Article 1

Measuring tree diameter through photogrammetry using mo- 2

bile phone cameras 3

Aakash Ahamed1,2*, John Foye1,3, Sanjok Poudel4, Erich Treischman1,5 and John Fike1,4 4

1 Working Trees Inc [2093 Philadelphia Pike #9249, Claymont, DE 19703, USA] 5
2 Stanford University Dept. of Geophysics [397 Panama Mall Mitchell Building, 3rd Floor Stanford, CA 94305] 6
3 Stanford University Emmett Interdisciplinary Program in Environment and Resources [ Suite 226, 473 Via 7
Ortega, Stanford, CA 94305, USA] 8
4 Virginia Polytechnic Institute and State University [185 Ag Quad Lane, Blacksburg, VA 24061, USA] 9
5 Stanford University Dept. of Statistics [390 Jane Stanford Way, Stanford, CA 94305, USA] 10
* Correspondence: aahamed@stanford.edu 11

Citation: Ahamed, A.; Foye, J.;
Poudel, S.; Treischman, E.; Fike, J.
Measuring tree diameter through
photogrammetry using mobile
phone cameras. Forests 2023, 14, x.
https://doi.org/10.3390/xxxxx
Academic Editor(s):
Abstract: Tree inventories are a cornerstone of forest science and management. Inventories are es- 12
sential for quantifying forest growth rates, determining biomass and carbon stock accumulation, 13
assessing species diversity, and evaluating impacts of both forest management and climate change. 14
Recent advances in digital sensing technologies on mobile phones have the potential to improve 15
traditional forest inventories through increased efficiency in measurement and transcription, and 16
potentially through increasing participation in data collection by non-experts. However, the degree 17
to which digital sensing tools (e.g. camera-enabled smartphone applications) can accurately de- 18
termine the tree parameters measured during forest inventories remains unclear. In this study, we 19
assess the ability of a smartphone application to perform a user-assisted tree inventory, and com- 20
pare digital estimates of tree diameter to measurements made using traditional forestry field sam- 21
pling approaches. Results suggest that digital sensing tools on mobile phones can accurately 22
measure tree diameter (R2 = 0.95; RMSE = 2.7 cm compared to manual measurements), while saving 23
time during both the data collection stage and data entry stage of field sampling. Importantly, we 24
compare measurements of the same tree across users of the phone application in order to deter- 25
mine the per-user, per-tree, and per-species uncertainty associated with each form of measure- 26
ment. Strong agreement between manual and app-based data suggests that digital technologies 27
have the potential to facilitate efficient collection of high-quality and auditable data, collected by 28

Received: date
non-experts, but with some important limitations compared to traditional tree measurement ap- 29
Accepted: date proaches. High quality, broadly accessible, and efficient field data collected through mobile phones 30
Published: date can in turn improve our understanding of tree growth and biomass accumulation, and the key 31
factors (e.g. climate change or management practices) that affect these processes. Thus, accurate 32
Publisher’s Note: MDPI stays neu-
and broadly accessible digital tree inventory tools stand to advance forest science and manage- 33
tral with regard to jurisdictional
claims in published maps and insti-
ment. 34
tutional affiliations.
1. Introduction and Background 35
Data collected during tree inventories play an essential role in understanding bio- 36
Copyright: © 2023 by the authors. mass fluctuations in forests, agricultural settings, and other environments (e.g. Brown et 37
Submitted for possible open access al., 1989; Smith, 2002; Corona et al., 2011). Traditionally, tree inventories are carried out 38
publication under the terms and using tape measures or calipers to estimate tree diameter (e.g. Nowak et al., 2003; Luoma 39
conditions of the Creative Commons et al., 2017), and notebooks or clipboards to record measurements, which are later logged 40
Attribution (CC BY) license in to computers. Accurate estimation of tree diameter is an essential component of (1) 41
(https://creativecommons.org/license determining biomass stocks and growth rates; and (2) understanding changes in stocks 42
s/by/4.0/). and growth rates over time, including understanding the impacts of factors such as cli- 43
mate change and management practices. In addition to quantifying biomass accumula- 44
tion in a variety of environments, tree inventories are also critical components of meas- 45
urement, reporting, and verification requirements associated with reforestation efforts 46

Forests 2023, 14, x. https://doi.org/10.3390/xxxxx www.mdpi.com/journal/forests

Forests 2023, 14, x FOR PEER REVIEW 2 of 18

and nature-based carbon markets (Fahey et al., 2010; Skutsch, 2012; Kerchner et al., 2015; 47
Hurtt et al., 2019). 48
Recent advances in sensing technologies present an opportunity to improve tree 49
inventories by facilitating simple, rapid, and auditable collection of digital measurements 50
of tree diameter. Existing camera sensors on most mobile phones can be leveraged to 51
cheaply acquire these measurements, in addition to other relevant information such as 52
images and geolocation (e.g. Fan et al., 2019; Wu et al., 2019; Proudman et al., 2022; 53
Tatsumi et al., 2022; Holcomb et al., 2023). If inventory data acquired from mobile phones 54
are accurate enough to replace manual measurements, mobile-enabled data collection 55
could significantly simplify the process of collecting the data required for a tree inven- 56
tory, permitting acquisition of high-quality tree measurements at a fraction of the time 57
and cost of current approaches. Moreover, the broad availability of smartphones can fa- 58
cilitate data collection from non-experts, with appropriate guidance and quality control 59
methods in place (Tatsumi et al., 2022). High quality field data can also be used to cali- 60
brate or improve the accuracy of remotely sensed estimates of biomass derived from 61
satellite or airborne platforms (e.g. Houghton et al., 2007; Kennedy et al., 2018; Yu et al., 62
2022). 63

