|
Validation of the Landsat 7 Long-term Acquisition Plan
Part 1
The launch of Landsat 7 on April 15, 1999 ushered in a new
era of managing global survey missions. For the first time (for
Landsat or any similar mission) an automated Long-Term Acquisition
Plan (LTAP) was employed to populate the US Archive at the USGS
EROS Data Center. The automated LTAP system was designed to
globally collect sunlit, essentially cloud-free, and seasonally
refreshed Landsat acquisitions for all land areas of the Earth.
The LTAP has successfully populated the archive such that most
users requiring data find it already in the archive, from the
right time of year, substantially cloud-free, and under the
right radiometric conditions, rather than having to order new
acquisitions. Requests for data acquisitions outside of the
LTAP have been lower than for any other Landsat mission. After
three years of operations, we have undertaken an exhaustive
analysis of the performance of the LTAP in terms of seasonality,
cloud contamination, gain setting strategy, and satisfaction
of science community requirements. This article addresses the
evolution of the LTAP during Landsat-7 operations, the validation
strategies that were designed, and the results of the validation
process. In addition, lessons learned from the validation exercise
and future updates to the LTAP strategy are discussed.
Terry Arvidson, Richard Irish, Brian Markham, Dr. Darrel Williams,
Jay Feuquay, John Gasch, and Dr. Samuel N. Goward
A new automated approach to Landsat mission operations was
introduced with the deployment of Landsat 7. A long-term acquisition
plan (LTAP) was developed prior to launch to systematically
populate the US archive, held at the USGS EROS Data Center,
with seasonally refreshed, essentially cloud-free observations
of all land areas of the planet (Arvidson, et al, 2001). The
LTAP system addresses terrestrial seasonality, cloud climatology
versus daily NOAA-predicted cloud cover, the two possible gain
states on the Enhanced Thematic Mapper plus (ETM+) sensor on
Landsat 7, and various technical and/or operational constraints
such as sensor duty cycle and battery reserve power. All of
these factors are merged in an Oracle-based information system
that resides in the Landsat mission operations center (MOC).
Information on predicted cloud cover and mission technical status
are fed to the MOC daily. The MOC then prepares a mission operations
plan, determining which Landsat WRS paths and rows will be acquired
for the next 37 hours. An updated operations plan is uploaded
to the satellite every 24 hours to command the payload data
acquisition subsystems.
Prior to launch, a team of researchers at NASA Goddard and the
University of Maryland developed the specific characteristics
of the LTAP system. As much as possible, the LTAP team depended
upon then-existing information about land seasonality from NOAA
Advanced Very High Resolution Radiometer (AVHRR) global observations,
and on cloud cover climatology, developed from AVHRR and GOES
observations by the International Satellite Cloud Climatology
Project (ISCCP) (Arvidson et al., 2001; Goward et al., 1996;
Goward et al., 1999). A similar approach was taken with ETM+
gain state considerations. This automated approach was reviewed
by the Landsat Science Team in 1998, prior to launch (Goward
& Arvidson, 1998).
The LTAP system has now predominately operated Landsat 7 for
the last 3 years, July 1999-July 2002. (There are times, such
as during calibration maneuvers and orbital realignments, when
mission operations are take over by ground control to accomplish
tasks other than the primary observation missions. Also during
national emergencies the system is targeted to meet other national
needs.) The LTAP development team is currently pursuing a validation
of the LTAP, to insure that it is accomplishing the goals of
the LTAP, initiated in 1995. The specific goals include:
-- All Land Areas: Observations will be collected for all Landsat
Worldwide Reference System (WRS) areas that contain any land
(including islands and coral reefs where possible).
-- Seasonally Refreshed: All land locations will be revisited
with sufficient temporal frequency to capture major seasonal
changes (e.g. spring, summer, fall, winter, dry versus wet season,
low versus high sun).
-- Essentially Cloud Free: To meet the previous requirements,
at least one observation per season should be as cloud free
as possible — this is an elusive goal in cloud-prone regions
of the globe.
-- Radiometry: The LTAP considers the radiometric dynamic range
across the optically reflective spectrum. The ETM+ sensor acquires
measurements in either “low” or “high” gain for each spectral
band. This is a trade-off between observation saturation and
radiometric precision. The desire was to minimize saturation
while maximizing scene radiometric contrast in the acquired
observations. The low light factor and the two gain states available
with the ETM+ sensor required a radiometric aspect to the LTAP
method.
