UEF//eRepository
DSpace https://erepo.uef.fi
Rinnakkaistallenteet Luonnontieteiden ja metsätieteiden tiedekunta
2020
Forest inventories for small areas using drone imagery without in-situ field measurements
Kotivuori, Eetu
Elsevier BV
Tieteelliset aikakauslehtiartikkelit
© 2019 Elsevier Inc.
CC BY-NC-ND https://creativecommons.org/licenses/by-nc-nd/4.0/
http://dx.doi.org/10.1016/j.rse.2019.111404
https://erepo.uef.fi/handle/123456789/24154
Downloaded from University of Eastern Finland's eRepository
1
Forest inventories for small areas using drone imagery without in-situ field measurements
Eetu Kotivuori1*, Mikko Kukkonen2, Lauri Mehtätalo3, Matti Maltamo4, Lauri Korhonen5, Petteri Packalen6
1University of Eastern Finland, School of Forest Sciences, P.O. Box 111, FI-80101 Joensuu, Finland. Email: eetu.kotivuori@uef.fi.
2University of Eastern Finland, School of Forest Sciences, P.O. Box 111, FI-80101 Joensuu, Finland. Email: mikko.kukkonen@uef.fi.
3University of Eastern Finland, School of Computing, P.O. Box 111, FI-80101 Joensuu, Finland. Email: lauri.mehtatalo@uef.fi.
4University of Eastern Finland, School of Forest Sciences, P.O. Box 111, FI-80101 Joensuu, Finland. Email: matti.maltamo@uef.fi.
5University of Eastern Finland, School of Forest Sciences, P.O. Box 111, FI-80101 Joensuu, Finland. Email: lauri.korhonen@uef.fi.
6University of Eastern Finland, School of Forest Sciences, P.O. Box 111, FI-80101 Joensuu, Finland. Email: petteri.packalen@uef.fi.
*Corresponding author.
2 Abstract
1 2
Drone applications are becoming increasingly common in the arena of forest management 3
and forest inventories. In particular, the use of photogrammetrically derived drone-based 4
image point clouds (DIPC) in individual tree detection has become popular. Use of an area- 5
based approach (ABA) in small areas has also been considered. However, in-situ field 6
measurements of sample plots substantially increase the cost of small area forest 7
inventories. Therefore, we examined whether small-scale forest management inventories 8
could be carried out without local field measurements. We used nationwide and regional 9
ABA models for stem volumes fitted with airborne laser scanning (ALS) data to predict stem 10
volumes using corresponding metrics calculated from DIPC data. The stem volumes were 11
predicted at the cell level (15 × 15 m) and aggregated to test plots (30 × 30 m). Height 12
metrics for the dominant tree layer from the DIPC data showed strong correlations with 13
similar metrics computed from the ALS data. The ALS-based models applied with DIPC 14
metrics performed well, especially if the ABA model was fitted in the same geographical area 15
(regional model) and the inventory units were disaggregated to coniferous and deciduous 16
dominated stands using auxiliary information from M 17
data (root mean square error at 30 × 30 m level was 13.1 %). The corresponding root mean 18
square error associated with the nationwide ABA model was 20.0 % with an overestimation 19
(mean difference 9.6 %).
20 21 22
3
Keywords: forest inventory; remote sensing; drone; unmanned aerial system; unmanned 23
aerial vehicle; remotely piloted aerial system; image point cloud; airborne laser scanning;
24
area-based approach.
25 26 27
4 1. Introduction
28 29
In recent years, the use of drones (also known as unmanned aerial systems (UAS), 30
unmanned aerial vehicles (UAV) or remotely piloted aerial systems (RPAS)) in remote sensing 31
has become popular (Colomina & Molina 2014). In particular, the use of drones as a tool for 32
forest inventories, mapping and monitoring has been highlighted (Tang & Shao 2015;
33
Torresan et al. 2017). The potential of drones to acquire auxiliary information for sampling 34
inventories has also been considered (Puliti et al. 2017). From an inventory point of view, 35
drones are used in the prediction of forest attributes (i.e. height, volume and biomass) based 36
on three-dimensional information extracted from aerial images (Lisein et al. 2013; Puliti et 37
al. 2015; Kachamba et al. 2016; Ota et al. 2017) or from drone-LiDAR (light detection and 38
ranging) data (Lin et al. 2011; Wallace et al. 2012; Vepakomma & Cormier 2017). Currently, 39
drone-based individual tree level inventories are studied extensively (e.g. Birdal et al. 2017;
40
Guerra-Hernández et al. 2017; Mohan et al. 2017; Panagiotidis et al. 2017; Alonzo et al.
41
2018; Goldbergs et al. 2018; Surový et al. 2018). In this paper, we use the 42
refer to all portable aerial systems without an onboard pilot.
43 44
Height information from digital aerial images can be extracted by photogrammetric 45
processing of overlapping images (Ullman 1979; Lisein et al. 2013). These so-called drone 46
image point clouds (DIPC) can be used in the same manner as airborne laser scanning (ALS) 47
data using the area-based approach (ABA) (Vauhkonen et al. 2014). In Norwegian forests, 48
Puliti et al. (2015) used height, density, and spectral metrics from DIPC for ABA prediction 49
(e.g. dominant height, basal area and stem volume), while height above ground for DIPC was 50
calculated relative to the position of the ground detected by ALS data from the same area. A 51
5
similar study was conducted by Ota et al. (2017) in Japan. For the prediction of stem 52
volumes, Puliti et al. (2015) and Ota et al. (2017) reported RMSE (root mean square error) 53
values of 14.9 % and 20.0 %, respectively.
54 55
An interesting application for DIPC is the updating of ALS-based inventories in small areas (<
56
100 ha) (Goodbody et al. 2016). For example, the repetition interval between ALS 57
inventories may be > 10 years, during which time the forest inventory data will certainly 58
become outdated. However, due to the small areal coverage of drone data, even the 59
measurement of a small number of sample plots from a drone inventory area can lead to 60
substantial costs per unit area. Also, the measurement of ground control points (GCP), often 61
required for georeferencing, further increases the costs. The optimal situation would be that 62
new field measurements (i.e. field sample plots, ground control points etc.) for a small area 63
inventory would not be needed at all.
64 65
DIPC-based inventories without new field measurements have not been considered in 66
studies published thus far. One option is to apply an ABA model previously fitted with ALS 67
data and field plots. For example, regional or nationwide ABA models from previous 68
inventories could be utilised (Kotivuori et al. 2016; Kotivuori et al. 2018). This scenario is 69
possible, if ABA metrics computed from DIPC data are comparable with the respective ALS 70
metrics. However, there are no published studies on this topic. As DIPC do not contain 71
observations from forest layers that are not visible in images (Lisein et al. 2013), we would 72
expect that the ALS metrics derived from the dominant tree layer ( 73
echo categories) would show the largest correlation with DIPC metrics. The selection of an 74
ALS inventory area from which the model is transferred may also influence the result, 75
6
because the properties of vegetation may differ between the areas, and ALS data from 76
different scanners may display systematic differences in the estimated canopy height 77
metrics (Næsset 2009, Kotivuori et al. 2016).
