Professional Documents
Culture Documents
1
Norwegian Institute for Bioeconomy Research (NIBIO), Division of Forest and Forest Resources. Høgskoleveien 8, 1433 Ås, Norway
2
Scion, Department of Forest Informatics. Sala Street, Private Bag 3020, Rotorua 3046, New Zealand
3
Scion, Department of Forest Informatics, 10 Kyle Street, Riccarton, Christchurch 8011, New Zealand
This study addresses the use of multiple sources of auxiliary data from unmanned aerial vehicles (UAVs) and air-
borne laser scanning (ALS) data for inference on key biophysical parameters in small forest properties (5–300 ha).
We compared the precision of the estimates using plot data alone under a design-based inference with model-
based estimates that include plot data and the following four types of auxiliary data: (1) terrain-independent
variables from UAV photogrammetric data (UAV-SfM); (2) variables obtained from UAV photogrammetric data
normalized using external terrain data (UAV-SfMDTM ); (3) UAV-LS and (4) ALS data. The inclusion of remotely
sensed data increased the precision of DB estimates by factors of 1.5–2.2. The optimal data sources for top
height, stem density, basal area and total stem volume were: UAV-LS, UAV-SfM, UAV-SfMDTM and UAV-SfMDTM .
We conclude that the use of UAV data can increase the precision of stand-level estimates even under intensive
field sampling conditions.
© Institute of Chartered Foresters, 2019. All rights reserved. For Permissions, please e-mail: journals.permissions@oup.com.
1
Forestry
commercial forest settings (Dash et al., 2015; Næsset, 2002; sample is available to fit a model, as in our case study, some of
Watt and Watt, 2013). Point clouds can be generated either the considered MB variance estimators can also be interpreted as
through LS (UAV-LS) (Jaakkola et al., 2010; Wallace et al., 2012) being DB model-assisted (Mandallaz, 2013).
or through the use of imagery captured from digital still cameras The comparison of different auxiliary data types from UAV
(Dandois and Ellis, 2013; Lisein et al., 2013). LS systems emit is clearly important for determining the contribution towards
2
Precision inventory of small-forest properties
Table 1 Description of datasets used to provide explanatory variables to model forest inventory attributes
3
Forestry
Table 2 Summary statistics for the ground reference data for top height (H), number of trees per hectare (N), basal area (G) and total stem volume
(TSV)
system and GNSS receiver. Data capture was completed using UAV photogrammetric data
a manned aircraft system. Campaign settings included a flying Full coverage imagery was acquired using a DJI ZenMuse X3 12
height of 700 m above ground level, pulse rate of 360 kHz and a MP camera (DJI, Shenzhen, China) mounted on a DJI Matrice 600
swath side overlap of 60 per cent. The scanning system operated
UAV. Dense overlapping images were acquired for the generation
in multiple return mode and had an effective laser footprint
of point cloud data using Agisoft PhotoScan (Agisoft, 2017). Flight
size of 0.35 m. The resulting dataset had a pulse density of
altitude was set at 300 m above ground level and image overlap
10.2 pulses/m2 and an on-ground point spacing of 0.31 m. Initial
to 95–90 per cent front and side overlap. Imagery was collected
processing was carried out by the supplier using Terrasolid and
under clear sky conditions within a single flight lasting 20 min.
included ground and noise classification. Subsequent process-
Five ground control points were established on the ground in open
ing was carried out in the LAStools software package (Rapid-
areas and/or canopy gaps and located using a Trimble Geo 7X
Lasso GMBH, Gilching, Germany) and included triangulation of
a terrain model using the ground classified points, further noise differential GNSS system.
removal and normalization of elevations of non-ground returns Photogrammetric processing of the UAV imagery resulted in
to heights above the local ground level. The same variables a dense point cloud (approximately 60 points/m2 ) and an ortho-
that were extracted for the UAV-LS data were extracted for the mosaic with a ground sampling distance of 10 cm. A total of 31
ALS data, including textural variables from the canopy height point cloud explanatory variables were extracted including height
model. percentiles.
