=Paper=
{{Paper
|id=Vol-3006/07_short_paper
|storemode=property
|title=Optical and microwave observation of forest restoration after abnormal fires
|pdfUrl=https://ceur-ws.org/Vol-3006/07_short_paper.pdf
|volume=Vol-3006
|authors=Aleksey V. Dmitriev,Tumen N. Chimitdorzhiev,Pavel N. Dagurov,Irina I. Kirbizhekova
}}
==Optical and microwave observation of forest restoration after abnormal fires==
https://ceur-ws.org/Vol-3006/07_short_paper.pdf
Optical and microwave observation of forest
restoration after abnormal fires
Aleksey V. Dmitriev1 , Tumen N. Chimitdorzhiev1 , Pavel N. Dagurov1 and
Irina I. Kirbizhekova1
1
Institute of Physical Materials Science, SB RAS, Ulan-Ude, Russia
Abstract
The paper discusses the possibility of a forest undergrowth detecting with help of polarimetric decompo-
sitions and time series analysis of ALOS-2/PALSAR-2 and Sentinel-1 radar images. The presented results
show principal possibility of confidently identifying the growth of forest stands.
Keywords
Reforestation, polarimetric decompositions, time-series analysis, ALOS-2/PALSAR-2, Sentinel-1/2.
1. Introduction
Modern climate changes are largely associated with the problem of carbon balance, the main
source of which is tropical and boreal forests. Deforestation and wildfires have significantly
reduced the area of the forest. Russian Federation has a federal project named “Forest Conser-
vation” currently, the purpose of which is reforestation. Territories suitable for reforestation
are preliminarily identified using data from multispectral optical sensors within the framework
of this project. Next, field observations and soil analysis are performed for further planting
of seedlings. However, due to the specifics of optical images, territories for planting forest
seedlings are often mistakenly allocated, which, as a result of self-seeding, are already covered
with forest. Identification of the natural restoration of the forest is difficult due to low-growing
young deciduous and rare coniferous trees are poorly distinguishable against the background
of dense grass and shrubs. This report examines the possibility of identifying undergrowth of
forest stands by analyzing multitemporal polarimetric radar images of ALOS-2 PALSAR-2. For
comparison, the results obtained from the time series of the Sentinel-1 C-band radar with dual
polarization are also presented.
2. Study area description
Study area is located at Eastern Siberia, Russia, east of the Lake Baikal in central part of Buryatia
Republic (see Figure 1). It is a mountainous hard-to-reach region covered by different types
of forest. A strong wildfire in the summer of 2007 destroyed the forest on a large territory
of this site. Four different sites were selected for detailed study within study area, which are
SDM-2021: All-Russian conference, August 24–27, 2021, Novosibirsk, Russia
" dmitriev.aleksey@gmail.com (A. V. Dmitriev)
© 2021 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
Workshop
Proceedings
http://ceur-ws.org
ISSN 1613-0073 CEUR Workshop Proceedings (CEUR-WS.org)
47
Aleksey V. Dmitriev et al. CEUR Workshop Proceedings 47–55
a b
Figure 1: Study area on the Google Maps and locations of the regions of interest (roi).
a b
Figure 2: Pictures of roi_1 (a) and roi_3 (b).
shown in the Figure 1, b as roi_1–roi_4. roi_1 is the overgrown territory of forest after a fire
with approximate area equaled to 12 Ha (see Figure 2, a). roi_2 is the rocky treeless site. Before
abnormal wildfire it was partially covered by cedar trees which were completely destroyed by
the strong fire. roi_3 represents wetlands with shrubs located in floodplain of the river (see
Figure 2, b). roi_4 is the mature pine forest not affected by wildfires.
Areas of burned forest are usually surrounded by surviving forest, due to which natural
reforestation takes place. Also, small groups of surviving mature trees are natural sources of
seeds in the burned areas.
3. Research methodology and data used
Several different methods were used to identify the undergrowth of forest stands.