78 79
The main objective of this study was to assess the performance of drone imagery forest 80
inventories without new in-situ field measurements. The study has two particular aims: 1) 81
determine which of the ABA metrics computed from the DIPC and ALS data are so analogous 82
that they can be used interchangeably with both data sources; 2) to construct ALS-based 83
nationwide and regional ABA models for stem volumes using these metrics, and evaluate the 84
associated error rates when using DIPC metrics instead of ALS metrics in a separate test 85
area. Our nationwide dataset consisted of 22 inventory areas from different parts of Finland.
86
The regional dataset was one of the inventory areas from the nationwide data, which was 87
located close to the test area and had similar forest characteristics than in the test area. The 88
DIPC results were compared with the results of the ALS-based ABA models fitted locally in 89
the test area.
90 91
7 2. Material and methods
92 93
2.1 Nationwide and regional field data 94
95
We used the same nationwide field datasets as Kotivuori et al. (2018) (Figure 1). Our sample 96
plots covered 22 ALS-based large-scale forest management inventory projects from various 97
parts of Finland. Field measurements were conducted between 2011 2015. The plots were 98
located using systematic or random stratified cluster sampling, or systematic sampling with 99
The plot radii were 5.64, 9.00, 12.62 or 12.65 m depending on tree size and 100
inventory area (Kotivuori et al. 2016, Kotivuori et al. 2018). Sample plots located in seedling 101
stands, and sample plots where the stem volume was < 3 m3 ha were omitted. The number 102
of sample plots varied among the inventory areas. Therefore, 301 sample plots were 103
selected from each area, with the exception of two inventory areas where only 233 104
(Savukoski) and 149 (Ilomantsi) plots were available (total of 6402 plots). The 301 sample 105
plots were chosen from larger, inventory area level samples so that the plot volume 106
distribution of an inventory area remained the same (see Kotivuori et al. 2018). Further 107
information regarding the nationwide field data is presented in Kotivuori et al. (2018). The 108
regional models were constructed using the plots from the Sulkava area, which is located 109
about 80 km south-west of the Liperi area where the test plots were located (Figure 1).
110
Sulkava is located close to the Liperi test area and has similar growth conditions and species 111
distributions.
112 113
The field measurement protocol differed slightly among the inventory areas. For example, 114
the minimum DBH (diameter at breast height) varied between areas, so we removed all the 115
8
trees from the dataset where DBH was < 5 cm. In the inventory areas of Tornio, Kolari and 116
Ranua, tree heights were measured for all trees. In the other areas, only the heights of the 117
sample trees were measured. Sample trees were selected from diameter classes that were 118
based on the measured species-specific diameter distribution of the plot. The heights for the 119
remainder of the trees were predicted with height models calibrated by plot (Eerikäinen 120
2009). Tree stem volumes (dm3) were predicted as a function of DBH and height using the 121
species-specific models described in Laasasenaho (1982), and aggregated to the plot level 122
(m3 ha ).
123 124
2.2. Field data for test plots 125
126
A total of 19 test plots were measured in Liperi in summer 2017 (Figure 1). The plot size was 127
30 × 30 m and the plots were all located within the respective forest stands. Test plot 128
locations were selected subjectively so that there would be a large variability in tree species 129
and age classes among the plots. Diameter and height were measured from all trees with a 130
DBH 5 cm. In addition, the location of every tree was determined utilising the method 131
presented in Korpela et al. (2007). The plots were divided into four (15 × 15 m) grid cells, 132
which comprised the basic inventory units (n = 76). The stem volumes of measured trees 133
were predicted as a function of diameter and tree height using the species-specific models 134
developed by Laasasenaho (1982) and then aggregated to 15 × 15 m cells (m3 ha ). Volumes 135
for the 30 × 30 m test plots were estimated as means of the 15 × 15 m cells. Stand attributes 136
for the 30 × 30 m test plots are provided in Appendix (Table A1).
137 138 139
9
140 Base maps: a) Municipality map of Finland / © National Land Survey of Finland 2011, Forest vegetation zones / © Finnish Environmental
141
Institute (SYKE) 2015, Europe coastline / © European Environment Agency 2015, b) General map 1:1 M / © National Land Survey of Finland
142 2018.
143 Figure 1. Liperi test plots (red points) at different geographical scales. a) Nationwide (purple 144
and green points) and regional (green points) sample plots within the forest vegetation 145
zones, b) spatial distribution of test plots in the Liperi test area, and c) grid cells (15 × 15 m) 146
10
inside the test plot (30 × 30 m) with a canopy height model constructed from a drone image 147
point cloud.
148 149
2.3 Airborne laser scanning data for nationwide and regional models 150
151
We used the same nationwide ALS datasets as Kotivuori et al. (2018). Datasets were 152
acquired for the same years as the corresponding field datasets (2011 2015), with the 153
exception of Ilomantsi where field data were collected 1-2 years later than the ALS data.
154
However, the time difference was taken into account when computing the field data using 155
the NFI-based growth modelling service of Luke (Natural Resources Institute Finland).
156
Datasets were acquired in leaf-on conditions. Pulse densities of the datasets were between 157
0.5 1.2 points m-2 and the half scan angles were either 15 or 20 degrees. Both 158
pulse 159
repetition frequencies (PRF) were 114.6 136.5 or 50.0 71.8 kHz, depending on the scanning 160
mode. Flying altitudes were 1675 2200 m. A total of 12 different sensor units by Optech and 161
Leica were used. Further information on scanner parameters of the individual inventory 162
areas are presented in Kotivuori et al. (2018).
163 164
2.4 Airborne laser scanning data for test plots 165
166
The ALS data for the test plots were acquired from Liperi during the summer 2016 (2 10 167
July) in leaf-on conditions. Data were collected with an Optech Titan multispectral LiDAR 168
system, which generates point clouds at the 1550 (channel 1), 1064 (channel 2) and 532 nm 169
(channel 3) wavelengths. In this study, we used channel 2 for ABA. Average pulse density of 170
11
the test plots was 12.4 points m-2. The half scan angle was 20 degrees and flying altitude was 171
850 m above ground level PRF of 250.0 kHz
172
was used.
173 174
2.5 Drone data for test plots 175
176
Drone images were collected from Liperi during 20 June 23 August 2017. Images were 177
collected in either sunny or cloudy weather conditions. Wind conditions varied slightly, but 178
imaging conducted in strong wind conditions was avoided. The images were captured with a 179
DJI Inspire 1 (v. 2.0) rotary-wing drone with Zenmuse x3 digital camera and Sony's image 180
sensor (IMX078CQK). Images were captured plot by plot using three flight lines with a 181
sidelap of 60 % and a forward overlap of 90 %. The centre of the test plot (30 × 30 m) was 182
always located in the middle of the centre flight line. Flying altitude of the drone was 75 m 183
and the nominal GSD (ground sampling distance) was 3.2 cm. Image point clouds were 184
generated with Agisoft Photoscan software (AgiSoft PhotoScan Professional 2017).
185 186
Point clouds were adjusted horizontally and vertically to ensure a better match with the ALS 187
data, and consequently the field data. Horizontal adjustment was required, because 188
georegistration of the DIPC data is inadequate without the use of GCP. In this study, a good 189
match between the DIPC data and field data is needed as the test plots are used for 190
validation. Horizontal adjustment was conducted by shifting each individual image point 191
cloud (i.e. plot) horizontally to ensure a maximum correlation between the DIPC based 192
canopy height model (CHM) and the ALS based CHM.