As the LTAP system was developed, members of the Landsat Science
Team, as well as other members of the Landsat user community,
pointed out that seasonality as defined by an AVHRR normalized
difference vegetation index (NDVI) would not necessarily meet
all their needs. Several such “niche” communities were identified,
and special provisions within the LTAP were incorporated to
meet their needs:
-- Glaciers & Ice Sheets
-- Volcanoes
-- Fire study sites
-- Ocean Islands
-- Calibration sites
-- Sea ice
-- Coral Reefs
-- Boreal forests
-- Rainforests
-- Agricultural areas of interest to the USDA Foreign Agricultural
Service
A validation plan to evaluate the basics of the Landsat-7 LTAP
approach was developed by the LTAP Working Group at Goddard
and the University of Maryland. The team proposed and developed
independent validation approaches for each of the unique aspects
of the LTAP:
-- Seasonality: In this case, the acquired coverage, at monthly
time steps, resident in the EROS Data Center archive was compiled
to evaluate how well the seasonality was achieving desired goals.
WRS-based maps of monthly coverage, as well as the lowest acquired
cloud cover, were examined.
-- Cloud Cover: There are multiple aspects to this methodology,
including the within-scene automated cloud cover assessment
(ACCA), the ISCCP cloud climatology, and the comparative performance
of the NOAA National Centers for Environmental Prediction (NCEP)
cloud forecast versus the ACCA results and the ISCCP climatology.
In effect: Are we better off using cloud forecasts versus climatology
versus nothing?
-- Radiometry: Quite early in the Landsat-7 mission, the Landsat
Science Team members reported observing measurement saturation
in the near-infrared band for full canopy agricultural fields
observed during the summer in the mid-latitudes. So began the
saga of changing gain strategies as a function of user complaints.
Thus the validation approach in this case has been predominately
based on reaction to user input. Beyond these on-going changes,
we also carried out a seasonal assessment of measurement saturation,
as a function of spectral
This LTAP validation exercise is significant for the following
reasons:
1 Refine LTAP-based mission operations for the remainder of
the Landsat-7 mission (through 2006 at the earliest).
2 Identify potential acquisition reduction strategies for the
purposes of extending mission life
3 Provide a refined assessment of the LTAP approach for the
benefit of the Landsat Data Continuity Mission (LDCM) and all
future Landsat missions.
Each of the validation analyses was complex and led to extensive
in-depth consideration of the questions raised. In this article
we will only present a high-level summary of the conclusions
we have reached without presenting many of the details that
underpin our conclusions. We hope that in the near future we
will be able to compile for publication more detailed manuscripts
on the various aspects of the Landsat-7 LTAP validation.
Land Definition
The Landsat-7 land database contains over 53,000 geographic
tags (Figure 1) identifying for each WRS land scene the geopolitical
extents, land-cover type, and other useful information. The
geographic tags were not included in the initial system design
but were added just before launch. They have proven extremely
useful for answering queries, selectively applying scheduling
rules, and conducting analyses.
These include:
-- Identification of nearly 17,000 Worldwide Reference System
(WRS) scenes which consist of land and shallow coastal waters.
Each of these scenes are acquisition candidates for the US archive
(Arvidson et al., 2000).
-- Land scenes also set limits for scheduling International
Ground Station (IGS) acquisitions, as IGS requests over water
are not allowed.
-- Geographic tags serve as keys to the scheduling software,
triggering special logic paths used during scheduling (e.g.,
high latitude overlap, night or ascending scene duty cycle constraints);
-- Geographic tags allow collection of statistics by interest
group, such as the various science niches (e.g., reefs, volcanoes)
(Arvidson et al., 2000).
Validation Strategies
The geographic tags were used to sort and display the data,
evaluate the tag versus the WRS, and identify errors.
Land scenes that are surrounded on two or more sides by water
were classified as “fringe coastal” scenes. Browse images of
these scenes were visually inspected in the context of their
neighboring scenes to determine whether the land portion of
each fringe scene was included within the surrounding scenes.
If so, then the scene could be removed from the land database
with no loss of coverage.
Finally, efforts are underway to compare the LTAP land database
with the land database generated by the Earth Satellite Corporation
(EarthSat) in their compilation of the Global Landsat 2000 Scientific
Data Purchase.
