; P9
Learning Center
Plans & pricing Sign in
Sign Out
Your Federal Quarterly Tax Payments are due April 15th Get Help Now >>



  • pg 1
									     Comparison of Methods and Approaches for Combining
      Multi-Temporal raster data in Geographic Information
       Gabe Emerson                                                                               Computer Science Department
   emers089 @ umn.edu                                                                               University of Minnesota
Computer Science Department                                     Chris Heuer
   University of Minnesota                                  heue0034 @ umn.edu
ABSTRACT                                                                     type of activity is evident in the differing backgrounds of
Analysis of remote sensing data often involves the comparison or             researchers, from human migration studies to geological programs
combination of imagery collected at different times. Depending on            and climate-change scientists.
the temporal resolution, season, end-use, and other factors, this             Because of the wide range of applications and approaches to this
data can be difficult to deal with and integrate in a meaningful             problem, classifying research into distinct subsets is not a simple
fashion. This paper investigates several methods of combining and            task. A number of possible classifiers and categorization schemes
using multi-temporal raster data. Projects using a variety of these          exist, including methodology, end-use, data acquisition and
methods are studied and described. Each method is evaluated for              source, data type, etc. We believe that while each scheme may
its overall performance and usefulness based on a number of                  have its merits, methodology is the most logical and effective
factors. In addition, potential ways to improve or combine                   classifier for this field.
existing methodologies are proposed.
                                                                             1.1       Novelty
1.         INTRODUCTION                                                      At present there do not exist any literature surveys for this topic,
When collecting raster data through satellite imaging or other               despite its recent rise in importance and number of applications.
remote sensing, a trade off occurs in the ground surface detail              Surveys and comparisons do exist for object and vector-oriented
available in different seasons. For example; the most data on                spatio-temporal data [9], but such studies are outside the scope of
vegetative cover can be obtained via infrared and near-infrared              this investigation. Our work presents a diversity of raster-based
sensing during leaf-on periods. Conversely, the highest level of             approaches, including applications currently underway and
imaging detail can be obtained during leaf-off periods, when                 proposed, and highlights the most important commonalities in
vegetative cover does not obscure surface features such as roads,            methodology and data structures.
streams, and structures. In addition, seasonal or temporal                   We evaluate each source for efficiency, cross-application
attributes of vegetative cover or other features may need to be              portability, and usefulness. We also investigate possible
compared, such as algal blooms, forest canopy changes, and                   improvements not currently identified in existing research, and
seasonal socioeconomic activities.                                           suggest possible combinations of methods which could lead to
While comparison of temporally-separated images traditionally                additional improvements.
requires side-by-side examination, it is possible that composite
multi-temporal imagery may provide the maximum amount of                     1.2       Validation
information. For example, a feature visible at one time (such as             The categorization and classification of multi-temporal research
algae blooms) may need to be viewed in context with features                 can be done in several ways. We show that while several such
visible in other times (pollution events, storm runoff, etc). Our            classifications are possible, the methodology-based classification
aim is to investigate current research into processing, combining,           appears most effective. An alternative classification scheme based
and visualizing multi-temporal raster imagery for the purposes of            on end-use is examined and shown to be less informative than the
increased data content and context.                                          methodology classification.
 The idea of combining raster aerial imagery in order to extract
maximum useful data has recently attracted attention from several            2.  CLASSIFICATION OF MULTI-
directions. The diversity of approaches and applications for this            TEMPORAL RASTER COMBINATION
 Permission to make digital or hard copies of all or part of this work for   Methodologies of the research projects illustrated here varied
 personal or classroom use is granted without fee provided that copies       greatly, but can be generally classified into four types. Those
 are not made or distributed for profit or commercial advantage and that     classes are defined as follows.
 copies bear this notice and the full citation on the first page. To copy
 otherwise, or republish, to post on servers or to redistribute to lists,    A. "Snapshot". This method presents raster data from multiple
 requires prior specific permission and/or a fee.                            times as a series of static images. This method was commonly
                                                                             used in early multi-temporal data sets, as it is the simplest and
                                                                             requires the least processing on the part of the GIS system. Data
                                                                             from different times can be stored as separate raster layers.
Georeferencing is carried out by matching constant features              coverage can be filled with estimations based on nearest temporal
between raster sets, via ground truthing, or between rasters and         neighbors.
vector base maps. Interpretation, overlays, and visualization are
                                                                         D. "Abstracting". This end use involves abstracting or simplifying
largely the responsibility of the user, although a few tools exist for
                                                                         multi-temporal data into a generalization about the feature or
change-detection and linear image subtraction.