193 194
12
Vertical adjustment was needed as the ALS DTM (digital terrain model) was used to scale 195
DIPC heights to above ground level, and the z-coordinates of the DIPC do not correspond 196
with the height system in the ALS data (we used Finnish Height System N2000, following the 197
principles of the EVRS 2000). This is quite a common situation with DIPC data when GCPs are 198
not used. Here, the DIPC altitude was corrected by computing the difference between the 199
ground observations of the DIPC and ALS data and then adding the difference to the DIPC 200
data. The ground level of DIPC data were interpolated from points located in open areas.
201
Open areas were determined as locations where the standard deviation of DIPC height was <
202
1 m within a cell size of 2 × 2 m. Because the algorithm assumes that the data are aligned in 203
a horizontal direction, the horizontal adjustment was performed first. The vertical 204
adjustment was done with open ALS data from the National Land Survey of Finland 205
(Airborne LiDAR data / © National Land Survey of Finland 2016).
206 207
After the adjustments, DIPC heights were scaled to above ground level (dZ) using an ALS- 208
based DTM from the National Land Survey of Finland (Airborne LiDAR data / © National Land 209
Survey of Finland 2016). Average point density of the DIPC data was 1148 points m-2. 210
211 212
13 213
Figure 2. (a) Airborne laser scanning (ALS) and (b) drone image point cloud (DIPC) data from 214
the same test plot (30 × 30 m) in the Liperi area. In this figure, point densities are scaled to 215
be similar.
216 217 218
14 2.6 Airborne laser scanning and DIPC metrics 219
220
Image point clouds can only provide reliable height measurements of the upper layer of the 221
canopy, because the lower layers are not visible in the images (Figure 2) (Lisein et al. 2013).
222
Therefore, ALS-metrics were calculated only from the first echoes (= original echo categories 223
and Corresponding metrics for DIPC were calculated using all point 224
observations. Calculated metrics were: maximum (hmax), median (hmed), mean (havg), 225
sample standard deviation (hsd), density percentages (veg0.5/2/5/10/15/20), and height 226
quantiles ( ). The density percentages (%) were calculated by dividing 227
the number of points over a certain height threshold (m) by the total number of points.
228
Percentages were calculated using the quantile function type 7 (Hyndman and Fan 1996) 229
available in the R software (R Core Team 2017). Height thresholds for metrics were not used.
230 231
2.7 Regression models 232
233
We constructed nationwide, regional and local ABA models to predict total stem volume (m3 234
ha ) using the ALS metrics described in section 2.6. The nationwide models were fitted using 235
sample plots from all 22 inventory areas, the regional models using sample plots from the 236
Sulkava inventory area, and the local models using 15 × 15 m cells from the Liperi test area.
237
First, we pre-selected the predictor variables for ABA by comparing the relationships 238
between the ALS and DIPC metrics in the 15 × 15 m cells from Liperi. The metric was 239
included as a candidate predictor variable in the ABA model (Table 1), if:
240
1) the r-squared (r2, where r = Pearson correlation coefficient) value between the ALS and 241
DIPC metric was > 0.9, and 242
15
2) estimates and of the relationship function DIPCmetric ALSmetric, as 243
estimated by OLS (ordinary least squares), were -1.0 1.0 and 0.9 1.1, respectively.
244
To evaluate whether the difference from the 0 1 line could result from random noise, we 245
also conducted an F-test for each DIPC-ALS relationship function for the following null 246
hypothesis: H0: = 0 and = 1.
247 248
Predictor variables were selected with a linear model form using the stem volume or its 249
square root transformation as the response variable. The OLS method was used to estimate 250
the parameters of a linear model. If square root transformation for stem volume was used, a 251
bias correction for the response variable was employed (Lappi 1993). A linear model was 252
used only in the selection of predictor variables; later models were fitted in a nonlinear 253
fashion whereupon bias correction for the response variable is irrelevant.
254 255
A comprehensive investigation was undertaken to select the predictor variables. The 256
variable combination that produced the smallest RMSE value between the observed and 257
predicted values was selected for the model. We also tested the square root, logarithm, 258
polynomial and inverse transformations for the predictor candidates. Kotivuori et al. (2018) 259
noted that the number of degree days (d.d.) (i.e. the effective temperature sum, °C) could be 260
used to describe the changes in forest structures as a function of growing conditions.
261
Therefore, the nationwide model contained three predictor variables (incl. d.d.). Details of 262
the d.d. data used here can be found in Kotivuori et al. (2018) and Ojansuu and Henttonen 263
(1983). Regional and local models contained only two predictor variables (d.d. was not 264
included).
265 266
16
Kotivuori et al. (2018) also noted that the proportion of birch trees (Betula spp.) in a stand is 267
an important variable in nationwide models. For this reason, we constructed alternative 268
nationwide, regional and local models by disaggregating the sample plots (or test plot cells) 269
as coniferous or deciduous dominated, and utilised this information as a dummy variable. In 270
model construction, field data were used to disaggregate plots as coniferous or deciduous 271
dominated.
272 273
The final model was fitted in a nonlinear form 274
275
, 276
277
where is the response variable, the vector of predictors and the residual error for 278
plot i, and includes the parameters of the mean function. The residual variance increased 279
with respect to the prediction, and therefore a power-type variance was 280
eters and h95 is the 95 % height 281
quantile. We used Maximum Likelihood for and nonlinear generalised least squares 282
(NGLS) for and in function gnls of package nlme in the R environment 283
(Pinheiro and Bates 2000; R Core Team 2017; Pinheiro et al. 2018; Mehtätalo & Lappi 2019).
284 285
2.8 Prediction for test plots 286
287
The models and methods described above were used to predict stem volume (m3 ha ) for 288
the 15 × 15 m cells in the Liperi test plots. We used six distinct scenarios:
289 290
17
NatDIPC: Nationwide stem volume (m3 ha ) model applied to cells using metrics calculated 291
from DIPC.
292
RegDIPC: Regional stem volume (m3 ha ) model applied to cells using metrics calculated 293
from DIPC.
294
LocALS: Local stem volume (m3 ha ) applied to 15 × 15 m cells where a model was also 295
fitted with ALS data using local field measurements.
296
NatDIPCD: Nationwide stem volume (m3 ha ) model applied to cells using metrics calculated 297
from DIPC and disaggregation of test plots to coniferous or deciduous dominated. Test plots 298
(30 × 30 m) were disaggregated to coniferous or deciduous dominated using the means of 299
species-specific (pine, spruce and birch) stem volumes taken from 300
Forest Inventory of Finland 2015 maps, which are based on satellite images, NFI 301
field plots, and other geographic information (© Natural Resources Institute Finland 2017).
302
The generation of the described in Tomppo and Halme (2004).
303
RegDIPCD: Regional stem volume (m3 ha ) model applied to cells using metrics calculated 304
from DIPC and disaggregation of test plots to coniferous or deciduous dominated (other 305
steps as above, NatDIPCD).
306
LocALSD: Local stem volume (m3 ha ) applied to 15 × 15 m cells where a model was also 307
fitted. Cells (15 × 15 m) were disaggregated to coniferous or deciduous dominated using 308
observed species- 309
disaggregation to coniferous or deciduous plots and local field measurements.
