140
of decline), we can use field data to present an accuracy assessment. Any accuracy assessment of classified image products should include overall accuracy as
well as users’ accuracy (percent of target pixels correctly classified;
inverse = errors of omission) and producers’ accuracy (percent of nontarget
pixels that are not classified as the target class; inverse = errors of commission;
Congalton 2001; Fassnacht et al. 2006). Splitting accuracy into users’ and producers’ values allows the end user to understand how false positives (saying a
stand is dead when it is not) and false negatives (saying a stand is healthy when
it is dead) can influence how the end product is used to inform management
activities. For example, if overall accuracy in classifying forest mortality is 70%
but almost all of the error results from false positives (many stands classified as
dead when they are actually alive), end users may decide to limit management
to locations with large clusters of predicted mortality or to clusters in higherdecline categories in order to avoid these common errors.
RS decline-detection products that result in ordinal classes of decline (e.g.,
healthy, degrees of decline, dead) can also be assessed for “fuzzy accuracy,”
which considers not only correct class assignments but also those within one
ordinal class of the correct class. Products that provide a continuous decline
metric can be used to produce more detailed accuracy metrics. Standard statistical regression techniques produce a coefficient of determination (r
2
) to describe
how well a statistical model fits the relationship between the input spectral variables and the output decline metric. Root mean square error, standard errors, and
prediction errors can be used to place confidence bounds on predicted values. We
can also examine how accuracy changes across the range of decline values predicted. For example, some models may be very good at quantifying severe
decline but may not be able to detect early decline symptoms. Some models may
overpredict early decline but underpredict severe decline. Standard statistical
methods can be useful to examine how well your model works, which is critical
to ensure that end users know how to best integrate your resulting RS products
into their decision-making process.
A Nested Approach No one sensor, field methodology, or scale is appropriate
for all applications. Different goals may require that you work at different scales
(Fig. 6.9). The most detailed and accurate information about specific stress agents
and response symptoms will always be obtained from on-the-ground field surveys
(Tier 1). Such location-specific studies allow researchers to directly measure
foliar chemistry, canopy structure, and spectral characteristics in situ. But these
studies are limited in their utility to inform management across the broader landscape. Aerial sensors are often used to collect RS imagery at the local scale (Tier
2). Typically, this scale allows for the use of high spatial and spectral resolution
imagery, ideally suited to detect forest stress conditions. However, such efforts
may still be limited in geographic extent due to the high cost and computing
needs. Most common is the use of broadband sensors at the regional-continental
scale (Tier 3). Landsat sensors have been widely used for such applications, with
sufficient spatial (30 m) and spectral resolution to prove useful in assessment of
J. Pontius et al.
of decline), we can use field data to present an accuracy assessment. Any accuracy assessment of classified image products should include overall accuracy as
well as users’ accuracy (percent of target pixels correctly classified;
inverse = errors of omission) and producers’ accuracy (percent of nontarget
pixels that are not classified as the target class; inverse = errors of commission;
Congalton 2001; Fassnacht et al. 2006). Splitting accuracy into users’ and producers’ values allows the end user to understand how false positives (saying a
stand is dead when it is not) and false negatives (saying a stand is healthy when
it is dead) can influence how the end product is used to inform management
activities. For example, if overall accuracy in classifying forest mortality is 70%
but almost all of the error results from false positives (many stands classified as
dead when they are actually alive), end users may decide to limit management
to locations with large clusters of predicted mortality or to clusters in higherdecline categories in order to avoid these common errors.
RS decline-detection products that result in ordinal classes of decline (e.g.,
healthy, degrees of decline, dead) can also be assessed for “fuzzy accuracy,”
which considers not only correct class assignments but also those within one
ordinal class of the correct class. Products that provide a continuous decline
metric can be used to produce more detailed accuracy metrics. Standard statistical regression techniques produce a coefficient of determination (r
2
) to describe
how well a statistical model fits the relationship between the input spectral variables and the output decline metric. Root mean square error, standard errors, and
prediction errors can be used to place confidence bounds on predicted values. We
can also examine how accuracy changes across the range of decline values predicted. For example, some models may be very good at quantifying severe
decline but may not be able to detect early decline symptoms. Some models may
overpredict early decline but underpredict severe decline. Standard statistical
methods can be useful to examine how well your model works, which is critical
to ensure that end users know how to best integrate your resulting RS products
into their decision-making process.
A Nested Approach No one sensor, field methodology, or scale is appropriate
for all applications. Different goals may require that you work at different scales
(Fig. 6.9). The most detailed and accurate information about specific stress agents
and response symptoms will always be obtained from on-the-ground field surveys
(Tier 1). Such location-specific studies allow researchers to directly measure
foliar chemistry, canopy structure, and spectral characteristics in situ. But these
studies are limited in their utility to inform management across the broader landscape. Aerial sensors are often used to collect RS imagery at the local scale (Tier
2). Typically, this scale allows for the use of high spatial and spectral resolution
imagery, ideally suited to detect forest stress conditions. However, such efforts
may still be limited in geographic extent due to the high cost and computing
needs. Most common is the use of broadband sensors at the regional-continental
scale (Tier 3). Landsat sensors have been widely used for such applications, with
sufficient spatial (30 m) and spectral resolution to prove useful in assessment of
J. Pontius et al.