107
information may be lost (Chen and Ho 2008) and these approaches were therefore
not considered further. Instead, five different scenarios where used:
1. Image bands and DTM layers such as slope and aspect only (known as Images in
Fig. 6);
2. Indices chosen on their recommendation from the EODHaM system and literature (known as EODHaM in Fig. 6);
3. 20 indices with the highest F scores from the separation analysis (ANOVA) per
class (known as ANOVA in Fig. 6);
4. Indices that have a low correlation value for each class (i.e., every observation
that has a correlation of 0.75 or above is discarded from further analysis (known
as Uncorrelated in Fig. 6);
5. 20 indices with the highest calculated importance for each class, as delineated
from the Random Forest (RF) algorithm (known as RF in Fig. 6).
Classification Structure and Validation
To preserve the hierarchical nature of the EODHaM system, a similar structure was
formulated where classes were determined by outputs from algorithms instead of
manually inputted thresholds (Fig. 8). This structure would still allow other datasets
Fig. 7 Schematic of the classifying process integrating machine learning algorithms
Mapping Coastal Habitats in Wales