310 311
Predictions for the 30 × 30 m test plots (m3 ha ) were estimated as means of the 15 × 15 m 312
cells (m3 ha ). Prediction error was assessed at the 30 × 30 m level. Differences between the 313
scenarios were compared using relative (%) RMSE (Equation 1) and MD (Equation 2) values:
314
18 315
where 316
i is the predicted value of attribute y in test plot i, 317
yi is the field-based estimate (observed) of attribute y in test plot i, 318
is the mean of field-based estimates of attribute y in test plot i, and 319
n is the number of sample plots.
320 321
Finally, the 95 % prediction intervals of stand volume were constructed, taking into 322
consideration the uncertainty of the model-based prediction at the level of the 30 × 30 m 323
test plots. Denoting the model-based prediction of mean volume per hectare for the target 324
stand i, based on the four 15 × 15 m cells, by , its variance was estimated as 325
326
327
where the row vector includes the column-specific means of the Jacobian , which 328
includes first derivatives of the assumed model function, evaluated at the parameter 329
estimates (Ståhl et al 2011), is the estimated variance-covariance matrix of 330
parameter estimates (Seber and Wild 1989), N is the number of 15 × 15 cells per test plot 331
19
(always 4 in this study), and is the estimated covariance of the residual errors of 332
the model for cells j and k within the same plot i. The covariance can further be written as 333
334
335
where the estimated variances are obtained from the estimated power-type variance 336
function. Our data did not allow estimation of the spatial correlation of the prediction error 337
because we had only one 30 × 30 m plot per forest stand. Therefore, we used correlation 338
coefficients based on the exponential correlogram of Breidenbach et al. (2016) although 339
they had larger cells size (20m x 20m) than in this study (15m x 15m). Their function gave 340
for adjacent cells with a centre point distance of 15 m and 341
for cells that touch each other in the corner and had a centre point 342
distance of 21.2 m. The confidence intervals of mean volume per hectare were thereafter 343
constructed as;
344 345
346
where is the th quantile of the normal distribution. We used and 347
. 348
349
20 3. Results
350 351
3.1 Comparison of DIPC and ALS metrics 352
353
The DIPC data describes the upper layer of the forest in more detail than the ALS data.
354
However, ALS data contains more observations close to the ground (Figure 2). Metrics 355
computed from the ALS and DIPC data for the 15 × 15 m cells are compared in Table 1 and 356
Figure 3. The upper row in Figure 3 shows two metrics with a strong relationship (r2 > 0.9) 357
and the lower row shows two metrics with a moderate relationship (r2 < 0.9) between the 358
ALS and DIPC data. The point cloud metrics are similar when the r2 value is close to 1 and the 359
estimates of and are close to 0 and 1, respectively. We accepted a metric if (1) r2 was >
360
0.9, (2) was > -1.0 and < 1.0, and (3) was > 0.9 and < 1.1. Because of the different 361
properties of the ALS and DIPC data, only 45 % of the calculated metrics fulfilled these 362
criteria. In particular, metrics that describe the height of the dominant tree layer (e.g. h60- 363
h95, havg, hmax) were similar across ALS and DIPC data. However, height metrics that 364
describe understorey height (e.g. h30), or canopy density metrics with a small threshold 365
value (e.g. d5) were less similar across ALS and DIPC data. The r2 values varied from 0.32 to 366
0.99, with a median of 0.88. Corresponding variations for of and were from -0.60 to 367
17.41 and 0.79 to 1.08, with medians of 1.18 and 0.98, respectively. The hypothesis that = 368
0 and = 1, with a significance level of 0.05, was rejected in 20 cases. It was not rejected 369
for the upper quantiles of the height distributions (h55-h95), which suggests that models 370
with these metrics are transferable between ALS and DIPC.
371 372
21
Table 1. Correlations (r2) between airborne laser scanning (ALS) and drone image point cloud 373
(DIPC) metrics.
374
Metric Definition r2 DIPCmetric ALSmetric Pr(>F)
havg Average height of points 0.91 = 0.1808 + 1.0680x 0.000 hmed Median height of points 0.78 = 1.8318 + 0.9019x 0.024 hstd Standard deviation of point heights 0.92 = -0.3161 + 0.9206x 0.000 hmax Maximum height of points 0.98 = 0.4512 + 1.0044x 0.000 d0.5 Proportion of points above 0.5 meters 0.83 = 17.4122 + 0.8893x 0.000 d2 Proportion of points above 2 meters 0.87 = 6.4357 + 1.0017x 0.000 d5 Proportion of points above 5 meters 0.88 = 3.8167 + 1.0373x 0.000 d10 Proportion of points above 10 meters 0.92 = 3.1848 + 1.0446x 0.000 d15 Proportion of points above 15 meters 0.92 = 0.0684 + 1.0609x 0.022 d20 Proportion of points above 20 meters 0.95 = -0.6004 + 1.0847x 0.011 h5 5th quantile of point heights 0.32 = 1.1839 + 1.0087x 0.000 h10 10th quantile of point heights 0.34 = 2.1064 + 0.9860x 0.000 h15 15th quantile of point heights 0.43 = 2.7248 + 0.8953x 0.000 h20 20th quantile of point heights 0.49 = 3.0713 + 0.8318x 0.000 h25 25th quantile of point heights 0.58 = 3.5343 + 0.8101x 0.000 h30 30th quantile of point heights 0.70 = 3.5384 + 0.8322x 0.000 h35 35th quantile of point heights 0.62 = 3.4116 + 0.7931x 0.000 h40 40th quantile of point heights 0.63 = 3.0534 + 0.8207x 0.000 h45 45th quantile of point heights 0.68 = 2.2581 + 0.8652x 0.022 h50 50th quantile of point heights 0.78 = 1.8318 + 0.9019x 0.024 h55 55th quantile of point heights 0.88 = 1.1662 + 0.9410x 0.063 h60 60th quantile of point heights 0.96 = 0.3071 + 0.9898x 0.341 h65 65th quantile of point heights 0.96 = 0.6888 + 0.9616x 0.134 h70 70th quantile of point heights 0.97 = 0.3832 + 0.9770x 0.467 h75 75th quantile of point heights 0.98 = 0.2758 + 0.9819x 0.513 h80 80th quantile of point heights 0.98 = 0.3196 + 0.9769x 0.196 h85 85th quantile of point heights 0.99 = 0.3646 + 0.9747x 0.102 h90 90th quantile of point heights 0.99 = 0.3253 + 0.9766x 0.063 h95 95th quantile of point heights 0.99 = 0.2331 + 0.9839x 0.256 Note: r2 is the r-squared between the ALS and DIPC metrics, column DIPCmetric
375
ALSmetric presents the relationship functions for corresponding ALS and DIPC metrics and 376
column Pr(>F) the p-value of an F-test of the null hypothesis = 0 and = 1.
377 378
22 379
Figure 3. Comparison of selected airborne laser scanning (ALS) and drone image point cloud 380
(DIPC) metrics in the 15 × 15 m cells. Note: havg = average of point heights, h95 = 95th 381
percentile of point heights, h30 = 30th percentile of point heights, d5 = proportion of points 382
above 5 m.
383 384 385
23 3.2 Nationwide and regional area-based models 386
387
Nationwide ABA models for stem volume (m3 ha ) were:
388 389
(6) 390
(7) 391
392
where decid is 1 in deciduous dominated plots and 0 in coniferous dominated plots.
