Comparative analysis of height-based vegetation segmentation methods: evaluating efficiency and accuracy
Height-based vector vegetation segmentation is one of the critical aspects of spatial analysis. This segmented data is used in radio propagation modeling, environmental monitoring, and vegetation mapping. Many studies on vector vegetation segmentation focus on delineating individual tree crowns, all...
Збережено в:
| Дата: | 2024 |
|---|---|
| Автори: | , |
| Формат: | Стаття |
| Мова: | English |
| Опубліковано: |
Інститут програмних систем НАН України
2024
|
| Теми: | |
| Онлайн доступ: | https://pp.isofts.kiev.ua/index.php/ojs1/article/view/651 |
| Теги: |
Додати тег
Немає тегів, Будьте першим, хто поставить тег для цього запису!
|
| Назва журналу: | Problems in programming |
Репозитарії
Problems in programming| id |
pp_isofts_kiev_ua-article-651 |
|---|---|
| record_format |
ojs |
| resource_txt_mv |
ppisoftskievua/29/ecf322275d95bf670659017f72c85629.pdf |
| spelling |
pp_isofts_kiev_ua-article-6512025-02-15T14:04:28Z Comparative analysis of height-based vegetation segmentation methods: evaluating efficiency and accuracy Розробка методів сегментаці рослинності за висотою та оцінка їх ефективност і точності Tsaryniuk, O.V. Glybovets, A.M. vegetation segmentation; spatial analysis; hexagonal grid; random points, convolution filters UDC 004.67 сегментація рослинності; просторовий аналіз; шестикутна сітка; випадкові точки; згорткові фільтри УДК 004.67 Height-based vector vegetation segmentation is one of the critical aspects of spatial analysis. This segmented data is used in radio propagation modeling, environmental monitoring, and vegetation mapping. Many studies on vector vegetation segmentation focus on delineating individual tree crowns, allowing detailed data sets to be obtained. However, the high level of detail results in a substantial data volume, making it impractical to use these datasets over large areas, such as an entire country. Segmentation of large vector data sets remains a significant challenge in geospatial data creation. In our study, we developed three different segmentation methods: hexagon segmentation, convolution segmentation, and random points method. A test data fragment was processed to compare the proposed methods and accuracy and volume metrics were calculated.Prombles in programming 2024; 2-3: 313-318 Cегментація векторної рослинності за висотою є одним із важливих етапів просторового аналізу. Цей тип даних використовується у побудові моделей розповсюдження радіосигналів, екологічному моніторингу та картографуванні рослинності. На сьогодні існує багато досліджень із сегментації векторної рослинності, що зосереджені на виокремленні індивідуальних крон дерев та дозволяють отримати деталізовані набори даних. Але наслідком високої деталізації є суттєвий обсяг, який унеможливлює використання цих даних на великих територіях, наприклад, в масштабі цілої країни. Сегментація великих масивів векторних даних досі є суттєвим викликом у сфері створення геопросторових даних. В процесі нашого дослідження ми розробили три різні методи сегментації: сегментація шестикутниками, сегментація за допомогою згорткових фільтрів та метод випадкових точок. Для порівняння запропонованих методів був опрацьований тестовий фрагмент даних та прораховані метрики точності та об’єму.Prombles in programming 2024; 2-3: 313-318 Інститут програмних систем НАН України 2024-12-17 Article Article application/pdf https://pp.isofts.kiev.ua/index.php/ojs1/article/view/651 10.15407/pp2024.02-03.313 PROBLEMS IN PROGRAMMING; No 2-3 (2024); 313-318 ПРОБЛЕМЫ ПРОГРАММИРОВАНИЯ; No 2-3 (2024); 313-318 ПРОБЛЕМИ ПРОГРАМУВАННЯ; No 2-3 (2024); 313-318 1727-4907 10.15407/pp2024.02-03 en https://pp.isofts.kiev.ua/index.php/ojs1/article/view/651/703 Copyright (c) 2024 PROBLEMS IN PROGRAMMING |
