Contrail Detection and Optical Properties Derived Using Infrared Satellite Data from MODIS
Sarah Bedka1, Patrick Minnis2, David P. Duda1, Rabindra Palikonda1, Robyn Boeke1 and Kristopher Bedka1
1 Science Systems and Applications Inc., Hampton VA2 NASA Langley Research Center, Hampton VA
Objectives
• Develop an automated Contrail Detection Algorithm (CDA) for identifying linear contrail features in MODIS data
• Determine appropriate brightness temperature difference (BTD) thresholds for CDA, and provide error estimates
• Retrieve contrail optical properties from MODIS observations and estimate the errors in the retrieved properties resulting from errors in the contrail mask
Contrail Detection Algorithm (CDA)
See talk by Dave Duda tomorrow morning for more details
10.8 μm BT10.8 – 12 μm BTD Contrail Mask
• Based on Mannstein et. al (1999)
• Contrails are minimally visible in the 10.8 μm image but quite visible in the 10.8-12 μm BTD image.
• Uses two IR channels (10.8 μm and 12 μm on MODIS) and applies a scene-invariant BTD threshold to identify possible contrail linear features
• Additional IR information from MODIS removes non-contrail linear features (e.g. cloud edges, surface features)
CDA Visual Analysis
Contrail Mask
• GUI-based tool allowed reviewers to examine MODIS IR and VIS information, as well as BTDs
• All cases contained contrails (223366 daytime, 68189 nighttime)
• A composite mask was created from the consensus of the 4 analyses, and is used as “truth”
Composite Contrail Mask
RED = Confirmed ContrailsGREEN = Added ContrailsBLUE = Deleted Contrails
Accuracy of the CDA was determined by visual analysis of 44 (21 daytime, 23 nighttime) MODIS granules by 4 reviewers
CDA Accuracy AssessmentProbability of Detection (POD): What fraction of the observed contrail pixels were correctly identified? Range = 0 -> 1 Perfect = 1
False Alarm Rate (FAR): What fraction of the predicted contrail pixels were incorrectly identified? Range = 0 -> 1 Perfect = 0
Frequency Bias: What is the relative frequency of predicted contrail pixels to observed contrail pixels? Range = 0 -> ∞ Perfect = 1
BIAS =hits + misses
hits + false alarms
FAR =hits + false alarms
false alarms
POD =hits + misses
hits hits: truth=Y mask=Ymisses: truth=Y mask=Nfalse alarms: truth=N mask=Y
POD/FAR/BIAS were calculated for each of 6 BTD thresholds
Probability of Detection/False Alarm RateDaytime Nighttime
cons00 cons01 cons02 cons03 cons04 cons050
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
POD - JanuaryPOD - AprilPOD - JulyPOD - October
cons00 cons01 cons02 cons03 cons04 cons050
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
POD - JanuaryPOD - AprilPOD - JulyPOD - October
cons00 cons01 cons02 cons03 cons04 cons050
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1FAR - JanuaryFAR - AprilFAR - JulyFAR - October
cons00 cons01 cons02 cons03 cons04 cons050
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1FAR - JanuaryFAR - AprilFAR - JulyFAR - October
# CasesJan = 8Apr = 9Jul = 1Oct = 3
# CasesJan = 6Apr = 5Jul = 7Oct = 5
more conservative more sensitivethresholds used thresholds used
POD =hits + misses
hits
hits + false alarmsFAR = false alarms
mask00 mask01 mask02 mask03 mask04 mask05mask00 mask01 mask02 mask03 mask04 mask05
mask00 mask01 mask02 mask03 mask04 mask05
more conservative more sensitivethresholds used thresholds used
mask00 mask01 mask02 mask03 mask04 mask05
Frequency Bias
cons00 cons01 cons02 cons03 cons04 cons050
0.5
1
1.5
2
2.5
3
3.5
Frequency Bias - JanuaryFrequency Bias - AprilFrequency Bias - JulyFrequency Bias - October
cons00 cons01 cons02 cons03 cons04 cons050
0.5
1
1.5
2
2.5
3
3.5
Frequency Bias - JanuaryFrequency Bias - AprilFrequency Bias - JulyFrequency Bias - October
Daytime Observations Only Nighttime Observations Only
# CasesJan = 8Apr = 9Jul = 1Oct = 3
# CasesJan = 6Apr = 5Jul = 7Oct = 5
A frequency bias of 1 indicates that the mask neither overestimates nor underestimates the number of contrails.