393
Subscript D denotes that the sample plots are disaggregated to coniferous or deciduous 394
dominated. The RMSE values between observed and predicted values in the training data 395
were 27.7 % and 24.3 % for models 5 and 6, respectively.
396 397
Regional ABA models for stem volume (m3 ha ) were:
398 399
(8) 400
(9) 401
402
The RMSE values between observed and predicted values in the training data were 26.3 % 403
and 22.0 % for models 7 and 8, respectively.
404 405
Local ABA models for stem volume (m3 ha ) were:
406 407
24
(10) 408
(11) 409
410
The RMSE values between the observed and predicted values in the training data were 24.6 411
% and 18.7 % for models 9 and 10, respectively. Estimated model parameters, associated 412
standard errors, and parameter estimates of power-type variance functions are presented in 413
Table 2. It should be noted that the use of disaggregation to coniferous or deciduous 414
dominated plots decreased the RMSE values in each scenario (nationwide by 3.4, regional by 415
4.3 and local by 5.9 percentage points).
416 417
Table 2. Estimated parameters of the mean function (Value), associated standard errors (SE), 418
and parameter estimates ( and ) of power-type variance function (Value) of nationwide 419
(Nat), regional (Reg) and local (Loc) models. Subscript D denotes that the sample plots are 420
disaggregated to coniferous or deciduous dominated.
421
VNat VNatD VReg VRegD VLoc VLocD
Value 4.89585 5.18609 11.38016 -0.52242 48.16397 9.27977
SE 0.28468 0.26133 5.31764 3.83188 6.30775 54.24493
Value 0.69467 2.80389 -4.06317 0.79935 1.70622 10.61760
SE 0.00721 0.15490 1.06491 0.03947 0.12107 2.69704
Value 2.11538 0.72222 4.83338 2.01073 -0.02091 -6.54612
SE 0.02324 0.01009 0.34838 0.18418 0.00748 1.14274
Value -1.34654 2.00006 0.00498 -2.79679 - 1.17704
SE 0.05340 0.04985 0.00089 0.24530 - 0.17336
Value - -3.18360 - - - -4.18874
SE - 0.10504 - - - 0.79109
Value - -2.16066 - - - -
SE - 0.05017 - - - -
Value 6.58947 4.16096 10.26694 3.24309 0.00590 0.61386
Value 0.98769 1.01740 0.95397 1.08674 2.16404 1.33882
422 423
424
25 3.4 Performance in the test plots
425 426
Table 3 shows the relative RMSE and MD values associated with stem volume predictions in 427
the different scenarios with DIPC (Nat, Reg) or ALS (Loc) data in the 30 × 30 m test plots. The 428
performance of the basic form of the nationwide model (NatDIPC) was similar to the 429
regional model (RegDIPC) in terms of RMSE (25.9 % vs. 24.1 %), but it had a larger MD value 430
(10.4 % vs. 5.7 %). Compared to the basic form, the disaggregation of plots to coniferous or 431
deciduous dominated specifically improved predictions in the regional model (RMSE 24.1 % 432
to 13.1 % and MD 5.7 % to 0.9 %). Disaggregation also decreased RMSE values in the 433
nationwide scenario (25.9 % to 20.0 %); however MD values remained relatively large (10.4%
434
to 9.6 %). Although the MD and RMSE values associated with the NatDIPCD scenario were 435
greater than the values associated with the RegDIPCD scenario, the relationships between 436
observed and predicted values at the test plots were quite similar (Figure 4).
437 438
Table 3. Relative (%) root mean squared error (RMSE) and mean difference (MD) values of 439
stem volume predictions for 30 × 30 m test plots.
440
Scenario NatDIPC NatDIPCD RegDIPC RegDIPCD LocALS LocALSD
RMSE 25.9 20.0 24.1 13.1 14.1 7.9
MD 10.4 9.6 5.7 0.9 0.2 0.3
441
The regional model with disaggregation (RegDIPCD) had a smaller RMSE value (13.1 %) than 442
the local ALS model (LocALS) without disaggregation (14.1 %). The smallest RMSE value was 443
obtained with the local ALS model, where disaggregation of cells was conducted using 444
observed species proportions (7.9 %). Using MS-NFI information, 95 % of the coniferous 445
dominated cells (n = 59) and 76 % of the deciduous dominated cells (n = 17) were correctly 446
classified.
447
26 448
In the nationwide scenarios NatDIPC and NatDIPCD, 84.2 % and 89.5 % of the 19 confidence 449
intervals included the field-measured volume (Figure 4). Corresponding values for the 450
regional scenarios were 84.2 % (RegDIPC) and 94.7 % (RegDIPCD). Observed values were 451
always inside the confidence intervals in the local scenarios (LocALS and LocALSD). Errors (m3 452
ha ) in each test plot are presented and discussed in Appendix A.
453 454 455
27 456
Figure 4. Observed volumes (m3 ha-1) against predicted stem volumes (m3 ha-1) at the test 457
plots using nationwide (Nat), regional (Reg) and local (Loc) area-based (ABA) scenarios with 458
and without disaggregation to coniferous or deciduous dominated plots (subscript D).
459
28 4. Discussion
460 461
We observed that the ABA metrics associated with the height of the dominant tree layer 462
were similar between ALS and DIPC data. We also showed that the ABA model fitted with 463
ALS data in another area performs well in stem volume prediction with metrics computed 464
from DIPC data (Table 3). Still, more tests are needed in different countries and types of 465
forests before we can recommend the use of the proposed approach outside of boreal 466
forests in the Nordic countries. Nevertheless, drone inventories without new in-situ field 467
measurements appear to be a cost-efficient solution for small- and medium-sized forest 468
holdings (< 100 ha). However, with the current technology, old ALS data are typically needed 469
for height normalisation of DIPC. In practical inventories, horizontal adjustment of DIPC may 470
not be needed if a small georegistration error in x/y-direction is acceptable.
471 472
Users should be aware of factors that cause uncertainty in the proposed approach. First, the 473
scanner model and data collection date (leaf-on vs. leaf-off) affect the ALS metrics used in 474
the models (Næsset 2005; Hopkinson 2007; Næsset 2009; Keränen et al. 2016). Second, 475
model parameters differ between the inventory areas because of variable growing 476
conditions and climate (Pretzsch 2009; Chave et al. 2014; Kotivuori et al. 2018). Therefore, 477
existing sample plots used as training data must be selected so that they represent the 478
target inventory area as much as possible. Third, the properties of DIPC vary within projects 479
because of the different sensors employed, data acquisition parameters, imaging conditions 480
and the algorithms used to construct point clouds from images. For these reasons, 481
prediction errors and MD values vary on a case-by-case basis. In this study, the regional 482
model resulted in smaller MD values than the nationwide model. We assume that this is due 483
29
to the fact that forest conditions are similar in the Sulkava and Liperi areas. Estimated 484
confidence intervals (Figure 4; Figure A1) further illustrate the uncertainties of our approach.