| institution |
Problems in programming |
| baseUrl_str |
https://pp.isofts.kiev.ua/index.php/ojs1/oai |
| datestamp_date |
2025-02-15T14:04:28Z |
| collection |
OJS |
| language |
English |
| topic |
vegetation segmentation; spatial analysis; hexagonal grid; random points convolution filters UDC 004.67 |
| spellingShingle |
vegetation segmentation; spatial analysis; hexagonal grid; random points convolution filters UDC 004.67 Tsaryniuk, O.V. Glybovets, A.M. Comparative analysis of height-based vegetation segmentation methods: evaluating efficiency and accuracy |
| topic_facet |
vegetation segmentation; spatial analysis; hexagonal grid; random points convolution filters UDC 004.67 сегментація рослинності просторовий аналіз шестикутна сітка випадкові точки згорткові фільтри УДК 004.67 |
| format |
Article |
| author |
Tsaryniuk, O.V. Glybovets, A.M. |
| author_facet |
Tsaryniuk, O.V. Glybovets, A.M. |
| author_sort |
Tsaryniuk, O.V. |
| title |
Comparative analysis of height-based vegetation segmentation methods: evaluating efficiency and accuracy |
| title_short |
Comparative analysis of height-based vegetation segmentation methods: evaluating efficiency and accuracy |
| title_full |
Comparative analysis of height-based vegetation segmentation methods: evaluating efficiency and accuracy |
| title_fullStr |
Comparative analysis of height-based vegetation segmentation methods: evaluating efficiency and accuracy |
| title_full_unstemmed |
Comparative analysis of height-based vegetation segmentation methods: evaluating efficiency and accuracy |
| title_sort |
comparative analysis of height-based vegetation segmentation methods: evaluating efficiency and accuracy |
| title_alt |
Розробка методів сегментаці рослинності за висотою та оцінка їх ефективност і точності |
| description |
Height-based vector vegetation segmentation is one of the critical aspects of spatial analysis. This segmented data is used in radio propagation modeling, environmental monitoring, and vegetation mapping. Many studies on vector vegetation segmentation focus on delineating individual tree crowns, allowing detailed data sets to be obtained. However, the high level of detail results in a substantial data volume, making it impractical to use these datasets over large areas, such as an entire country. Segmentation of large vector data sets remains a significant challenge in geospatial data creation. In our study, we developed three different segmentation methods: hexagon segmentation, convolution segmentation, and random points method. A test data fragment was processed to compare the proposed methods and accuracy and volume metrics were calculated.Prombles in programming 2024; 2-3: 313-318 |
| publisher |
Інститут програмних систем НАН України |
| publishDate |
2024 |
| url |
https://pp.isofts.kiev.ua/index.php/ojs1/article/view/651 |
| work_keys_str_mv |
AT tsaryniukov comparativeanalysisofheightbasedvegetationsegmentationmethodsevaluatingefficiencyandaccuracy AT glybovetsam comparativeanalysisofheightbasedvegetationsegmentationmethodsevaluatingefficiencyandaccuracy AT tsaryniukov rozrobkametodívsegmentacíroslinnostízavisotoûtaocínkaíhefektivnostítočností AT glybovetsam rozrobkametodívsegmentacíroslinnostízavisotoûtaocínkaíhefektivnostítočností |
| first_indexed |
2025-07-17T09:36:55Z |
| last_indexed |
2025-07-17T09:36:55Z |
| _version_ |
1838499720238989312 |
| fulltext |
313
Великі дані (Big Data) та Аналітика даних (Data Scienсe)
УДК 004.67 http://doi.org/10.15407/pp2024.02-03.313
O.V. Tsaryniuk, A.M. Hlybovets
DEVELOPMENT OF HEIGHT-BASED VEGETATION
SEGMENTATION METHODS: EVALUATING EFFICIENCY
AND ACCURACY
Height-based vector vegetation segmentation is one of the critical aspects of spatial analysis. This
segmented data is used in radio propagation modeling, environmental monitoring, and vegetation mapping.