1.1. Background and Previous Work 64

Many recent studies have evaluated the ability of handheld sensors and postpro- 65
cessing algorithms to accurately measure tree diameter, the key parameter measured 66
during tree inventories. We note that a review of terrestrial laser scanning and postpro- 67
cessing approaches to estimate diameter are outside the scope of this paper, because de- 68
ploying such approaches at scale with non-experts is currently prohibitively expensive. 69
In general, diameter estimation approaches are performed in real-time, or via 70
post-processing applied to images, point clouds, or other sensor data. Both hardware 71
(e.g. Proudman et al., 2022) and software (e.g. Tatsumi et al., 2022) have been developed 72
for this purpose, and various studies have sought to measure diameter in real-time (i.e. 73
the user can view the measurement while they are still at the location of the tree, as in 74
Holcomb et al., 2023), or via offline postprocessing algorithms to extract diameter from 75
photos or other data after the user has left the location (e.g. Wu et al., 2019). Proudman et 76
al. (2022) developed a handheld Light Detection and Ranging (LiDAR) unit to estimate 77
tree diameters by modeling each tree trunk as a cylinder through a least-squares opti- 78
mization within a Random Sample Consensus (RANSAC) loop, producing inventory 79
results that can be viewed in the field, in ~¼ the time of a traditional inventory. Errors 80
associated with this approach ranged from 7 - 14 cm. However, despite the much im- 81
proved efficiency of the approach, the LiDAR unit used in the study was custom-made 82
and not readily available, and errors of 7 to 14 cm may not be sufficient for applications 83
requiring high accuracies and low uncertainties. Other studies have used post processing 84
techniques (applied after data collection) to photos or point clouds in order to determine 85
tree diameter. Wu et al. (2019) developed an approach to estimate tree diameter from 86
photos through image segmentation and depth extraction, obtaining a RMSE of 0.21 cm, 87
which corresponds to 2.32% at imaging distances between 2-10m. However, the ap- 88
proach must be carried out offline, after data collection has concluded, which potentially 89
limits utility in field inventories, when immediate feedback can be essential to identifying 90
issues and acquiring correct measurements. 91
In terms of real-time mobile sensors, Tatsumi et al. (2022) developed a mobile 92
phone-based LiDAR scanning approach that measures tree diameters through real-time 93
instance segmentation and circle fitting, finding a strong agreement (R 2 = 0.96) between 94
LiDAR-based and manual measurements. However, the LiDAR hardware required for 95
this application is only available on iPhones from 2020 onwards, which prohibits use of 96
the approach on older phones which make up the majority of mobile devices in most 97
areas of the world. Fan et al. (2019) developed an algorithm based on Simultaneous Lo- 98
calization and Mapping (SLAM; Durrant-Whyte and Bailey, 2006) applied to a Time of 99
Forests 2023, 14, x FOR PEER REVIEW 3 of 18

Flight (TOF) camera that uses infrared light, in order to estimate tree diameters. Their 100
results suggest a relative RMSE of 1.26 cm, or 6.4%. Although promising, this approach 101
requires an infrared light source, which is not typically included with mobile phone 102
camera units. Holcomb et al. (2023) developed an algorithm which uses LiDAR and 103
camera data in order to estimate tree diameter via segmentation, filtering, depth estima- 104
tion, and trunk boundary identification. The approach produced a RMSE of 3.7 cm and a 105
R2 of 0.97 when evaluated against manually measured diameter, and has the advantage 106
of adequate performance even in occluded conditions. 107
Each of the studies described in this section have either relied on Light Detection 108
and Ranging (LiDAR) sensors (e.g. Proudman et al., 2022; Tatsumi et al., 2022; Holcomb 109
et al., 2023), infrared sensors (Fan et al., 2019), or offline post-processing of images or 110
point clouds (Wu et al., 2019) in order to determine tree diameter. Because most 111
smartphones currently being used are not equipped with LiDAR sensors, these ap- 112
proaches cannot be readily applied, effectively limiting field applications in most areas of 113
the world. 114
Recent advances in computer vision and photogrammetric imaging processing 115
software have made it possible to easily apply the SLAM algorithm (Durrant-Whyte and 116
Bailey, 2006) with exclusively optical data in order to create a 3-D map and accurately 117
measure distances from a series of images or video stream. Notably, the technique does 118
not require the use of LiDAR, which potentially permits broad application across mobile 119
phones, including the use of older smartphone models (i.e. those not equipped with 120
LiDAR sensors) to conduct tree inventories. However, the accuracy and measurement 121
uncertainties associated with such an approach are still not well understood, which cur- 122
rently limits the practical use of distance measurement via SLAM on mobile phones as 123
suitable substitutes for traditional forestry approaches. 124

1.2. Study Goals 125

In this study, we present a new approach to estimate tree diameter using optical 126
SLAM and user-specified anchor points, which permits the use of most existing 127
smartphones to obtain accurate estimates of tree diameter. In order to determine the ef- 128
ficacy of measurements made using the developed approach, we compare estimates of 129
diameter acquired via a mobile phone application (on 3 iPhone models) to measurements 130
of diameter acquired using a measuring tape, the standard method of measurement in 131
forest inventories. This comparison includes a detailed statistical analysis of different 132
species, including the variance among measurements collected through the mobile ap- 133
plication by different users, as well as traditional approaches. 134

2. Methods 135
Here we describe the phone application, measurements collected, field data collec- 136
tion campaigns, and statistical analyses performed to assess the fidelity of the proposed 137
approach to estimate tree diameter via mobile phones. 138
Forests 2023, 14, x FOR PEER REVIEW 4 of 18

139
Figure 1. Screen captures of the mobile phone application workflow, demonstrating in order, 140
measurement of tree (A) height and (B) diameter, (C) selection of species, (D) recording of location, 141
and (E) summary screen of collected data. To measure diameter, the user is prompted to place 142
‘anchor points’ at the left and right side of the tree trunk, respectively. The application uses Apple’s 143
ARKit library to detect surfaces and place anchor points. 144

