139
metrics into one summary metric (Pontius and Hallett 2014). This may be easier
and more efficient than creating models to assess each of the various decline
symptoms you expect to see in your target system or having to pick one decline
metric to use.
3. Calibrate imagery with field data. In an ideal world, we would be able to develop
one model that could be automated and applied to imagery over time and space
regardless of sensor, acquisition condition, or location. Several automated RS
tools currently available (see Sect. 6) have proven incredibly useful for monitoring large areas over time. But automated applications are limited in their ability
to detect subtle, early decline, which requires careful calibration between the
imagery acquired and ground conditions at the time that imagery was collected
to make it possible to identify the targeted stress response while controlling for
other sources of spectral variability. Ideally, field calibration data can be collected within several weeks of imagery acquisition (or at least before conditions
on the ground change). GPS locations of field calibration sites link field data to
the spectra of the associated pixel or pixels to calibrate the larger image.
Various proprietary software modules exist for spectral calibration, modeling,
and analyses. These modules can range from simple classification techniques
that match pixels to various stages of decline based on your calibration spectra,
to more complex spectral unmixing algorithms that approximate the proportion
of “healthy to declining” spectra contained within each pixel. Even without specialized RS software, simple statistics can be used to quantify relationships
between spectral reflectance and derived vegetation indices using field calibration data. A common approach is to use correlations between individual vegetation indices and decline metrics to qualitatively assess canopy condition across
the landscape. Another approach uses multivariate statistical models to identify
the best combination of bands or vegetation indices to quantify the decline metric of interest. Regardless of the mathematical approach, accuracy and detail are
ultimately determined by the quality and range field calibration data available for
model development. This type of targeted calibration to match the timing, location, and sensor characteristics for each decline assessment maximizes accuracy
and detail of the final products.
4. Validate and assess accuracy to inform interpretation. One of the dangers inherent in linking RS products with management applications is overconfidence in
the RS products. There is error inherent in each component of the RS process,
from incorrect sensor calibration, to the variability introduced by atmospheric,
topographic, and georegistration errors. However, when presented with a RS
product, many end users develop their plans without consideration of how accurate the product may be or how inaccuracies can be avoided.
Any RS product should include some measure of accuracy as well as any
caveats that should be considered in its use. In some cases (e.g., the use of a
vegetation index to qualitatively describe relative states of decline), it is sufficient to remind users that the scale presented is intended to be relative and does
not necessarily identify stands in specific states of decline or resulting from
specific stress agents. In other cases (e.g., the classification of pixels into levels
6 Remote Sensing for Early, Detailed, and Accurate Detection of Forest Disturbance…
metrics into one summary metric (Pontius and Hallett 2014). This may be easier
and more efficient than creating models to assess each of the various decline
symptoms you expect to see in your target system or having to pick one decline
metric to use.
3. Calibrate imagery with field data. In an ideal world, we would be able to develop
one model that could be automated and applied to imagery over time and space
regardless of sensor, acquisition condition, or location. Several automated RS
tools currently available (see Sect. 6) have proven incredibly useful for monitoring large areas over time. But automated applications are limited in their ability
to detect subtle, early decline, which requires careful calibration between the
imagery acquired and ground conditions at the time that imagery was collected
to make it possible to identify the targeted stress response while controlling for
other sources of spectral variability. Ideally, field calibration data can be collected within several weeks of imagery acquisition (or at least before conditions
on the ground change). GPS locations of field calibration sites link field data to
the spectra of the associated pixel or pixels to calibrate the larger image.
Various proprietary software modules exist for spectral calibration, modeling,
and analyses. These modules can range from simple classification techniques
that match pixels to various stages of decline based on your calibration spectra,
to more complex spectral unmixing algorithms that approximate the proportion
of “healthy to declining” spectra contained within each pixel. Even without specialized RS software, simple statistics can be used to quantify relationships
between spectral reflectance and derived vegetation indices using field calibration data. A common approach is to use correlations between individual vegetation indices and decline metrics to qualitatively assess canopy condition across
the landscape. Another approach uses multivariate statistical models to identify
the best combination of bands or vegetation indices to quantify the decline metric of interest. Regardless of the mathematical approach, accuracy and detail are
ultimately determined by the quality and range field calibration data available for
model development. This type of targeted calibration to match the timing, location, and sensor characteristics for each decline assessment maximizes accuracy
and detail of the final products.
4. Validate and assess accuracy to inform interpretation. One of the dangers inherent in linking RS products with management applications is overconfidence in
the RS products. There is error inherent in each component of the RS process,
from incorrect sensor calibration, to the variability introduced by atmospheric,
topographic, and georegistration errors. However, when presented with a RS
product, many end users develop their plans without consideration of how accurate the product may be or how inaccuracies can be avoided.
Any RS product should include some measure of accuracy as well as any
caveats that should be considered in its use. In some cases (e.g., the use of a
vegetation index to qualitatively describe relative states of decline), it is sufficient to remind users that the scale presented is intended to be relative and does
not necessarily identify stands in specific states of decline or resulting from
specific stress agents. In other cases (e.g., the classification of pixels into levels
6 Remote Sensing for Early, Detailed, and Accurate Detection of Forest Disturbance…