Many studies on vector vegetation segmentation focus on delineating individual tree crowns, allowing
detailed data sets to be obtained. However, the high level of detail results in a substantial data volume,
making it impractical to use these datasets over large areas, such as an entire country. Segmentation of large
vector data sets remains a significant challenge in geospatial data creation. In our study, we developed three
different segmentation methods: hexagon segmentation, convolution segmentation, and random points
method. A test data fragment was processed to compare the proposed methods and accuracy and volume
metrics were calculated.
Keywords: vegetation segmentation, spatial analysis, hexagonal grid, random points, convolution filters.
О.В. Царинюк, А.М. Глибовець
РОЗРОБКА МЕТОДІВ СЕГМЕНТАЦІ РОСЛИННОСТІ ЗА
ВИСОТОЮ ТА ОЦІНКА ЇХ ЕФЕКТИВНОСТ І ТОЧНОСТІ
Cегментація векторної рослинності за висотою є одним із важливих етапів просторового аналізу. Цей
тип даних використовується у побудові моделей розповсюдження радіосигналів, екологічному
моніторингу та картографуванні рослинності. На сьогодні існує багато досліджень із сегментації
векторної рослинності, що зосереджені на виокремленні індивідуальних крон дерев та дозволяють
отримати деталізовані набори даних. Але наслідком високої деталізації є суттєвий обсяг, який
унеможливлює використання цих даних на великих територіях, наприклад, в масштабі цілої країни.
Сегментація великих масивів векторних даних досі є суттєвим викликом у сфері створення
геопросторових даних. В процесі нашого дослідження ми розробили три різні методи сегментації:
сегментація шестикутниками, сегментація за допомогою згорткових фільтрів та метод випадкових
точок. Для порівняння запропонованих методів був опрацьований тестовий фрагмент даних та
прораховані метрики точності та об’єму.
Ключові слова: сегментація рослинності, просторовий аналіз, шестикутна сітка, випадкові точки,
згорткові фільтри.
Introduction
Integrating diverse datasets is a pivotal
challenge in geospatial data production,
particularly in vegetation analysis, where
combining vector-based vegetation cover with
canopy height models (CHM) is essential for
depth-enhanced segmentation. This study
tackles such integration, aiming to segment
vegetation based on height — a crucial step
for comprehensive environmental and
geographical analyses. Through the lens of
satellite and aerial imagery, vegetation
segmentation unlocks insights into vegetation
distribution, health, and variety across vast
areas. We introduce and assess three
segmentation methods: Hexagon
Segmentation, Convolution Segmentation,
and Random Point Method, prioritizing their
applicability to large-scale datasets,
potentially encompassing entire countries.
This comparative evaluation showcases the
method's precision and practicality and
advances our methodological toolkit for
environmental studies.
© O. Tsaryniuk, A. Hlybovets, 2024
ISSN 1727-4907. Проблеми програмування. 2024. №2-3
314
Великі дані (Big Data) та Аналітика даних (Data Scienсe)
1. Literature review
Image segmentation is one of the most
challenging tasks in image processing.
Currently, there are numerous approaches and
methods for image segmentation, such as the
hexagon segmentation method Hofmann &
Tiede (2014) and the Point Initialization
Approach Mueller & Corcoran (2021). Most
of the research in vegetation segmentation has
focused on identifying individual tree crowns.
This direction has been instrumental in
detailed studies of forest ecosystems, as
exemplified by the works of Douss et al.
(2022), Li et al. (2014), Lindberg et al.
(2021), and Jakubowski et al. (2013). These
studies have significantly advanced our
understanding of individual tree
characteristics, forest structure, and biomass
distribution.
In contrast to the detailed focus on
individual tree crowns, our research aims to
develop a method for generalized
segmentation that represents large arrays of
vegetation with similar (or nearly identical)
heights. This approach is well-suited for
segmenting vegetation over vast areas, such
as entire countries, addressing the need for
macro-level vegetation analysis. Such
analysis is essential for regional and national
environmental assessments, land use
planning, and large-scale conservation efforts.