485
It should be noted that the model-based confidence intervals get wider as the number of 486
sample plots in the training data decreases (Figure 4). For this reason, the local scenarios, 487
which represent the best-case scenarios achievable by remote sensing, have the widest 488
confidence intervals despite exhibiting the smallest RMSE and MD values. However, almost 489
every inventory procedure contains uncertainties that may not be quantified by the 490
estimated confidence intervals. In this study, for example, such uncertainties include the 491
effects of different plot sizes in the modelling data and the different cell size used in the 492
prediction (Packalen et al. 2019), positioning errors of sample and test plots (Gobakken and 493
Næsset 2009), differences in field measurement protocols, use of (partly) predicted tree 494
heights in the field data (Eerikäinen 2009), use of diameter and height related stem volume 495
models (Laasasenaho 1982), errors in ground level interpolations from the ALS and DIPC 496
data, and the fact that the field data were updated using a growth model in one ALS 497
inventory area.
498 499
In this study, we used ALS DTM for the normalisation of drone-based DIPC heights. We also 500
tested the use of DIPC DTM, but it was not applied in the final analyses, because the terrain 501
level was elevated by several meters in one very dense spruce-dominated test plot (Table 502
A1, ID 18). Thus, ALS DTM is useful, especially in stands with a dense canopy. An alternative 503
is to use DTM-independent metrics, such as proposed by Giannetti et al. (2018). However, 504
ALS DTM is likely the best option when the canopy is very dense in large areas.
505 506
30
The DIPC data in this study were collected in either sunny or cloudy weather conditions. It is 507
fair to assume that differences in illumination (e.g. due to clouds or solar angle) affect the 508
DIPC data. However, we did not observe that illumination had a noticeable effect on the 509
results, as also reported by Dandois et al. (2015). Wind conditions were rather stable when 510
DIPC data were collected, and, therefore, we assume that wind did not have a major impact 511
on the results.
512 513
The large MD values associated with our ALS based nationwide model predictions are in line 514
with the findings of Kotivuori et al. (2016) and Kotivuori et al. (2018). In contrast to these 515
previous studies, however, we found that the disaggregation of plots to deciduous or 516
coniferous dominated decreased the RMSE values associated with the nationwide and 517
regional models. This observation is logical, and in an ABA context, the benefits of 518
disaggregation have also been reported for tree species and site quality (Næsset 2002), 519
forest type (Bouvier et al. 2015) and dominant tree type (Ota et al. 2017). Furthermore, 520
Maltamo et al. (2016) reported that accounting for the proportions of spruce and deciduous 521
trees provided the greatest improvements in biomass predictions.
522 523
We also tested the sensitivity of predictions to wrongly predicted deciduous or coniferous 524
trees dominance (i.e. disaggregation error). In this study, we used the dominance 525
information calculated from the MS-NFI data. When the correct disaggregation was used, 526
the RMSE and MD values associated with the NatDIPCD scenario improved by 1.2 (20.0 % vs.
527
18.8 %) and 0.7 percentage points (9.6 % vs. 8.9 %), respectively. Correspondingly, the RMSE 528
value associated with the RegDIPCD scenario improved by 0.6 percentage points (13.1 % vs.
529
12.5 %) and the MD value went to zero (0.9% vs. 0.0%). As only minor improvements were 530
31
achieved with the correct disaggregation, we can conclude that the MS-NFI layers available 531
for the whole country are suitable for deciduous vs. coniferous disaggregation as applied in 532
this study.
533
32 5. Conclusions
534 535
In the boreal forests of the Nordic countries, the total stem volume can be predicted with 536
small error rates without in-situ field measurements using DIPC data, by applying ALS-based 537
ABA models that have been fitted elsewhere, as long as the models are constructed using 538
upper height percentiles computed from ALS echo elevations. The results were especially 539
promising if the model was fitted nearby in a similar type of forests. However, more tests in 540
different areas, datasets and forest types are needed before more detailed conclusions can 541
be drawn. The ABA metrics that are related to upper canopy height (e.g. h95) correlate 542
strongly between DIPC and ALS data. However, the use of DIPC data usually requires ALS- 543
based DTM for height normalisation, especially in areas with a dense canopy. Disaggregation 544
of the inventory area to coniferous and deciduous further decreased the error rates.
545 546 547
33 Acknowledgements
548 549
The authors thank the Finnish Forestry Centre, the National Land Survey of Finland, Blom 550
Kartta Oy, TerraTec Oy and Arbonaut Oy Ltd for providing airborne laser scanning (ALS) and 551
field data for the creation of nationwide and regional models. The authors also thank the 552
Finnish Forest Centre and the Academy of Finland for supporting data collection in the Liperi 553
test area with two projects: ALS4D (The Research Council for Natural Sciences and 554
Engineering, Grant No. 295341) and FORBIO (The Strategic Research Council, Grant No.
555
314224). Finally, we thank the anonymous reviewers for their comments that helped us to 556
greatly improve the manuscript.
557 558
34 References
559 560
AgiSoft PhotoScan Professional (2017). Version 1.3. Retrieved from 561
http://www.agisoft.com/downloads/installer/.
562 563
Alonzo M., Andersen H-E., Morton D. C., Cook B. D. (2018). Quantifying boreal forest 564
structure and composition using UAV structure from motion. Forests 9 (3). 15 p. DOI:
565
10.3390/f9030119.
566 567
Birdal A. C., Avdan U., Türk T. (2017). Estimating tree heights with images from an unmanned 568
aerial vehicle. Geomatics, Natural Hazards and Risk 8 (2): 1144 1156. DOI:
569
10.1080/19475705.2017.1300608.
570 571
Bouvier M., Durrieu S., Fournier R. A., Renaud J-P. (2015). Generalizing predictive models of 572
forest inventory attributes using an area-based approach with airborne LiDAR data. Remote 573
Sensing of Environment 156: 322 334. DOI: 10.1016/j.rse.2014.10.004.
574 575
Breidenbach J., McRoberts R. E., Astrup R. 2016. Empirical coverage of model-based variance 576
estimators for remote sensing assisted estimation of stand-level timber volume. Remote 577
Sensing of Environment 173: 274 281. DOI: 10.1016/j.rse.2015.07.026.
578 579
umayo E., Colgan M. S., Delitti W. B.C., Duque 580
581 582
35
P., Ryan C. M., Saldarriaga J. G., Vieilledent G. (2014).
583
Improved allometric models to estimate the aboveground biomass of tropical trees. Global 584
Change Biology 20(10): 3177 3190. DOI: 10.1111/gcb.12629.
585 586
Colomina I., Molina P. (2014). Unmanned aerial systems for photogrammetry and remote 587
sensing: A review. ISPRS Journal of Photogrammetry and Remote Sensing 92: 79 97. DOI:
588
10.1016/j.isprsjprs.2014.02.013.
589 590
Dandois J. P., Olano M., Ellis E. C. (2015). Optimal altitude, overlap, and weather conditions 591
for computer vision UAV estimates of forest structure. Remote Sensing 7: 13895 13920.
592
DOI: 10.3390/rs71013895.
593 594
Eerikäinen K. (2009). A multivariate linear mixed-effects model for the generalization of 595
sample tree heights and crown ratios in the Finnish National Forest Inventory. Forest Science 596
55(6): 480 493.
597 598
European Environment Agency. (2015). Europe coastline.
599
https://www.eea.europa.eu/ds_resolveuid/06227e40310045408ac8be0d469e1189. [Cited 600
27 March 2018]
601 602
Finnish Environmental Institute (SYKE). (2015). Forest vegetation zones.
