properly executing field spectroradiometric work in remote sensing lies in identifying
pitfalls and problems and selecting appropriate solutions in advance (McCoy, 2004).
The variability of the ground measurements factors, together with the natural
variability in land cover such as vegetation, constrain the results from hyperspectral
image vegetation analysis studies. Support of ground-collected spectral libraries are to
be interpreted only relative to their specific collection phenomenology/methodology
and study area (Salvaggio et al., 2005). That makes the linkage and comparison of
in situ data with other remote sensing data complex. If in situ spectra measurements
are significantly linked with the image end members, only then should the field data
serve to validate the identity of the vegetation type mapped.
15.4 FINAL REMARKS
The rapid development of remote sensing sensor technology makes it increasingly
and steadily more and more feasible to derive land cover information from hyperspectral Earth observation data acquired at different observational scales (Ben-Dor
et al., 2009; Herold, 2009; Underwood et al., 2007).
Although a novel technology, field spectroradiometry land cover study, especially
for vegetation, has led to remarkable achievements and has proven the significant
capability in land cover analysis domain over the past 15 years. Both multi- and
univariate statistics proved their greatest ability to discriminate vegetation land cover
reflectance, as determined by the reflectance analysis of variance, in bands of the VIS
and especially in the NIR spectrum, including the red edge between 680 and 730 nm.
Single bands yielded by field scale dimension–reduction studies for discrimination
between targets of a certain land cover such as vegetation species are reported within
the NIR, but findings remain to be confirmed on satellite imagery in future
investigations. In addition, experiences gained by vegetation discriminant analyses
at the field scale so far indicate an increasing importance of the SWIR spectrum,
considering its relative ability to characterize water content in vegetation cover.
Many studies have been performed in highly heterogeneous or vegetation fragmented regions, aiming to derive information on land use/cover distribution. It is thus
expected land cover to cluster on satellite image to satisfactory classification accuracy
values if the imagery scale is smaller than or equal to the FoV used on the ground. The
full potential of field spectroradiometry is yet to come.
Indeed, various studies conducted in dissimilar ecosystems and implementation
settings have already demonstrated that very accurate maps of land cover can be
derived from hyperspectral systems mounted in both airborne and satellite platforms.
Different methods combined with such data via the computation of standardized
approaches based on error matrix statistics have generally reported, high overall
accuracies and kappa coefficients, reaching or exceeding in some cases 85% and
0.800. These are generally regarded as very satisfactory and accurate estimates for
many practical applications on which regional maps of land cover might be required.
However, field scale remote sensing such as field spectroradiometry to hyperspectral imaging remote sensing analysis of land cover has not yet been fully
FINAL REMARKS
311
pitfalls and problems and selecting appropriate solutions in advance (McCoy, 2004).
The variability of the ground measurements factors, together with the natural
variability in land cover such as vegetation, constrain the results from hyperspectral
image vegetation analysis studies. Support of ground-collected spectral libraries are to
be interpreted only relative to their specific collection phenomenology/methodology
and study area (Salvaggio et al., 2005). That makes the linkage and comparison of
in situ data with other remote sensing data complex. If in situ spectra measurements
are significantly linked with the image end members, only then should the field data
serve to validate the identity of the vegetation type mapped.
15.4 FINAL REMARKS
The rapid development of remote sensing sensor technology makes it increasingly
and steadily more and more feasible to derive land cover information from hyperspectral Earth observation data acquired at different observational scales (Ben-Dor
et al., 2009; Herold, 2009; Underwood et al., 2007).
Although a novel technology, field spectroradiometry land cover study, especially
for vegetation, has led to remarkable achievements and has proven the significant
capability in land cover analysis domain over the past 15 years. Both multi- and
univariate statistics proved their greatest ability to discriminate vegetation land cover
reflectance, as determined by the reflectance analysis of variance, in bands of the VIS
and especially in the NIR spectrum, including the red edge between 680 and 730 nm.
Single bands yielded by field scale dimension–reduction studies for discrimination
between targets of a certain land cover such as vegetation species are reported within
the NIR, but findings remain to be confirmed on satellite imagery in future
investigations. In addition, experiences gained by vegetation discriminant analyses
at the field scale so far indicate an increasing importance of the SWIR spectrum,
considering its relative ability to characterize water content in vegetation cover.
Many studies have been performed in highly heterogeneous or vegetation fragmented regions, aiming to derive information on land use/cover distribution. It is thus
expected land cover to cluster on satellite image to satisfactory classification accuracy
values if the imagery scale is smaller than or equal to the FoV used on the ground. The
full potential of field spectroradiometry is yet to come.
Indeed, various studies conducted in dissimilar ecosystems and implementation
settings have already demonstrated that very accurate maps of land cover can be
derived from hyperspectral systems mounted in both airborne and satellite platforms.
Different methods combined with such data via the computation of standardized
approaches based on error matrix statistics have generally reported, high overall
accuracies and kappa coefficients, reaching or exceeding in some cases 85% and
0.800. These are generally regarded as very satisfactory and accurate estimates for
many practical applications on which regional maps of land cover might be required.
However, field scale remote sensing such as field spectroradiometry to hyperspectral imaging remote sensing analysis of land cover has not yet been fully
FINAL REMARKS
311