Our study on vegetation segmentation
will leverage canopy height model (CHM)
data with a 10-meter resolution, as developed
by Liu et al. (2023). This CHM data is crucial
for our methodology as it provides a detailed
representation of vegetation height across
large areas. Using a 10-meter resolution
matrix allows for a fine-grained analysis of
vegetation structure, making it manageable
for large-scale applications like country-wide
segmentation.
2. Methodology
We developed three distinct methods
to address the challenge of segmenting
vegetation based on height. We aimed to
understand the complexity of accurately
determining vegetation at different altitudes
on large datasets. A series of specific metrics
were selected to assess the effectiveness and
appropriateness of these approaches. These
metrics serve as a foundation for evaluating
each method's performance, ensuring a
balanced analysis between the innovative
aspects of our methodologies and their
practical outcomes.
The following metrics were used for
comparison:
Accuracy (1). This is the ratio of
correctly identified pixels, TruePixels (2) to
the total number of pixels. It is a
straightforward measure of how accurately a
model classifies or segments pixels.
𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴 = 𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇
𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 (1)
Where: Total Number of Pixels is the
sum of all pixels within all vegetation
segments.
𝑇𝑇𝐴𝐴𝐴𝐴𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 = ∑(|ℎ𝑇𝑇𝑖𝑖𝑖𝑖𝑇𝑇𝑇𝑇(𝑝𝑝) − ℎ𝑇𝑇𝑇𝑇𝑇𝑇𝑖𝑖𝑇𝑇𝑇𝑇(𝑝𝑝)| ≤ 3) (2)
𝑖𝑖
𝑇𝑇=0
Where: ℎ𝑇𝑇𝑖𝑖𝑖𝑖𝑇𝑇𝑇𝑇(𝑝𝑝) is the height
associated with pixel p in the input data,
ℎ𝑇𝑇𝑇𝑇𝑇𝑇𝑖𝑖𝑇𝑇𝑇𝑇(𝑝𝑝) is the height associated with pixel p
in the output data, as determined by the
segmentation process.
Volume. This metric is expressed in
the number of vertices after segmentation. It
reflects the segmentation's complexity and
detail. A more significant number of vertices
usually implies a more detailed segmentation
but negatively affects the display speed and
processing.
The Hexagon segmentation method
involves creating a hexagonal grid with
uniform hexagons (each side is 100 meters
long) and generalizing the height matrix to a
3-meter interval. The vegetation vector is
clipped according to the hexagon grid to form
segments. Heights from the height matrix are
then assigned to each segment, with the most
frequent height value in the segment being
selected (using the MODE function).
Adjacent segments with the same height are
merged.
Statistics are computed for each height
value and the number of coordinates
necessary for comparing the methods.
315
Великі дані (Big Data) та Аналітика даних (Data Scienсe)
Fig. 1. Result of the hexagon
segmentation method
Like the first method, the
Convolution Segmentation Method also
generalizes the height matrix to a 3-meter
interval. The matrix is then generalized using
a convolutional filter. Several iterations with
different convolutions (7x7, 9x9) are
conducted using the "Majority" operation,
selecting the most frequently occurring value,
as in the first method. The generalized matrix
is then converted into vector polygons and
intersected with the vegetation vector.
Final statistics, including accuracy and
volume, are calculated similarly to the first
method.
Fig. 2. Result of the convolution
segmentation method
The Random Point Method is based
on creating random points within a vegetation
polygon using several approaches: 1)
Generation of random points across the
bounding box of the polygon; 2) Generating
points along the central line of the polygon; 3)
Extracting the central point of the polygon.
Utilizing different approaches for point
generation ensures an even coverage of all
types of polygons with points. The next step
involves using the ArcGIS procedure
'Generate Subset Polygons' to construct
Thiessen polygons for a given set of points.
The methodology for assigning
elevations to segments follows the approach
established in previous methods. Each
segment intersects with a generalized
elevation matrix up to 3 meters. The elevation
assigned to each segment is determined by the
most frequently occurring pixel values within
that intersection. This technique ensures
consistency in elevation assignment across
different segments, leveraging the established
practices from prior methodologies for
effective elevation mapping.
