Title: Landsat Change Detection of Forests with a Modified Enhancement Classification Methodology
1Landsat Change Detection of Forests with a
Modified Enhancement Classification Methodology
- Jean Beaubien, Nick Walsworth and Don Leckie
- Canadian Forest Service
- Natural Resources Canada
- Presented by Don Leckie
Third International Workshop on the Analysis of
Multi-temporal Remote Sensing Images 16-18 May
2005 Biloxi, Mississippi USA
2Next Generation Forest Measuring and Monitoring
System for Canada
NFIS
3NFI Photo Plot Design
- Photo Plots
- Provide estimates of basic attribute data
- Sample units are 2 x 2 km on a 20 km grid
- Located at network points
- Air photos for Southern
- Satellite data for Northern
4(No Transcript)
5NFI Remeasurement Overview
- NFI report is expected in 2005 or 2006
- Remeasure on 10-year cycle
- Southern Canada measure one photo plot panel
one ground plot panel/year - Northern Canada measure all plot panels every 5
years - Update all photo plots annually for significant
disturbances - Produce annual estimates and publish periodic
reports
6Earth Observation for Sustainable Development of
Forests (EOSD) National Forest Inventory Product
- Core NFI data in sections (e.g., north)
- Confirm/Verify the Inventory
- Assess Change
- Indicate update priorities
- Accuracy assessment
- 2 x 2 km photo plot based on the centroid of a
20 x 20 km grid
7Overview, EOSD Land Cover Quebec
8EOSD Land Cover Product
- Land cover tile over Chibougamau
9EOSD Biomass Product (preliminary)
10 EOSD Change Product (preliminary)
11Deforestation Overview
- Pilots
- Methods
- EOSD change
- Carbon modeling
- Kyoto Protocol reporting
12Deforestation and EOSD Change Pilot Project and
RD Overview
13(No Transcript)
14Deforestation and EOSD Change Pilot Project and
RD Overview
15Saskatchewan Pilot Project
16(No Transcript)
17Meadow Lake Study Area
18Prince George Pilot Change
19Prince George Pilot
Stand Initialization
1990 Landsat TM
1990 Inventory
1990 Carbon stocks
Growth yield
Modeling parameters
1990-1999 Change Map
CBM
Disturbance list
1999 Carbon stocks
1999 Landsat TM
1999 Inventory
Landsat images T1 August 12, 1990 T2
September 12, 1999
20Prince George, British ColumbiaPilot Project
- CBM-CFS3 used to calculate carbon stock change
21EOSD National Change Product
- Map of major changes in the forests of Canada
- complete coverage two date change
- 5 -10 year time frame (e.g., 2000-2007-2012)
- clearcuts, major burns, deforestation, some
regeneration or lack thereof
22Annual Change Product
- A sample of annual change (e.g. PS2Ps)
- monitoring or history of change on the sample
- partial cut, regeneration, roads, landings, soil
disturbance, blowdown, some ID - composition change, normal growth, succession
- no change
23EOSD Change Approach Overview(Two Date)
241. Data selection/Compilation
Image Collection and Assembly
Top of Atmosphere Reflectance, Orthorectification,
Radiometric Normalization
2. Pre-processing
Evidence for Change
3. Has a pixel or polygon changed?
Yes! - Aggregate
No! - Assign T2 class
4. Spatial amalgamation
Aggregation
Spatial Change Unit
Pixels not changed and not included in change
units
5. What is the change?
Evidence for Change Type
Manual Vetting
Quality Control
6. Change Output and Validation
T1-T2 Change map
Accuracy Assessment
7. What is the T2 land cover?
Change Type, Evidence, Focused T2 Classification
T1 Classification for Non Change Pixels
T2 Landcover
25Recognition of Change - Two Date Change
Clustering
- A main change evidence element is a two date
unsupervised methodology using normalized images
change mask - Currently a two-date x 6 band (12 bands total)
clustering under a generous change mask is used - Tasseled-cap has also shown promise.
- Issues
- differentiating some change types, especially if
subtle - eliminating false alarms
- capturing small classes a common problem in
change mapping - Enhancement Classification Method (ECM) -
effective in single date land cover
classification could potentially solve some of
these difficulties ability to impose a classes
useful
26Two Date Change Enhancement for ECM
- Principle is to select band combinations such
that in a 3 band colour rendition, change classes
are distinct widely separated colours use this
for the clustering
27Enhancement Rendition Petawawa, Ontario
R (MIR (T1) MIR (T2)) /2 2(NIR (T2) -
NIR (T1) 128) -256 G MIR (T2) B MIR
(T1) RED (T1)
T11984 and T21988
28Two Date Unsupervised ECM Classification
29Two Date Unsupervised Enhancement and Cluster
Emplacement
30BAND4 BAND5 BAND3
BAND4 BAND5 BAND3
Changes image 123 classes
1988
1984
1
1
1
2
2
2
3
3
3
4
4
4
5
5
5
31Clearing Recognition
- Clearing and cuts were well recognized from
stable areas. - Some confusion occurs with subtleties of time 1
forest and time 2 ground compositions, in part
due to the two date nature of the clustering.
