Detection and analysis of forest cover dynamics with Landsat satellite imagery Application in the Romanian Carpathian Ecoregion Steven Vanonckelen Public PhD defense – 5 March 2014 Supervisor: Prof. A. Van Rompaey ARENBERG DOCTORAL SCHOOL Faculty of Science Detection and analysis of forest cover dynamics with Landsat satellite imagery 1/38
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Detection and analysis of forest cover dynamics with Landsat satellite imagery
Application in the Romanian Carpathian Ecoregion
Steven Vanonckelen
Public PhD defense – 5 March 2014
Supervisor: Prof. A. Van Rompaey
ARENBERG DOCTORAL SCHOOL
Faculty of Science
Detection and analysis of forest cover dynamics with Landsat satellite imagery 1/38
• Large dynamics
Introduction Conclusions 4 Research Questions
Forests
• Large dynamics
• Worldwide: deforestation and afforestation
• Provision of ecosystem services
Detection and analysis of forest cover dynamics with Landsat satellite imagery 2/38
Need for reliable data: – Definition of forest
– Up and downward trends
– Patchy structures
– Validation data
– Remote areas
Introduction Conclusions 4 Research Questions
Forests
Need for reliable data: – Definition of forest
– Up and downward trends
– Patchy structures
– Validation data
– Remote areas
Remote sensing data: + Recurrence time, costs, accessibility, …
- Distortions, calibration, mosaicking, …
(USGS, 2013)
Detection and analysis of forest cover dynamics with Landsat satellite imagery 3/38
Forest cover dynamics
Introduction Conclusions 4 Research Questions
Detection and analysis of forest cover dynamics with Landsat satellite imagery 4/38
Vergelijking bosbestand in centraal Europa: 900 AD en. 1900 AD (Schlüter, 1952)
Jaarlijkse snelheid van verandering in landgebruik in de Karpatan tussen 1760 -2010 (Munteanu et al., 2014)
.
Conceptual forest transition curve.
Introduction Conclusions 4 Research Questions
Forest transition
“the process in which the forested area in a given region or country changes from decreasing to expanding” (Mather and Needle, 1998)
Detection and analysis of forest cover dynamics with Landsat satellite imagery 5/38
Estimated deforestation, by type of forest and time period.
(Williams, 2002; FAO, 2010).
Forest transition
(FAO, 2010)
Introduction Conclusions 4 Research Questions
Detection and analysis of forest cover dynamics with Landsat satellite imagery 6/38
Forest transition
Recent forest transitions (Meyfroidt and Lambin, 2011).
Introduction Conclusions 4 Research Questions
Detection and analysis of forest cover dynamics with Landsat satellite imagery 7/38
Scientific challenges?
Reliable forest assessments
• standardized procedures at regional scale, • link with ecosystem services, • allow policies and financial compensations.
Detection of forest cover dynamics
• fluctuating forest cover trends and patchy landscapes, • in remote and inaccessible areas, • inclusion of translocation processes.
Use of satellite data (remote sensing tools)
• different remote sensors, • improvement of mapping procedures, • allows mapping of large areas, • trade-off: complexity automation.
(FAO, 2010)
(FAO, 2010)
Introduction Conclusions 4 Research Questions
Detection and analysis of forest cover dynamics with Landsat satellite imagery 8/38
4 Research questions
1. Do atmospheric and topographic corrections improve Landsat satellite imagery in mountain areas?
2. Does image preprocessing (corrections) lead to more accurate land cover classification?
Introduction 4 Research Questions Conclusions
(Riaño et al., 2003)
(ASTRIUM, 2013)
Detection and analysis of forest cover dynamics with Landsat satellite imagery 9/38
4 Research questions
3. Does topographic correction and pixel-based image compositing improve large area change mapping? 4. What are the controlling factors of forest cover dynamics in the Romanian Carpathians?
Introduction 4 Research Questions Conclusions
Detection and analysis of forest cover dynamics with Landsat satellite imagery 10/38
1. Do atmospheric and topographic corrections improve Landsat satellite imagery in mountain areas?
Problems with satellite imagery:
• Sensor calibration,
• Orbital drift,
• Atmospheric and topographic distortions correction
decomposition and evaluation
Atmospheric correction:
Introduction Research Question 1 Conclusions
Process of solar (ir)radiance entering a sensor: (a) path radiance, (b) solar direct irradiance, (c) sky diffuse irradiance, and (d) adjacent terrain reflected irradiance.