2.1. Description of Mobile Phone Application 145

The Working Trees phone application prompts a user to record measurements of 146
tree diameter (at 1.3 m height) using a phone’s camera and SLAM photogrammetry. Be- 147
yond measurement, the application provides an interface to establish sampling plots, 148
view sample locations on an interactive map (including species, photos, timestamps, and 149
units of sampled trees) and export data collected from multiple phones via an application 150
programming interface (API). 151
The application currently runs on Apple iOS; an Android operating system version 152
is currently in development. The iOS version (available on the Apple App Store at: 153
https://apps.apple.com/us/app/working-trees/id1621922176) uses the ARKit library 154
(short for Augmented Reality kit, see: https://developer.apple.com/documentation/arkit), 155
an open source software library which provides image processing and computer vision 156
functionality for iOS devices. Specifically, within ARKit, the application uses the SLAM 157
algorithm; images from the camera sensor are used in conjunction with data from the 158
gyroscope and accelerometer to model surfaces, and ray casting is used to measure dis- 159
tances. Importantly, ARKit does not require LiDAR, thus permitting use of older phone 160
models which use cameras exclusively, as well as newer models which can make use of 161
LiDAR sensors. 162
Distance can be determined via the application through a three step process. (1) 163
Objects within the camera’s field of view are identified and tracked from frame to frame, 164
(2) Visual odometry (Nistér et al., 2004) is used to determine the location of the camera 165
with respect to the objects in the field of view. This involves using sensor data from the 166
gyroscope and the accelerometer to determine the location of the camera relative to the 167
tracked objects through optical flow (Lucas and Kanade, 1981). (3) SLAM is then used to 168
construct a 3-D surface based on the tracked objects and known imaging geometries. 169
Distances can then be measured by placing anchor points to mark the ‘start’ and ‘end’ 170
coordinates on a surface and an underlying ray casting algorithm (e.g. Ro et al., 2019) to 171
determine distance (see Figure 1 for examples, a demo video is available at: 172
https://youtu.be/n0P-UASCf-c). 173
The analysis described herein focuses on the measurements of tree diameter, be- 174
cause diameter is the most important property recorded in tree inventories. However, the 175
mobile application developed for this study facilitates the collection of various ancillary 176
Forests 2023, 14, x FOR PEER REVIEW 5 of 18

data: (1) height of tree at sampling time (2) location in latitude and longitude coordinates, 177
(3) species of tree being measured (user-selected from a menu of species before meas- 178
urement), (4) date and time of measurement, and (5) photos of the height and diameter 179
measurements. A follow-on study will assess the efficacy of tree height measurements 180
collected using the application. The next section describes the field trials conducted to 181
evaluate the mobile phone-based diameter measurements. 182

2.2. Description of Field Trials 183

In order to assess the fidelity of the mobile application to accurately measure tree 184
diameter, two double-blind sampling campaigns were conducted at research farms op- 185
erated by Virginia Tech in May and October of 2022. Digital measurements of diameter 186
were collected using the mobile application while manual measurements were obtained 187
using a diameter tape measure. Three individuals acquired manual measurements; three 188
individuals used the mobile application, and one individual acquired both manual and 189
mobile phone-based measurements. The iphones used to collect tree diameter measure- 190
ments were an iPhone 8+, an iPhone 12 mini, and an iPhone 13 pro. 191
The first field trial was undertaken at the Catawba Sustainability Center, Catawba, 192
VA (Figure 2) on May 5th, 2022. The application was tested on 20 black walnut (Juglans 193
nigra), 14 black locust (Robinia pseudoacacia), and 13 pitch-loblolly pine (Pinus rigida x 194
taeda), all aged 7 years, and planted in rows. During the first field trial, two individuals 195
recorded diameter measurements of 47 trees using tape measures, while one individual 196
used the mobile application to measure the diameter and height of the same set of 47 197
trees. 198
The second field trial was performed at a long-term agroforestry research site at 199
Kentland Farm (Blacksburg, VA) between October 15th and October 31st, 2022. For the 200
study, measures were made on 58 honeylocust (Gleditsia triacanthos), and 50 black walnut 201
trees. Trees were 27 years old, planted in ~13 m x ~13 m configuration, and boles had 202
been pruned to facilitate site access with farm equipment and livestock and to support 203
light to the understory. During the second field trial, one individual acquired both 204
manual and phone-based measurements, while one other individual acquired 205
phone-based measurements. 206
Tree diameter was defined at 1.3 m. While not strictly enforced at the time of sam- 207
pling, each participant in the field trials was instructed to measure each tree at breast 208
height for relative consistency. This inconsistency is unlikely to have a substantial effect 209
on the results - Chojnacky et al. (2013) showed that in most cases, tree diameter varies 210
relatively little between the base of the tree and the diameter at breast height. For the case 211
of a split tree trunk (approximately 3 total trees), diameter was measured directly below 212
the split using both manual and phone-based methods. For every tree, measurements 213
were acquired using a diameter tape measure and the mobile phone application. 214
Measurements were collected independently and double-blind, without the indi- 215
viduals sharing knowledge of the results from the other measurement approach. During 216
the second field trial, all trees were measured at least twice, using both the manual and 217
phone-based approach, but only 47 trees were measured a third time due to the time 218
constraints of field researchers. Repeated measurements collected using both the mobile 219
application and manual approaches allowed for an analysis of uncertainty for each tree 220
measured multiple times using both approaches, as well as an assessment of uncertainty 221
associated with different users of the mobile application. These analyses are described in 222
detail in Section 3.6. 223

2.2.1. Preprocessing of Field Data 224

In order to utilize all of the collected data from the Kentland site, we constructed all 225
unique combinations of manual and phone-based measurements for each tree. This was 226
done because (as previously noted), all trees at Kentland were measured exactly twice 227
manually and either two or three times using the mobile application (47 trees measured 228
Forests 2023, 14, x FOR PEER REVIEW 6 of 18

thrice, 61 trees measured twice). The following example explains the procedure: For trees 229
that were measured twice manually and twice using the mobile application, there are 230
four possible combinations of data (1st phone and 1st manual; 1st phone and 2nd manu- 231
al; 2nd manual and 2nd phone; 1st manual and 2nd phone). Similarly, there were six 232
combinations for trees measured thrice through the mobile application. Iterating through 233
all combinations provided a total of 567 samples ([47 * 6] + [61*4]). However, because 234
some diameter measurements collected for a given tree matched exactly, we eliminated 235
duplicate points from the dataset. This final step in the procedure resulted in 367 paired 236
samples describing the 108 total trees at the Kentland site. When combined with the data 237
collected at Catawba (measured only once using each method), the total size of paired 238
data describing both manual and phone-based measurements was 414 samples. 239