                                                                         features being studied. The data can be used to make statements
B. "Temporal Averaging". In this approach, data from multiple            about overall change trends, patterns can be derived, or statistical
times is averaged to portray a baseline mean or median for the           data can be produced that does not require detailed knowledge of
entire time period being studied. Raster sets can be averaged            the original data elements
either on a pixel-by-pixel value level, or on a larger scale by
                                                                         As we show in section five, this classification scheme is too
averaging the spatial coverage of like-valued regions.
                                                                         limiting in describing individual research projects.
C. "Compositing". This method involves the combination of
multi-temporal data sets for maximum imagery detail. Images
                                                                         3.        PROJECTS BY CLASSIFICATIONS
from different times or seasons are combined and represented in a
way that allows features visible at one time to be portrayed along       3.1       Snapshot
with features which may obscure them in another time. This can           Remote Sensing and GIS methods for dynamics studies in the
be done by vectorizing one or more of the feature sets, by               Caspian Sea Coastal Zone by E.A. Baldina and I.A Labutina. [1]
animating multiple images, or by creating an new raster image to         This research focuses on analysis of changing environmental
portray combined data.                                                   conditions in Europe's Caspian Sea, specifically the Volga Delta
D. "Visualization". In the most simple form, this can be an              wetlands. This area has been designated as internationally
animation of a spatial region where each frame is from a different       important waterfowl habitat. The authors utilized aerial and
sequential time. More complex visualizations can take existing           satellite imagery from 1951, 1977-1981, 1989-1992, and 1996-
data and predict future values, or can highlight and modify images       1999 to study seasonal and yearly variation in sea levels and
to show temporal cause-and-effect relationships and dynamics not         vegetation cover.
obvious from a simple animation.                                         The operations performed on multi-temporal data were mainly to
                                                                         co-reference imagery with varying resolution, both spatial and
2.1       Alternate Classification                                       temporal. The more recent ERS SAR satellite data provided a
We identified one other major classification scheme as most likely       much finer temporal resolution (more frequent overflights), but
to meet the needs of our readers. Out of the range of possible           older aerial photography provided additional data on historical
classifications, that of final end-use seems to be the second-most       coastlines and land cover. Raster sets from different years and
effective for multi-temporal data sets. This classification would        capture sources had significant variation in angular distortion,
divide up research based on its eventual use and application,            spatial resolution, and coverage extent, which made
ignoring methods or approaches used to reach the final product.          georeferencing difficult. The changing sea levels in the area also
Categories for this type of classification could include the             complicated georeferencing, as there was not always a common
following data uses:                                                     topography for the region. In addition, some data sets offered
                                                                         insufficient spatial coverage to allow identification of control
A. "Tracking". At the basic level, multi-temporal data is mainly         points for georeferencing.
about tracking change over time. Whether it be feature attributes,
feature location, or other changes, the basic metric being               The authors overcame these problems by creating a common
produced is a record of change. While such a record is often             vector base map expressing the most constant and most easily
useful by itself, it can also be used to produce more detailed           identifiable geographic features of the study region. Raster images
information such as the following examples.                              were then referenced from control points identifiable on both the
                                                                         rasters and the vector base. For the small-extent aerials, it was
B. "Predicting". In this case, the goal is to show probable future       found that combining historic surface topography maps with the
values based on the history of previous values. A change over            more recent vector base allowed control points to be identified in
time in raster cell values or spatial coverage can be traced through     small areas. Afterwards the small-extent aerials were stitched
time to show patterns and dynamics, and can then be extrapolated         together into a mosaic image.
into the future based on known or inferred cause-and-effect
relationships. For example, urban sprawl can be measured as the          Once the common frame of reference and multi-temporal images
rate of development spread from an urban source, and can be used         had been integrated, it was found that historical maps predating
to predict the loss of agricultural land surrounding a city.             aerial imagery could also be added to the system. Maps from the
                                                                         late 19th and early 20th century were scanned into raster images
C. "Completing". Similar to the "predicting" goal, this use              and referenced into the common coordinate system.
involves filling in the gaps for areas obscured or invisible at one
time by inference from other time. As a trivial example, nighttime       While the display of multi-temporal data as separate "snapshots"
ground cover can be inferred from daytime ground cover (this             is one of the simplest, this research illustrates how it is not always
does not hold for all day/night attribute patterns, such as land use     easy to create such snapshots and coordinate them into a common
or socioeconomic activity). In a more realistic example, monsoon-        and correct georeferencing system.
season ground cover which is obscured by clouds can be derived
from dry-season imagery of the same area. Gaps in temporal
3.2       Visualization                                                different times, and can include "continue existence" (between
GIS and Remote Sensing techniques for the assembly of a                time A and time B), create, eliminate, destroy, and reincarnate.
database characterizing the spatio-temporal character of the           Additional operations deal with historically existing object state
South Australian continental shelf environments by Brett A.            transitions.
