high r
2 values sometimes are reported, the relationships “work” only because the
“non-optical” variable is correlated in the water bodies used to develop the relationship with an optical variable that affected reflectance. For example, the bacterial
indicator of fecal contamination, E. coli, is found in waters contaminated by human
activities, and correlation of E. coli abundance with TSS and turbidity might be
expected. Similarly, phosphorus often is the limiting nutrient for algal growth in
lakes, and correlations between chlorophyll and TP or SD and TP thus are common
(e.g., [12]). Such relationships cannot be applied reliably beyond the database from
which they were derived because no intrinsic or causative relationship exists
between the non-optical and optical variables or between the non-optical variable
and reflectance.
The use of empirical relationships that depend on secondary correlations has led
to criticisms that remote sensing scientists are “overselling” their technology (e.g.,
[8, 26]). A more transparent and defendable approach is to develop relationships
between reflectance data and variables that directly affect reflectance and then
separately determine whether a sufficiently close relationship exists between the
optical variable retrieved from imagery and a non-optical variable of interest.
Applications of such empirical relationships still should be limited to the data
sets on which they are based, but situations exist in which useful information can
be obtained by this approach. For example, evaluation of a suite of environmental
conditions retrieved from satellite imagery was found useful in predicting outbreaks
of waterborne diseases even though the disease-causing microorganisms do not
directly affect satellite imagery signals [27]. Similarly, atmospheric scientists have
estimated transport of specific pollutants like mercury (Hg) from Asia across the
Pacific Ocean to North America by tracking atmospheric dust using satellite
imagery and independent measurements of the pollutant (e.g., Hg) concentrations
in atmospheric dust over the Pacific.
Relationships between DOC (dissolved organic carbon) and reflectance also
have been reported (e.g., [28]), but insofar as DOC per se is not an optical variable
and does not itself affect reflectance, these also are the results of indirect correlations. To the extent that such relationships work, they rely on the fact that a
fraction of DOC (CDOM) affects reflectance. For some waters good correlations
exist between CDOM and DOC, but as Brezonik et al. [29] recently showed, no
single DOC-CDOM relationship applies across a broad spectrum of surface waters.
Some sources of DOC, e.g., autochthonous organic matter and anthropogenic
organic matter derived from wastewater, have low color per unit of carbon. As a
result, CDOM and DOC are poorly correlated in many natural waters. For example,
Spencer et al. [30] found that r
2 values for DOC-CDOM relationships were 0.5 in
11 of 30 large North American rivers, and four rivers (the Colorado, Columbia, Rio
Grande, and St. Lawrence) had r
2
< 0.2. Factors giving rise to poor DOC-CDOM
relationships include the extent to which the DOC is autochthonous or anthropogenic and the extent to which allochthonous DOC has been photodegraded.
118
L.G. Olmanson et al.
2 values sometimes are reported, the relationships “work” only because the
“non-optical” variable is correlated in the water bodies used to develop the relationship with an optical variable that affected reflectance. For example, the bacterial
indicator of fecal contamination, E. coli, is found in waters contaminated by human
activities, and correlation of E. coli abundance with TSS and turbidity might be
expected. Similarly, phosphorus often is the limiting nutrient for algal growth in
lakes, and correlations between chlorophyll and TP or SD and TP thus are common
(e.g., [12]). Such relationships cannot be applied reliably beyond the database from
which they were derived because no intrinsic or causative relationship exists
between the non-optical and optical variables or between the non-optical variable
and reflectance.
The use of empirical relationships that depend on secondary correlations has led
to criticisms that remote sensing scientists are “overselling” their technology (e.g.,
[8, 26]). A more transparent and defendable approach is to develop relationships
between reflectance data and variables that directly affect reflectance and then
separately determine whether a sufficiently close relationship exists between the
optical variable retrieved from imagery and a non-optical variable of interest.
Applications of such empirical relationships still should be limited to the data
sets on which they are based, but situations exist in which useful information can
be obtained by this approach. For example, evaluation of a suite of environmental
conditions retrieved from satellite imagery was found useful in predicting outbreaks
of waterborne diseases even though the disease-causing microorganisms do not
directly affect satellite imagery signals [27]. Similarly, atmospheric scientists have
estimated transport of specific pollutants like mercury (Hg) from Asia across the
Pacific Ocean to North America by tracking atmospheric dust using satellite
imagery and independent measurements of the pollutant (e.g., Hg) concentrations
in atmospheric dust over the Pacific.
Relationships between DOC (dissolved organic carbon) and reflectance also
have been reported (e.g., [28]), but insofar as DOC per se is not an optical variable
and does not itself affect reflectance, these also are the results of indirect correlations. To the extent that such relationships work, they rely on the fact that a
fraction of DOC (CDOM) affects reflectance. For some waters good correlations
exist between CDOM and DOC, but as Brezonik et al. [29] recently showed, no
single DOC-CDOM relationship applies across a broad spectrum of surface waters.
Some sources of DOC, e.g., autochthonous organic matter and anthropogenic
organic matter derived from wastewater, have low color per unit of carbon. As a
result, CDOM and DOC are poorly correlated in many natural waters. For example,
Spencer et al. [30] found that r
2 values for DOC-CDOM relationships were 0.5 in
11 of 30 large North American rivers, and four rivers (the Colorado, Columbia, Rio
Grande, and St. Lawrence) had r
2
< 0.2. Factors giving rise to poor DOC-CDOM
relationships include the extent to which the DOC is autochthonous or anthropogenic and the extent to which allochthonous DOC has been photodegraded.
118
L.G. Olmanson et al.