48
Aleksey V. Dmitriev et al. CEUR Workshop Proceedings 47–55
1. Analysis of multitemporal polarimetric decompositions that distinguish the dominant
scattering mechanisms. For example, volume scattering from the forest stands, surface
from the rough surface, double-bounce from natural targets. Since the forest vegetation is
destroyed after abnormal wildfires, it is assumed that there is a double-bounce scattering
(for example, scattering from “dihedral” reflector formed by the earth’s surface and the
remains of a tree trunk) and a surface scattering. After self-seeding and the growth of new
trees, it is assumed that the volume scattering component will increase and the surface
and double-bounce scattering components will decrease. Model-based decomposition,
proposed by Freeman and Durden [1] and Claude – Pottier eigenvalues/eigenvectors
decomposition [2] were used for analysis.
2. Analysis of temporal variations of radar backscattering amplitude and interferometric
coherence at different polarizations and in different bands. It is expected that the growth
of forest stands will lead to an increase in the amplitude of the radar signal due to an
increase in the effective “roughness” of the reflecting surface. At the same time, there
should be a decrease in the interferometric coherence.
3. Normalized difference vegetation index (NDVI) time series analysis of Sentinel-2A/B
satellites to obtain dynamics of overgrowing forestless areas. The analysis was made with
help of Google Earth Engine cloud platform.
We used data of Japanese L-band synthetic aperture radar (SAR) PALSAR-2 aboard the ALOS-
2 satellite which was launched in May, 2014. ALOS-2/PALSAR-2 has different observation modes
including two modes with full (quad) polarization. Unfortunately, only two fully polarimetric
scenes were available for the study area. Therefore, seven images with dual polarization
(HH/HV) were selected for processing (see Table 1).
Also, we used data of Sentinel-1B SAR. Sentinel-1B is a European radar imaging satellite
launched on 25 April 2016. It is the second of two satellites in the Sentinel-1 constellation.
The satellite carries a C-band sensor (wavelength is equal 5.6 cm) operating in four exclusive
imaging modes with different resolution and coverage. It provides dual polarization capability
(VV/VH) and 12-days revisit time. We chose data imaged from relative orbit 62 for regions of
interest. 111 IW SLC images from February 25, 2017 to February 28, 2021 were processed totally.
Sentinel-2 optical multispectral images were used to create multitemporal NDVI time series
for test sites. The Sentinel-2 mission is a constellation with two twin satellites, Sentinel-2A
Table 1
ALOS-2/PALSAR-2 data.
Mode HBQ (Quad pol.) FBD (Dual pol.)
ALOS2018861050-140928
ALOS2062331050-150719
ALOS2116151050-160717
ALOS2159621050-170507
Scene IDs ALOS2169971050-170716
ALOS2283821050-190825
ALOS2223791050-180715
ALOS2277611050-190714
ALOS2331431050-200712
49
Aleksey V. Dmitriev et al. CEUR Workshop Proceedings 47–55
(launched on June 23, 2015) and Sentinel-2B (launched on March 7, 2017). The satellites each
carry a single multi-spectral instrument (MSI) with 13 spectral channels in the visible/near
infrared (VNIR) and short wave infrared spectral range (SWIR) for surveying with a resolution
of 10 to 60 m. The presence of two satellites in the mission allow repeated surveys every 5 days.
More than one hundred cloudless images were used for NDVI time series calculation.
All SAR’s data were processed in ESA SNAP 8.0, free and open source software [3]. Calculation
of vegetation index was performed with help of Google Earth Engine cloud platform [4].
4. Results and discussion
4.1. ALOS-2/PALSAR-2
4.1.1. Fully polarimetric data
Let us consider polarimetric decomposition, proposed by Freeman and Durden [1]. It is based
on the three-component scattering mechanism model. The span P of radar return is assumed to
be the sum of three independent terms: the contribution of a single (surface), double-bounce (or
even-bounce) and a volume (multiple) scattering components:
𝑃 = |𝑆𝐻𝐻 |2 + 2|𝑆𝐻𝑉 |2 + |𝑆𝑉 𝑉 |2 = 𝑃𝑠 + 𝑃𝑑 + 𝑃𝑣 ,
where 𝑠, 𝑑 and 𝑣 indexes denote single, double-bounce and volume scattering respectively.