Bryan, Martine Kinloch, and David B. Gerner. [3]                       Once this model has been defined, the author describes a set of
The authors of this paper developed a GIS database for part of         operators to support temporal zooming. This concept essentially
Australia's near-coast ocean, in order to address what they saw as     allows abstraction or articulation (increased detail) of an object's
a lack of data for this biologically important region. Some of the     history. For example, the change in regional identity of "USSR"
main features studied were monthly and yearly water temperature,       to "Former USSR" is an abstraction of many events that made up
biological activity, and seasonal current patterns. Temporal           the dissolution of the Soviet Union, and many resulting new (or
resolution for each data set varied based on the sensor platforms      reincarnated) regional entity identities. The first operator which
used and the availability of sampling data from various satellites     supports this type of zooming is "uncover", which returns prior
and sea-surface observations. A spatial referencing system was         states linked to a current state, and refines the level of detail to
established based on location and the relatively constant sea          expose past states of an object. "Extend" returns the following
bottom topography.                                                     state of a current state, and can be used with uncover to move
                                                                       from a course-grained temporal view to a more detailed one. The
The authors are very meticulous about outlining their data             other two temporal operators are "Expand", which shows related
collection, storage, and analysis methods. Details of precision        sequences of an object's history, and "Collapse", which returns a
control and error factors are carefully and systematically covered.    less detailed sequence of an objects' history.
This is important because several of the features being studied are
observable only via second-hand effects or related factors. For        While this research is not directly related to a specific field of
example, biological activity is not directly measurable via raster     inquiry, it is a unique method of visualizing multi-temporal data
imagery, but the related quantity of chlorophyll a in the water        which is applicable to a wide range of uses. The ability to change
column has an affect on the color of the water surface. Similarly,     from temporal views at different levels of detail can lead to better
undersea temperatures were calculated based on observed surface        understanding of the overall temporal dynamics of a data set,
infrared data and current data, to predict water column mixing.        while retaining all of the original data in a form that can be easily
                                                                       accessed and visualized along with the abstracted overview.
The visualization method comes into play in the processing of
chlorophyll a data. The researchers created animated images of
yearly chlorophyll data in order to show the spatio-temporal
                                                                       3.3       Temporal Averaging
                                                                       GIS and Remote Sensing techniques for the assembly of a
dynamics of the feature set. An animation of sea-surface
                                                                       database characterizing the spatio-temporal character of the
temperature between 1997 and 2002 was also created. Data
                                                                       South Australian continental shelf environments by Brett A.
averaging performed on these sets is discussed further under the
                                                                       Bryan, Martine Kinloch, and David B. Gerner. [3]
"Temporal Averaging" subsection below.
                                                                       As mentioned in the previous subsection, Bryan et al collected a
                                                                       large amount of ocean data via repeated remote sensing over
Temporal Zooming by Kathleen Hornsby. [8]                              several years. The authors performed monthly averaging on sea-
                                                                       surface color data to produce yearly composites of cloud-free
This paper deals specifically and directly with the issues of
                                                                       imagery. This allowed the authors to study the spread and change
portraying and visualizing temporal data, and does not apply the
                                                                       in chlorophyll a, used as an indicator of biological activity.
methodology to any specific use other than example cases.
Hornsby's main idea is that temporal data can be displayed at          Sea-surface temperature was also averaged before being processed
varying levels of detail or resolution in the same way that spatial    into a visualization (discussed above). This study draws it's
data can be zoomed in and out to show different extents and            temperature source data from the NASA AVHRR scanning
resolutions. The concept of changing temporal granularity via          radiometer. The visible and near-infrared bands are useful for
zooming was first discussed in this work.                              cloud detection and the thermal infrared bands are used to
                                                                       measure surface temperature. To increase accuracy for sea-surface
Hornsby develops a new data model and a series of operations to
                                                                       temperature, the authors processed raw data into 15-day averages,
support this concept. The model allows descriptions of change for
                                                                       considering only valid non-cloud data for each cell's average
identifiable spatial (or feature) objects at different times.
