Effectively monitoring deforestation is a crucial component for the success of REDD (Reducing Emissions from Deforestation and forest Degradation). In this presentation, Florian Reimer from South Pole Carbon compared techniques for monitoring small-scale deforestation in Juma reserve, exploring three classification methods: PRODES, CLASlite and ImageSVM. He found that using one model over another could avoid underdetection worth roughly US$1 million over four years, a compelling argument for careful selection of techniques depending on the characteristics of the region.
Florian Reimer gave this presentation on 8 March 2012 at a workshop organised by CIFOR, ‘Measurement, Reporting and Verification in Latin American REDD+ Projects’, held in Petropolis, Brazil. Credible baseline setting and accurate and transparent Measurement, Reporting and Verification (MRV) of results are key conditions for successful REDD+ projects. The workshop aimed to explore important advances, challenges, pitfalls, and innovations in REDD+ methods — thereby moving towards overcoming barriers to meeting MRV requirements at REDD+ project sites in two of the Amazon’s most important REDD+ candidate countries, Peru and Brazil. For further information about the workshop, please contact Shijo Joseph via s.joseph (at) cgiar.org
Small-scale deforestation monitoring in Juma REDD project
1. Small-
Small-scale deforestation
monitoring in Juma REDD project
Comparing PRODES, CLASlite and ImageSVM
CIFOR MRV Workshop
8th March 2012, Petrópolis
Florian Reimer
South Pole Carbon
2. Study Area
• 5890 km² of which roughly 4720 km²
are old-growthrainforest
• Sustainable Development Reserve (RDS)
created in 2006
• 2008 start of Bolsa Floresta and
CCBS Validation by TÜV Süd
as REDD project
• Land-use by traditional non-indigenous
communities legalized
• 339 families of traditional shifting
cultivators
• Low historic Deforestation pressure inside
the reserve, but high in southern area of
extensive cattle-ranching
4. Comparisson of Small-scale Deforestation Classifiers
• PRODES (Projeto Desmatamento) of INPE operational Wall-to-Wall
Mapping since 1988. Landsat-Pixels (30m) aggregated to 60m Pixels.
Minimal-Mapping Area > 3 ha.
• CLASlite 2.3 developed by Carnegie Institute for Science, Standford
University, Asner et al. 2009. Minimal-Mapping Area < 1 ha.
• imageSVM – Support Vector Machine developed by Rabe, van der Linden,
& Hostert of TU Berlin Geomatics. Minimal-Mapping Area < 1 ha.
• Data Basis: Landsat 5 TM images 1997, 2006, 2007, 2008, 2009, 2010
• Training & Validation data: SPOT 2,4, & 5, HRC CBERS2 & Field data
• SPOT images, ENVI & ArcGIS Licenses provided by
5. imageSVM
Pros
Can map any landcover and forest type
Can be adjusted if erroneous
Highest accuracy over all years
Cons
Needs GPS ground-truthing data
Needs high input in time, expertise and
computer power
PRODES
CLASlite v2.3 Pros
Pros Easy download
Semi-Automated Needs only GIS
Needs only georeferenced image
Works for entire Amazon Cons
Only for Brazilian Amazon
Cons Maps nothing smaller 3 ha
Maps only Deforestation Sometimes 1-2 years behind
Degradation not very reliable
Closed system, no adjustments
6. Time Step 1 Time Step 2
Single-Year Layerstacks
Multi-Year
Layerstack
e.g.
2007
& 2008
7. Every Image gets classified for Forest / Non-Forest and the maps
compared to get change „Deforestation“
8. Results
3.
1. 2.
4.
1. Deforestation in Reserve much lower than in Buffer
2. Sharp decline from 2008 to following years – stop of single large cattle ranch clearing
3. Overdetection CLASlite 2008 due to seasonally drying wetlands
4. Underdetection PRODES 2009 due to small-scale deforestation dominance
9.
10.
11. • PRODES detected only 33 % of the
deforestation detected by ImageSVM in the
Juma reserve over the observation period
(214 ha vs. 655 ha).
• The average margin of error of ImageSVM
‘New Deforestation Class’ was +/- 8.7 %
(average accuracy 91.3 %)
• 655 ha – 214 ha – (655 ha*8.7%) = 384.1 ha;
Multiply emission factor 529.43 tCO2 / ha =
203,354 tCO2
• Multiply with a price of 5 US$ / tCO2 =
1,016,505 US$.
12. Conclusion and synthesis I
• Deforestation inside the Juma Reserve has historically been low
• Trend continued during our study period, possibly effect of Reserve, PES &
REDD project
• High predicted baseline for future deforestation in the REDD project reflects
external pressure from extensive pastures south of the Reserve
• The drop of deforestation in the buffer zone after 2008 related to a slow
down in the expansion of a single large-scale land clearing east to the reserve
• External actors (e.g. cattle ranchers of Apuí) do not benefit from the REDD
project
13. Conclusion and synthesis II
• In low-deforestation, small-holder deforestation settings like the
Juma Reserve, classification must have adequate minimal-mapping
area.
• In the specific case of the Juma Reserve, using ImageSVM instead
of PRODES, would have avoided underdetection worth roughly US$
1 million over four years
• Supervised classifications fulfill the VCS REDD methodology
requirement of monitoring various forest types and land uses
• Advanced image classification (SVM) using all satellite bands can
better differentiate similar forest types or landcovers than simpler
approaches like Maximum Likelihood or using 3 bands only