Professional Documents
Culture Documents
Iecon
Iecon
Abstract— The article provides a scientific investigation of were investigated using computer vision methods [13-15].
the processing of UAV data on forest lands in Russia's European Figure 1 and Figure 2 show aerial images, which presents a
North taiga. forest area with a resolution of 5 cm/pix and 10 cm/pix,
A technique for automatic identification of tree crowns, respectively. Maintaining the Integrity of the Specifications
computation of forest parameters is discussed using the trial plot
as an example. The methodology incorporated morphological
methods for processing digital images, geographic information
tools for expressing and processing spatial data, and the findings
of top forestry experts' statistical observations.
The methodology may be used to automatically analyse
orthorectified aerial images with a resolution of 5 to 10 cm/pix.
The studies discussed in this text were conducted on images in
Russia's northern European region. The findings of the study
are utilized to automate the update of information on forest
resources on a regular basis.
I. INTRODUCTION
Fig. 1 - An example of a UAV digital imagery (5 cm per pixel)
Many scientists have published papers that used remote
sensing data to analyze taiga forests across time [1-3]. The
majority of work in detecting forest stand features using
remote sensing data has focused on medium and high
resolution satellite photos [4-8]. Furthermore, data from
spacecraft cannot be used to precisely determine forest
characteristics.
There are increasing efforts to develop new approaches for
preparing and updating forest data [9-11]. In terms of the
cartographic and attributive material collected, data
processing from UAVs is one of the most essential and cost-
effective fields. Furthermore, manually evaluating aerial
photographs to address information updating issues is labor-
intensive task. As a result, new automated methods and
processing algorithms are required.
Fig. 2 - An example of a UAV digital imagery (10 cm per pixel)
II. METHODOLOGY
First, confirm that you have the correct template for your A. Highlighting of contours
he most essential interpretative factors in aerial image analysis Consider the process of highlighting the contours of
are the size, tone, and form of tree tops [1]. These properties crowns on the example of the image shown in Figure 2.
An ASF filter (alternating sequential filtering) was applied P[n] = {(s, t)|G(s, t) < n}
to the image with a window size of 2 pixels. This image
filtering method belongs to raster morphological processing P[n] is the set of points where the surface G (x,y) lies
algorithms. The essence of this algorithm is to use the below the plane G(x,y) = n. At the moments when the relief is
morphological operations “opening” and “closing” in relation filled with water, the level rises in the form of discrete
to the processed image. increments from n = min + 1 to n = max + 1.
As is known, the operation of “opening” binary image A Let Cn (Mi) be the set of points of the basin with a local
by a structural element B is mathematically represented as a minimum Mi, which are filled with water at step n. In view of
combination of operations of “growing” and “eroding” a the above, Cn (Mi) is considered as a binary image, given by
raster: the relation:
In the above formula, the operation ⊖ “eroding” has the Let С[n] be the water-filled parts of all pools at step n,
following mathematical meaning: then:
K
A ⊖ B = {z ∈ A | Bz ⊆ A},
C[n] = ⋃ Cn (Mi )
operation "growing": i=1
A ⊕ B =⋃b∈B Ab . The execution of the watershed algorithm begins with the
initialization C[min+1] = P [min+1]. The sequence of actions
Similarly, the morphological operation “closing” has the
is performed recurrently, while it is assumed that at the n-th
following mathematical formulation:
step the set C [n-1] has already been built. To obtain the set
A • B = (A ⊕ B) ⊖ B С[n] from С[n-1], the following actions are performed. Let
Q[n] be the set of connected components of the set P [n]. Then
Based on the presented formulas, the ASF filtering for each connected component q in Q[n] there are three
algorithm for the image S is implemented as follows: possible values. The way C[n] is constructed from C[n-1]
𝐴𝑆𝐹(𝑆)𝑛𝐵 = (((((S • B) ◦ B) • 2B) ◦ 2B) . . . • nB) ◦ nB, depends on which of the conditions is true:
Using this type of filtering, we got rid of small “peaks” and a) q ∩ C[n-1] gives the empty set; this means that a new
“troughs”, i.e. flattened the original image. local minimum has been met; in this case, to construct the set
C[n], the component q is added to C[n-1].
The next step was to determine the local maxima of the
raster. Figure 3 shows the result of processing the original b) q ∩ C[n-1] contains a single connected component of
image with local maxima superimposed on it. the set C[n-1]. This condition q lies inside the basin of some
local minimum; in this case, to construct the set C[n], the
component q is also added to C[n-1].
c) q ∩ C[n-1] contains more than one connected
component of the set C[n-1]. This condition occurs when the
points of a ridge separating two or more basins meet. In this
case, the further rise of the water will lead to the merging of
the basins, therefore, inside the connected component q, a
partition must be built to prevent the basins from merging
together. The result of applying the watershed algorithm is
shown in Figure 4.
Fig. 3 – Result of ASF filtering and calculation of local maxima of the image
Let С(Mi) be the set of points in the basin corresponding The application of the watershed segmentation algorithm
to the local minimum of M i . We denote by min and max the often leads to the effect of excessive segmentation caused by
smallest and largest values of the image G (x,y). The notation noise and other local irregularities in the image. This means a
P [n] means the set of points (s, t) for which G(s,t) < n, i.e. huge number of areas allocated during segmentation. Over-
segmentation can be so significant that it renders the result
almost useless.
The approach used to manage over-segmentation is based
on the idea of markers. The marker is a connected component
belonging to the image. In our case, as markers, we set the
previously selected local maxima of the tree vertices. As a
result, we obtain a set of contours shown in Figure 5.
B. Species identification
The following step of image analysis determined if the
species of trees was deciduous or coniferous.
Model trees were marked and their locations were mapped
in the field. The average intensity of pixels in a image was
measured for a variety of species, enabling contours to be
Рис. 5 – The result of the watershed algorithm with markers classified. (Figure 8).
The result of the proposed methodology will be a set of
contours of tree crowns in the forest, obtained automatically
using the methodology described in the article.
After obtaining an image with highlighted contours of
crowns, the procedure of vectorization of the obtained
contours was carried out for the purpose of integration and
subsequent processing in a geographic information system.
The technique was applied on UAV images with a
resolution of 5 and 10 cm/pix. Vector layers of crown
contours, superimposed on the original UAV images, are
shown in Figure 6 and Figure 7.
4S
Dcr = ,
π
Several research shown a correlation between tree tops and
other taxing features in various categories of forests. The
following formula is supplied for the transition from the
diameter of the crown to the diameter of the tree trunk at a
Fig. 6 – Highlighted contours (10 cm per pixel)
height of 1.3 m (d1,3):