                                                                       value. The temperature data was derived from the infrared bands
Operations defining object creation, destruction, "reincarnation",
                                                                       through an application of a linear function and correlation with
splits, joins, and other possible behaviors of objects can be
                                                                       ground-truthing data from ships and ocean-observing buoys.
expressed and stored in Hornsby's model. Examples of temporal
events or operations include the creation of a nation-state, a split
into multiple countries, and the possible destruction of a country
                                                                       Mapping the land cover of the forested area of Canada with
as an independent entity (such as through absorption into another
                                                                       Landsat data by Mike Wulder. [5]
country). Previous spatio-temporal data models had no way of
storing such relationships and time-based status of objects.           This paper presents a proposed large-scale land cover
Hornsby's model stores such data as "Object Identity states". An       classification program for assessing the forest resources of
object can exist, be non-existing without history (never existed),     Northern Canada. The author lays out the general spatial
or can be non-existing with history (existed in the past).             classification methods to be used by the program, including
Operations describe transitions in the object state between
standard K-means spatial clustering on a relatively small set of         Comparison of Methods for Estimation of Kyoto Protocol
interesting categories. While the methodology for temporal               Products of Forests From Multitemporal Landsat by David G.
classification is not fully explained, Wulder states that a mosaic of    Goodenough, A.S. Bhogal, Hao Chen, and Andrew Dyk. [7]
images combining leaf-on and leaf-off should be produced in
                                                                         This research aimed to calculate estimated forest harvesting for
order to aid seasonal land-cover classification. He also proposes
                                                                         determining adherence to international environmental treaties.
that ranging imagery over seasons and years will help avoid
                                                                         The authors used remote-sensing data from multiple years in
erroneous or missing data caused by clouds or unseasonal snow
                                                                         imitation of the methods assumed to be used by nations
cover. By averaging the set of images over time, the short-term
                                                                         monitoring each other. Initially, this research was done with
effects of weather can be reduced, and a greater number of cells
                                                                         single-season leaf-on remote sensing images from three years of
can be assigned near-true values.
                                                                         Landsat flights. In the most recent version, the authors combine
                                                                         multi-seasonal leaf-on and leaf-off images in order to separate
                                                                         mature deciduous trees from under story and 2nd-growth
Spatio-Temporal Aggregates over Raster Image Data by Jiue
                                                                         vegetation (which would appear after logging).
Zhang, Michael Gertz, and Demet Aksoy [10]
                                                                         The study co-referenced imagery into a common spatial area (a
Zhang et al. propose that a continuous stream of raster data (such
                                                                         region of Alberta, Canada), using GIS techniques. Image sets
as a live video or image stream transmitted by a remote sensor)
                                                                         were classified by systematic sampling and ground truthing from
can be averaged into discrete time or duration-based values.
                                                                         government harvest maps, then by using spatial clustering on both
While their main focus is on the summation and aggregation of
                                                                         the leaf-on and "leaf-all" (both leaf-on and leaf-off) data. It was
such a data stream (see further details in next section), they also
                                                                         found that the combined leaf-all data sets were much more
discuss applications which take averages based on the stream
                                                                         accurate when compared to traditional forest inventory methods,
aggregation. For example, their method supports such queries as
                                                                         offering a ground-vegetation classification accuracy of 91-95%,
an average ground-cover or surface temperature value for a region
                                                                         compared to 81-90% for leaf-on only data sets.
during a certain interval of time. As the ability to average is a
relatively minor feature based upon the creation of aggregate sets,
we further discuss this paper's features in the Compositing
                                                                         Photogrammetric and GIS techniques for the development of
                                                                         vegetation databases of mountainous areas: Great Smokey
                                                                         Mountains National Park. by Roy Welch, Marguerite Madden,
3.4       Compositing                                                    and Thomas Jordan. [6]
Integration of multi-seasonal remotely-sensed images for
improved landuse classification of a hilly watershed using               While the primary focus of this paper is the combination of
geographical information systems by J. Adinarayana and N. Rama           techniques used to orthorectify imagery in the case of severe