more conservative more sensitivethresholds used thresholds used
more conservative more sensitivethresholds used thresholds used
mask00 mask01 mask02 mask03 mask04 mask05 mask00 mask01 mask02 mask03 mask04 mask05
Monthly CDA Performance Statistics
Jan Apr Jul Oct# Granules 6 5 7 5
% Contrail 0.371 0.491 0.181 0.277
POD 0.535 0.589 0.590 0.616FAR 0.455 0.345 0.389 0.413Frequency Bias 0.981 0.900 0.965 1.049% Added 47.43 45.61 42.54 36.59% Deleted 45.48 34.51 38.89 41.27
Jan Apr Jul Oct# Granules 8 9 1 3
% Contrail 0.070 0.086 0.156 0.148
POD 0.519 0.550 0.500 0.790FAR 0.681 0.539 0.458 0.574Frequency Bias 1.625 1.193 0.922 1.854% Added 29.59 37.76 54.18 11.33% Deleted 68.06 53.93 45.78 57.39
Daytime Observations Only Nighttime Observations OnlyJan Apr Jul Oct
# Granules 6 5 7 5
% Contrail 0.371 0.491 0.181 0.277
POD 0.535 0.589 0.590 0.616FAR 0.455 0.345 0.389 0.413Frequency Bias
0.981 0.900 0.965 1.049% Added 47.43 45.61 42.54 36.59% Deleted 45.48 34.51 38.89 41.27
Jan Apr Jul Oct# Granules 6 5 7 5
% Contrail 0.371 0.491 0.181 0.277
POD 0.535 0.589 0.590 0.616FAR 0.455 0.345 0.389 0.413Frequency Bias 0.981 0.900 0.965 1.049% Added 47.43 45.61 42.54 36.59% Deleted 45.48 34.51 38.89 41.27
Jan Apr Jul Oct# Granules 8 9 1 3
% Contrail 0.070 0.086 0.156 0.148
POD 0.519 0.550 0.500 0.790FAR 0.681 0.539 0.458 0.574Frequency Bias
1.625 1.193 0.922 1.854% Added 29.59 37.76 54.18 11.33% Deleted 68.06 53.93 45.78 57.39
Jan Apr Jul Oct# Granules 8 9 1 3
% Contrail 0.070 0.086 0.156 0.148
POD 0.519 0.550 0.500 0.790FAR 0.681 0.539 0.458 0.574Frequency Bias 1.625 1.193 0.922 1.854% Added 29.59 37.76 54.18 11.33% Deleted 68.06 53.93 45.78 57.39
In general, a higher percentage of contrails were added during the daytime than at night. A higher percentage of contrails were deleted at night.
The contrail mask slightly underestimates the number of contrails during the day, and slightly overestimates it at night.The Probability of Detection (POD) was slightly lower in the winter than in the summer, due to lower contrast of contrails over a colder and/or possibly snow-covered surface. False Alarm Rate (FAR) was also slightly higher in the winter, and was significantly higher at night than during the day.
Jan Apr Jul Oct# Granules 6 5 7 5
% Contrail 0.371 0.491 0.181 0.277
POD 0.535 0.589 0.590 0.616FAR 0.455 0.345 0.389 0.413Frequency Bias
0.981 0.900 0.965 1.049% Added 47.43 45.61 42.54 36.59% Deleted 45.48 34.51 38.89 41.27
Jan Apr Jul Oct# Granules 8 9 1 3
% Contrail 0.070 0.086 0.156 0.148
POD 0.519 0.550 0.500 0.790FAR 0.681 0.539 0.458 0.574Frequency Bias
1.625 1.193 0.922 1.854% Added 29.59 37.76 54.18 11.33% Deleted 68.06 53.93 45.78 57.39
• Uses 9 ice cloud models (Minnis et.al 1998).
• Accurate contrail temperature and clear-sky BTDs are key to accurate retrievals
• 2-channel retrieval (11, 12 μm) may capture the data better than the traditional 3-channel retrieval
predicted clear-sky value
Retrieval of Contrail Optical PropertiesTechnique is to minimize the difference between the observed and calculated BTDs for 3 IR bands (3.9, 11, and 12 μm).
Contrail Retrieved Optical PropertiesOverall Averages – 23 Daytime Granules
Confirmed Added DeletedMean τ (2-channel IR) 0.46 0.70 0.78
Mean τ (3-channel IR) 0.40 0.93 0.85
Mean De (2-channel IR) 77.1 92.3 89.6
Mean De (3-channel IR) 57.6 67.6 46.4
• 223366 total contrail observations (116789(52%) confirmed, 82118(37%) added, 24459(11%) deleted)
• Total τ (Pc*τc+Pa*τa-Pd*τd) is 0.58 for the 2-channel IR retrieval and 0.64 for the 3-channel
• Total De is 84.1 (60.1) for the 2-channel (3-channel) retrieval
Contrail Retrieved Optical PropertiesOverall Averages – 21 Nighttime Granules
Confirmed Added DeletedMean τ (2-channel IR)Mean τ (3-channel IR) 0.90 0.81 0.84
Mean De (2-channel IR)
Mean De (3-channel IR) 40.02 71.41 40.60
• Average of 21 nighttime cases• 4-reviewer composite mask used• 68189 total contrail observations (31941
confirmed, 18159 added, 19240 deleted)
Future Work• Improve contrail optical property retrievals by:
better characterizing background temperature improving contrail temperature estimates expanding model parameterizations (from P. Yang)
• Apply CDA to a larger data set to get an estimate of annual contrail coverage and trends over the Northern Hemisphere
• Use CDA performance statistics from visual analysis to estimate errors in long term data sets