603
http://www.syke.fi/en-US/Open_information/Spatial_datasets. [Cited 27 March 2018]
604 605
36
Giannetti F., Chirici G., Gobakken T., Næsset E., Travaglini D., Puliti S. (2018). A new approach 606
with DTM-independent metrics for forest growing stock prediction using UAV 607
photogrammetric data. Remote Sensing of Environment 213: 195 205. DOI:
608
10.1016/j.rse.2018.05.016.
609 610
Gobakken T., Næsset E. (2009). Assessing effects of positioning errors and sample plot size 611
on biophysical stand properties derived from airborne laser scanner data. Canadian Journal 612
of Forest Research 39(5): 1036 1052. DOI:10.1139/X09-025.
613 614
Goldbergs G., Maier S. W., Levick S. R., Edwards A. (2018). Efficiency of individual tree 615
detection approaches based on light-weight and low-cost UAS imagery in Australian 616
savannas. Remote Sensing 10(2). 19 p. DOI: 10.3390/rs10020161.
617 618
Goodbody T. R. H., Coops N. C., Tompalski P., Crawford P., Day K. J. K. (2016). Updating 619
residual stem volume estimates using ALS- and UAV-acquired stereo-photogrammetric point 620
clouds. International journal of remote sensing 38(8 10): 2938 2953. DOI:
621
10.1080/01431161.2016.1219425.
622 623
Guerra-Hernández J., González-Ferreiro E., Monleón V. J., Faias S. P., Tomé M., Díaz-Varela R.
624
A. (2017). Use of multi-temporal UAV-derived imagery for estimating individual tree growth 625
in Pinus pinea stands. Forests 8(8), 300. 19 p. DOI: 10.3390/f8080300.
626 627
37
Hopkinson C. (2007). The influence of flying altitude, beam divergence, and pulse repetition 628
frequency on laser pulse return intensity and canopy frequency distribution. Canadian 629
Journal of Remote Sensing 33(4): 312 324. DOI: 10.5589/m07-029.
630 631
Hyndman R.J., Fan Y. (1996). Sample quantiles in statistical packages. The American 632
Statistician 50(4): 361 365. Published by: Taylor & Francis, Ltd. on behalf of the American 633
Statistical Association. DOI: 10.2307/2684934.
634 635
Kachamba D. J., Ørka H. O., Gobakken T., Eid T., Mwase W. (2016). Biomass estimation using 636
3D data from unmanned aerial vehicle imagery in a tropical woodland. Remote Sensing 637
8(11). 18 p. DOI: 10.3390/rs8110968.
638 639
Keränen J., Maltamo M., Packalen P. (2016). Effect of flying altitude, scanning angle and 640
scanning mode on the accuracy of ALS based forest inventory. International Journal of 641
Applied Earth Observation and Geoinformation 52: 349 360. DOI:
642
10.1016/j.jag.2016.07.005.
643 644
Korpela I., Tuomola T., Välimäki E. (2007). Mapping forest plots: An efficient method 645
combining photogrammetry and field triangulation. Silva Fennica 41(3): 457 469. DOI:
646
10.14214/sf.283.
647 648
Kotivuori E., Korhonen L., Packalen P. (2016). Nationwide airborne laser scanning based 649
models for volume, biomass and dominant height in Finland. Silva Fennica 50(4). 28 p. DOI:
650
10.14214/sf.1567.
651
38 652
Kotivuori E., Maltamo M., Korhonen L., Packalen P. (2018). Calibration of nationwide 653
airborne laser scanning based stem volume models. Remote Sensing of Environment 210:
654
179 192. DOI: 10.1016/j.rse.2018.02.069.
655 656
Laasasenaho J. (1982). Taper curve and volume functions for pine, spruce and birch.
657
Communicationes Instituti Forestalis Fenniae 108. 74 p.
658 659
Lappi L. (1993). Metsäbiometrian menetelmiä (Methods of forest biometrics). Silva Carelica 660
24 University of Joensuu, Faculty of Forestry. Gummerus Kirjapaino Oy, Jyväskylä. 182 p.
661 662
Lin Y., Hyyppä J., Jaakkola A. (2011). Mini-UAV-borne LiDAR for fine-scale mapping. IEEE 663
Geoscience and Remote Sensing Letters 8(3): 426 430. DOI: 10.1109/LGRS.2010.2079913.
664 665
Lisein J., Pierrot-Deseilligny M., Bonnet S., Lejeune P. (2013). A photogrammetric workflow 666
for the creation of a forest canopy height model from small unmanned aerial system 667
imagery. Forests 4(4): 922 944. DOI: 10.3390/f4040922.
668 669
Maltamo M., Bollandsås O.M., Gobakken T., Næsset E. (2016). Large-scale prediction of 670
aboveground biomass in heterogeneous mountain forests by means of airborne laser 671
scanning. Canadian Journal of Forest Research 46(9): 1138 1144. DOI: 10.1139/cjfr-2016- 672
0086.
673 674
39
Mehtätalo L., Lappi J. (2019). Forest biometrics with examples in R. Textbook manuscript in 675
preparation for Chapman & Hall. http://cs.uef.fi/~lamehtat/textbook.htm.
676 677
Mohan M., Silva C. A., Klauberg C., Jat P., Catts G., Cardil A., Hudak A. T., Dia M. (2017).
678
Individual tree detection from unmanned aerial vehicle (UAV) derived canopy height model 679
in an open canopy mixed conifer forest. Forests 8(9). 17 p. DOI: 10.3390/f8090340.
680 681
Næsset E. (2002). Predicting forest stand characteristics with airborne scanning laser using a 682
practical two-stage procedure and field data. Remote Sensing of Environment 80(1): 88 99.
683
DOI: 10.1016/S0034-4257(01)00290-5.
684 685
Næsset E. (2005). Assessing sensor effects and effects of leaf-off and leaf-on canopy 686
conditions on biophysical stand properties derived from small-footprint airborne laser data.
687
Remote Sensing of Environment 98: 356 370. DOI: 10.1016/j.rse.2005.07.012.
688 689
Næsset E. (2009). Effects of different sensors, flying altitudes, and pulse repetition 690
frequencies on forest canopy metrics and biophysical stand properties derived from small- 691
footprint airborne laser data. Remote Sensing of Environment 113(1): 148 159. DOI:
692
10.1016/j.rse.2008.09.001.
693 694
National Land Survey of Finland. (2011). Municipality map of Finland.
695
https://tiedostopalvelu.maanmittauslaitos.fi/tp/kartta?lang=en. [Cited 27 March 2018]
696 697
40
National Land Survey of Finland. (2016). Airborne LiDAR data.
698
https://tiedostopalvelu.maanmittauslaitos.fi/tp/kartta?lang=en. [Cited 24 May 2018]
699 700
National Land Survey of Finland. (2018). General map 1:1 M.
701
https://tiedostopalvelu.maanmittauslaitos.fi/tp/kartta?lang=en. [Cited 4 May 2018]
702 703
Natural Resources Institute Finland. (2017). Volume of pine, spruce and birch.
704
http://kartta.luke.fi/opendata/valinta-en.html. [Cited 27 March 2018]
705 706
Ojansuu R., Henttonen H. (1983). Estimation of local values of monthly mean temperature, 707
effective temperature sum and precipitation sum from the measurements made by the 708
Finnish Meteorological Office. Silva Fennica 17(2): 143 160. DOI: 10.14214/sf.a15099.