Krishna. [2]                                                             elevation changes without the aid of ground truthing, the authors
                                                                         also set forth some interesting compositing methods for multi-
Adinarayana and Krishna deal with the issue of partially-obscured        seasonal raster data. The study focuses on land cover and
temporal raster data. Their work aims at developing landuse              vegetation classification for a mountainous and largely
classifications for a region of India affected by seasonal monsoon       inaccessible national park, in which over story (tall trees) and
rains. These rains and the associated cloud cover cause surface          under story (brush and shrubs) both play a role in determining the
features to be obscured throughout most of the "Kharif"                  potential fire hazard of forest stands. The researchers used aerial
(Monsoon) season when some important cultivation takes place.            photos from leaf-on and leaf-off periods to develop a detailed
In contrast, relatively clear data can be obtained during the "Rabi"     database of vegetation cover, which allows dynamic calculation of
(post-Monsoon) season, but this season alone does not provide            feature attributes based on multi-seasonal data and interactive data
complete agricultural data.                                              input. For example, once an area's over story and under story
                                                                         vegetation classes have been recorded, a user can enter the current
The authors developed a set of classification rules for assigning
                                                                         rainfall conditions (wet, drought, etc), and the database can return
values to land cover in the region, based on a supervised
                                                                         an estimation of fire danger and fuel conditions for the chosen
classification raster set from the rainy Kharif season, the raster set
for the cloud-free Rabi season, and other raster feature sets for
soil properties. An expert system was designed to use these              Classification was made based on leaf-on aerial photography from
classification rules, in order to derive a landuse map for the           Fall of 1997 and 1998, and leaf-off photography from early
regions obscured by cloud and shadow during the monsoons.                Spring of 1998. A combination of digital and manual hard-copy
Training data for the expert system was collected via ground             image analysis allowed these sets to be coreferecned and
truthing for eight classes of land cover, including forest, pastures,    converted into vector coverage zones. The resulting products
open land, several types of agriculture, and cloudy or shadowed          included the combined understory/overstory classifications
cells. A two-dimensional cross table of knowledge-based rules            mentioned before, as well as seasonal snapshots of separate under
was developed, such as "(If Kharif + Rabi = vegetation, Drainage         and over story conditions which could be incorporated into other
Density = low, Soil = Chromusterets, and Slope =< 5 degrees,             analyses. Such techniques could easily be applied to pure raster
Then Class = Cropped pasture)". Various other logical                    cell data sets, although complex operations may take additional
combinations of Kharif and Rabi seasonal data were used to               time and resources than when working with vector conversions.
produce both an improved landuse map for the obscured areas,
and an improved overall multi-season GIS dataset.
A Web-based browsing and spatial analysis system for regional          wide range of data for a multitude of purposes, or is it narrowly
natural resource analysis and mapping by Ranga Raju Vatsavai,          application and topic-specific?
Thomas E. Burk, B. Tyler Wilson, and Shashi Shekhar. [4]
                                                                       Baldina and Labutina's [1] Caspian Sea dynamics study is very
This paper typifies the standard change detection analysis method      efficient, in that the only processing done is georeferencing raster
of temporal raster processing. Subsequent raster sets in an archive    sets from different times into a common spatial reference frame.
can have cell-by-cell subtraction performed on them to show the        This snapshot-type processing requires few computational
amount and patterns of change over time. The authors of this           resources. The automation potential depends largely on the quality
paper point out that such operations are very resource-intensive       of data collected. Skewed or distorted images are difficult for
when performed on a GIS server, and are bandwidth-intensive            algorithmic methods to match, and are more suited to human
when performed on clients which must download each raster from         analysis. If archived raster sets from sources with varying
the server. They propose that a pre-processing or "pre-realization"    resolution and distortion are being processed, then the potential
of common spatial and temporal analysis functions such as              for automation is low with today's technology. The result is
change-detection can reduce the overall per-request server-side        immediately useful in that it provides side-by-side comparisons of
and client network resource usage. Since preprocessing of all          feature change over time, but the potential exists for further
possible change-detection combinations would be prohibitive for        processing to enrich the final product. Accuracy is high when
large data sets, the authors propose a set of criteria for             feature points are collocated on a vector base map, especially
determining if pre-realization is beneficial.                          when large scale but small-extent maps are combined to provide a
                                                                       greater number of reference points. This simple method is
                                                                       applicable to a wide range of projects and is relatively
Spatio-Temporal Aggregates over Raster Image Data by Jiue              application-unspecific. On the whole, this method is simple and
Zhang, Michael Gertz, and Demet Aksoy [10]                             fast, but is somewhat outdated in modern use and provides limited
                                                                       initial results.