Fig. 3. Result of random point
segmentation method
3. Evaluation of the quality of the
proposed approaches
For this study, a test site covering an
area of 430 square kilometers in the western
Czech Republic was selected as the primary
focus. The data concerning vegetation heights
was sourced from a detailed 10-meter Canopy
Height Model (CHM), as elaborated in the
research conducted by Liu et al. (2023). The
vegetation data itself was derived from a
comprehensive vector dataset. This dataset
was generated through machine learning
techniques to automatically analyze high-
resolution satellite imagery, a process
meticulously carried out by the Visicom
company.
316
Великі дані (Big Data) та Аналітика даних (Data Scienсe)
Fig. 4. Research area location
The methods discussed in this article,
as well as the analysis of the results, were
implemented on PC using the Feature
Manipulation Engine (FME). The obtained
Accuracy and Volume results are shown in
Tables 1,2,3.
Table 1
Hexagon method statistics
Vegetation
Height
Accuracy
%
Total pixels
in CHM Volume
0 66.22 980
558338
3 58.92 1020
6 87.51 4485
9 94.05 28114
12 93.19 80631
15 89.95 145203
18 82.01 219782
21 80.75 343390
24 82.07 512259
27 85.62 749204
30 88.73 905916
33 90.17 517650
36 90.29 94701
39 84.83 3723
Table 2
Convolution method statistics
Vegetation
Height
Accuracy
%
Total pixels
in CHM Volume
0 76.29 949
752412
3 58.6 1256
6 69.77 8657
9 79.43 44705
12 87.72 98143
15 92.41 156412
18 95.2 238215
21 96.38 360859
24 97.26 534555
27 98.17 741339
30 98.92 836759
33 99.39 485668
36 99.65 94605
39 99.73 5176
Table 3
Random point method statistics
Vegetation
Height
Accuracy
%
Total pixels
in CHM Volume
0 65.57 909
737853
3 58.26 1567
6 83.38 6361
9 88.01 41188
12 88.55 87758
15 82.67 141607
18 82.01 213139
21 80.75 360787
24 82.07 542611
27 85.62 780707
30 88.73 905794
33 90.17 516782
36 90.29 93784
39 84.83 4905
42 86.36 374
To evaluate the segmentation's
accuracy, 3-meter height ranges were
selected. After testing various height range
options (1m, 3m, and 5m), the 3-meter range
was chosen as optimal. This selection was
based on its ability to accurately reflect the
vegetation's true height while minimizing the
amount of "noise" from individual pixels with
varying heights. This compromise ensures a
balance between precision and the reduction
of outliers, providing a more reliable
assessment of segmentation performance.
We did not consider the performance
evaluation of the segmentation methods
within the scope of this study. This decision
was based on the understanding that
performance assessments conducted on a
limited test dataset would not yield
representative results.
317
Великі дані (Big Data) та Аналітика даних (Data Scienсe)
Conclusion
The findings emphasize the potential
of integrating high-resolution satellite
imagery and LiDAR data with advanced
segmentation techniques to enhance
understanding of forest ecosystems and
vegetation distribution. The hexagon
segmentation method provides detailed
insights through a hexagonal grid,
convolution segmentation leverages
convolutional filters for generalized analysis,
and the random points method introduces a
novel segmentation approach through random
point generation and Thiessen polygons.
The research contributes to
environmental science by proposing a
scalable and efficient methodology for
vegetation analysis over large geographical
areas. Utilizing canopy height model data
with a 10-meter resolution demonstrates the
feasibility of these methods for country-wide
vegetation segmentation, highlighting their
potential for regional and national
environmental assessments, land use
planning, and conservation efforts.
The comparative analysis reveals that
each method has its merits in terms of
accuracy and volume of the final segmented
vector. The choice of method may depend on
specific research needs, available
computational resources, and the scale of the
analysis. Future work should focus on
refining these methodologies, exploring their
application in different ecological contexts,
and integrating additional data sources to
enhance the accuracy and utility of vegetation
segmentation for environmental monitoring
and management.
