3.1. Increasing Spectral Bands Involved a Better Discrimination of Coral-Dominated Assemblages and Health State
Irrespective of their health state, the three coral-dominated assemblages were correctly classified (CA > 80%) by the three spectral bands (
QB2) for three out of four classifiers (
Figure 5(A)). Although the RF showed a very moderate performance (CA = 51%), results inherent to the three other models all exceeded 80%. When developed from the five spectral bands (
WV2), the CA consistently provided higher values, ranging from 82% to 96%, stemming from the RF and the SVM algorithms, respectively. These results demonstrated the interspecific spectral variability of corals that could be quantified even with only red, green and blue predictors. However, the addition of two spectral bands,
i.e., “coastal” and yellow, allowed for an increase in the power of discrimination of the three coral-dominated assemblages.
Comparing both spectral datasets’ performances revealed that the
WV2 dataset systematically outperformed the
QB2 dataset against the four classifiers for distinguishing health states of the three coral-dominated assemblages (
Figure 5(B)). Three out of the four algorithms (
i.e., RF, kNN and SVM) showed that the value added of two bands was significant in terms of classification performance (
pZ-test < 0.01,). As for the NB algorithm, albeit superior, no substantial contribution of the two added bands was detected (
pZ-test = 0.018). While RF, NB, SVM and kNN gradually improved the
QB2-derived classification, the RF, NB, kNN and SVM gradually refined the
WV2-derived coral separation. Developing a model trained by three spectral bands for coral classification provided satisfactory results when using the kNN classifier (CA = 84%). On the other hand, analyzing coral spectra through five visible spectral bands rated the kNN and SVM to provide the best discernments (CA = 91%, and CA = 93%, respectively). This latter classifier incidentally displayed the highest difference between
WV2 and
QB2.
Classifying the health states of corals (i.e., six classes) resulted in a loss of the classification performance compared to classifying only the coral-dominated assemblages (i.e., three classes). Using the QB2 dataset, a reduction in CA, albeit slight (topped at 4%), appeared for the four classifiers. When employing the WV2 dataset, the decline in CA was slight for SVM, kNN and NB (2%, 4% and 6%, respectively) but exacerbated for RF (i.e., 18%). Irrespective of the number of spectral bands included in the classification scheme, scaling up the ecological accuracy in partitioning the coral-dominated assemblages as a function of their health states involved a decrease of the distinction, corroborating theory that underlies image classification.
Both the coral-dominated assemblages and their health states (healthy or unhealthy) were successfully predicted from multispectral remote sensing, even though the optical properties of the water column were assumed to remain homogeneous across the two study areas. Quadrat sampling was deliberately collected at different water depths, ranging from approximately 0.5 to 3 m, so that pattern recognition of coral-dominated assemblages and health state was not skewed by spectral changes induced by variation of depths. However, comparing spectra retrieved from sharply contrasted water depths (e.g., 0.5
vs. 20 m) and clarities (varying on short a distance in reef slope) could raise substantial issues when considering the diffuse attenuation coefficient,
Kd, as constant over the lagoon. Expressed in the form of an exponential (
Equation (1)), constant
Kd could lead to strong discrepancies between water-corrected and actual reflectance in over- or under- estimating attenuation of the light in the water column (corresponding to a lower or higher
in situ Kd, respectively). The problem of
Kd assessment could be addressed, at least partially, by means of quantifying water clarity and pigment concentration during the field survey [
38]. As it would be fragmented and time-consuming, such a survey could not bring enough information to be included in reef health monitoring procedure, demanding a high time/area ratio. Remote sensing of coral-dominated assemblages and health state will attain expectations of multi-scale managers and decision-makers as soon as the
Kd can be resolved directly from the spaceborne imagery, as it can be done from the hyperspectral airborne imagery [
39]. Recent applications of the spaceborne sensor QB2 [
12] are very promising to retrieve the coefficient for each pixel and will be applied to fine-tune the WV2 radiative transfer model.