As mentioned in the section on spatio-temporal averaging, Zhang
and co's work is based on the ability to combine a continuous          Adinarayana and Krishna's [2] classification of hilly watersheds
stream of raster data into a single composite or sum. For a given      has a medium efficiency, as it requires extensive rule-based
time, all previous data contributing to the current situation can be   processing by an expert system. When initial setup and training
combined into a single summary set. This is useful for analyzing       time for the expert system is taken into account, overall efficiency
potentially unbounded sets (where data is still coming in), for        is rather low. Accuracy was reported to be high based on the
identifying maximums of coverage or attribute values, and for          authors' own estimations and ground-truthing data. The results of
reducing huge amounts of raw data (such as remote sensing video        the multi-seasonal combined vegetation dataset were immediately
or imagery) into a more compact storage format. Problems can           useful, providing new data about a time period not directly
arise in existing approaches to this problem, since missing pixels,    measurable due to atmospheric conditions. Cross-platform
breaks in the stream, or not-yet-scanned imagery may skew the          applicability seems very promising, as an expert system could be
value of the desired region. The paper solves this by computing a      trained to take any set of temporal rasters and combine them into a
number of answers for segments of the query region in which            new raster with a known or trained rule set. Automation is already
some image data only contributes partially. This segmentation          a major factor, although the system's training takes a good deal of
assures that each object (image) only provides data for a particular   human input. Overall this is a very good method for processing
segment. The segmentation is part of a two-part algorithm, the         large amounts of data and producing accurate renderings of
second component of which computes the aggregate value for             hidden features, but takes more time to set up and perform than
each section using a BA-tree approach. This allows for simplified      other methods.
summation-based raster cell output such as counts, sums, and
                                                                       Bryan, Kinloch and Gerner's [3] animated visualization sets are
averages (see previous section).
                                                                       quite useful to the end user and provide a highly-understandable
                                                                       abstraction of feature change over time which could be applied to
4.        COMPARISON                                                   almost any investigation of change over time. The accuracy of
We compare the effectiveness of each method proposed in the            their method depends upon the accuracy of input data, the data
reviewed papers based on a number of metrics. First, the               used in this particular research seemed to be mainly second-hand
efficiency of the method at reaching an end-use product should be      effects of invisible biological and climatological factors.
high, that is, the method should not take undue time in processing     Automation potential and efficiency are both high, as the creation
and outputting it's final product. Second, the ability to automate     of simple and moderately complex animations from multiple
the method is investigated. If the method in question is wholly or     raster frames is a very simple operation.
partially manual, can it be translated into an algorithmic
                                                                       Vatsavai et al's [4] proposals for improving resource use in multi-
operation? Thirdly, how applicable is the result of the method to
                                                                       temporal GIS systems are very good for improving efficiency. The
what people need? Is the result immediately useful in its field of
                                                                       method is applicable to any application which utilizes change-
inquiry, or does it require further post-processing before a
                                                                       detection as an analysis operation, but the final product may not
conclusion can be drawn from the data? Fourth, how accurate are
                                                                       be immediately useful without further processing. Automation is
the results? What degree of accuracy does the method or proposed
                                                                       supported through rules for data set preprocessing selection.
approach offer when compared to traditional ground truthing or
                                                                       Innate accuracy is high, as the data is not altered beyond simple
other previous work? Finally, how well does the method apply to
                                                                       raster subtraction.
other fields of investigation? Is it something that can be used on a
Wulder's [5] proposal to remove short-term imagery anomalies via         Project Efficiency Automation      Usefulness   Accuracy    Cross-
long-term averaging shows promise in the area of cross-platform                                                                       field
applicability. However, he does not detail any algorithmic
methods or discuss the potential efficiency of such an approach,                              dependent
so the effectiveness in these areas is questionable. Accuracy and
value to the user depends on the degree of trust placed in the data,     [2]        Medium    Yes          immediate High            Yes
a large number of cloudy or snowy datasets could degrade the
overall average value away from the ideal.                               [3]        High      Yes          immediate data-     Yes
Welch, Madden, and Jordan's [6] national park vegetation
database is fairly inefficient in it's data collection and integration   [4]        High      Yes          after      High           Yes
methods, and does not lend itself well to automation. These                                                additional
failings are due largely to the inaccessibility of the study region                                        processing
and the difficulty of conducting georeferencing on a uniformly
tree-covered image. It is possible that a different data set would       [5]        Unknown Unknown        data-     data-     Yes
not experience the same issues. Their user-friendly situational-                                           dependent dependent
dependant database provides very useful data output, and could be
applied to a wide range of uses with minor modification.                 [6]        Low       No           data-     High            Yes
Accuracy was shown to be acceptably high despite the lack of                                               dependent
ground-truthing availability.
                                                                         [7]        High      Yes          immediate High            Yes
Goodenough et al. [7] show that their method for forest product
use determination is more efficient that single-season multi-year        [8]        Low       Some         data-     data-     Yes
data analysis, and is in fact nearly as accurate as traditional                                            dependent dependent
ground-based resource-estimation studies. The output produced is
immediately useful for determining statistical data on forestry          [10]       High      Yes          data-     High            Some
harvests. Analysis is already done largely by automated systems,                                           dependent
and could be applied to other studies which use multi-seasonal
data for enhanced ground-cover classification.                           Table 1.

