LandTrendr and TimeSync: Leveraging the Landsat archive to monitor landscape dynamics

1
LandTrendr and TimeSync Leveraging the Landsat
archive to monitor landscape dynamics
October 16, 2008
Robert E. Kennedy, Zhiqiang Yang, Warren B.
Cohen, Peder Nelson, Eric Pfaff

• LandTrendr and TimeSync

2
Change detection Motivation
Accurate observation of change is critical to
management
Processes
Future Landscape
Landscape
Changed Landscape
Time 1
Time 2
Infer process from observed changes in landscape
3
Change detection with Landsat

• Landsat is a good all-around tool for resource
  management at landscape scale

• Appropriate spatial, temporal, and spectral
  properties
• Relatively inexpensive
• Maps of stand-replacing forest disturbance by
  Cohen, Healey

AND SOON FREE!
4
Standard change detection approach

• Known issues
• 3-5 year intervals too coarse
• Subtle disturbances missed, especially in drier
  forests with sparse canopy
• No capture of succession or regrowth
• Change detection must occur within pre-determined
  forest area, often poorly defined
• Basic strategy The Two-date approach
• Subtract spectral values from two dates
• Large changes in spectral values indicate change
• Threshold of difference used to separate change
  from no-change

5
Trajectory-based approach

• Rather than look for disturbance EVENTS, look for
  disturbance TRAJECTORIES
• One example use idealized trajectories
• Build stacks of Landsat images from every year to
  examine trajectories over time
• Capture both events and processes
• Facilitated by cheap computers, data storage, and
  imagery

Kennedy, R.E., Cohen, W.B., Schroeder, T.A.
(2007). Trajectory-based change detection for
automated characterization of forest disturbance
dynamics. Remote Sensing of Environment, 110,
370-386
6
LandTrendr Landsat-based Detection of Trends in
Disturbance and Recovery
Steps in the LandTrendr process
Evaluate veracity of selected events
Prepare stack of yearly imagery
Extract summary information from segments
Extract spectral trajectories for pixels
Statistically identify and fit segments with
consistent trends
7
LandTrendr Steps
Scan archive for appropriate imagery
Normalize stack to single reference image
Develop cloud masks using thermal and SWIR bands
Prepare stack of yearly imagery
Trajectory-based Change Detection Algorithms
Vertex images
Extract and fit spectral trajectories for pixels
Maps
Extract summary information from segments
Post-processing batch files
Cover Models
Type II Validation Compare with other datasets
TimeSync
Evaluate veracity of selected events
Error Assessment
8
Uses Capture the same types of disturbance more
often
Yearly resolution commensurate with economic,
climate drivers and with habitat indicators
Forestlands north of Cle Elum
9
Uses Capture subtler disturbances
LandTrendr Year
LandTrendr Magnitude
Existing Approach Year
Increased signal-to-noise ratio of trajectory
approach allows detection of thinning and other
partial harvest
Map area Between Cle Elum and Cashmere, WA, on
public and private lands
Most areas missed by two-date approach are
partial harvests (magnitude maps)
10
Uses Capture slow disturbances
Multi-year insect mortality and recovery captured
at pixel scale
1985
1996
2002
FHM Overflight Data 1997 - 2006
LandTrendr Disturbance
FHM locational accuracy?
Spatial pattern of forest mortality consistent
over time and space
Alpine Lakes Wilderness, Wenatchee NF
FHM Forest Health Monitoring insect
observations taken from fixed-wing aircraft yearly
11
Uses Observe growth and recovery
Example growth trajectory
Recovery of vegetation after fire allows
distinction between full and partial regrowth
1970
1987
1988
Map area includes Van Creek (1987), Dinkleman
(1988), and Gold Ridge (1970) fires, west of
Entiat WA
A)
A) Rapid increase in cover on lands burned in
1970 B) Slow increase in cover near for
forests near complete cover
B)
1985
1996
2005
Tasseled cap imagery red, orange tones for
sparse cover, blue tones for conifer cover, cyan
and yellow for mixed forest
12
Uses Build temporally-robust forestland masks
Method Train stack median NDVI and LandTrendr
fitting statistics with NLCD 2001 cover
Resultant map captures lands that have been
forestland during 20 year period
13
Uses Create temporally-smooth yearly
pseudo-imagery
Classified images over time based on stable
spectral space
Use a single index to capture meaningful segments
in the trajectory
Band 5
Build models, classifiers at Time T
Band 4
Apply those segments to other indices/bands
Smooth other bands
Band 1
LandTrendr-smoothed
Recombine bands to create new smoothed images
14
Side benefit Clouds
Original tasseled-cap image
LandTrendr-smoothed image
15
Uses Utilize poor-quality images
Olympic Peninsula
Multiple images with clouds or with gaps (Landsat
7) fed into LandTrendr algorithms
1996
1998
1997
Algorithm picks best date on pixel-by-pixel basis
16
Uses Utilize poor-quality images
Allows use of multiple cloud images, as well as
Landsat 7 images with gaps
17
Landscape dynamics
Mapping all disturbance and recovery processes
reveals a dynamic landscape more consistent with
our ecological understanding than the simple
snapshots we have used until now
18
Validation New Challenges

• LandTrendr outputs require new approaches for
  validation/corroboration
• Outputs now at yearly time step from mid 1980s to
  present for large areas
• Processes mapped include a variety of subtle
  effects
• No single independent data source available to
  retroactively validate at the appropriate grain
• Our strategy
• Type 1 Direct interpretation of imagery using
  photo-interpretation rules
• Development of new software tool TimeSync
• Type 2 Link with other validation sets
  opportunistically to validate subsets of
  time/space
• Airphotos, ground plots, DOQs, etc.

