken, 2008, Bright et al., 2012). A possibility to improve predictive accuracy could be including additional information, for example on species composition, in the estimation process. This could be achieved by various techniques. One is combining LiDAR information with optical data, but results have been mixed. Whereas some improvements could be obtained (e.g., Popescu et al., 2004), these were occasionally reported to be only marginal (Kulawardhana et al., 2014), particularly in case of pure deciduous stands (Tonolli et al., 2011). Previous studies using predictors from LiDAR-based biomass models (Packalén and Maltamo, 2006, Packalén and Maltamo, 2007, Breidenbach et al., 2010a, Breidenbach et al., 2010b) show promising results for predicting biomass on species level. Further refinements have been reported by incorporating hyperspectral metrics (e.g., Sarrazin et al., 2011). However, in many cases (e.g., in highly mixed stands) a realistic biomass prediction at tree species level will be severely restricted by factors such as spectral mixture due to tree crown overlaps. In such cases, a coarser division (i.e., post-stratification) into species groups (or communities) or into major strata of coniferous, deciduous and mixed stands is a compromise to retrieve strata-specific estimates (e.g., Eckert, 2012, Latifi et al., 2012). A practical example under which a similar stratification approach is applied is the Forest Inventory and Analysis program of the US, where remote sensing data are used to stratify sample plots from a nation-wide regular grid to subpopulations. The proportionally-allocated samples of each subpopulation are eventually inventoried in the field (e.g., Reams et al., 2005).

A superiority of species (or strata) – specific biomass models to those predicting the entire units at once has been found in a number of previous reports (Breidenbach et al., 2010a, Breidenbach et al., 2010b, Latifi et al., 2012). In case of LiDAR data, this may be related to the differing interactions of the laser pulse signals with the architecture of broadleaved and coniferous trees, as stated by Heurich and Thoma, (2008) who suggested the stratification into deciduous, coniferous and mixed strata for LiDAR-assisted forest parameter estimation.

There are several examples on comparisons between modeling approaches while predicting area-based biomass (e.g., Breidenbach et al., 2010a, Latifi et al., 2010, Powell et al., 2010, Main-Knorn et al., 2015, Gagliasso et al., 2014). However, studies addressing the general issue of post-stratification of the input data for remote sensing-based estimates are still scarce (see Heurich and Thoma, 2008, Dahlke et al., 2013). It has been suggested that classifying inventory plots information to forest types or districts may improve the precision of forest attribute estimation (Reams et al., 2005, Nelson, 2010, Latifi and Koch, 2012), particularly when the aim is to design a multi-level forest inventory for large area estimations (Katila and Tomppo, 2002, Andersen et al., 2011). However, recent reports also state an existing shortage of statistical analysis on post-stratified estimation of forest attributes to be a function of restriction in the sample size in small scale domains (McRoberts et al., 2012), who also provided examples on regional inferences of standing timber volume (McRoberts et al., 2013). Yet in order to draw reliable conclusions on the effect of stratification on forest biomass estimates, stratification approaches are needed to be examined in interaction with several other parameters which are known to influence remote sensing-based biomass estimates (e.g., sensor type, prediction method, sample size).

Here, we explore the question of whether stratification of sampling units into major forest types can influence the predictive quality of area-based forest biomass modeling. We based the models on a number of pre-selected predictors from sets of LiDAR and hyperspectral data. We based the models on a number of pre-selected predictors from sets of LiDAR and hyperspectral data. We did not consider building models based on combined LiDAR and hyperspectral predictors due to the previously-available reports on the fairly similar performance of LiDAR and combined LiDAR + Hyperspectral data for the examined dataset (e.g., Latifi et al., 2012, Fassnacht et al., 2014).

Commonly applied parametric and non-parametric prediction methods were used on bootstrapped subsamples of the data to obtain a relative accuracy measure (RMSE) as well as the degree of variance explained by the models (r2) under cross-validation. Two subsequent analyses of Variance (ANOVA) were used to compare the differences in RMSE and r2 (a) between the strata (b) between the stratified and the non-stratified case with differences in predictive accuracy from other factors (prediction method, input data type and sample size). This allows us to systematical assess the importance of factors which typically occur when modeling stratified forest biomass by means of remote sensing data.

Section snippets

Study site

The study site consists of nearly 900 ha of managed pure and mixed stands located in the vicinity of the southwestern German city of Karlsruhe (8°24′09′′E, 49°03′37′′N to 8°25′49′′E, 49°01′15′′N). The dominant tree species is scots pine (Pinus sylvestris L., with 56.3% of the total timber volume), occurring with other species such as European Beech (Fagus sylvatica L., with 17.8% of the total volume), Sessile Oak (Quercus petraea Liebl.) and Pedunculate Oak (Quercus robur L.) (jointly 14.9% of

Model performances

Fig. 3, Fig. 4, Fig. 5, Fig. 6 summarize the results of model performances. For the strata-specific models, the median r2 values varied between 0.17 (kNN for broadleaves) and 0.48 (RF for coniferous) for the hyperspectral predictors. For the LiDAR predictors they ranged between 0.42 (SVM for coniferous) and 0.6 (RF for coniferous). No explicit trend was observed among the applied modeling approaches. When applying hyperspectral metrics, lower r2 rates were returned by the broadleaved stratum

Discussion

Regional forest inventories are often based on systematic sampling grids, and thus are constrained to previously-defined sampling designs and intensities (McRoberts et al., 2012). Since a stratified sampling can often not be integrated into forest inventory plans prior to the inventory, a post-stratification of the previously- recorded plots into major forest types have been stated to partially enable more precise estimations of forest attributes (Westfall et al., 2011, McRoberts et al., 2012).