240
Figure 2. Map data from Google, NASA. (A) Map of the study location in Blacksburg (red star), 241
Virginia (black shading), USA is shown. (B) Aerial photo of the Kentland Study site comprised of 242
58 Honey Locusts and 50 Black Walnuts aged 27 years, (C) Aerial Photo of the Catawba Study Site, 243
comprised of 20 Black Walnut, 14 Black Locust, and 13 Loblolly Pine, all aged 7 years. 244

2.3. Statistical Analyses 245

To assess the accuracy of the phone-based approach, six metrics were calculated to 246
compare the phone-based measurements to manual measurements. These are described 247
below. In this section, the manual measurements are referred to as the “true values” 248
while the phone-based measurements are the “predicted values”. 249
Mean Error (ME) was used to assess the average difference between phone-based 250
and manual measurements: 251

(1) 252

where 𝑦𝑖 are the true values and 𝑥𝑖 are the predicted values, and n is the number of 253
samples. 254
Mean Percent Error (MPE) is similar to ME as defined above, but expressed in per- 255
centages: 256
Forests 2023, 14, x FOR PEER REVIEW 7 of 18

(2) 257
where 𝑦𝑖 are the true values and 𝑥𝑖 are the predicted values, and n is the number of 258
samples. 259
The R2 correlation coefficient measures the strength of correlation between two 260
variables. It is defined: 261

(3) 262
where 𝑦𝑖 are the true values, 𝑥𝑖 are the predicted values, and ȳ is the mean of the true 263
values. 264
Root Mean Square Error (RMSE) measures the difference in values predicted by a 265
model or an estimator and observed values. It is defined: 266

(4) 267
where 𝑦𝑖 are the true values, 𝑥𝑖 are the predicted values, and n is the number of 268
samples. 269
The Concordance Correlation Coefficient (𝜌𝑐) measures the degree of agreement 270
between two variables which rate, code, or assess the same phenomenon. It is defined: 271

(5) 272
where p is the Pearson Correlation (i.e. √R2); σx and σy are the variances of x and y 273
respectively; and µx and µy are the means of x and y respectively. In general, a higher 𝜌𝑐 274
denotes a stronger relationship between the data. 275
The Intraclass Correlation Coefficient (r) is used to assess the consistency, or con- 276
formity, of measurements made by multiple observers or methods measuring the same 277
quantity (Shrout and Fleiss, 1979). It is defined: 278

(6) 279
where xi are the true values, yi are the predicted values, N is the total number of 280
samples, and 281

(7) 282
and 283

. (8) 284
Generally, a higher r also denotes a stronger relationship between the data. 285
The above metrics were computed in order to assess the level of agreement between 286
phone-based and manual measurements of tree diameter. Table 1 shows each of the sta- 287
tistics for the entire set of measurements across both field trials (N = 414). 288
Forests 2023, 14, x FOR PEER REVIEW 8 of 18

2.4. Graphical Statistical Analysis 289

The Bland-Altman plot, also known as the Tukey Mean Difference Plot, is a widely 290
used graphical statistical inference tool which depicts the agreement between two quan- 291
titative measurements of the same parameter (Bland and Altman, 1999). Often used in 292
clinical trials to assess the fidelity of different assays or treatment techniques, it is suitable 293
for comparing estimates of diameter derived from the manual approach with measure- 294
ments made using the phone application. 295
These metrics were computed for diameter in order to assess the level of agreement 296
between phone-based and manual measurements for the entire population (N = 414) of 297
measurements across both field trials. 298

2.5. Analysis of Uncertainty Among Users and Species 299

Repeated measurements of 108 trees were collected using both a tape measure and 300
the phone application (described in Section 3.2). This allowed for an analysis of uncer- 301
tainty both at the tree level (capturing the uncertainty in using each approach to measure 302
a given tree as well as the user level (capturing the uncertainty between different users 303
measuring the same subsets of trees). The tree-specific uncertainty analysis evaluated the 304
per-tree mean and standard deviation produced by both the application-based and 305
manual forms of measurement. The results are depicted as a scatterplot with error bars 306
which correspond to standard deviations. The user-specific uncertainty analysis evalu- 307
ated the population means and standard deviations of each tree species for each separate 308
user of the application, and for each of the separate manual measurements collected. 309

2.5.1. Statistical Tests 310

For the 108 trees located at the Kentland site where multiple measurements of di- 311
ameter were available from both the manual and mobile approach, we applied a number 312
of statistical tests in order to compare the manually collected measurements with the 313
phone-based measurements. These included (1) a paired T-Test to test the null hypothesis 314
that the mean between two groups is equal; (2) Analysis of Variance (ANOVA) which 315
includes computation of an F statistic in order to test the null hypothesis that samples in 316
all groups are drawn from populations with the same mean values, and (3) Levene’s test 317
(Levene, 1960) which is used to test the null hypothesis that two populations have equal 318
variances. All statistical tests were implemented using the stats module of the scipy 319
package (Virtanen et al., 2020) in the Python programming language. 320

3. Results and Discussion 321

The metrics described in the Methods section were computed across both field trials 322
and for all species in the study areas in order to assess the level of agreement between 323
mobile phone-based and manual measurements of tree diameter. The results of the sta- 324
tistical analysis and species-level analysis are described in Section 4.1; the results of 325
graphical statistical tests are described in Section 4.2; and the results of the user-level 326
uncertainty analysis are described in Section 4.3. Limitations of the approach are dis- 327
cussed in Section 4.4, while the implications of the study and goals of follow on studies 328
are presented in Section 4.5. 329