19
TimeSync Sample design

• Stratification by Thiessen polygon
• Approximately 200 randomly chosen plots per full
  polygon

20
TimeSync

• Image window allows for PI, like with an airphoto
  (high spatial resolution), but applied to high
  temporal resolution/domain with moderate spatial
  resolution data

21
TimeSync

• Google Earth interface to examine near-current DOQ

22
TimeSync Observations

• Observations
• Cover type start and change, if any
• Disturbance segments
• start year
• duration
• agent relative intensity
• Regrowth segments
• start year
• duration
• Stable segments
• start year
• duration
• Tree and open cover ()

Develop point-based database of observations to
link with LandTrendr outputs
23
TimeSync Percent cover
Y axis space-for-time simulation of change in
cover
Photo-interpreted percent vegetative cover linked
with tasseled-cap wetness
Y axis percent vegetative cover
Note Models for tree cover, conifer cover, etc.
are not as strong as vegetative cover because
only a single spectral band is used
24
Existing new maps vs. human
New maps capture disturbance better than existing
maps, particularly at medium and low intensities
25
TimeSync
Developing approaches to quantify segment
agreement between LandTrendr and TimeSync
26
LandTrendr Mapping status

• LandTrendr mapping
• Northwest Forest Plan (NWFP) update - western WA
  and OR (Dec 2008)
• ORCA II -- Carbon modeling and integration -
  Southern California (Dec 2008)
• National Parks (2008-2009)
• Idaho and Colorado insect outbreaks (Spring 2009)
• OWEB - link disturbance to coupled
  forest/hydrological models (Spring 2009)

27
Summary

• Paradigm shift in use of Landsat imagery for
  mapping of landscape change
• Leverage entire historical archive
• New information from old data
• Better temporal frequency and connectivity
• Reduction in false positives and false negatives
• Improved ability to incorporate poor imagery
• Capture familiar processes more often, capture
  entirely new processes
• Broad implications for forest management, habitat
  monitoring, ecological modeling, carbon
  accounting

28
Thank you.
29
Interpretation and validation
TimeSync Interpretation
Interpretation Database
Segmentwise Agreement Algorithms
Image Stack
Validation Tables
LandTrendr Algorithms
Vertex Year, Value Maps
30
Disturbance mapping
LandTrendr Algorithms
Vertex Year Value Maps
Unfiltered Maps Disturbance Year Absolute cover
change Pre-Dist. Cover Relative Cover Change
Merge segments, Identify abrupt disturbance,
Incorporate cover eqns
Image Stack
Photointer- pretation and Regression
Cover Eqns
NAIP Imagery GoogleEarth
31
Spatial filtering
Disturbance Mapping (prior slides)
Unfiltered Disturbance Maps
Image Stack
Mask by relative change and starting cover
Combine
Temporal variability and Median NDVI Maps
Cover Mask Map
Temporally consistent Forest/Non-Forest Mask
Supervised Classification
NLCD 2001 Cover Map
Spatially Filtered Disturbance Maps
32
Insect Fire

• Identify insect mortality that precedes fire
• Feedbacks among multiple interacting disturbance
  processes
• Uses Carbon, climate change effects, resource,
  recreation

Frequency Sensitivity Trajectory
1985
1995
2006
Projects NIP
33
Insects and fire
Identify disturbance (duration gt 1 yr) and (1986
lt onset year lt 1992)
34
Insects and fire
BB Complex Fire, Fall 2003
Identify disturbance (duration 1 yr) and (onset
year 2004))
35
LanDTrenDR Insect vs. Fire
1985
1995
2006
Fire causes drop in reflectance from charring of
already dead stands)
FHM Observations Spruce budworm (1 to 4 trees /
acre)
FHM Observations Mtn. Pine Beetle (.25 to 5
trees /acre)
36
Example trajectories
37
(No Transcript)
38
TimeSync Point-based interpretation

• Interpretation / validation tool for TM/MSS
  stacks

39
Prescribed burns Sequoia National Park

• Underburns in Sequoia groves not discernible with
  two-date change approaches (dNBR)
• Post-fire trajectory may be diagnostic, however
  (B)
• Post-fire effects also apparent in wildfire (A)
• Multiple underburns detectible

Frequency Sensitivity Trajectory
Sequoia National Park Giant Forest Prescribed
and Natural Fires
Sequoia National Park Grant Grove Prescribed
Fire
Projects SIEN, NIP
40
Existing new vs. GIS database

• Example sparse eastside conifer forest
• Compare with Forest Service FACTS database
  (black lines)
• Existing map many false negatives ( )
• New good matches capture of other disturbance
  ( ) and recovery ( ) processes

41
Comparisons with Deschutes FACTS database
42
(No Transcript)