Hornsby's [8] Temporal Zooming concept is very useful for any            Table 1 shows a cross-project comparison of the 5 evaluation
large or complex temporal data set in which different levels of          metrics. From this table, it is apparent that the most difference
detail can enhance understanding. The initial setup and                  between methods lies in their degree of usefulness (immediate vs
assignment of object states and transitions are likely to be slow        post-processing), and accuracy (high vs data-dependent). Every
and resource-demanding. Accuracy is less of a factor in this             method seemed to have potential for wide application, most were
methodology, as data is not actually altered, but the accuracy of        at least as efficient as existing methods, and more than half
the visualization should be high to avoid misinterpretation. As          allowed some degree of automation.
there is no real-world implementation and only theoretical
examples given by the author, the potential accuracy and end-use         5.         VALIDATION
functionality are not immediately clear.                                 We feel that the Methodology-based classification system is the
                                                                         best choice for this survey. The end-use-based classification
Zhang et al's. [10] Spatio-temporal aggregation approach is
                                                                         scheme which we present as an alternative falls short in several
applicable mainly to streaming raster data, but could potentially
                                                                         areas. Firstly, data is often in a specific format and only certain
be applied to historical raster collections in order to observe
                                                                         tools are available to work with it. By choosing a system based on
patterns, averages, or the results of change over time. The
                                                                         end-use, the techniques utilized to reach that goal may be
proposed method of segmenting data before aggregating was
                                                                         incompatible with available data. It is true that in cases where the
shown to be highly efficient, contributing no significant delay to
                                                                         research begins from scratch with no data, or with cross-platform
the overall aggregation computations used in previous methods.In
                                                                         raw data, choosing a system based on end-use may be convenient.
addition, the accuracy of data aggregates returned by this method
                                                                         However, our second point is that choosing a system based on
is higher than that of previous methods, as the partitioning
                                                                         end-use can also limit functionality of the system. Generally, a
approach reduces the skewing effects of neighbors, outliers, and
                                                                         user would have to choose a single system that could complete
missing data. Automation takes the form of a two-step algorithm
                                                                         their primary goal. If the user wanted to do other things with their
using tree-based summation. Depending on the desired end-
                                                                         data, or if the data and end products were to be shared between
product, the output is either immediately useful (such as an
                                                                         users, a proprietary or specialized system based on end-use may
averaged value for an attribute), or may require some additional
                                                                         not be able to handle a wider range of operations.
                                                                         As such, we feel that a classification system based solely on end-
                                                                         use does not meet the requirements needed to evaluate and
Project Efficiency Automation       Usefulness   Accuracy     Cross-     categorize research projects. Users attempting to tailor a system to
                                                               field     their needs would be limited to only a few choices, and may not
                                                                         be presented with the full range of possible options when using
[1]      High         data-        immediate High            Yes         such a classification system. In addition, readers and reviewers
                                                                         using such a system to classify research will be forced to pigeon-
hole projects into one or more categories, which is inappropriate       most room for improvement. Depending on the data set and the
for widely-applicable research such as Hornsby's Temporal               desired product, it may be useful to create a temporal averaging or
Zooming [8].                                                            abstraction of the snapshot data. It may also be beneficial to
                                                                        combine snapshots into composites to show feature sets common
Our chosen classification scheme allows a higher degree of
                                                                        to multiple times or conditions. In the specific case of this
abstraction when comparing papers, so that multiple projects can
                                                                        research, users could possibly have benefited from an averaged
potentially be accurately classified in multiple ways. By using
                                                                        plot of sea-level conditions as compared to the average health and
methodology as a classifier, we believe that we can effectively
                                                                        extent of vegetation.
show multiple types of research in a wide range of end uses, as
well as a range of cross-method approaches. For example, we             While Bryan, Kinloch and Gerner's [3] visualization method does
show how Bryan, Kinloch, and Gerner [3] use both Visualization          an adequate job of portraying a long temporal span, it may be
and Temporal averaging to meet the needs of their research. We          useful to combine this with Hornsby's [8] zooming visualization
also discuss abstracted uses of each approach in ways that could        or even Zhang's aggregate functions. Large-scale changes or
increase the potential range of uses for each method. We feel that      aggregates of periodic changes in sea conditions deemed
this is the most useful categorization for researchers trying to        significant to the research could be used as low-detail sets, with
combine raster spatio-temporal data. By offering a wider and more       zooming available to produce more detailed histories on
open classification, we hope that readers will have a better idea of    command.
the choices available to meet their needs.
