This document summarizes a case study comparing different methods and data sources for estimating deforestation rates in Moramanga, Madagascar. The study found that reported deforestation rates can vary widely depending on the resolution of the imagery and methods used, from 26% per year using Landsat data to 11.9% using Sentinel-2. Higher resolution RapidEye imagery estimated a rate of 15.7% per year. Accuracy assessments showed higher resolution data had lower error rates. The study concludes that higher resolution imagery is needed to accurately estimate deforestation at local scales and that standard definitions and reporting practices would help compare estimates for Madagascar.
Horizon Net Zero Dawn – keynote slides by Ben Abraham
Demystifying Deforestation Rates in Madagascar
1. DEMYSTIFYING
DEFORESTATION IN
MADAGASCAR, CASE STUDY
AT MORAMANGA
Andriambolantsoa
RASOLOHERY, helene
RALIMANANA, tianjanahary
RANDRIAMBOAVONJY, mamy
tiana RAJAONAH, stuart CABLE,
franck RAKOTONASOLO, david
RABEHEVITRA, mihajamalala,
andotiana ANDRIAMANOHERA
3. BUT …
Green and Sussman, 1990 200,000ha/year
Harper et al, 2000 90,000ha/year
MEF et al, 2009 40,000ha/year
Rakotomalala et al, 2013 100,000ha/year
Vieilledent et al, 2017100,000ha/year
GFW/WRI, 2017500,000ha/year
…
4. AND IN THE MEDIA
GFW reports 510,000ha of cover loss in 20
8. • Eastern Madagascar
• 15kmx20km(300
sqkm)
• Ununderstandably
high deforestation
rate
• Part of a REDD+
carbon project
• Next to a protected
areas
AREA OF
INTEREST
Rakotomalala et al, 2017
9. DATA
SOURCE
S
• Landsat 8 (30m)
• Sentinel 2 (10m)
• RapidEye (5m)
• WRI/GFW forest
cover loss
• GRID of points
every 500m as
references
• 2015 - 2017
10. REFERENCES
• 1218 points
• Evenly spaced every 500m
• 50% (green) to be used as
training points
• 50% (red) to be used as
validation points
• Land use value looked up
in very high resolution
images (google earth)
11. EVALUATION OF DEFORESTATION
Supervised classification
Random Forest Algorithm
Script in R
Manual digitisation and checking for the 1200 reference points,
50% used for training and 50% used for validation
Output classes : Forest, Non Forest, Deforestation
Evaluation of accuracy for each land use change map : emission error,
commission error, kappa, error margin
15. COMPARISON OF DEFORESTATION
Landsat 8
Defor 26% per year
Omission 0.34
Commission 0.64
Kappa 0.48
Area defor 2847ha
Sentinel 2
Defor 11.90% per year
Omission 0.45
Commission 0.31
Kappa 0.65
Area defor 1113ha
Rapideye
Defor 15.66% per year
Omission 0.26
Commission 0.27
Kappa 0.71
Area defor 1707ha
Reality
Defor 10.50% per year
Omission N/A
Commission N/A
Std error 0.005
Area defor 1206±234 ha
16. CONCLUSIONS
- Reported mapped areas can be overestimated (or underestimated) by
up to 50% using medium resolution imagery
- Kappa is a good measure overall but not very good for small minority
classes such as deforestation (missing all deforestation in a map and
can still have reasonable Kappa value)
- Great enhancement of accuracy is observed by using higher resolution
images
- Always seek the underlaying data for reference when reporting
deforestation rates (or mapped area), like what percent of tree cover
constitutes “forest”, 30% (GFW) or 70% (other), also the error margin
should always accompany the number being reported
- Madagascar SHOULD have a standard definition of forest,
deforestation, land use classes … that should settle the differences
17. LIMITATIONS AND NEXT STEPS
- Do not extrapolate nationally, this was in a deforestation hotspot
- High resolution imagery limited us to the area of interest, could
have done better
- No fieldwork were done, but very high resolution images from
google. The reference data could have benefitted from a few points
from the ground.
- Other classification methods and other image sources can be used
- Enhancement in classification accuracy can greatly help REDD+
project establish their baseline, and revenue
- Upscaling the analysis to a regional or national scale can be
imagined
18. THANK YOU
MISAOTRA
This study used commercial imagery courtesy of PLANET.com
(RapidEye)
Also used free imagery (Landsat 8 ) from NASA
Also used free imagery (Sentinel 2) from ESA
Also used high resolution imagery from Google Earth
Contacts:
Andriambolantsoa Rasolohery
arasolohery@ileiry.com
+261 3403 77177
Helene RALIMANANA, tianjanahary RANDRIAMBOAVONJY,
Mamy Tiana RAJAONAH, Stuart CABLE, Franck
RAKOTONASOLO, David RABEHEVITRA, Mihajamalala,
Andotiana ANDRIAMANOHERA
Editor's Notes
Wanna go old school and use the most used pick up lines ever in the history, for over 30 decades, many … a lot of publications, reports, started with it
Green and Sussman,
Harper et al,
MEF et al,
Rakotomalala et al,
Vieilledent et al
WRI/GFW
….
Reported in the media, the internet, all the social media
The problem with information getting out is that it is next to impossible to stop/change them, the 2000.000hecatres of deforestation, from study in the 1990 was used thorough 2010, and sometimes people still use that figures
L8 = 7 bands used (30m)
SE = 4 bands used (10m), the 2015 images was at the early launch of Sentinel so not possible to find cloud free images for 2015
RE = 5 bands used