Validation Outcome
As expected with such a large database, we found 64 errors in
the geographic tags and in the WRS assigned to some of the tags,
and we found 32 WRS scenes in the land database with no entries
in the seasonality file. These were corrected in August 2002.
The analysis of the list of fringe coastal scenes identified
352 scenes that could be deleted from the land and seasonality
databases. These were deleted in August 2002.
Future analysis of the large land database to refine the geographic
tags and correct any reported errors in location is planned.
As an example, future work will include adding a detailed reef
definition to the land database, using inputs from the reef
community to correct the location of reefs based on the imagery,
and adding the names of the reefs to the database instead of
using the generic “reef niche” tag. We have discovered islands
in the EarthSat database that had not been included in the LTAP
database, and reefs in the LTAP database that are not included
in the EarthSat database. The comparison is ongoing and results
will be exchanged with EarthSat.
Seasonality
The basic goal of the seasonality portion of the LTAP is to
insure that Landsat 7 acquires new observations for each new
variation in land surface conditions. In general this is with
respect to the condition of local vegetation cover, although
there are other reasonable definitions as noted by the various
niche communities discussed later in this article. For vegetative
cover, the typical mid-latitude perspective is “spring, summer,
fall, winter,” approximate quarterly (3-month) sectors of the
annual 12-month Earth orbital cycle. However, seasonality varies
by region of the planet. For example, wet and dry cycles in
lower latitude locations where temperature is not the primary
determinant of seasonality, or high latitude locations where
there is a very short summer and a very long winter. Our effort
to characterize these regional variations in “seasonality” led
us to employ the AVHRR observation record of land in the visible
and near-infrared, converted to a measure of green vegetation
presence with the normalized difference vegetation index (NDVI)
(Arvidson et al., 2001; Justice et al., 1985). Our analysis
produced monthly WRS location assessments to acquire-once or
acquire-every opportunity (i.e. ~ twice) during that month (Figure
2).
Operational acquisition performance has been closely monitored
throughout operations and has been fine-tuned through database,
parameter, and software updates to conform to this definition
of seasonality.
Validation Strategies
Our primary interest was to evaluate whether the LTAP was acquiring
the basic spatio-temporal coverage we would have expected. Landsat
7’s sensor, limited by a 16% maximum duty cycle, should optimally
be able to acquire nearly the entire Earth’s land areas every
two 16-day orbital cycles. We therefore began with examination
of WRS maps consisting of observations collected over a 32-day
(2-cycle) period (Figure 3). Following our initial analysis
of these 2-cycle maps, we determined that analysis of seasonal
quarterly maps would also be important, where the seasons were
defined to approximately coincide with the Earth’s orbital solstices
and equinoxes (i.e. Dec-Feb, Mar-May, Jun-Aug., Sept.-Nov.)
In support of this analysis, we also employed the newly implemented
Landsat global visualization (GLOVIS) data exploration tool,
available at the EROS Data Center website (http://glovis.usgs.gov/).
This system provides interactive access to the Landsat-7 metadata,
including the browse images. This facilitated rapid and qualitative
evaluation of specific examples of the results shown in the
global WRS maps noted previously.
Validation Outcome
Based upon the 2-cycle and quarterly WRS maps, we concluded
that the LTAP is operating the Landsat-7 mission to acquire
better than 90% global coverage for each quarter of the year.
Thus for most locations, we are accomplishing the desired seasonal
coverage originally specified as a goal of the LTAP.
However, we do note that the geographic distribution of the
observations is not entirely in line with our expectations.
Our conclusions concerning regional geographic coverage include:
-- Too many “desert” scenes are being acquired currently. Our
estimate is that on average 8.1 scenes are acquired annually
of each WRS desert scene. Of these 5 are cloud-free and 3.1
have cloud cover that averages over 43%! Our seasonality analysis
was clearly too conservative (or too generous, depending on
how you look at it) in addressing desert coverage. We do have
to worry about desert-based irrigated agriculture, for example
in central Saudi Arabia and Sudan. This activity appears to
be rapidly expanding. Further analysis of the phenomenon is
warranted.
-- The mission is consuming substantial acquisition resources
in tropical, cloud-prone regions (as noted in Figure 3). A quick
examination of the browse imagery suggests a) there are only
selected seasons of the year when low cloud cover images are
acquired, and b) the only probable way of producing “cloud-free”
coverage in these locations is to employ image composition methods,
such as those used with AVHRR and MODIS measurements today.
