transformations may occur within minutes, rather than weeks, years, or decades, as is
the case with land-based changes. This basic potential for swift change makes in situ
validation data even more important than for land-based analyses. Increased repeat
sensor coverage also becomes more important to gain a better understanding of the
more minute changes occurring over a shorter period of time. The greater the time
period between data acquisitions, the greater the number of assumptions that are
typically made, resulting in increased errors that could therefore directly influence final
results and conclusions.
Other types of error may be introduced in the data integration stage. In processing
multi-sensor, multi-stage satellite images for change detection, for example, there are
several factors that must be addressed. The first is geometric processing so that each
image matches the other in terms of coverage extent. If the image registration is not
done correctly, determining where and how much change has occurred is impossible
(Igbokwe, 1999). Another important potential source of error in change detection
analysis is correction for atmospheric effects. Since different dates of imagery are
required, atmospheric effects will not be the same for each scene. This topic is
thoroughly explained and discussed in Conghe et al. (2001).
A fourth element of data integration error occurs when mathematically evaluating
and comparing satellite-derived products to one another. Such errors occur, for
example, when comparing thematic maps based on map values to decide which map is
more appropriate for an application. Several comparative statistics may be employed to
help assess whether the derived maps are significantly different, or no better than a
random result. These evaluative methods include the error matrix, kappa statistic, and
Z statistic. These statistics can be very helpful to the novice remote sensing researcher
and are discussed at length in Congalton and Green (1999), Stehman (1999), and
Congalton (2001).
The integration of remote sensing and in situ data can also lead to serious error. In
terms of coastal land cover or land use change, for example, it is important to note
whether the researcher starts from survey data and links it to landscape change, or
begins from remotely sensed data of the land and links it to survey data (Rindfuss et al.,
2001). In either case, the investigator must accurately identify and georeference the
parcel of land under investigation. This can be very difficult, especially in less
developed countries lacking cadastral surveys. An exceedingly time consuming
approach involves going to the field with a GPS unit. However, this will usually result
in smaller sample sizes. After the parcel is georeferenced, it must be co-registered with
the remotely sensed image. Co-registration in most cases is fairly uncomplicated.
However, in some instances registration may be difficult due to lack of ground control
points, thereby introducing spatial uncertainty. This spatial uncertainty would
potentially increase the errors of any subsequent research using the two datasets, such
as overlay analysis in a GIS (Rindfuss and Stern, 1998; Evans and Moran, 2002).
Additionally, a spatial mismatch between plot size and pixel size may exist. The remote
sensing spatial resolution may be larger than the plot sizes. Higher resolution imagery
or aerial photography may be a solution in some instances, but for most studies their
purchase will be outside the means of investigators (Evans and Moran, 2002).
There is a progressive trend to integrate remote sensing and social science data.
There are two ways of approaching data integration with these two data types. The first
is to take social science data and create a grid based on socioeconomic data, thus
matching it with the format of earth science data. This approach has been coined
“pixelizing the social” in coastal change (Geoghegan et al., 1998). One method to
pixelize the social has been developed by the Center for International Earth Science
193
Integration of New Data Types
the case with land-based changes. This basic potential for swift change makes in situ
validation data even more important than for land-based analyses. Increased repeat
sensor coverage also becomes more important to gain a better understanding of the
more minute changes occurring over a shorter period of time. The greater the time
period between data acquisitions, the greater the number of assumptions that are
typically made, resulting in increased errors that could therefore directly influence final
results and conclusions.
Other types of error may be introduced in the data integration stage. In processing
multi-sensor, multi-stage satellite images for change detection, for example, there are
several factors that must be addressed. The first is geometric processing so that each
image matches the other in terms of coverage extent. If the image registration is not
done correctly, determining where and how much change has occurred is impossible
(Igbokwe, 1999). Another important potential source of error in change detection
analysis is correction for atmospheric effects. Since different dates of imagery are
required, atmospheric effects will not be the same for each scene. This topic is
thoroughly explained and discussed in Conghe et al. (2001).
A fourth element of data integration error occurs when mathematically evaluating
and comparing satellite-derived products to one another. Such errors occur, for
example, when comparing thematic maps based on map values to decide which map is
more appropriate for an application. Several comparative statistics may be employed to
help assess whether the derived maps are significantly different, or no better than a
random result. These evaluative methods include the error matrix, kappa statistic, and
Z statistic. These statistics can be very helpful to the novice remote sensing researcher
and are discussed at length in Congalton and Green (1999), Stehman (1999), and
Congalton (2001).
The integration of remote sensing and in situ data can also lead to serious error. In
terms of coastal land cover or land use change, for example, it is important to note
whether the researcher starts from survey data and links it to landscape change, or
begins from remotely sensed data of the land and links it to survey data (Rindfuss et al.,
2001). In either case, the investigator must accurately identify and georeference the
parcel of land under investigation. This can be very difficult, especially in less
developed countries lacking cadastral surveys. An exceedingly time consuming
approach involves going to the field with a GPS unit. However, this will usually result
in smaller sample sizes. After the parcel is georeferenced, it must be co-registered with
the remotely sensed image. Co-registration in most cases is fairly uncomplicated.
However, in some instances registration may be difficult due to lack of ground control
points, thereby introducing spatial uncertainty. This spatial uncertainty would
potentially increase the errors of any subsequent research using the two datasets, such
as overlay analysis in a GIS (Rindfuss and Stern, 1998; Evans and Moran, 2002).
Additionally, a spatial mismatch between plot size and pixel size may exist. The remote
sensing spatial resolution may be larger than the plot sizes. Higher resolution imagery
or aerial photography may be a solution in some instances, but for most studies their
purchase will be outside the means of investigators (Evans and Moran, 2002).
There is a progressive trend to integrate remote sensing and social science data.
There are two ways of approaching data integration with these two data types. The first
is to take social science data and create a grid based on socioeconomic data, thus
matching it with the format of earth science data. This approach has been coined
“pixelizing the social” in coastal change (Geoghegan et al., 1998). One method to
pixelize the social has been developed by the Center for International Earth Science
193
Integration of New Data Types