You are on page 1of 2

There is a resurgence in the art and science of visual image interpretation as the digital remote sensor

systems provide increasingly higher spatial resolution imagery. Many people are displaying high spatial
resolution GeoEye-1 and WorldView-2 imagery on the computer screen and then visually interpreting the
data. The data are also often used as a base map in GIS projects (Jensen and Jensen, 2013)

Analog (hard-copy) and digital remotely sensed data are used operationally in many Earth science, social
science, and planning applications (e.g., Jensen and Jensen, 2013). Analog remotely sensed data such as
positive 9 × 9 in. aerial photographs are routinely analyzed using the fundamental analog elements of
image interpretation (e.g., size, shape, shadow) and optical instruments such as stereoscopes and zoom-
transferscopes. Digital remote sensor data are analyzed using a digital image processing system that
consists of computer hardware and special-purpose image processing software.

Ehlers et al. (1990) used this methodology to merge SPOT 20 × 20 m multispectral and SPOT panchromatic
10 × 10 m data. The resulting multiresolution image retained the spatial resolution of the 10 × 10 m SPOT
panchromatic data, yet provided the spectral characteristics (hue and saturation values) of the SPOT
multispectral data. The enhanced detail available from merged images was found to be important for
visual land-use interpretation and urban growth delineation (Ehlers et al., 1990). In a similar study, Carper
et al. (1990) found that direct substitution of the panchromatic data for intensity (I) derived from the
multispectral data was not ideal for visual interpretation of agricultural, forested, or heavily vegetated
areas. They suggested that the original intensity value obtained in step 1 be computed using a weighted
average (WA) of the SPOT panchromatic and SPOT multispectral data; that is, WA = {[(2 × SPOT Pan) +
SPOT XS3]/3}.

Chavez

The high resolution image

A human interpreter performing visual image interpretation may label an image-derived polygon
incorrectly.

. In addition, the person(s) collecting ground reference information in the field often make mistakes (e.g.,
mislabeling). Therefore, it is imperative to use the term ground refer

ence information rather than ground-truth because almost all ground reference information will contain
some error (Stehman and Wickham, 2011).

Some image analysts perform an accuracy assessment based only on the training pixels (or training
polygons if the study is based on human visual interpretation or object-based image analysis—OBIA) used
to train classification algorithm. Unfortunately, the locations of these training sites are usually not
random. They are biased by the analyst’s a priori knowledge of where certain land-cover types exist in the
scene. Because of this bias, an accuracy assessment based on how well the training class data were
classified usually results in higher classification accuracies than an accuracy assessment based on unbiased
ground reference test information (e.g., Muchoney and Strahler, 2002).
The

The ideal situation is to locate ground reference test pixels (or polygons) in the study area. These sites
are not used to train the classification algorithm and therefore represent unbiased reference
information. It is possible to collect some ground reference test information prior to the classification,
perhaps at the same time as the training data. But the majority of test reference information is often
collected after the classification has been performed using a random sample to collect the appropriate
number of unbiased observations per category (to be discussed).

Landscapes often change rapidly. Therefore, it is best to collect both the training and ground reference
test information as close to the date of remote sensing data acquisition as possible. This is especially
important where the land use or land cover changes rapidly within a single season (Foody, 2010). For
example, agriculture and illicit drug crops often have relatively short growing seasons or are harvested
rapidly and then repeatedly during a season. If a scientist waits too long to obtain the ground reference
test information, it may not be possible to tell what was actually in the field the day of the remote
sensing data acquisition. This is a serious condition because it becomes very difficult to populate an
error matrix with accurate information.

Ideally, the ground reference test data are obtained by visiting the site on the ground and making very
careful observations that can then be compared with the remote sensing–derived information for the
exact location (McCoy, 2005). Unfortunately, sometimes it is difficult to actually visit all the sites
identified in the random sample. Some sites selected in the random sample may be completely
inaccessible due to extremely rugged terrain. Others may be inaccessible because private land owners,
government agencies, or even criminals (e.g., illegal drug cultivators) deny access.

When this occurs, scientists sometimes obtain higherspatial-resolution remotely sensed data, interpret
it, and use it as a surrogate for ground reference test information (e.g., Morisette et al., 2005). In such
cases, the general rule of thumb is that the imagery used to obtain the ground reference test
information should be substantially higher in spatial and/or spectral resolution than the imagery used to
derive the original thematic information. For example, many studies have used high spatial resolution
aerial photography (e.g., nominal spatial resolution <1 × 1 m) to extract ground reference test
information that is then compared with thematic data produced using Landsat Thematic Mapper
imagery (30 × 30 m). Admittedly, this is not the best solution. However, sometimes it is the only
alternative if one wants to populate an error matrix and perform an accuracy (error) assessment.

You might also like