                                                                        7.        REFERENCES
6.        CONCLUSION                                                    [1] Baldina and Labutina, (2002) “Remote sensing and GIS
6.1       Common Elements                                                   methods use for dynamics studies in the Caspian Sea coastal
While commonalities between approaches in each category should              zone” in Geoscience and Remote Sensing Symposium, 2002.
be obvious, a few projects deserve having their similarities                IGARSS '02. IEEE International. Volume 5. Pages 2838-
highlighted. All three papers classified as "Temporal Averaging"            2840
attempt to reduce the effect of imaging outliers on the eventual        [2] Adinarayana and Krishna. (1996) “Integration of multi-
raster cell assignments. Both Wulder and Bryan et al. deal with             seasonal remotely-sensed images for improved landuse
the issues of "raw" weather imagery, in which clouds, cloud                 classification of a hilly watershed using geographical
shadow, rain, and other "null-value" or "obscured value" cells              information systems” In International Journal of Remote
appear. In the Compositing class, Goodenough et al [7],                     Sensing, Volume 17, Issue 9, Pages 1679 – 1688
Adinarayana and Krishna [2], and Welch et al [6] are all working        [3] Bryan, Kinloch, and Gerner (2003) “GIS and Remote
towards improving data context from dissimilar sets acquired in             Sensing techniques for the assembly of a database
different seasons. The creation of a "all-season" classifier for land       characterizing the spatio-temporal character of the South
cover based on leaf-on and leaf-off requires not only linear                Australian continental shelf environments” In Proceedings of
combinations of pixels, but often complex rule-based analysis of            Coastal GIS 2003 7 July 2003, University of
the effects of different types and families of vegetation visible at        Wollongong,Wollongong, Australia.
different times.
                                                                        [4] Ranga Raju Vatsavai, Thomas E. Burk, B. Tyler Wilson,
Despite the fact that we have placed it into a separate class,              Shashi Shekhar. (2000) “A Web-based browsing and spatial
Wulder's [5] paper shows some marked similarities to the works              analysis system for regional natural resource analysis and
by Goodenough et a [7] and Welch et al. [6] Each of these                   mapping”, In Proceedings of the 8th ACM international
methods are designed to handle combinations of leaf-on/leaf-off             symposium on Advances in geographic information systems,
raster data, but the averaging approach used by Wulder [5] is               Pages 95-101.
closer to that of Bryan et al. [3].
                                                                        [5] Wulder, M. (2002) "Mapping the land cover of the forested
                                                                            area of Canada with Landsat data", In Geoscience and
6.2       Potential Areas for Improvement                                   Remote Sensing Symposium, Volume 3, Pages 1303 - 1306.
We feel that the methods described in several of these papers
                                                                        [6] Welch, R. M. Madden, and T Jordan. (2002).
could be enhanced by combination with other methods
                                                                            "Photogrammetric and GIS techniques for the development
investigated. For example, the simplified change-detection
                                                                            of vegetation databases of mountainous areas: Great Smokey
algorithm discussed by Vatsavai et al [4] might be improved
                                                                            Mountains National Park". In ISPRS Journal of
through visualization. Hornsby's [8]Temporal Zooming could in
                                                                            Photogrammetry and Remote Sensing. Volume 57, Pages 53-
fact be used server-side to preprocess common views for a
temporal sequence, thus further improving the efficiency of query
responses. Change detection products could be replaced by               [7] Goodenough, D. G., A.S. Bhogal, H. Chen, and A. Dyk.
animations which would allow users to directly view change over             (2001). "Comparison of Methods for Estimation of Kyoto
time, and may provide details otherwise overlooked in a single              Protocol Products of Forests From Multitemporal Landsat".
raster change map.                                                          In Geoscience and Remote Sensing Symposium, 2001. IEEE
                                                                            International.Volume 2, Pages 764 - 767
Baldina and Labutina's [1] snapshot-based approach could also
benefit from combination with one or both of the visualization          [8] Hornsby, K. (2001). "Temporal Zooming". In Transactions
methods investigated here. As the snapshot technique is one of the          in GIS Volume 5, Issue 3, Pages 255-272
simplest ways to deal with multi-temporal data, it has perhaps the
[9] Jeong, S-H, A. Fernandex, N. W. Paton, and T. Griffiths      [10] Zhang, J. , M. Getz, and D. Aksoy (2004). "Spatio-Temporal
    (2005). "An Experimental Performance Evaluation of Spatio-        Aggregates over Raster Image Data" In Proceedings of the
    Temporal Join Strategies" In Transactions in GIS Volume 9         12th annual ACM international workshop on Geographic
    Issue 2. Pages 129-156                                            information systems. Pages 39-46.

To top