Considering the rapid development
and high efficiency of machine learning
methods, future development of this research
aims to incorporate AI-based approaches
alongside the methods already compared. The
introduction of the Segment Anything Model
(SAM) is planned. SAM, an innovative AI-
driven method, promises to enhance
segmentation accuracy and efficiency by
leveraging advanced machine learning
algorithms capable of adapting to various
vegetation and height delineation tasks. This
expansion will comprehensively evaluate
traditional segmentation techniques against
AI-powered models, potentially setting a new
benchmark in vegetation segmentation
methodologies.
Additionally, plans are underway to
apply the described segmentation methods to
large countrywide datasets. In this context, it
would be prudent to analyze each method's
performance speed and calculate the
computational resources required for its
implementation. This comprehensive
evaluation will ensure the methods' scalability
and efficiency when applied to extensive data
sets.
Authorship Contribution Statement
A. Hlybovets: Selection of metrics and
assessment of the complexity of the proposed
algorithms.
O. Tsaryniuk: Development and
implementation of segmentation methods.
References
1. R. Douss, I.R Farah, Extraction of
individual trees based on Canopy
Height Model to monitor the state of
the forest. Trees, Forests and People 8,
2022, doi: 10.1016/j.tfp.2022.100257.
2. P. Hofmann, D. Tiede,Image
segmentation based on hexagonal
sampling grids, South‐Eastern
European Journal of Earth
Observation and Geomatics 3, 2014
pp. 173-177.
3. M.K. Jakubowski, W. Li, Q. Guo, M.
Kelly, Delineating individual trees
from lidar data: A comparison of
vector- and raster-based segmentation
approaches. Remote Sensing 5(9),
2013, pp. 4163–4186. doi:
10.3390/rs5094163.
4. W. Li, Z. Niu, S. Gao, N. Huang, H.
Chen, Correlating the horizontal and
vertical distribution of LiDAR point
clouds with components of biomass in
a Picea crassifolia forest. Forests 5(8),
2014, pp. 1910–1930. doi:
10.3390/f5081910.
5. E. Lindberg, J. Holmgren, H. Olsson,
Classification of tree species classes in
318
Великі дані (Big Data) та Аналітика даних (Data Scienсe)
a hemi-boreal forest from
multispectral airborne laser scanning
data using a mini raster cell method.
International Journal of Applied Earth
Observation and Geoinformation 100,
2021, doi: 10.1016/j.jag.2021.102334.
6. S. Liu, et al, The overlooked
contribution of trees outside forests to
tree cover and woody biomass across
Europe. Science Advances 9(37),
2023, doi: 10.1126/sciadv.adh4097.
7. L. Ma, Y. Gao, T. Fu, L. Cheng, Z.
Chen, M. Li, Estimation of Ground
PM2.5 Concentrations using a DEM-
assisted Information Diffusion
Algorithm: A Case Study in China.
Scientific Reports 7(1), 2017, doi:
10.1038/s41598-017-14197-z.
8. J.N. Mueller, J.N. Corcoran, A
Random Point Initialization Approach
to Image Segmentation with
Variational Level-sets. 2021,
http://arxiv.org/abs/2112.12355.
9. R. Weibel, Using Vector and Raster-
Based Techniques in Categorical Map
Generalization, 1999.
Одержано: 17.04.2024
Внутрішня рецензія отримана: 25.04.2024
Зовнішня рецензія отримана: 30.04.2024
Про авторів:
Царинюк Олександр Васильович,
Phd Комп'ютерні науки,
аспірант.
https://orcid.org/0000-0003-1394-2040
Глибовець Андрій Миколайович,
доктор технічних наук,
професор,
декан факультету
https://orcid.org/0000-0003-4282-481X
Місце роботи авторів:
Національний університет
«Києво-Могилянська академія»
https://www.ukma.edu.ua/
|