dividing the total number of pixels that were correctly classified (i.e., assigned to
the correct class; principal axis in Table 14.1), by the total number of validation
pixels or reference points used to construct the error matrix (Congalton 1991).
Understanding the overall, user’s and producer’s accuracy values is essential for
interpreting thematic maps, determining if they are useable for a specific application, and understanding which image map classes are mapped more accurately
than others.
Although the overall accuracy is the most commonly used accuracy measure, it
does not take into account both errors of commission and omission for all individual classes. As a result, overall accuracy tends to overestimate accuracy (Ma
and Redmond 1995). As a result, adjustments to the overall accuracy are often
made to account for the chance agreement or accuracy resulting from a completely
random assignment of pixels to mapping categories (Lillesand et al. 2008). This is
referred to as ‘adjusted’ overall accuracy and is represented by Kappa (Congalton
and Mead 1983) or Tau (Ma and Redmond 1995) values. Although adjusted
accuracy measures were commonly used in past remote sensing studies, recent
research advises that they are not as useful as thought due to significant flaws in
their assumptions and methods and it is now advised to report only overall
accuracy (Foody 2011; Pontius Jr and Millones 2011).
14.2.3 Accuracy of Continuous Maps
Continuous maps, in comparison to discrete maps, require alternative accuracy
measures as they do not consist of a set of fixed thematic classes, but rather
represent a biophysical property (e.g., depth) that falls within a given range for
every pixel (e.g., 0–30 m water depth). In general, continuous mapping methods
are applied for mapping biophysical properties that vary across an image, examples of which include percentage cover, depth, and water column properties such
as chlorophyll biomass and light attenuation. The most commonly used accuracy
measures for continuous maps are the root-mean square error (RMSE) and
regression coefficient (i.e., coefficient of determination), which are both used to
quantify the co-variance between reference data and mapped data.
Because RMSE and linear regression have been commonly applied to various
coral reef datasets, map users can easily compare results with the published literature to evaluate the relative quality of the map product. As with thematic
accuracy assessment data, the reference sample set for continuous maps should
also be a statistically sound representation of the continuous parameter.
Effective validation of continuous variables requires collection of field measurements coinciding with acquisition of the remote sensing data, which is often
challenging in dynamic environments, or impractical in remote or very large coral
reefs. As a result, point-to-point comparison of field measures and image data can
be limited. Therefore, alternative statistical parameters (e.g., distribution functions) need to be estimated to assess the accuracy of dynamic continuous variables.
14 Validation
381
the correct class; principal axis in Table 14.1), by the total number of validation
pixels or reference points used to construct the error matrix (Congalton 1991).
Understanding the overall, user’s and producer’s accuracy values is essential for
interpreting thematic maps, determining if they are useable for a specific application, and understanding which image map classes are mapped more accurately
than others.
Although the overall accuracy is the most commonly used accuracy measure, it
does not take into account both errors of commission and omission for all individual classes. As a result, overall accuracy tends to overestimate accuracy (Ma
and Redmond 1995). As a result, adjustments to the overall accuracy are often
made to account for the chance agreement or accuracy resulting from a completely
random assignment of pixels to mapping categories (Lillesand et al. 2008). This is
referred to as ‘adjusted’ overall accuracy and is represented by Kappa (Congalton
and Mead 1983) or Tau (Ma and Redmond 1995) values. Although adjusted
accuracy measures were commonly used in past remote sensing studies, recent
research advises that they are not as useful as thought due to significant flaws in
their assumptions and methods and it is now advised to report only overall
accuracy (Foody 2011; Pontius Jr and Millones 2011).
14.2.3 Accuracy of Continuous Maps
Continuous maps, in comparison to discrete maps, require alternative accuracy
measures as they do not consist of a set of fixed thematic classes, but rather
represent a biophysical property (e.g., depth) that falls within a given range for
every pixel (e.g., 0–30 m water depth). In general, continuous mapping methods
are applied for mapping biophysical properties that vary across an image, examples of which include percentage cover, depth, and water column properties such
as chlorophyll biomass and light attenuation. The most commonly used accuracy
measures for continuous maps are the root-mean square error (RMSE) and
regression coefficient (i.e., coefficient of determination), which are both used to
quantify the co-variance between reference data and mapped data.
Because RMSE and linear regression have been commonly applied to various
coral reef datasets, map users can easily compare results with the published literature to evaluate the relative quality of the map product. As with thematic
accuracy assessment data, the reference sample set for continuous maps should
also be a statistically sound representation of the continuous parameter.
Effective validation of continuous variables requires collection of field measurements coinciding with acquisition of the remote sensing data, which is often
challenging in dynamic environments, or impractical in remote or very large coral
reefs. As a result, point-to-point comparison of field measures and image data can
be limited. Therefore, alternative statistical parameters (e.g., distribution functions) need to be estimated to assess the accuracy of dynamic continuous variables.
14 Validation
381