4
Precision inventory of small-forest properties
Two sets of variables were calculated from the UAV pho- data and wall-to-wall coverage of UAV-SfM, UAV-SfMDTM , UAV-LS
togrammetric data, namely the UAV-SfM (i.e. DTM-independent) or ALS data. The DB estimate (i.e. using only plot data) was then
and the UAV-SfMDTM variables (from the normalized point used to assess the efficiency of using additional remotely sensed
cloud). The former included a total of 163 explanatory variables data. Within this study, we considered an efficient estimator
extracted from the raw point cloud without any operations to one that had a smaller variance estimate compared to that of
alter the height values (i.e. normalization). The use of these purely field-based estimates, while not showing obvious signs of
variables mainly aims at describing the variations in height and bias. Precision and accuracy for the four biophysical parameters
spectral response of the top of the canopy. Among the most were compared among the possible combinations of data source
relevant DTM-independent variables indicated by Giannetti et al. (UAV-SfM, UAV-SfMDTM , UAV-LS and ALS). The bias of the estima-
(2018) were textural variables calculated based on the digital tors of the mean was assessed by: (1) scrutinizing the model fit
surface model (i.e. top of the canopy), local maxima counts and graphically and by the mean of differences between observed
spectral variables. The UAV-SfMDTM variables included a total of 31 and predicted values; (2) comparing the first two moments of the
variables including height percentiles (p10 , p20 , . . ., p100 ), density distributions (for the sample and for the whole population) of the
variables (d10 , d20 , . . ., d100 ), the mean and standard deviation selected explanatory variables and (3) checking whether the MB
of textural variables calculated on the canopy height model and estimates were within DB confidence intervals. Furthermore, the
the mean of the red, green and blue digital number values (R, G uncertainty of the variance was assessed by comparing analyti-
and B). cal and bootstrap variance estimators.
Methods
Inference
Overview This section describes the methodology used to estimate the
In this study we estimated four forest biophysical parameters μ̂)) for each biophysical variable
mean (μ̂) and variance (Var(
(i.e. H, N, G and TSV) according to MB inference using field plot using both DB and MB inference. In the latter case the variance
5
Forestry
was expressed as the mean squared error (MSE) for a population element is equal to 1/N and X U is a N × (r + 1) matrix of
U composed of j = {1, . . . , N} sampling units. explanatory variables from the remotely sensed auxiliary data.
For the DB estimate we assumed the field data (S) to represent The mean square error (MSE) estimator is:
a simple random sample of size n and a direct estimator of the
mean is:
p μ̂MB = ιU T X U Cov(
β̂ S )X U T ιU
1
n
μ̂DB = yi ((1)) where the subscript p indicates that this estimator incorporates
n
i=1
only the uncertainty of the estimates of the regression parame-
ters. In our case study, where the models are fitted on internal
where yi is equal to the ground reference value of the biophysical data (field plots from within the area of interest) based on a
variable of interest for the ith field plot (i = 1, . . ., n). The DB probability sample, this estimator can also be interpreted as
(μ̂DB )) was estimated assuming simple
variance of the mean (Var being model-assisted (design-based) and suitable for internal
random sampling by the variance estimator: models (Mandallaz,
2013). The covariance matrix of the model
parameters Cov β̂ S is estimated according to
2
μ̂DB = sY
Var ((2))
n
ε̂TS ε̂S T
β̂ S =
Cov XS XS + ((6))
s2Y
where is the sample variance. n−r−1
For the MB inference, we fitted separate multiple linear
regression models for the different combinations of ground where ε̂S is a n-length vector of estimated residuals over the
reference biophysical parameters of interest and explanatory sample S and is a matrix that describes the variance and
variables extracted from the remotely sensed data (UAV- correlation of the residuals (Pinheiro and Bates, 2006, p. 203). The
SfM, UAV-SfMDTM , UAV-LS, or ALS data). Prior to using these gls function of the R-package (R Core Team, 2017) nlme (Pinheiro
models for the estimation of population parameters, their et al., 2019) was used to estimate the model parameters. An esti-
validity was evaluated according to a visual assessment of the mator of the mean squared error (MSE (μ̂MB )) that incorporates
scatterplots of observed vs predicted values. This assessment the uncertainty due to presence of heteroscedasticity and auto-
aimed at identifying potential signs of heteroscedasticity and correlation has been previously adopted by Breidenbach et al.
non-linearity. Further evaluation of the model included the (2016), McRoberts (2006) and McRoberts et al. (2018). Based on
assessment of the model fit (R2 Adj ), root mean square error the formulation by Breidenbach et al. (2016) such an estimator
(RMSE), mean difference (MD) and RMSE and MD values as can be expressed as:
percentage of the observed mean (RMSE% and MD% ). The RMSE
and MD were determined by leave-one-out cross validation as
the square root of the mean of the squared residuals and the prh μ̂MB
MSE = p μ̂MB + ιU T Cov
MSE σ̂ 2 ιU
ε
mean of the residuals, respectively.