Surface (single) scattering is a typical mechanism of scattering from soil or water. Double-
bounce scattering occurs when there are vertically oriented targets on a horizontal surface and
is often found in built-up areas (reflections from the ground and walls), as well as in forests
(reflections from trunks and soil). Canopy (volume) scattering occurs where there is a cloud of
randomly oriented, very thin, cylinder-like scatterers — for example, tree branches.
The Table 2 shows the results of three-component Freeman – Durden decomposition applied to
fully polarimetric ALOS-2/PALSAR-2 data. The values in the table are expressed as a percentage.
First value in cells corresponds to May 7, 2017 (ALOS2159621050-170507), second one to May
25, 2019 (ALOS2283821050-190825). All results are mean values calculated for respective roi.
It is evident, that volume scattering increased at the sites roi_1, roi_2, roi_3. The maximum
increase about 20% is observed for overgrown area of forest after fire (roi_1) and rocky treeless
area (roi_2). In the first case, the forest undergrowth grew up by 0.5–1.5 m over two years with
a corresponding increase in the total number of scatterers, such as tree trunks and branches.
The increased forest biomass leads to surface “screening” effect resulting in decreased surface
Table 2
Results of three-component Freeman – Durden decomposition.
roi_1 roi_2 roi_3 roi_4
𝑃𝑑 , % 5.9/5.3 5.4/4.5 5/7.9 7.6/6.2
𝑃𝑠 , % 49.2/29 52/33.5 66/49.8 7.3/8.8
𝑃𝑣 , % 44.9/65.7 42.6/62 29/42.3 85.1/85
50
Aleksey V. Dmitriev et al. CEUR Workshop Proceedings 47–55
scattering component by the same 20%. Growth of volume scattering component at roi_3
mainly associated with shrubs density increasing. The scattering components variations are
not significant for the mature forest.
4.1.2. Dual polarimetric data
Unfortunately, due to small amount of fully polarimetric data (only 2 scenes) we can’t analyze
dynamics of overgrowing. But there are more images made in FBD mode (HH/HV polarizations)
which we can analyze (see Table 1).
For dual polarimetric data we can only perform Cloude – Pottier decomposition [3] which
is based on the analysis of the coherency matrix eigenvalues (𝐻𝛼 decomposition). The 𝐻
parameter called entropy, describes the randomness degree of scattering. High entropy values
usually indicate the presence of vegetation. The angle 𝛼 indicates the radar backscattering
mechanism.
Figure 3 shows entropy and 𝛼 time series calculated from the data of ALOS-2/PALSAR-2 in
FBD observation mode. The data obtained was averaged within the test site to get one value
per roi for each date.
The plots in Figure 3 show that neither the entropy nor the 𝛼 parameter estimated from dual
polarization data underwent changes significantly over 5 years. So, they cannot serve as the
indicators of the reforestation process.
4.1.3. Amplitude and coherence time series
Since the polarimetric decomposition of the dual polarimetric data did not reveal considerable
changes in the test areas, it was decided to analyze other characteristics of the scattered signal
such as amplitude and coherence. Figure 4 shows the temporal dependencies of radiometrically
calibrated and normalized amplitude and sequential interferometric coherence (first date with
second, second with third etc.) for different polarization states.
The amplitude plots show that it is almost impossible to distinguish forest undergrowth from
other areas. For example, the correlation between the plot for roi_1 (forest undergrowth) and
the plot for roi_2 (rocky treeless area) is 0.92 for HH polarization and 0.96 for HV. The only area
that can be identified by the amplitude plots is the mature forest (roi_4). The radar signal at
cross-polarization is on average 4.5 dB higher than at other sites due to scattering on large tree
branches. The situation is similar in terms of coherence. The sequential coherence on roi_1 has
a b
Figure 3: Temporal dependencies of entropy (a) and a angle (b) for dual polarization data
of ALOS-2/PALSAR-2.