3.1. Results of Statistical Analysis and Species-Level Analysis 330

Across all species, the mobile app underpredicted diameter; the mean error aver- 331
aged -1.9 cm which represented about 8.3% underestimate (Table 1). Greatest underpre- 332
diction (-2.45 cm) occurred with black walnut trees. The overall RMSE (2.7 cm) was 333
comparable to previous studies: Holcomb et al. (2023) reported an overall RMSE of 3.7 cm 334
and a RMSE of 2.7 cm for trunks under 100 cm in diameter (all trunks in this study). 335
Tatsumi et al. (2022) reported an RMSE of 2.27 cm using LiDAR enabled mobile phone 336
Forests 2023, 14, x FOR PEER REVIEW 9 of 18

and Ipad, while Fan et al. (2019) reported an RMSE of 1.26 cm with an offline post pro- 337
cessing approach. 338

Table 1. Summary of Error Statistics for phone-based estimates of diameter. 339

Concordance Intra Class

Mean Mean Per-
RMSE R² (dimen- Correlation Correlation
Error cent Error
(cm) sionless) (dimension- (dimension-
(cm) (%)
less) less)
Honey Lo-
-1.33 -5.50 2.18 0.84 0.87 0.87
cust (N=207)
Black Walnut
-2.45 -9.62 3.19 0.95 0.94 0.93
(N=180)
Black Locust
-1.29 -15.42 1.54 0.95 0.89 0.89
(N=14)
Pitch-Lobloll
-1.58 -17.49 1.65 0.99 0.94 0.94
y Pine (N=13)
All (N=414) -1.90 -8.27 2.7 0.90 0.91 0.91

340
Figure 3. Scatterplot showing the tree diameter measured manually (x-axis) and through the mo- 341
bile application (y-axis). Results are color coded by species, and results for the Catawba study area 342
are shown as crosses, while the results for the Kentland study area are shown as circles. The mean 343
error between phone-based and manual measurements was -1.89 cm, and the RMSE was 2.7 cm. 344

Strong correlation (R2 = 0.95) between phone-based and manual tree diameter 345
measurements (Figure 3) were comparable to R2 values reported in past studies that uti- 346
Forests 2023, 14, x FOR PEER REVIEW 10 of 18

lized LiDAR or offline postprocessing (e.g. Tatsumi et al., 2022; Holcomb et al., 2023). The 347
line of best fit (shown in red in Figure 3, equation shown in legend) and slope of 0.91 also 348
suggests that the phone-based approach underpredicts the true mean diameter for most 349
trees. The analysis included trees with diameters between 0 and 50 cm. 350

351
Figure 4. Violin plots showing the mean error for each species is shown as a white dot, while error 352
distributions around the mean are shown for each species. 353

Lower magnitudes of percent errors were observed with honeylocust and black 354
walnut trees (Figure 4). Larger errors with lack locust and pitch-loblolly pine likely reflect 355
the much smaller sample sizes. The variability among estimates of diameter for each tree 356
species were further evaluated through analysis of histograms (Figure 5). Similar distri- 357
butions for the manual and phone-based measurements again suggest that the mean 358
diameter estimated with the phone application is comparable, but slightly lower than the 359
mean diameter estimated through manual measurements. 360
Forests 2023, 14, x FOR PEER REVIEW 11 of 18

361
Figure 5. Histograms showing the distribution and variance of manual diameter measurements 362
(orange), and phone-based diameter measurements (blue) for each species. The histograms and 363
fitted distributions show that diameter and height measured using the application are conservative 364
relative to manual measurements, and both approaches have similar distributions. 365

3.2. Graphical Statistical Analysis 366

The Bland Altman plot (FIgure 6) provides a different means of visualizing the 367
comparison, showing more than 98% of the collected paired data fell within the 2 stand- 368
ard deviation confidence range. The application tends to underestimate diameter by -1.9 369
cm, and the results suggest a weak underestimation trend as diameter increases, espe- 370
cially for black walnut (shown in blue). 371
Forests 2023, 14, x FOR PEER REVIEW 12 of 18

372
Figure 6. Bland-Altman Plots for Tree Diameter showing that phone-based measurements of 373
heights and diameters are conservative, and >95% of all measurements all within a two-standard 374
deviation range. 375

3.3. Results of User and Species Uncertainty Analysis 376

Collection of repeat measurement of individual trees allowed us to assess the us- 377
er-specific uncertainty associated with both the manual and phone-based measurements. 378
The results of repeated surveys for 108 trees at Kentland Farm are shown in Figures 7 and 379
8. Figure 7 shows a scatterplot with the manual measurements on the x-axis and the 380
phone-based measurements on the y-axis. Each point represents the mean of all meas- 381
urements collected using multiple measurements from the phone application and a tape 382
measure. The vertical error bars denote the phone-based standard deviation, while the 383
horizontal error bars (which are so small they are difficult to detect) denote the manual 384
standard deviation. Uncertainty associated with phone-based measurements was larger 385
than that of manual measurements (Figure 7), but the phone-based approach still pro- 386
vided a strong level of agreement (R2 = 0.95; RMSE = 2.5 cm). 387
Forests 2023, 14, x FOR PEER REVIEW 13 of 18

388
Figure 7. Scatterplot showing the tree diameter measured manually (x-axis) and through the mo- 389
bile application (y-axis). Points correspond to the mean value for each tree, while error bars corre- 390
spond to the standard deviation of each tree. Results are color coded by species, and results for the 391
Catawba study area are shown as crosses, while the results for the Kentland study area are shown 392
as circles. The mean error between phone-based and manual measurements was -1.89 cm, and the 393
RMSE was 2.7 cm. 394