N
N
The linear models had the form: = p μ̂MB + 1
MSE σ̂j σ̂k ˆ jk ((7))
N2
j=1 k=1
yS = X S β S + εS , εS ∼ N 0, σε2 W ((3))
σ̂ 2 is the N × N covariance matrix of estimated grid
where Cov
where X S is a n × (r + 1) matrix, r is the number of selected ε
explanatory variables, β S is a column vector of model parameters cell residual variances with the elements σ̂j σ̂k ˆ jk and the sub-
of length r + 1, and ε S is a column vector of random errors with script prhs indicates that the estimator accounts for parameter
zero expectation, of length n. The residual variance is expressed uncertainty, residual heteroscedasticity and spatial correlation.
as the product of the mean squared residual (σε2 ) and a n × In equation (7), the first component is estimated according to
n weight matrix W (Breidenbach et al., 2016). W is an identity equation (5) and incorporates the uncertainty due to model
matrix under homoscedasticity, a diagonal matrix with weights parameter estimation while the second component incorporates
based on a variance model under heteroscedasticity, or a matrix the uncertainty due to residual error. Specifically, σ̂j and σ̂k repre-
where the off-diagonal elements contain weights that result sent the residual variation estimated for the grid cells j = k = {1,
from a model describing the correlation pattern in the residuals . . ., N} while ˆ jk is the estimated correlation between two grid cell
in the case of autocorrelation. predictions j and k due to autocorrelation. See Breidenbach et al.
The population mean of a small forest property is estimated (2016) for more information on the estimation of ˆ jk . For each of
as: the possible combinations of variables of interest and auxiliary
data we fitted generalized least squares (GLS) models with and
without the correlation structure and with and without weights to
μ̂MB = ιU T X U β̂ S ((4))
accommodate heteroscedasticity. An improvement in the model
over OLS regression was defined if the inclusion of autocorrelation
where β̂ S is a vector of estimated coefficients and, as defined and heteroscedasticity in the GLS models resulted in a decrease
by Saarela et al. (2016a) ιU is a vector of length N where each of two AIC points.
6
Precision inventory of small-forest properties
In those cases where homoscedasticity (σ̂j σ̂k = σ̂ε2 ) and null Results
autocorrelation (ˆ jk = 1) are assumed by the model, the estima-
tor in equation (7). reduces to: Models
The visual assessment of the observed against predicted values
using the fitted OLS models (Figure 3) revealed that there was no
we estimated the bootstrapping mean and variance according of biophysical parameters and auxiliary data (see Table 4). This
to: suggested a lack of bias in the estimator or the parameter
estimates for half of the studied combinations. For the remaining
half, the results indicated that for H and N the analytical
SE
1
nboot
estimator over-estimated the MSE ( SE analytic ranging from 1.3 to
bootstrap
μ̂boot = μ̂b ((9))
nboot 1.6). On the contrary, for G and TSV there may have been some
b=1 SE
small underestimation of the MSE ( SE analytic = 0.8). This indicates
bootstrap
that the analytical MSE estimators are generally applicable
nboot
with the utilized linear models. Interestingly, the indication of
1 2
μ̂boot ) =
Var( μ̂b − μ̂boot ((10)) potential presence of bias in the analytical MSE estimators was
nboot − 1 always associated with UAV-SfMDTM data. Because of the small
b=1
7
Forestry
differences between analytical and bootstrapping estimates, noticeable for the H-model using UAV-SfM data and theN-model
only the analytical estimates were used for further comparison. using UAV-LS data (see Figure 4).
The inclusion of RS data increased the precision of population For the same models as in Figure 4 (H-model using UAV-SfM
parameter estimates compared to the DB approach (i.e. field data andN-model using UAV-LS data), the mean estimate was
data alone) in 10 out of the 16 (62 per cent) combinations of either outside or close to the edge of the 95 per cent confidence
response variables and data sources (Table 4). Overall, for 14 intervals of the DB estimate (see Figure 5). Further potential bias
of the 16 combinations of biophysical parameters and auxiliary was noticeable for the MB estimators of G and TSV using UAV-
data, the estimated mean was within the 95 per cent confidence SfM data, for which the estimated means were smaller than
interval of the DB estimate. In contrast to the assessment of the lower bound of the 95 per cent confidence interval for the
the MD (Table 3), which suggested no presence of significant DB estimate (see Figure 5). No differences in the sample and
systematic errors in the model predictions, for some explanatory population distributions of explanatory variables were found for
variables the mean and variance varied substantially between these models.