- Labels associated with land usage defaulted to
the dominant classes associated with forestry
cutting activities. - Deforestation associated with urban development
agriculture were included in clearcuts. - Small events (roads and landings showed some
confusion with stable due to mixed pixel edges.
32BAND4 BAND5 BAND3
BAND4 BAND5 BAND3
Changes image 123 classes
1988
1984
1
1
1
2
2
2
3
3
3
4
4
4
5
5
5
33Partial Disturbance / Removal Results
- Partial cuts are characteristically a mixture of
stable, clearcut and partial cut pixels - The process causing depletion is not well
discriminated (e.g., hail typically in the same
clusters as partial cuts) - Removal intensity and age
- not well differentiated
- Some confusion existed
- with burns and wetlands
34Partial cut preliminary two date results
35Spatial Aggregation into Change Units
- Individual pixel changes may be
- smaller than a minimum mapping unit
- classes may be contextually linked
hierarchically i.e. partial cut - or simply be an aberrant pixel
- Methods
- Sieving small polygons ( e.g., all conifer to
shrub lt 3 pixels no change) - Modal filters, can be iterative to build up edges
in highly fragmented areas - Use of secondary class probabilities to merge
neighboring pixels - Rules based system
MIR Pixel Difference
2 Date Unsupervised
Rule based aggregation
If ((diff gt 30) or (diff gt20 and neighbor diff
gt30) and polygon has 60 partial cut class
36(No Transcript)
37Change Typing
- Change clusters, labeling and classification
cant identify all types of change - After spatial aggregation, evidence identifies
some further change types - Need manual tools to help change typing
- Will need vetting / proofing
38Evidence and Vetting
Vetting includes all the information used in the
process, and is directed to checking and
upgrading delineations and type calls.
1990 Image
1999 Image
MIR Comp w. Inv.
New cut not updated on inventory
Not forest at T1
Inventory with Change Units
1995 Orthophoto
1998 Roads
2 Date Unsupervised
Quarry Expansion
Residential
39(No Transcript)
40(No Transcript)
41Conclusions
- ECM differentiated some classes better
(regeneration) than a 6 band two date
classification - Still unable to differentiate hail, deforestation
detailed partial cut intensity - ECM change enhancement effective at highlighting
change - ECM made cluster labeling easier
- ECM captures some subtler changes but still has
false alarms - Mechanically the methodology improves clustering
and improves identification of smaller classes - Three band enhancement makes identifying the need
for imposing a new cluster and creating them
easier - Evident that clustering techniques will not
differentiate all change types desired - Need new tools to help identify change type of
selected events
42(No Transcript)
43(No Transcript)
44(No Transcript)
45Test Area Recognition
46 Earth Observation for Sustainable Development
of Forests (EOSD) Overview, land cover
products
- Paletted GeoTiff file
- FGDC Compliant Metadata
- Accuracy information
- Shape file of mosaic lines
- User Agreement
47Interim Approach
Remote Sensing Mapping
GIS Systems Integration
- change is an event caused by a process no
change is an important product
48(No Transcript)
49(No Transcript)
50BAND4 BAND5 BAND3
BAND4 BAND5 BAND3
Changes image 33 classes
1988
1984
1
1
1
1
2
2
2
2
3
3
3
3
4
4
4
4
5
5
5
5
51Acknowledgements
- The authors would like to thank the staff of the
Pacific Forestry Center in Victoria, B.C. for
there inputs and support, as well as the Canadian
Space Agency and the Canadian Forest Service for
funding the temporal change research component of
the Earth Observation for Sustainable Development
project.
52Landsat Temporal Sequence
53Role of Low Resolution Imagery
- assume if change large enough for detection at
low resolution, it will be detected at 5 year
Landsat cycle - some ephemeral changes not detectable on 5-10
year Landsat-type cycle will be detectable - a main role is to add evidence to occurrence of
change and change type observed on Landsat - sometimes will help time stamping event for some
time sensitive changes - to confirm regional trends representativeness
(vice versa) - benefit from a higher repeat cycle increased
temporal sampling intensity
54Modis Simulation
55Information Needs
- Earth systems science (processes, carbon, climate
modeling, biophysical parameters, ----) - Forest inventory and/or management
- national and regional mapping and statistics
- local mapping, inventory update
- application specific change (e.g., damage,
deforestation, burns) - carbon accounting/inventory
- confirmation of currency of inventories
- aid in forest classification/mapping
56(No Transcript)