Figure 8 shows violin plots of the mean diameter (y-axis) for each species (x-axis), 395
color coded by each user who participated in the Kentland study. In the study, one user 396
acquired measurements of each tree twice using the mobile application. These repeated 397
measurements are termed “User 1” and “User 1 repeat” and those repeat distributions 398
were nearly identical, indicating high repeatability within users, both for honeylocust 399
and black walnut. For both sets of phone-based measurements acquired by User 1, di- 400
ameter was slightly underestimated during both sampling sessions. User 2 tended to 401
underestimate the diameter for both species more significantly, although this may be an 402
effect of lower sample size. 403
Forests 2023, 14, x FOR PEER REVIEW 14 of 18

404
Figure 8. Violin plots depicting the mean and distribution of diameter (cm, y-axis), estimated by 405
each user (colors), for the two species at the Kentland study site (x-axis). 406

Table 2 shows results of the statistical tests performed on the repeat measurement 407
data acquired at Kentland. The values in each cell depict the results of the statistical tests 408
applied to the row and column combination, with p values listed in parentheses. The first 409
two columns of Table 2 show the results of a paired T test. Because the T statistic in each 410
case is larger in magnitude than the critical T value (1.35 < Tcrit < 1.65), and the p values 411
are < 0.05, the paired T test suggests that we fail to reject the null hypothesis that there is 412
no difference between the means of the two samples at a 95% confidence level. 413
A significant, negative T statistic in each case shows that the mean of the samples 414
measured by users of the mobile application are lower compared to the mean of the 415
samples measured manually. The middle two columns of Table 2 show the results from 416
the one-way ANOVA, which tests the null hypothesis that samples in both groups are 417
drawn from populations with the same mean values. The F statistics in the test are larger 418
than the critical values for a 95% confidence interval (1.37 < F crit < 1.55), and the p values 419
are all <0.05, which suggests that we can reject the null hypothesis that samples in both 420
groups are drawn from populations with the same mean values. These results intuitively 421
make sense, because we observe in the data that the mean diameter values determined 422
using the mobile application are generally lower than those produced by manual meas- 423
urements. This supports the finding that the sample means and the means of the two 424
underlying distributions are also not the same. 425

Table 2. Results of statistical tests applied to the field data collected at Kentland. The columns de- 426
pict the two sets of measurements collected manually, while columns depict the three sets of 427
measurements collected using the mobile application. 428

(A) Paired T Test (B) One Way ANOVA (C) Levene's Test
T statistic (p value) F statistic (p value) W Statistic (p value)
manual 1 manual 2 manual 1 manual 2 manual 1 manual 2
User 1 -3.27 (0.0014) -3.45 (0.0008) 6.81 (0.0097) 7.46 (0.0068) 1.21 (0.2735) 1.03 (0.312)
Forests 2023, 14, x FOR PEER REVIEW 15 of 18

User 1 re- 1.39

-2.82 (0.0057) -2.98 (0.0036) 5.7 (0.0178) 6.29 (0.0129) 1.6 (0.2078)
peat (0.2391)
6.8 0.04
User 2 -3.04 (0.0039) -3.06 (0.0037) 6.31 (0.0131) 0.07 (0.7928)
(0.01) (0.8398)

Levene’s test produced a low test statistic with high p values, (all p>0.2), meaning 429
that we cannot reject the null hypothesis that two populations have equal variances. This 430
again supports our finding that while the means of mobile phone-based measurements 431
and manual measurements of diameter are significantly different, the distribution of 432
values about those means is similar (as presented in Figure 5). 433

3.4. Limitations and Assumptions 434

While the approach presented in this paper shows promising results for accurately 435
estimating tree diameter using sensors on mobile phones, it is not without limitations. 436
First and foremost, the mobile phone-based estimates of diameter have slight negative 437
bias relative to manual estimates, as indicated by regression slopes <1 (Figures 3 and 438
Figure 7), as well as the Mean Error and Mean Percent Error results shown in Table 1. 439
While the conservative bias can be a positive attribute in the context of estimating carbon 440
sequestration when reporting nature-based carbon projects, it may be alleviated by ap- 441
plying a calibration factor to mobile phone-based measurements, something that will be 442
explored in future work. One potential source of the negative bias could be the lack of 443
strict enforcement of diameter acquired at a height of 1.3 m (~4 ft). While diameter is un- 444
likely to vary substantially between where manual and phone-based data were collected, 445
this is nevertheless a source of uncertainty. 446
Another important consideration is that the approach is not fully algorithmic - it still 447
requires a human to subjectively place anchor points on the ‘start’ and ‘end’ locations of a 448
tree trunk (see Figure 1 for an example). Because the mobile application also stores pho- 449
tographs of the line segments that are drawn on the trees, they can be visually inspected 450
for quality, though no quality control was undertaken for the data in this study. Future 451
work will compare algorithmic estimation of tree diameter from the acquired images to 452
the diameters captured by users placing anchor points, as was done in this study. 453
Context is another limitation of this study. The mobile application was tested in an 454
agroforestry research site. Closed canopy forests with a dense understory can present 455
challenging conditions related to surface detection and anchor point placement. In addi- 456
tion to dense understory and canopies, the detection of surfaces through the SLAM al- 457
gorithm and the measurement of distances through ray casting can also be affected by 458
various conditions such as fog/haze, and a low sun angle. Moreover, homogeneous sur- 459
faces (uncommon in natural settings) can require a user to point the camera at a flat sur- 460
face and move the phone around until enough points are identified to create a surface. 461
This procedure might take several seconds for a setting that is homogenous in color and 462
texture. 463