Conclusion

In the context of remote sensing-assisted estimation of forest biomass, we examined the importance of sensor type, statistical prediction method, sample size as well as the influence of stratification of the sample units in two experiments. The results lead us to the conclusion that the sensor type (hyperspectral/LiDAR) showed to be the essential source of impact on the yielded predictive performance of the models. This was followed by the effect of prediction method, while sample size turned

Acknowledgements

This study was partly funded by the German Aerospace Center (DLR) and the German Federal Ministry of Economy and Technology based on the Bundestag resolution50EE1025 and 50EE1265-66. The authors would like to acknowledge the valuable suggestions of Prof. Dr. Carsten Dormann concerning the applied methodology and the visualization of the results.

References (64)

  • P. Packalén et al.

    The k-MSN method for the prediction of species-specific stand attributes using airborne laser scanning and aerial photographs

    Remote Sens. Environ.

    (2007)
  • S.L. Powell et al.

    Quantification of live aboveground forest biomass dynamics with Landsat time-series and field inventory data: a comparison of empirical modeling approaches

    Remote Sens. Environ.

    (2010)
  • J.O. Sexton et al.

    Comparison of lidar, radar, and field measurements of canopy height in pine and hardwood forests of southeastern north America

    For. Ecol. Manage.

    (2009)
  • P.S. Thenkabail et al.

    Hyperion, IKONOS, ALI and ETM+ sensors in the study of African reinforests

    Remote Sens. Environ.

    (2004)
  • P.S. Thenkabail et al.

    Accuracy assessments of hyperspectral waveband performance for vegetation analysis applications

    Remote Sens. Environ.

    (2004)
  • S. Tonolli et al.

    Fusion of airborne LiDAR and satellite multispectral data for the estimation of timber volume in the Southern Alps

    Remote Sens. Environ.

    (2011)
  • O.W. Tsui et al.

    Using multi-frequency radar and discrete-return LiDAR measurements to estimate above-ground biomass and biomass components in a coastal temperate forest

    ISPRS J. Photogramm. Remote Sens.

    (2012)
  • X. Yu et al.

    Predicting individual tree attributes from airborne laser point clouds based on the random forests technique

    ISPRS J. Photogramm. Remote Sens.

    (2011)
  • J. Verrelst et al.

    Machine learning regression algorithms for biophysical parameter retrieval: opportunities for Sentinel-2 and -3

    Remote Sens. Environ.

    (2012)
  • G. Mountrakis et al.

    Support vector machines in remote sensing: a review

    ISPRS J. Photogramm. Remote Sens.

    (2011)
  • H.-E. Andersen et al.

    Using multilevel remote sensing and ground data to estimate forest biomass resources in remoteregions: a case study in the boreal forests of interior Alaska

    Can. J. Remote Sens.

    (2011)
  • J. Breidenbach et al.

    Comparison of nearest neighbor approaches for small area estimation of tree species-specific forest inventory attributes in central europe using airborne laser scanner data

    Eur. J. For. Res.

    (2010)
  • B.C. Bright et al.

    Predicting live and dead tree basal area of bark beetle affected forests from discrete-return lidar

    Can. J. Remote Sens.

    (2012)
  • J.M. Chambers et al.
  • W.G. Cochran

    Sampling Techniques

    (1977)
  • T. Cocks et al.

    The HyMap airbornehyperspectral sensor: the system, calibration and performance

  • C. Cortes et al.

    Suppirt-vector networks

    Mach. Learn.

    (1995)
  • M. Dahlke et al.

    Nonparametric endogenous post-stratification estimation

    Stat. Sinnica

    (2013)
  • S. Eckert

    Improved forest biomass and carbon estimations using texture measures from Worldview-2 satellite data

    Remote Sens.

    (2012)
  • F. Fassnacht et al.

    Comparison of feature reduction algorithms for classifying tree-species with hyperspectral data on three central-european test sites

    IEEE J. Select. Top. Appl. Earth Obs. Remote Sens.

    (2014)
  • D. Gagliasso et al.

    A comparison of selected parametric and non-parametric imputation methods for estimating forest biomass and basal area

    Open J. For.

    (2014)
  • M. Heurich et al.

    Estimation of forestry stand parameters using laser scanning data in temperate, structurally rich natural European beech (Fagus sylvatica) and Norway spruce (Picea abies) forests

    Forestry

    (2008)
  • Cited by (0)

    View full text