Detection and analysis of forest cover dynamics with Landsat satellite imagery 11/38
Selection of 2 atmospheric correction (AC) methods (Vanonckelen et al., 2013).
complex
Introduction Research Question 1 Conclusions
Topographic correction:
15 combined corrections
Detection and analysis of forest cover dynamics with Landsat satellite imagery 12/38
1. Do atmospheric and topographic corrections improve Landsat satellite imagery in mountain areas?
Selection of 4 topographic correction (TC) methods (Vanonckelen et al., 2013).
complex
AC/TC No TC Band Ratio Cosine PBM PBC
No AC
DOS
TF
Study area
One Landsat footprint in the Romanian Carpathians (185 x 185 km, 2009).
(a) Mixed and broadleaved forests, (b) coniferous forests, (c) small vegetation, and (d) grasses above the tree line.
Introduction Research Question 1 Conclusions
Detection and analysis of forest cover dynamics with Landsat satellite imagery 13/38
1. Do atmospheric and topographic corrections improve Landsat satellite imagery in mountain areas?
Elevation range (SRTM, 2010).
Methodology
Overview of the methodology (Vanonckelen et al., accepted).
Introduction Research Question 1 Conclusions
Detection and analysis of forest cover dynamics with Landsat satellite imagery 14/38
1. Do atmospheric and topographic corrections improve Landsat satellite imagery in mountain areas?
Evaluation methods
a. Differences in reflectance values between shaded
and illuminated slope groups,
b. Coefficient of variation,
c. Correlation between reflectance and cos β.
Average reflectance (%) calculated in the forest class as a function of spectral band (Vanonckelen et al., accepted).
Introduction Research Question 1 Conclusions
Detection and analysis of forest cover dynamics with Landsat satellite imagery 15/38
1. Do atmospheric and topographic corrections improve Landsat satellite imagery in mountain areas?
True color composite images (RGB: band 3, 2 and 1) of the study area (Vanonckelen et al., accepted).
(a) no AC or TC (b) TF with cosine (c) TF with PBC
Evaluation methods
a. Differences in reflectance values between shaded and illuminated slope groups,
b. Coefficient of variation of reflectance values,
c. Correlation between reflectance values and cos β.
Yes!
Introduction Research Question 1 Conclusions
(a) no AC or TC (b) DOS without TC (c) DOS with band ratio
(d) TF with cosine (e) TF with PBM (f) TF with PBC
True color composite images (RGB: band 3, 2 and 1) of the study area (Vanonckelen et al., accepted).
Detection and analysis of forest cover dynamics with Landsat satellite imagery 16/38
1. Do atmospheric and topographic corrections improve Landsat satellite imagery in mountain areas?
2. Does image preprocessing lead to more accurate land cover classification?
Methodology
2009 and 2010
Overview of the methodology (Vanonckelen et al., 2013).
Introduction Research Question 2 Conclusions
Detection and analysis of forest cover dynamics with Landsat satellite imagery 17/38
2. Does image preprocessing lead to more accurate land cover classification?
Evaluation methods
a. Separability of reflectance values,
b. Overall accuracy,
c. Class accuracy,
d. Accuracy in 3 illumination conditions.
Introduction Research Question 2 Conclusions
TF-P
BM
ref
lect
ance
(%
)
Wavelength (µm)
0
5
10
15
20
25
30
35
40
0.49 0.56 0.66 0.82 1.65 2.25
0
5
10
15
20
25
30
35
40
0.49 0.56 0.66 0.82 1.65 2.25
Un
corr
. ref
lect
ance
(%)
(a) (b)
Wavelength (µm)
Detection and analysis of forest cover dynamics with Landsat satellite imagery 18/38
Average reflectance values per wavelength and land cover type : uncorrected and corrected ( Vanonckelen et al., 2013).
2. Does image preprocessing lead to more accurate land cover classification?
Evaluation methods
a. Separability of reflectances,
b. Overall accuracy (average kappa),
c. Class accuracy,
d. Accuracy in 3 illumination conditions.
Introduction Research Question 2 Conclusions
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
No
TC
Ban
d ra
tio
Co
sin
e
PB
C
PB
M
No
TC
Ban
d ra
tio
Co
sin
e
PB
C
PB
M
No
TC
Ban
d ra
tio
Co
sin
e
PB
C
PB
M
No AC DOS TF
Ave
rage
kap
pa
Detection and analysis of forest cover dynamics with Landsat satellite imagery 19/38
Average kappa coefficients of the 2009 and 2010 images (Vanonckelen et al., 2013).