These observations indicate that a more in-depth analysis of
tropical coverage is warranted. This problem may in part be
addressed with a new 80% cloud cover threshold cut-off rule
that has recently been implemented in the LTAP. That is, if
the NCEP cloud cover forecast is for 80% cloud cover or worse,
acquisition will not be scheduled.
-- There appears to be excessive coverage of the high mid-latitudes
(>45 degrees) in the winter. Apparently our conservative
(generous) seasonality file prescribes acquisition every opportunity
during the winter and the low cloud cover makes such scenes
good candidates for acquisitions. However, having seen the frozen,
snow-covered boreal and tundra regions once, there seems little
value to acquiring them again.
-- The reverse is true in these mid- to high-latitude regions
during the summer. As with the tropical forest regions, these
boreal and tundra zones are cloud-prone during the summer. However,
only one scene is acquired during the summer (June-August) timeframe.
There may be a need to set a high summer priority on acquisitions
for these locations during the summer, particularly for Russian
Siberia and Northern China, where there is no international
ground station coverage to augment the LTAP.
We now believe that the binary NDVI interpretation (acquire-once,
acquire-every opportunity) we deployed in this version of the
LTAP does not take full advantage of the LTAP system functions.
In fact, use of a continuous variable extract of the NDVI record
may well be far more successful in addressing a number of the
shortcomings we have already discovered, specifically with respect
to desert acquisitions. Implementation of such a new seasonality
definition would require an additional commitment to development,
testing and validation but we believe that such an approach
would produce a substantial improvement in LTAP operations.
Clouds
To achieve the least cloud contamination in acquired imagery,
the LTAP considers cloud conditions in the past, the present,
and nominal:
-- Cloud forecasts: The National Centers for Environmental Prediction
furnish daily global weather forecasts. These forecasts are
used as a scheduling aid to avoid acquiring scenes likely to
be cloud contaminated.
-- Cloud climatology: Acquisition scheduling decisions employ
the forecasted cloud conditions normalized against the historical
average cloud cover. Priority of a candidate acquisition appreciates
if the forecasted cloud cover is better than the historical
average or depreciates if the forecast is worse than average.
(Gasch & Campana, 2000)
-- Cloud cover assessment: Image processing assesses the amount
of cloud contamination in each acquired image. This automated
cloud cover assessment (ACCA) score (Irish, 2000) is conveyed
back to the scheduling system as an indicator of past success.
Future acquisition decisions consider the results of prior acquisitions
as defined by the ACCA.
Cloud avoidance is a major goal of the LTAP and is accomplished
by using all three types of data. In some areas of the world,
cloud avoidance is not attempted. These are the 50 United States,
where data is required every opportunity regardless of cloud
cover, and Antarctica where the reliability of the cloud predictions
has been questionable due to persistent ground fog.
Validation Strategies
Inter-comparisons were made among the different types of cloud
data used to understand the relationships, compare results,
and identify any biases: climatological versus predicted, climatological
versus assessed, assessed versus predicted.
To evaluate ACCA performance, a blind study was conducted on
192 images selected randomly across 8 latitudinal zones, 21
locations in each zone. 160 of the 192 images (for 20 of the
21 locations) were selected from the time interval when the
Normalized Difference Vegetation Index (NDVI) was within 10%
of its peak value. For the other location, four seasonal images
were chosen for each latitudinal zone. Various Adobe Photoshop
tools were used to generate binary cloud masks from each of
the browse images; histograms of the cloud masks yielded cloud
mask percentages for comparison with the calculated cloud cover
scores.
To determine the regional success of cloud avoidance, maps showing
the best cloud cover score for each WRS scene and the influence
of cloud cover on acquisition frequency were generated and analyzed
(Figure 4).
Validation Outcome
Our analysis has demonstrated that the ACCA algorithm is performing
quite well (Figure 5). As expected, the ACCA algorithm did under-report
very thin high cirrus clouds (Irish 2000), especially where
water dominates a scene. Also, commanded changes
to the sensor gain during image acquisition presents a problem
for the algorithm, which can over-estimate clouds in the image
area following the gain change.