709 710
Ota T., Ogawa M., Mizoue N., Fukumoto K., Yoshida S. (2017). Forest structure estimation 711
from a UAV-based photogrammetric point cloud in managed temperate coniferous forests.
712
Forests 8(9). 11 p. DOI: 10.3390/f8090343.
713 714
Packalen P., Strunk J., Packalen T., Maltamo M. and Mehtätalo L. (2019). Resolution 715
Dependence in an Area-based Approach to Forest inventory with Airborne Laser Scanning.
716
Remote Sensing of Environment 224: 192 201. DOI: 10.1016/j.rse.2019.01.022 717
718
Panagiotidis D., Abdollahnejad A., Surový P., Chiteculo V. (2017). Determining tree height 719
and crown diameter from high-resolution UAV imagery. International journal of remote 720
sensing 38(8 10): 2392 2410. DOI: 10.1080/01431161.2016.1264028.
721
41 722
Pinheiro J. C., Bates D. M. (2000). Mixed-Effects Models in S and S-PLUS. Springer-Verlag 723
New York, Inc. 2000. DOI: 10.1007/b98882.
724 725
Pinheiro J., Bates D., DebRoy S., Sarkar D., EISPACK authors., Heisterkamp S., Willigen B. V., 726
R-core. (2018). Linear and nonlinear mixed effects models. https://cran.r- 727
project.org/web/packages/nlme/nlme.pdf.
728 729
Pretzsch H. (2009). Forest dynamics, growth and yield. From measurement to model.
730
Springer - Verlag Berlin Heidelberg: 664 p. DOI: 10.1007/978-3-540-88307-4.
731 732
Puliti S., Ene L. T., Gobakken T., Næsset E. (2017). Use of partial-coverage UAV data in 733
sampling for large scale forest inventories. Remote Sensing of Environment 194: 115 126.
734
DOI: 10.1016/j.rse.2017.03.019.
735 736
Puliti S., Ørka H. O., Gobakken T., Næsset E. (2015). Inventory of small forest areas using an 737
unmanned aerial system. Remote Sensing 7(8): 9632 9654. DOI: 10.3390/rs70809632.
738 739
R Core Team (2017). R: A language and environment for statistical computing. R Foundation 740
for Statistical Computing, Vienna, Austria. https://www.R-project.org/.
741 742
Seber G. A. F., Wild C. J. (1989). Nonlinear Regression. Wiley Series in Probability and 743
Statistics. John Wiley & Sons, Inc. DOI: 10.1002/0471725315.
744 745
42
Ståhl G., Holm S., Gregoire T. G., Gobakken T., Næsset E., Nelson R. (2011). Model-based 746
inference for biomass estimation in a LiDAR sample survey in Hedmark County, Norway.
747
Canadian Journal of Forest Research 41(1): 96 107.
748 749
Surový P., Ribeiro N. A., Panagiotidis D. (2018). Estimation of positions and heights from 750
UAV-sensed imagery in tree plantations in agrosilvopastoral systems. International Journal of 751
Remote Sensing 39(14): 4786 4800. DOI: 10.1080/01431161.2018.1434329.
752 753
Tang L., Shao G. (2015). Drone remote sensing for forestry research and practices. Journal of 754
Forestry Research 26(4): 791 797. DOI: 10.1007/s11676-015-0088-y.
755 756
Torresan C., Berton A., Carotenuto F., Di Gennaro S. F., Gioli B., Matese A., Miglietta F., 757
Vagnoli C., Zaldei A., Wallace L. (2017). Forestry applications of UAVs in Europe: a review.
758
International Journal of Remote Sensing 38(8 10): 2427 2447. DOI:
759
10.1080/01431161.2016.1252477.
760 761
Tomppo E., Halme M. (2004). Using coarse scale forest variables as ancillary information and 762
weighting of variables in k-NN estimation: a genetic algorithm approach. Remote Sensing of 763
Environment 92(1): 1 20. DOI: 10.1016/j.rse.2004.04.003.
764 765
Ullman S. (1979). The interpretation of structure from motion. Proceedings of the Royal 766
Society of London. Series B, Containing papers of a Biological character. Royal Society (Great 767
Britain) 203(1153): 405 426. DOI: 10.1098/rspb.1979.0006.
768 769
43
Vauhkonen J., Maltamo M., McRoberts, R. E., Næsset E. (2014). Introduction to forest 770
applications of airborne laser scanning. In: Maltamo M., Næsset E., Vauhkonen J. Forestry 771
applications of airborne laser scanning - concepts and case studies. Managing Forest 772
Ecosystems. Springer Science+Business Media Dordrecht. p. 1 16. DOI: 10.1007/978-94- 773
017-8663-8.
774 775
Vepakomma U., Cormier D. (2017). Potential of multi-temporal UAV-borne LiDAR in 776
assessing effectiveness of silvicultural treatments. The International Archives of the 777
Photogrammetry, Remote Sensing and Spatial Information Sciences XLII(2/W6): 393 397.
778
International Conference on Unmanned Aerial Vehicles in Geomatics, 4 7 September 2017, 779
Bonn, Germany. DOI: 10.5194/isprs-archives-XLII-2-W6-393-2017 780
781
Wallace L., Lucieer A., Watson C., Turner D. (2012). Development of a UAV-LiDAR system 782
with application to forest inventory. Remote Sensing 4(6): 1519 1543. DOI:
783
10.3390/rs4061519.
784 785
44 Appendix A
786 787
The stem volume errors (m3 ha ) of individual test plots are presented for the NatDIPCD, 788
RegDIPCD and LocALSD scenarios in Figure A1. Only scenarios where plots are disaggregated 789
by the dominance of coniferous or deciduous trees are presented. Figure A1 also shows the 790
differences (m3 ha ) between the stem volume observations and the upper and lower 791
bounds of the confidence intervals. These errors can be contrasted against the stand 792
attributes presented in Table A1. Note that test plot IDs shown in Figure A1 correspond to 793
the IDs shown in Table A1.
794 795
When the different scenarios were compared at the level of individual plots, we noted that 796
prediction errors associated with the DIPC-based regional model were closer to zero than 797
errors associated with the corresponding nationwide model in 68.4 % of the test plots. It 798
should also be noted that the overestimation associated with the nationwide ABA (Table 3) 799
is more clearly shown in Figure A1 than in Figure 4. The errors associated with the regional 800
and local models were more evenly distributed (Figure A1). Overall, the nationwide 801
approach provided the smallest errors in 21.1 %, the regional approach in 47.4 % and the 802
local area-based approach in 31.6 % of the individual test plots.
803 804
In scenarios NatDIPCD and RegDIPCD, there was one test plot (ID 18) where the confidence 805
interval was clearly underestimated (Figure A1). Plot 18 was very dense (G > 40 m2) (Table 806
A1), which may explain the large error associated with it. In plot 18, the understorey points 807
of DIPC were rare due to a lack of visible targets, which affected the ABA metrics. However, 808
45
the ALS pulses were better able to penetrate the canopy, resulting in more reliable ABA 809
metrics and final predictions (LocALSD).
810 811 812
813
Figure A1. Differences (m3 ha ) (black dots) between observed and predicted stem volumes 814
in each 30 × 30 m test plot using the nationwide (Nat), regional (Reg) and local (Loc) area- 815