2. Does image preprocessing lead to more accurate land cover classification?
Evaluation methods
a. Separability of reflectance values,
b. Overall accuracy,
c. Class accuracy (δkappa values),
d. Accuracy in 3 illumination conditions.
OVERALL
Average 2009-2010 δkappa per class between uncorrected and corrected images (Vanonckelen et al., 2013).
Introduction Research Question 2 Conclusions
Detection and analysis of forest cover dynamics with Landsat satellite imagery 20/38
2. Does image preprocessing lead to more accurate land cover classification?
Evaluation methods
a. Separability of reflectance values,
b. Overall accuracy,
c. Class accuracy,
d. Accuracy in 3 illumination conditions (average kappa).
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
No
TC
Ban
d ra
tio
Co
sin
e
PB
C
PB
M
No
TC
Ban
d ra
tio
Co
sin
e
PB
C
PB
M
No
TC
Ban
d ra
tio
Co
sin
e
PB
C
PB
MNo AC DOS TF
Ave
rage
Kap
pa
Introduction Research Question 2 Conclusions
Detection and analysis of forest cover dynamics with Landsat satellite imagery 21/38
Average kappa coefficients of the 2009 and 2010 images for three different illumination characteristics (Vanonckelen et al., 2013).
2. Does image preprocessing lead to more accurate land cover classification?
Evaluation methods
a. Separability of reflectance values,
b. Overall accuracy,
c. Class accuracy,
d. Accuracy in 3 illumination conditions.
Yes!
Introduction Research Question 2 Conclusions
Detection and analysis of forest cover dynamics with Landsat satellite imagery 22/38
True color composite (RGB: band 3, 2 and 1) and ML classification of 2009 (Vanonckelen et al., 2013).
(c) TF with PBC
(a) no AC or TC
(b) TF with cosine
3. Does topographic correction and pixel-based image compositing improve large area change mapping?
Study area: Romanian Carpathian Ecoregion
Introduction Research Question 3 Conclusions
Detection and analysis of forest cover dynamics with Landsat satellite imagery 23/38
Elevation range (Vanonckelen et al., submitted)
Pixel-based compositing:
- Selection of the most suitable pixels from available imagery,
- For three years: 1985, 1995 and 2010,
Topographic correction: PBM,
Two classification methods: ML and SVM.
(Vanonckelen et al., submitted)
3. Does topographic correction and pixel-based image compositing improve large area change mapping?
Introduction Research Question 3 Conclusions
Detection and analysis of forest cover dynamics with Landsat satellite imagery 24/38
(Vanonckelen et al., submitted)
3. Does topographic correction and pixel-based image compositing improve large area change mapping?
Methodology
Introduction Research Question 3 Conclusions
Detection and analysis of forest cover dynamics with Landsat satellite imagery 25/38
Overview of the methodology.
a. Overall accuracy for different scenarios:
- uncorrected and corrected,
- 2 classifiers: ML and SVM,
- number of classes: 8 or 4.
Yes!
b. Land cover maps,
c. Land cover change maps.
3. Does topographic correction and pixel-based image compositing improve large area change mapping?
Introduction Research Question 3 Conclusions
Detection and analysis of forest cover dynamics with Landsat satellite imagery 26/38
Overall accuracies between uncorrected (uncorr.) and corrected (corr.) data.
1985
3. Does topographic correction and pixel-based image compositing improve large area change mapping?
Introduction Research Question 3 Conclusions
Detection and analysis of forest cover dynamics with Landsat satellite imagery 27/38
2010 1995 b.
1985-1995 1995-2010
3. Does topographic correction and pixel-based image compositing improve large area change mapping?
Introduction Research Question 3 Conclusions
(foto: Deliever, 2011)
Detection and analysis of forest cover dynamics with Landsat satellite imagery 28/38
c.
1995-2010
3. Does topographic correction and pixel-based image compositing improve large area change mapping?
Introduction Research Question 3 Conclusions
2000-2010
(Hansen et al., 2013)
vs.
Detection and analysis of forest cover dynamics with Landsat satellite imagery 29/38
c.
a. Accessibility, distance to • primary roads, • secondary roads, • nearby settlement.
b. Demography
• demographic evolution.
c. Land use policy
• protection level (SCI or SPA).
d. Biophysical environment
• slope gradient, • elevation, • soil type (7 types).