Considering the 7 scenes that differed by more than 15%:
-- Two scenes under-reported (17% and 39%) due to thin cirrus
not visible to the ACCA algorithm
-- One scene under-reported (22%) due to weaker cirrocumulus
cloud signature over water
-- One scene over-reported (17%) due to gain change
Three scenes with apparent over-reporting (26%, 38%, 30%) proved
not true because the remaining clouds are not visible in the
spectral bands shown in the browse images.
If the mask scores are adjusted for the three over-reporting
cases where ACCA appeared to work, then ACCA is within 10% of
truth for 95% of all scenes examined. Considering the 160 scenes
at peak NDVI, ACCA is within 10% of truth 97% of the time. ACCA
under-performed for peak NDVI scenes in the mid-latitudinal
zone (30-45 degrees North), with results within 10% of truth
only 85% of the time, however several of the over-reporting
cases fell into this zone. A seasonal bias is suggested in the
Boreal South zone (45-60 degrees South), but 2 of the 4 seasonal
scenes were contaminated by gain changes and were thus over-estimated.
We concluded that ACCA is performing as designed in most cases
and is providing reliable feedback to the scheduling system.
Comparisons among the cloud file types revealed a slight bias
(approximately 3%) in the cloud predictions to the high side;
this value is considered to be in the noise. The predictions
are also more accurate at the extremes (0% and 100%) as shown
in Figure 6. Based on this, we recently decided to implement
a cloud cover cut-off of 80%; if the forecasted cloud cover
for a scene is worse than 80%, the scene is not considered a
candidate for scheduling. This rule is applied only to scenes
outside the 50 US states. Being a recent change, analysis of
the consequences of this rule change is pending. Modeling runs
predict that a tiny percentage of clear scenes will be missed
and a significant amount of duty cycle will be reallocated to
acquiring less cloudy scenes.
There was no apparent degradation due to the staleness of cloud
forecasts out to 36 hours. Degradation of cloud forecasts at
higher latitudes was evident during the campaigns over Antarctica.
This combined with the ACCA algorithm’s difficulty in discriminating
clouds from snow and ice (Irish 2000) led us to not consider
cloud avoidance over Antarctica and to rely on the feedback
of manual cloud cover assessments to guide the scheduler for
future scene acquisitions over the continent.
Two interesting phenomena were observed in our analysis of browse
scenes and the associated cloud assessments:
1. Island effect: For isolated oceanic islands and reefs, an
assessment of low cloud cover is misleading because the few
reported clouds are almost always over the island or reef, rendering
the scene useless; conversely, for islands and reefs that are
close to continental coasts, an assessment of high cloud cover
is also misleading because the clouds are almost always over
the continental landmass and the offshore islands and reefs
are in the clear.
2. Western continental coastal cloud/fog banks: We found poor
correlation between usefulness of the image and the ACCA score
for these scenes, where the morning cloud/fog banks are hanging
off the coast over the water, and the land is clear. These scenes
are reported at 50 or 70% cloud cover but the land area is clear.
This was primarily seen in North and South America.
Both of these phenomena highlight the importance for users to
check the browse images and ascertain that their areas of interest
are not under the clouds. Also, in their queries, it is important
to keep the cloud cover threshold wide open in order to not
exclude some of the scenes as described above that have high
cloud cover assessments but generally clear land areas.
Future plans include investigating the feasibility of employing
a land mask in the cloud assessment process, resulting in a
scene utility score as well as a cloud cover score. We will
consider manual cloud assessment in other areas of the world
where automated cloud assessment is unreliable, and explore
the compositing of images to construct cloud-free images in
areas that are persistently cloud-covered. We will also investigate
a cloud-confidence factor based on our assessment of latitudinal
and regional reliability of the cloud forecasts.
(End of Part One—To be continued)
About the Authors
Terry Arvidson is a Senior Systems Engineer for Lockheed
Martin.
Richard Irish is a Senior Systems Engineer for SSAI.
Brian Markham is a Landsat Calibration Scientist for NASA/Goddard
Space Flight Center.
Dr. Darrel Williams is a Landsat Project Scientist for NASA/Goddard
Space Flight Center.
Jay Feuquay is a Landsat 7 Program Data Acquisition Manager
for USGS, EROS Data Center.
John Gasch is a Landsat 7 Mission Analyst for Emalico LLC, Goddard
Space Flight Center.
Dr. Samuel N. Goward is a Geography Professor for the University
of Maryland.
Back
|