the sample and the population distributions, thus suggesting a When averaged across the four RS data sources, the model
possible presence of bias of the estimator. This was particularly based estimates resulted in an average RE for H, N, G and
8
Precision inventory of small-forest properties
Table 3 Summary of the developed models. The models that resulted in the most precise model-based estimates for each of the biophysical
parameters are highlighted
an
intensity1Q : number of points within the first quartile of intensity value (see Giannetti et al., 2018), ntrees: average number of local maxima,
crown: average crown size, maxgreen : maximum green value, blue: mean blue value, p10 -p100 : height percentiles, d10 -d100 : density variables, mean,
homogeneity, second moment, dissimilarity, variance: mean textural variables, SDdissimilarity : standard deviation textural variables.
TSV of 1.4, 1.2, 1.2 and 1.5, respectively (see Table 4). When plots (for a total of 47 plots). The largest REs in this study, which
excluding models leading to potentially biased estimators, the were achieved using of remotely sensed auxiliary data, ranged
largest RE for H, N, G and TSV were obtained using the UAV- from 2.1 for total stem volume using UAV-SfMDTM data to 2.2 for
LS (RE = 2.2), UAV-SfM (RE = 1.5), UAV-SfMDTM (RE = 1.6) top height using UAV-LS data. This indicates more than twice
and UAV-SfMDTM data (RE = 2.1), respectively. For N the largest as many field plots would be required to provide an equivalent
RE was actually found for UAV-LS data, however the estimate level of precision to that obtained using auxiliary data from the
was determined as potentially biased (see Figure 4) and thus UAV. In addition to the economic value of these data, UAV-LS
excluded from further comparison. Gains in RE for H were largest in regions with little to no terrain data would offer substantial
using either the UAV-LS (RE = 2.2) or ALS data (RE = 1.5) but additional benefits such as improved information for harvest
non-existent when relying on the UAV photogrammetric data planning, while UAV imagery would offer a high-resolution map
(RE = 1.0). Other than for TSV, UAV-LS resulted in consistently of the area that may further aid in harvest planning and iden-
larger RE than ALS. The use of predictor variables extracted tifying areas of damage or abnormal yield (e.g. wind damage)
from a normalized photogrammetric point cloud (i.e. UAV- that may be missed using field plots alone. Imprecise forest
SfMDTM ) resulted in consistently larger RE than the use of inventory estimates remain a significant problem in the study
DTM-independent variables (i.e. UAV-SfM) for almost all biophysi- context where measurement of a larger sample size can be
cal parameters. Using UAV-SfMDTM data provided marked RE gains deemed too expensive for smaller forests. This can lead to sub-
over UAV-SfM for G (1.6 vs 1.0) and TSV (2.1 vs 1.1) (Table 4). optimal transactions during forest sales or inefficient market-
ing and logistics following harvest. Integrating UAV data in the
manner we have investigated offers higher precision forest inven-
tory estimates alongside valuable novel information for forest
Discussion managers.
Our results highlight the advantages of using UAVs to support Despite these benefits, the gains reported here using MB infer-
inventory data acquisition for small forests, in which it may not ence were relatively small compared to previous research that
be economically feasible to undertake traditional manned aerial has shown an RE of 2.5– 8.0 (Ene et al., 2012; Næsset et al., 2013;
survey. Field plots and UAV data for this study were measured Puliti et al., 2017) for volume using ALS data as the auxiliary vari-
by using a well-established commercial contractor. Using these able. These modest gains in RE reflect the relatively homogenous
costs as a basis for comparison, the cost of acquiring the UAV nature of the sampled stand and the relatively larger sampling
data could have paid for the collection of an additional 17 field intensity of field data compared to previous studies. The smaller
9
Forestry
Table 4 Estimated biophysical parameters according to the studied inferential frameworks (DB and model-based) and data sources in terms of
μ̂DB )) or mean squared error (MSE(
estimated mean (μ̂), variance (Var( μ̂MB )) of the mean, standard error (SE), standard error as a percentage of
μ̂DB ) and MSE(
the estimated mean for case A (SE% ), relative efficiency as a ratio between Var( μ̂MB ). The models that resulted in the most precise MB
estimates for each of the biophysical parameters are highlighted
Figure 4 distribution of the explanatory variables for the sample and for the entire population models for two example models. The first row represents
the H model fitted using UAV-SfMDTM data and the second row the N model fitted using UAV-LS data.
