attribution of change agents based on all available ...why classifying change agent? •to better...
TRANSCRIPT
Change Agent Classification Based on All Available Landsat Data
Zhe Zhu Texas Tech University
Zhiqiang Yang Oregon State University
Landsat Science Team Meeting, 01/11/2017, Boston
Why classifying change agent?
• To better understand change, it is important to know the cause of change.
• Different type of change agent has quite different impacts to the environment.
Mathematical prediction models fit to clear observations
Reference: Zhu, Z. and C.E. Woodcock. 2014. Continuous change detection and classification of
land cover using all available Landsat data. Remote Sensing of Environment 144:152–171.
Continuous Change Detection and
Classification (CCDC) “Breaks”
CCDC Breaks vs Change agents
CCDC breaks indicate occurrence of spectral changes, but not all spectral changes are real change or meaningful change!
o Ephemeral break (i.e., moisture change, aerosols, clouds, shadows)
o Recovery break (i.e., break between re-growing stage to mature stage)
Wet
Dry
Regrowth
Mature
Training “Breaks”: Ephemeral and Recovery Breaks from USFS
• Cohen et al., Forest disturbance across the conterminous United States from 1985–2012: The emerging dominance of forest decline (2016).
• Simple random of 7,200 pixels from 180 individual frames that provide time segments of stable, recovery, and other disturbances.
• Breaks in stable segments for training ephemeral breaks.
• Breaks in recovery segments for training recovery breaks.
Training “Breaks”: Change agents from USGS LANDFIRE project
Change agents from USGS LANDFIRE project
Confidence Prescribed Fire Wildland Fire Wildland Fire Use Planting Reforestation Seeding Biological Chemical Herbicide Insecticide Low 10032 1 2 384 0 301 216223 51 18 0
Low/Moderate 10 0 0 0 0 0 0 0 0 0 Moderate 35574 3 124 135 0 136 5 208 33 0
Moderate/High 0 2 0 0 0 0 0 0 0 0 High 3688 4 13 19882 180 496 0 1978 52 0
Unchanged 0 0 0 0 0 0 0 0 0 0
Confidence Thinning Harvest Clearcut Development Mastication Other Mechanical Weather Insects Insects/Disease Disease Wildfire Low 26474 514 0 0 306 184 5046 333 111 0 1925
Low/Moderate 0 0 0 0 0 0 651 0 0 0 1 Moderate 38936 13615 306 0 219 20687 1567 1824 0 431 2177
Moderate/High 0 0 0 0 0 0 123 0 0 0 0 High 21 4890 8374 706 2593 13747 423 18019 92 0 1690
Unchanged 0 0 0 0 0 0 0 0 0 0 785 Agent Harvest Mechanical Weather Insets/disease fire
Extract breaks randomly for each category and subcategory
• 1,000 breaks per category
• Ephemeral (500) + Recovery (500) -> Others
• Harvest (500) + Mechanical (500) -> Mechanical
• Weather (500) + disease/insect (500) -> Nonmechanical
• Fire (1000) -> Fire
How to use CCDC outputs to classify different breaks?
Pre-change curves
Post-change curves
During-change vector
10 repeated cross validation 80% training & 20% validation
Change Agents Others Mechanical Nonmechanical
(insects/disease + weather) Fire Total Users Others 1796 143 14 46 1999 90%
Mechanical 119 1801 109 92 2121 85% Nonmechanical 29 11 1804 8 1852 97%
Fire 72 32 0 1904 2008 95% Total 2016 1987 1927 2050 7980
Producers 89% 91% 94% 93% Overall 91.54%
Change Agents Others Mechanical Insect/disease Weather Fire Total Users Others 1757 134 10 3 39 1943 90%
Mechanical 111 1912 32 122 74 2251 85% Insect/disease 1 6 947 2 0 956 99%
Weather 20 0 0 850 0 870 98% Fire 85 39 0 0 1836 1960 94%
Total 1974 2091 989 977 1949 7980 Producers 89% 91% 96% 87% 94% Overall 91.50%
Variables No DEM No Thermal No Thermal No DEM DEM & Thermal
Overall 89.40% 90.63% 90.88% 91.50%
Conclusion
• The CCDC algorithm can classify change agent with high accuracies.
• The insect/disease and weather related change can be well separated by the CCDC algorithm.
• Both DEM and thermal band are helpful for change agent classification.
Back up slides