51
Aleksey V. Dmitriev et al. CEUR Workshop Proceedings 47–55
a b
c d
Figure 4: Temporal dependencies of radiometrically calibrated and normalized amplitude (a, b) and
sequential coherence (c, d).
a significant positive correlation (greater than 0.55) with all other sites. Thus, the L-band data
with dual polarization is not suitable for monitoring the reforestation process. This is mainly
due to the high penetration of L-band waves. A low growing forest does not attenuate the radar
backscattering significantly at a time interval of 7 years, even despite its relative density nor
does it lead to visible depolarization effects.
4.2. Sentinel-1B
Due to the inability to reveal the forest restoration in the L-band, it was decided to use the freely
available data of the European Sentinel-1B all-weather radar operating in C-band (wavelength
5.6 cm). Sentinel-1B survey frequency for the territory of Siberia is 12 days. Regular survey of
the study area has been conducted since 2017. Sentinel-1 radar emits vertically polarized waves
(VV and VH) unlike the ALOS-2 PALSAR-2, which mainly operates at horizontal polarization
on radiation (HH and HV).
4.2.1. 𝐻/𝛼 decomposition
Figure 5 shows the calculated time series of entropy and 𝛼 angle for each roi.
a b
Figure 5: Temporal dependencies of entropy (a) and 𝛼 angle (b) for Sentinel-1B data.
52
Aleksey V. Dmitriev et al. CEUR Workshop Proceedings 47–55
In this case, the 𝐻/𝛼 polarimetric decomposition allows us to separate areas with vegetation
from a treeless area by the entropy value and the scattering mechanism. However, as before,
the vegetation areas are not divided among themselves. It should also be noted that there is a
distinct seasonal dependence of both the entropy and the 𝛼 angle.
4.2.2. Amplitude and coherence time series
Let us consider temporal dependencies of radiometrically calibrated and normalized backscatter-
ing amplitude. Due to the smaller wavelength in C-band the radar backscattering is very sensitive
to changes in the underlying media, for example, in variations of soil moisture, freezing/thawing
processes, etc. All these factors strongly influence to the magnitude of the amplitude, which is
especially noticeable in the summer and autumn-spring periods (see the seasonal dependence
in Figure 5), and are expressed in a large variance of the measurement results. Therefore, to
assess the reforestation processes, it was decided to consider more stable winter data. A total
of 38 winter images were used, taken between the beginning of November of one year until
the end of February of the year following it. The obtained results are presented in Figure 6.
The plots in Figure 6, a for the vertical co-polarization shows continuous increasing of the
backscattered signal for roi_1 and roi_3. A comparison of the average amplitude value for the
winter period of 2017–2018 with the same period of 2020–2021 shows it growth up by 11% and
12% for roi_1 and roi_3 respectively.
It clearly visible that the amplitude trend lines are almost parallel for these areas. It indicates
similar scattering processes. The radar backscattering also increases for roi_2 and roi_4, but it
is significantly less in quantitative terms (<2% for roi_2 and 6% for roi_4). The dependencies for
cross-polarization (Figure 6, b) show a similar trend for roi_1 and roi_3, but are less expressed
a b
c d
Figure 6: Temporal dependencies of radiometrically calibrated and normalized amplitude (a, b) and
sequential coherence (c, d) during winter.
53
Aleksey V. Dmitriev et al. CEUR Workshop Proceedings 47–55
quantitatively (<4% for both sites). roi_4 should be noted individually because it demonstrates
the stability of the average reflective properties over 4 years (variations are equal to −0.6%).
The behavior of radar backscattering on roi_2 is decreasing in contrary to co-polarization case.
It is interesting to compare roi_2 and roi_3 within the same winter season. As can be clearly
seen for the seasons 2017/2018, 2018/2019 and 2020/2021, the radar backscattering consistently
increases during the winter. Obviously, this is due to snow cover growth (which can exceed 1 m
in a given area) and an increase in the volume scattering component in the total reflected signal.
However, if roi_2 is a rocky treeless area, then roi_3 is the shrubs in the floodplain of river.