3.5. Discussion of Results and Future Work 464

Results of the field trials suggest a strong ability to accurately measure tree diameter 465
through use of the mobile phone application as validated against manual measurements 466
for each of the species being measured (Table 1; Figure 2 - Figure 4). Many widely-cited 467
allometric equations rely exclusively on the use of diameter measurements (e.g. Jenkins 468
et al., 2003; Chave et al., 2014; Chojnacky et al., 2014) to estimate standing biomass. Strong 469
agreement between the phone application and traditional measures (R2 > 0.95; p<0.01; 470
RMSE =2.7 cm; concordance correlation > 0.9; Intra Class Correlation > 0.9) indicate this 471
technology can provide reasonable, rapid, and conservative estimates of woody biomass 472
in orchard systems (Figure 3 - Figure 6.). Additionally, the phone application has the 473
advantage of automatically logging measurements (as well as photos and metadata) to an 474
Forests 2023, 14, x FOR PEER REVIEW 16 of 18

online database, saving time that technicians would use for data entry, and preventing 475
transcription errors. These data can also be used for post hoc quality control and removal 476
of any erroneous measurements. 477
Results suggest that diameter measurements estimated with the mobile application 478
vary more than those gathered with a tape measure (Figures 7 and 8), but this may also 479
be user dependent. For example, Figure 8 demonstrates that a single user who acquired 480
fewer measurements with the mobile application significantly underestimated diameter 481
relative to both another mobile user as well as to manual diameter measurements. These 482
findings are corroborated by results of statistical tests (Table 2) that suggest the means of 483
manual and mobile phone-based measurements are different (with mobile measure- 484
ments being biased low), but their distributions are similar. More analysis of the per-user 485
accuracy and uncertainties is needed in order to confidently adopt any sensing technol- 486
ogy for long-term use. 487
Ongoing and future work seeks to (1) develop the approach for android operating 488
systems to permit use on more mobile phones; (2) accommodate common edge cases, 489
such as split trunks, sloping ground, and leaning, curved, or buttressed trees; (3) better 490
understand the limitations of digital measurements including the influence of different 491
users of the application, lighting and shadows, foliage, seasonality, and other factors; (4) 492
algorithmically identify tree diameter (e.g. as done in Holcomb et al., 2023) in order to 493
compare data collected by different users of the application, and (5) develop, refine, and 494
test the capabilities of mobile-based photogrammetric techniques to estimate tree height. 495

4. Conclusions 496
Accurate measurements of tree diameter are essential for quantifying biomass 497
stocks, determining changes in tree growth, and understanding the impacts of factors 498
such as climate change and management practices. Digital sensors on mobile phones 499
have recently been gaining traction as viable alternatives to tape measures, calipers, and 500
clipboards, the tools traditionally used for tree inventories. Previous studies of sen- 501
sor-based systems have focused on LiDAR or require offline postprocessing, which limits 502
the number of compatible phones and thus the global reach of those sensing approaches. 503
This study developed a mobile application that can run in real time and does not require 504
LiDAR, which means it can be widely deployed on existing phones. This study also an- 505
alyzed the uncertainty between repeat samples acquired by multiple individuals using 506
the mobile application, and multiple acquisitions of manual diameter measurements. 507
Such an exercise is essential to validate the application for broadscale use, but this type of 508
work has not been widely tested in previous studies. While mobile-based measurements 509
of diameter varied more than manual measurements, the relatively low errors (Mean 510
Error = - 1.9 cm) and high correlation (R2 = 0.95) between mobile and manual estimates of 511
diameters suggest that with tweaking this system can be robust, accurate, and more rapid 512
than traditional measurement techniques. Although more studies are needed to better 513
understand the limitations and uncertainties of mobile phone sensing technology, mo- 514
bile-enabled estimates of diameter may ultimately help facilitate widespread data collec- 515
tion by non-experts. This stands to improve data collection practices, increase available 516
data, and afford a richer understanding of the environmental and management factors 517
that influence tree growth. 518

Acknowledgments: All data and code used to generate the figures and statistics reported here can 519
be found at: https://github.com/WorkingTrees/SamplingPaper. Thanks to Austin Unruh, Carole 520
Bankhead, and Brenda Graff, who assisted with field sampling. Thanks to Wendy Silverman from 521
the Virginia Tech Cooperative Extension Master Gardeners Program, who recruited field assis- 522
tants. Thanks to Professor John Munsell who provided measurement equipment for these experi- 523
ments. Thanks to Vickesh Solanki for mobile application development services. 524

Conflicts of Interest: The authors claim that there are no conflicts of interest or financial interests 525
beyond that expressed by the fact that Lead Author Aakash Ahamed and co-author John Foye are 526
Forests 2023, 14, x FOR PEER REVIEW 17 of 18

employed by Working Trees, and Erich Treischman and John Fike are advisors of Working Trees, 527
which could use the proposed measurement approach in their business model. 528

Data Availability Statement: All collected data and code used to generate the figures and statistics 529
reported in this paper can be found at: https://github.com/WorkingTrees/SamplingPaper. 530

Author Contributions: Conceptualization: AA, JF, SP, ET, JF; Data curation: AA, JF, SP; Formal 531
analysis: AA; Funding acquisition: AA, JF, JF; Investigation: AA; Methodology: AA, SP, ET; Project 532
administration: AA, JF, SP, JF; Resources: AA, JF, SP, JF; Software: AA; Supervision: SP, JF; Valida- 533
tion: AA; Visualization: AA; Writing - original draft: AA; Writing - review & editing: AA, JF, SP, 534
ET, JF 535