variation in stand biophysical variables, associated with a rather With such limited number of plots the uncertainty of the DB mean
large sampling intensity, resulted in great precision of the DB estimate is likely to increase substantially compared to this study.
estimate and reduced model performance. In such conditions, the MB framework may be attractive as it
In contrast to this study where a large number of field plots enables the use of external data to augment small local field
were measured to obtain a relatively small uncertainty of the data acquisitions for modelling purposes. Within the UAV realm,
field-based estimates, in operational conditions only a portion Puliti et al. (2017) showed how the use of external data in UAV MB
of these would be measured (i.e. one day of work corresponding and hybrid inference can substantially increase the RE. Thus, fur-
to approximately eight field plots in the study forest conditions). ther research should address possibilities to develop large scale
10
Precision inventory of small-forest properties
models using field and UAV data acquired over different areas predictive accuracy, particularly for H, but also to a lesser extent
and apply them to newly acquired UAV data and a small sample for all other variables including TSV. Given that UAV-SfM inven-
of field data. Such a possibility could be particularly interesting tories are not limited to areas where ALS DTMs are available,
for single tree species plantations as it could enable substantial further research should explore the use of this approach for the
reductions in the costs for field data acquisition. estimation of biophysical parameters across a range of forest
The use of UAV photogrammetric data (i.e. UAV-SfM and UAV- types and stand conditions.
SfMDTM ) provided more accurate model predictions than LS data The use of UAV-LS was notable for producing the largest
(i.e. UAV-LS and ALS) for all biophysical parameters and the RE for H and models that were nearly identical to the most
largest RE for all variables butH. This highlights the advantages accurate model obtained from UAV-SfMDTM . This broadly agrees
of fine resolution UAV imagery that can capture both structural with findings by Wallace et al. (2016) who showed UAV-LS to
information through SfM as well as the spectral and textural be more accurate (RMSE = 0.92 m) than UAV-SfMDTM (RMSE =
properties of the forest canopy. The most precise estimates were 1.30 m) for prediction of height. Results from this study extended
found when the SfM data were height normalized to a DTM the comparison between these two UAV data sources to N, G
obtained from ALS data even though for H, we found indica- and TSV which, in contrast to H, showed larger RE for models
tions of bias in the estimators. Giannetti et al. (2018) demon- developed from UAV-SfMDTM than those developed from UAV-
strated how UAV-SfM DTM-independent variables could be used LS data. The UAV-LS sensor used in this study was developed
to model growing stock volume with similar accuracy to ALS data. primarily for automotive applications. Consequently, the laser
Our study confirmed the results by Giannetti et al. (2018) and operates at less power and with larger beam divergence than
extended this approach to a broader range of inventory variables. many survey-grade ALS scanners, which strongly restricts the
Interestingly, for estimating N and discarding the estimator using ability of the laser beam to penetrate through small canopy gaps.
UAV-LS because potentially biased, UAV-SfM data led to the most Furthermore, the UAV-LS scanner used in this study only provided
precise estimates among the remaining alternatives. However, a single return; a multi-return sensor may better characterize
when using UAV-SfM to estimate G and TSV for the entire forest the lower canopy improving the depiction of some biophysical
area, our results revealed that not only the RE was close to one parameters. In terms of our results, the large precision found
but that the estimators also indicated potential negative bias. for H estimates is, therefore, likely to be due to the increased
Thus, it appears that there is need to better understand how probability of resolving actual tree top heights because of the
well DTM-independent variables (i.e. UAV-SfM) perform beyond very large point density. Further research should examine how
assessing their model predictive accuracy. When compared to estimates made using UAV-SfMDTM compare to those from more
UAV-SfMDTM , we found models using UAV-SfM to have reduced advanced survey-grade UAV-LS (e.g. Riegl VUX-1UAV), which offer
11
Forestry
12
Precision inventory of small-forest properties
McRoberts, R.E. 2006 A model-based approach to estimating forest area. Saarela, S., Holm, S., Grafström, A., Schnell, S., Næsset, E., Gregoire, T.G. et
Remote Sens. Environ., 103 (1), 56–66. al. 2016a Hierarchical model-based inference for forest inventory utilizing
McRoberts, R.E., Næsset, E. and Gobakken, T. 2013 Inference for lidar- three sources of information. Ann. Forest Sci., 73 (4), 895-910.
assisted estimation of forest growing stock volume. Remote Sens. Environ. Saarela, S., Schnell, S., Tuominen, S., Balázs, A., Hyyppä, J., Grafström, A.
128, 268–275. et al. 2016b Effects of positional errors in model-assisted and model-
13