where the statistically-based D-S method becomes useful. The D-S method utilizes
two fundamental components: hypothesis and proposition. An hypothesis is a
fundamental statement about nature (e.g., a pixel or object is a reef). A proposition
may be either a hypothesis or a combination of hypotheses, which in turn may
contain overlapping or conflicting hypotheses. For example, ‘proposition 1’ = the
seafloor class is sand, ‘proposition 2’ = the seafloor class is sand or reef, and
‘proposition 3’ = the seafloor type is reef. When these propositions are input to the
D-S method, along with probabilities for each proposition, output consists of a
single integrated classification image (Fig. 7.10c).
Examining the MLC image from the hyperspectral-derived seafloor reflectance
in Fig. 7.10a, there is a large patch of deeper water that is misclassified as ‘hard
bottom type 3’. Similarly, there are ‘mid sand’ areas in the south misclassified as
‘reef’, and there are ‘hard bottom type 1’ areas in the south misclassified as ‘hard
bottom type 2’. The MLC image from the LiDAR-derived seafloor reflectance in
Fig. 7.10b appears to have more detail and sharper boundaries between classes, but
nonetheless it also has several misclassified areas. For example, most ‘hard bottom
type 1’ areas were misclassified as ‘channel sand’, and some ‘reef areas’ were
misclassified as ‘hard bottom type 3’. Areas of misclassification in Figure 7.10 are
indicated by yellow polygons.
In the fusion classification image in Fig. 7.10c, many areas were re-assigned
different classes based on the propositions used in the D-S method. For example,
the ‘hard bottom type 2’ areas in the hyperspectral MLC image and the ‘channel
sand’ areas in the LiDAR MLC image were both correctly re-classified as ‘hard
bottom type 1’. The large area that was classified as ‘hard bottom type 3’ in the
hyperspectral MLC image was re-classified as ‘deep sand’, ‘reef’, or ‘hard bottom
type 2.’ However, the ‘shallow sand’ areas in the hyperspectral MLC image and
some ‘mid sand’ and ‘channel sand’ areas in the LiDAR MLC image were
incorrectly re-classified as ‘reef’ (as again indicated by the yellow polygons),
suggesting there is still room for improvement. Nevertheless, the D-S method
produces better overall results compared with either of the MLC images alone.
7.4 Summary and Discussion
This chapter introduced the concept of data fusion and presented a model developed specifically for integrating LiDAR data and hyperspectral imagery to
improve benthic classification. Examples demonstrated the extraction of seafloor
reflectance and water column attenuation from LiDAR waveforms, and then how
this information can be used to improve the processing of hyperspectral data to
spectral seafloor reflectance using modified depth-correction and model inversion
techniques. An approach for combining LiDAR-derived information with spectral
data in a decision tree classifier was also presented, along with a higher-level data
fusion technique known as the Dempster-Shafer method for combining independent classification images into an integrated product.
188
J. M. Wozencraft and J. Y. Park
two fundamental components: hypothesis and proposition. An hypothesis is a
fundamental statement about nature (e.g., a pixel or object is a reef). A proposition
may be either a hypothesis or a combination of hypotheses, which in turn may
contain overlapping or conflicting hypotheses. For example, ‘proposition 1’ = the
seafloor class is sand, ‘proposition 2’ = the seafloor class is sand or reef, and
‘proposition 3’ = the seafloor type is reef. When these propositions are input to the
D-S method, along with probabilities for each proposition, output consists of a
single integrated classification image (Fig. 7.10c).
Examining the MLC image from the hyperspectral-derived seafloor reflectance
in Fig. 7.10a, there is a large patch of deeper water that is misclassified as ‘hard
bottom type 3’. Similarly, there are ‘mid sand’ areas in the south misclassified as
‘reef’, and there are ‘hard bottom type 1’ areas in the south misclassified as ‘hard
bottom type 2’. The MLC image from the LiDAR-derived seafloor reflectance in
Fig. 7.10b appears to have more detail and sharper boundaries between classes, but
nonetheless it also has several misclassified areas. For example, most ‘hard bottom
type 1’ areas were misclassified as ‘channel sand’, and some ‘reef areas’ were
misclassified as ‘hard bottom type 3’. Areas of misclassification in Figure 7.10 are
indicated by yellow polygons.
In the fusion classification image in Fig. 7.10c, many areas were re-assigned
different classes based on the propositions used in the D-S method. For example,
the ‘hard bottom type 2’ areas in the hyperspectral MLC image and the ‘channel
sand’ areas in the LiDAR MLC image were both correctly re-classified as ‘hard
bottom type 1’. The large area that was classified as ‘hard bottom type 3’ in the
hyperspectral MLC image was re-classified as ‘deep sand’, ‘reef’, or ‘hard bottom
type 2.’ However, the ‘shallow sand’ areas in the hyperspectral MLC image and
some ‘mid sand’ and ‘channel sand’ areas in the LiDAR MLC image were
incorrectly re-classified as ‘reef’ (as again indicated by the yellow polygons),
suggesting there is still room for improvement. Nevertheless, the D-S method
produces better overall results compared with either of the MLC images alone.
7.4 Summary and Discussion
This chapter introduced the concept of data fusion and presented a model developed specifically for integrating LiDAR data and hyperspectral imagery to
improve benthic classification. Examples demonstrated the extraction of seafloor
reflectance and water column attenuation from LiDAR waveforms, and then how
this information can be used to improve the processing of hyperspectral data to
spectral seafloor reflectance using modified depth-correction and model inversion
techniques. An approach for combining LiDAR-derived information with spectral
data in a decision tree classifier was also presented, along with a higher-level data
fusion technique known as the Dempster-Shafer method for combining independent classification images into an integrated product.
188
J. M. Wozencraft and J. Y. Park