274
overview]. Continuum removal is another technique used to target absorption
features. For each pixel reflectance, a convex hull is fit over the top of the spectrum, absorption features are normalized to that hull, and the depth of a specific
absorption feature (e.g., leaf water content) can be quantified. PCA is a linear
transformation method that maximizes the variance of the data. When applied to
a hyperspectral image, it produces a series of components that correspond to
linear combinations of the original bands aligned to represent the variation
within the original data set, with the first component being the plane responsible
for the most variation. This allows for determining the most significant characteristics within an image that relate to classes. Minimum noise fraction transformation (MNF) rescales the noise in the data (a process called noise whitening),
enabling the analyst to eliminate bands containing too much sensor noise and
leaving only coherent image data.
Commonly used classification techniques include random forest, a supervised
machine learning algorithm that constructs many decision trees and utilizes their
outputs to get an accurate class prediction based upon training data, and maximum likelihood estimation (MLE), a supervised classification method in which
parameter values of a statistical model are determined that maximize the chance
that the process described by the model was actually observed. All of these data
enhancement and classification methods can be performed using open-source
software, such as R (https://www.r-project.org/) and Python (https://www.
python.org/), where many packages are available to use, or in commercial software, such as ENVI (https://www.harrisgeospatial.com/).
4. Assess accuracy. One of the most important considerations is accuracy assessment following mapping. Depending on the objectives of the study, some types
of error may be acceptable, while some may not. Typical accuracy metrics for
image classification include overall accuracy, user’s accuracy, producer’s accuracy, and Kappa coefficient. Overall accuracy is the probability that an image
classifier will correctly classify a pixel. This metric does not account for the
number of validation pixels per class and may be misleading if a similar number
is not used for each class. User’s accuracy and producer’s accuracy may be better
metrics for assessing the classification. User’s accuracy (error of commission) is
the fraction of correctly classified pixels with regard to all pixels classified.
Producer’s accuracy (errors of omission) is the fraction of correctly classified
pixels with regard to all ground reference validation pixels. In some situations,
such as automated weed management in agriculture, overall accuracy and producer’s accuracy may not be as much of a concern as user’s accuracy because
identifying small amounts of weeds (IAS) as crops may be okay, but spraying
crops misidentified as IAS could be more damaging to crop yields than the IAS
themselves. An example where maximizing producer’s accuracy may be more
important would be in mapping IAS to understand species spread and the invasion process; any omitted species data as changes are monitored over time could
affect process understanding and spread predictions. The last metric, the Kappa
coefficient, can be useful for comparing multiple classification methods within
the same data set. The Kappa coefficient is a measure of how closely the resulting
E. A. Bolch et al.
overview]. Continuum removal is another technique used to target absorption
features. For each pixel reflectance, a convex hull is fit over the top of the spectrum, absorption features are normalized to that hull, and the depth of a specific
absorption feature (e.g., leaf water content) can be quantified. PCA is a linear
transformation method that maximizes the variance of the data. When applied to
a hyperspectral image, it produces a series of components that correspond to
linear combinations of the original bands aligned to represent the variation
within the original data set, with the first component being the plane responsible
for the most variation. This allows for determining the most significant characteristics within an image that relate to classes. Minimum noise fraction transformation (MNF) rescales the noise in the data (a process called noise whitening),
enabling the analyst to eliminate bands containing too much sensor noise and
leaving only coherent image data.
Commonly used classification techniques include random forest, a supervised
machine learning algorithm that constructs many decision trees and utilizes their
outputs to get an accurate class prediction based upon training data, and maximum likelihood estimation (MLE), a supervised classification method in which
parameter values of a statistical model are determined that maximize the chance
that the process described by the model was actually observed. All of these data
enhancement and classification methods can be performed using open-source
software, such as R (https://www.r-project.org/) and Python (https://www.
python.org/), where many packages are available to use, or in commercial software, such as ENVI (https://www.harrisgeospatial.com/).
4. Assess accuracy. One of the most important considerations is accuracy assessment following mapping. Depending on the objectives of the study, some types
of error may be acceptable, while some may not. Typical accuracy metrics for
image classification include overall accuracy, user’s accuracy, producer’s accuracy, and Kappa coefficient. Overall accuracy is the probability that an image
classifier will correctly classify a pixel. This metric does not account for the
number of validation pixels per class and may be misleading if a similar number
is not used for each class. User’s accuracy and producer’s accuracy may be better
metrics for assessing the classification. User’s accuracy (error of commission) is
the fraction of correctly classified pixels with regard to all pixels classified.
Producer’s accuracy (errors of omission) is the fraction of correctly classified
pixels with regard to all ground reference validation pixels. In some situations,
such as automated weed management in agriculture, overall accuracy and producer’s accuracy may not be as much of a concern as user’s accuracy because
identifying small amounts of weeds (IAS) as crops may be okay, but spraying
crops misidentified as IAS could be more damaging to crop yields than the IAS
themselves. An example where maximizing producer’s accuracy may be more
important would be in mapping IAS to understand species spread and the invasion process; any omitted species data as changes are monitored over time could
affect process understanding and spread predictions. The last metric, the Kappa
coefficient, can be useful for comparing multiple classification methods within
the same data set. The Kappa coefficient is a measure of how closely the resulting
E. A. Bolch et al.