4. What are the controlling factors of forest cover dynamics in the Romanian Carpathians?
Introduction Research Question 4 Conclusions
(foto’s: Deliever, 2011) Afforestation (AFFOR)
Detection and analysis of forest cover dynamics with Landsat satellite imagery 30/38
Deforestation (DEFOR)
1985-1995 1995-2010
a. Accessibility, distance to • primary roads + corr. for AFFOR, - corr. for DEFOR • secondary roads - corr. for AFFOR and DEFOR • nearby settlement + corr. for AFFOR, - corr. for DEFOR
b. Demography
• demographic evolution.
c. Land use policy
• protection level (SCI or SCA).
d. Biophysical environment
• slope gradient, • elevation, • soil type (7 types).
4. What are the controlling factors of forest cover dynamics in the Romanian Carpathians?
Introduction Research Question 4 Conclusions
Primary and secondary roads (European Commission, 2013b).
Detection and analysis of forest cover dynamics with Landsat satellite imagery 31/38
a. Accessibility, distance to • primary roads, • secondary roads, • nearby settlement.
b. Demography
• demographic evolution + corr. for AFFOR and DEFOR
(2nd per.)
c. Land use policy • protection level (SCI or SPA).
d. Biophysical environment
• slope gradient, • elevation, • soil type (7 types).
4. What are the controlling factors of forest cover dynamics in the Romanian Carpathians?
Introduction Research Question 4 Conclusions
(foto’s: Deliever, 2011)
Detection and analysis of forest cover dynamics with Landsat satellite imagery 32/38
Change in inhabitants per km2 between 1986-2000 (NIS, 2013).
a. Accessibility, distance to • primary roads, • secondary roads, • nearby settlement.
b. Demography
• demographic evolution.
c. Land use policy
• protection level (SCI or SPA) + corr. for DEFOR (2nd per.)
d. Biophysical environment
• slope gradient, • elevation, • soil type (7 types).
4. What are the controlling factors of forest cover dynamics in the Romanian Carpathians?
Introduction Research Question 4 Conclusions
Protected area s (European Commission, 2013b).
Detection and analysis of forest cover dynamics with Landsat satellite imagery 33/38
a. Accessibility, distance to • primary roads, • secondary roads, • nearby settlement.
b. Demography
• demographic evolution.
c. Land use policy
• protection level (SCI or SPA).
d. Biophysical environment
• slope gradient + corr. for AFFOR and DEFOR • elevation + corr. for AFFOR and DEFOR • soil type (7 types) Podzols + corr. for DEFOR
4. What are the controlling factors of forest cover dynamics in the Romanian Carpathians?
Introduction Research Question 4 Conclusions
Detection and analysis of forest cover dynamics with Landsat satellite imagery 34/38
Elevation range (SRTM, 2010).
Main soil types (FAO/UNESCO, 1998).
Conclusions
1. Do atmospheric and topographic corrections improve Landsat satellite imagery in mountain areas?
Yes, homogeneity of pixels increased and dependency of reflectance values on terrain illumination was reduced after correction.
2. Does image preprocessing lead to more accurate land cover classification? Yes, corrected images resulted in higher overall classification accuracies than uncorrected images, especially on weakly illuminated slopes.
Introduction Conclusions 4 Research Questions
Detection and analysis of forest cover dynamics with Landsat satellite imagery 35/38
Conclusions
3. Does topographic correction and pixel-based image compositing improve large area change mapping? Yes, highest classification accuracies for corrected scenario with 4 classes and observation of a steady greening between 1985 and 2010. 4. What are the controlling factors of forest cover dynamics in the Romanian Carpathians? Mainly positively related to the biophysical environment (elevation and slope gradient) for both periods.
Introduction Conclusions 4 Research Questions
Detection and analysis of forest cover dynamics with Landsat satellite imagery 36/38
Further research
• Improvement of forest maps and monitoring of ecosystem services,
• Basis for policy instruments and conservation measures,
• Comparative analyses of corrections,
• Evaluation of trade-offs between correction complexity and automation,
• Topographic correction before pixel-based compositing,
• Addition of extra parameters in the pixel-based compositing technique.
Introduction Conclusions 4 Research Questions
Detection and analysis of forest cover dynamics with Landsat satellite imagery 37/38