References 536
1. Bland: J.M. and Altman, D.G., 1999. Measuring agreement in method comparison studies. Statistical methods in medical re- 537
search, 8(2), pp.135-160. 538
2. Brown, S., Gillespie, A.J. and Lugo, A.E., 1989. Biomass estimation methods for tropical forests with applications to forest in- 539
ventory data. Forest science, 35(4), pp.881-902. 540
3. Chave, J., Réjou-Méchain, M., Búrquez, A., Chidumayo, E., Colgan, M.S., Delitti, W.B., Duque, A., Eid, T., Fearnside, P.M., 541
Goodman, R.C. and Henry, M., 2014. Improved allometric models to estimate the aboveground biomass of tropical trees. 542
Global change biology, 20(10), pp.3177-3190. 543
4. Chojnacky, D.C., Heath, L.S. and Jenkins, J.C., 2014. Updated generalized biomass equations for North American tree species. 544
Forestry, 87(1), pp.129-151. 545
5. Corona, P., Chirici, G., McRoberts, R.E., Winter, S. and Barbati, A., 2011. Contribution of large-scale forest inventories to bio- 546
diversity assessment and monitoring. Forest ecology and management, 262(11), pp.2061-2069. 547
6. Durrant-Whyte, H. and Bailey, T., 2006. Simultaneous localization and mapping: part I. IEEE robotics & automation magazine, 548
13(2), pp.99-110. 549
7. Fahey, T.J., Woodbury, P.B., Battles, J.J., Goodale, C.L., Hamburg, S.P., Ollinger, S.V. and Woodall, C.W., 2010. Forest carbon 550
storage: ecology, management, and policy. Frontiers in Ecology and the Environment, 8(5), pp.245-252. 551
8. Fan, Y., Feng, Z., Mannan, A., Khan, T.U., Shen, C. and Saeed, S., 2018. Estimating tree position, diameter at breast height, and 552
tree height in real-time using a mobile phone with RGB-D SLAM. Remote Sensing, 10(11), p.1845. 553
9. Holcomb, A., Tong, L. and Keshav, S., 2023. Robust Single-Image Tree Diameter Estimation with Mobile Phones. Remote 554
Sensing, 15(3), p.772. 555
10. Houghton, R.A., Butman, D., Bunn, A.G., Krankina, O.N., Schlesinger, P. and Stone, T.A., 2007. Mapping Russian forest bio- 556
mass with data from satellites and forest inventories. Environmental Research Letters, 2(4), p.045032. 557
11. Hurtt, G., Zhao, M., Sahajpal, R., Armstrong, A., Birdsey, R., Campbell, E., Dolan, K., Dubayah, R., Fisk, J.P., Flanagan, S. and 558
Huang, C., 2019. Beyond MRV: high-resolution forest carbon modeling for climate mitigation planning over Maryland, USA. 559
Environmental Research Letters, 14(4), p.045013. 560
12. Jenkins, J.C., Chojnacky, D.C., Heath, L.S. and Birdsey, R.A., 2003. National-scale biomass estimators for United States tree 561
species. Forest science, 49(1), pp.12-35. 562
13. Kennedy, R.E., Ohmann, J., Gregory, M., Roberts, H., Yang, Z., Bell, D.M., Kane, V., Hughes, M.J., Cohen, W.B., Powell, S. and 563
Neeti, N., 2018. An empirical, integrated forest biomass monitoring system. Environmental Research Letters, 13(2), p.025004. 564
14. Kerchner, C.D. and Keeton, W.S., 2015. California's regulatory forest carbon market: Viability for northeast landowners. Forest 565
Policy and Economics, 50, pp.70-81. 566
15. Levene, H., 1960. Robust tests for equality of variances. Contributions to probability and statistics, pp.278-292. 567
16. Lucas, B.D. and Kanade, T., 1981, August. An iterative image registration technique with an application to stereo vision. In 568
IJCAI'81: 7th international joint conference on Artificial intelligence (Vol. 2, pp. 674-679). 569
17. Luoma, V., Saarinen, N., Wulder, M.A., White, J.C., Vastaranta, M., Holopainen, M. and Hyyppä, J., 2017. Assessing precision 570
in conventional field measurements of individual tree attributes. Forests, 8(2), p.38. 571
18. Nistér, D., Naroditsky, O. and Bergen, J., 2004, June. Visual odometry. In Proceedings of the 2004 IEEE Computer Society 572
Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004. (Vol. 1, pp. I-I). Ieee. 573
19. Nowak, D.J., Crane, D.E., Stevens, J.C. and Hoehn, R.E., 2003. The urban forest effects (UFORE) model: Field data collection 574
manual. US Department of Agriculture Forest Service, Northeastern Research Station: Syracuse, NY, USA, pp.4-11. 575
20. Proudman, A., Ramezani, M., Digumarti, S.T., Chebrolu, N. and Fallon, M., 2022. Towards real-time forest inventory using 576
handheld LiDAR. Robotics and Autonomous Systems, 157, p.104240. 577
21. Ro, H., Byun, J.H., Park, Y.J., Lee, N.K. and Han, T.D., 2019. AR pointer: Advanced ray-casting interface using laser pointer 578
metaphor for object manipulation in 3D augmented reality environment. Applied Sciences, 9(15), p.3078. 579
22. Shrout, P.E. and Fleiss, J.L., 1979. Intraclass correlations: uses in assessing rater reliability. Psychological bulletin, 86(2), p.420. 580
23. Skutsch, M. ed., 2012. Community forest monitoring for the carbon market: opportunities under REDD. Routledge. 581
Forests 2023, 14, x FOR PEER REVIEW 18 of 18

24. Smith, W.B., 2002. Forest inventory and analysis: a national inventory and monitoring program. Environmental pollution, 116, 582
pp.S233-S242. 583
25. Tatsumi, S., Yamaguchi, K. and Furuya, N., 2022. ForestScanner: A mobile application for measuring and mapping trees with 584
LiDAR-equipped iPhone and iPad. Methods in Ecology and Evolution. 585
26. Virtanen, P., Gommers, R., Oliphant, T.E., Haberland, M., Reddy, T., Cournapeau, D., Burovski, E., Peterson, P., Weckesser, W., 586
Bright, J. and Van Der Walt, S.J., 2020. SciPy 1.0: fundamental algorithms for scientific computing in Python. Nature methods, 587
17(3), pp.261-272. 588
27. Wu, X., Zhou, S., Xu, A. and Chen, B., 2019. Passive measurement method of tree diameter at breast height using a smartphone. 589
Computers and Electronics in Agriculture, 163, p.104875. 590
28. Yu, Y., Saatchi, S., Domke, G.M., Walters, B., Woodall, C., Ganguly, S., Li, S., Kalia, S., Park, T., Nemani, R. and Hagen, S.C., 591
2022. Making the US national forest inventory spatially contiguous and temporally consistent. Environmental Research Letters, 592
17(6), p.065002. 593