So, we can conclude that in winter, when the shrub sheds its foliage and the soil freezes, the
backscattering from the site with deciduous undergrowth becomes similar to the backscattering
from a treeless site, i.e. deciduous young growth becomes “radio-transparent”. Coniferous forest
undergrowth on roi_1 does not show such behavior. This can serve as an indirect indicator for
the classification of forest undergrowth by vegetation type.
Figures 6, c and d show the coherence time series. The highest coherence values are observed
in the treeless area as expected. However, no unambiguous conclusions can be drawn from
these plots for the rest of the forest sites.
Thus, the analysis of the temporal series of radiometrically calibrated and normalized backscat-
tering amplitude in C-band allow us to reliably detect forest undergrowth of both coniferous and
deciduous species and distinguish it from mature forest and treeless areas. Forest undergrowth
can also be divided by species composition with help of indirect features.
We must point out that all these conclusions are true for the low height undergrowth only. If
the trends shown in Figures 6, a and b will proceed, then saturation effect will occur in 5–6 years,
when the coniferous undergrowth reaches a height of 2.5–3 m and the undergrowth will be
impossible to distinguish from the mature forest. However, the increased forest biomass should
affect the backscattering in L-band in this case.
4.3. Sentinel-2A/B
As an additional step we investigated normalized difference vegetation index (NDVI) obtained
from Sentinel-2A/B satellites data. Calculation of NDVI was performed with help of Google
Earth Engine cloud platform. More than one hundred cloudless images were used for NDVI
time series calculation. (see Figure 7).
Figure 7: Temporal dependencies of NDVI.
54
Aleksey V. Dmitriev et al. CEUR Workshop Proceedings 47–55
It is evident that NDVI can’t identify young growth. But it can help separate different types
of undergrowth. For example, shrubs at roi_3 have a higher value of NDVI than young growth
at roi_1. Rocky treeless area (roi_2) has NDVI less than 0.3 during summer period. It means
that increased volume scattering, as was shown by polarimetric decomposition (see Table 2),
not essentially related with vegetation. Vegetation index at roi_1 almost follows NDVI values
at roi_4, where pine forest is located. So, we can conclude that undergrowth at roi_1 contains
predominantly young pines.
5. Conclusion
The paper discusses the possibility of a forest undergrowth detecting with help of time series
analysis of ALOS-2/PALSAR-2 and Sentinel-1 radar images. It is shown that model-based
decompositions of fully polarimetric data in L-band (ALOS-2/PALSAR-2) allow us to confidently
identify the growth of forest stands. However, data with double polarization are no longer able
to separate forest undergrowth from other types of vegetation. On the contrary, the C-band
(Sentinel-1) data with a vertical co-polarization taken in the winter period make it possible to
determine the forest undergrowth by virtue of time series analysis of radiometrically calibrated
and normalized amplitude. The use of optical sensors data, in particular Sentinel-2A/B, allows
us to classify the identified undergrowth by species composition and eliminate classification
errors in treeless areas, which show increased volume scattering on model-based polarimetric
decompositions.
Acknowledgments
The authors would like to thank Japanese aerospace agency (JAXA) for ALOS-2/PALSAR-2 data
granted under ALOS-2 RA6 (PI-3092). The research was carried out within the state assignment
of Ministry of Science and Higher Education of the Russian Federation and partially supported
by the Russian Foundation for Basic Research (grant No. 18-47-030001-r_a).
References
[1] Freeman A., Durden S.L. A three-component scattering model for polarimetric SAR data //
IEEE Trans. Geosci. Remote Sens. 1998. Vol. 36. No. 3. P. 963–973.
[2] Cloude S.R., Pottier E. An entropy based classification scheme for land applications of
polarimetric SAR // IEEE Trans. Geosci. Remote Sens. 1997. Vol. 35. P. 68–78.
[3] ESA SNAP. URL: https://step.esa.int (accessed June 23, 2021).
[4] Google Earth Engine. URL: https://earthengine.google.com (accessed June 23, 2021).
55