Adding the “coastal” and yellow bands to spectral variables underpinned that the WV2, as a very high spatial and enhanced spectral resolution spaceborne sensor, has the spectral capability to discriminate the ma or reef builder’s types at the colony scale (
Figure 6 and [
40]). Coral colors are recognized to be closely linked with absorption related to zooxanthellae (dinoflagellate endosymbiont of scleractinian corals). Examining the absorption spectrum of “brown” corals modeled by concentration-weighted pigments included in endosymbiotic zooxanthellae [
41] pointed out that, while (1) conventional blue, green and red bands adequately recorded the last (
i.e., third) peak of chlorophyll-a, the first peak, the start of the second peak and the end of the third peak of chlorophyll-c, the peak of the peridinin, the second and the third peaks of diadinoxanthin, as well as the second and the third peaks of β-caroten, (2) the “coastal” and the yellow bands allowed the first and the second peaks of chlorophyll-a, the second and third peak of chlorophyll-c, the first peak of diadinoxanthin, and the first peak of β-caroten to be detected. The spectral continuity conveyed by the adjacency of “coastal” and blue bands ensured the coverage of the first peak of chlorophyll-c, the spectrum of diadinoxanthin and of β-caroten. The association of the “coastal”, blue and green bands offered the complete detection of peridinin. The addition of the yellow and red bands enabled the sensing of both chlorophylls (
i.e., their last two peaks). The specific contribution of the bands or combinations of bands to presence/absence coral-associated pigments will be further discussed in the light of spectral indices in the next section. Increasing the spectral resolution of a very high spatial resolution sensor encouraged the thought that this new tool would favorably enhance classification of coral reefs’ features, spanning other reef builders (e.g.,
Pocillopora sp.,
Montipora sp.), algae (macroalgae and coralline red algae), as well as sediment classes (mud, sand and reef flat) [
20].
3.2. Spectral Diversity Indices Enhanced Classification of Coral-Dominated Assemblages and Health State
Beyond the efficient distinction of coral-dominated assemblages and health state retrieved from the original
WV2 dataset, classifications modeled by the diversity index-boosted
WV2 datasets produced high CA for all classifiers except RF (
Figure 7). The “dart board-like”
Figure 7 synoptically plotted 52 classification performances resulted from diversity index-enhanced
WV2 combinations as a function of the four classifiers while remaining readable. At first sight, a hierarchy between classifiers, consistent across spectral combinations, could be outlined. Showing the lowest CA in the entire analysis (
i.e., 49%), the RF embodied the worst classifier, followed by the NB, whose lowest value reached 75%, followed by the kNN and SVM, whose lowest values were close to 90%. According to this classifier’s sorting (from RF to SVM), the amplitude between the highest and the lowest CA for each classifier tended to diminish (namely, 28, 9, 4 and 5%, respectively), as visually represented by the circle linearity (
Figure 7). Through distributions of the four circles against the CA gradient, two clusters emerged on either side of a numerical gap: The first group composed of the SVM and kNN was characterized by high and stable results, the second group encompassing the NB and the RF produced moderate and variable accuracies. CA appeared close enough to those issued from the
WV2 dataset so that in-depth analyses regarding changes were conducted.
The contributions of both series of diversity indices brought distinctness among the coral classes to be refined despite the counter-intuitive negative contributions somewhat disseminated across the spectral combinations. While the best contribution for the RF occurred with E12 and P12 (gain of 13%), contributions of E24 and P24 substantially lowered the classification (loss of 15%). As for the NB, kNN and SVM, contributions of both best and worst diversity indices, albeit significant, were further reduced (gains of 4, 2 and 3%; losses of 5, 2 and 2%, respectively). While the best contributions for the NB, kNN and SVM stemmed from E34 and P34; E35, E2345, P345 and P1345; E35, E45 and E123, respectively, the worst contributions resulted from E45 and P45; E15; E234, respectively. Given the multiple similar contributions, we chose to further analyze the highest gain (i.e., RF) and the highest CA (i.e., SVM).
Changes in CA issued from the contribution of the best spectral indices were analyzed using confusion matrices (CM), allowing for the change detection at the class level. Despite the lowest CA among the four classifiers, the RF showed the highest gain using either the
E12 or the
P12. We could deduce that the knowledge of both “coastal” and blue reflectance could substantially enhance the discrimination among coral-dominated assemblages and health states. Confronting the CM derived from the
WV2 and the
E12-enhanced
WV2 (CM of
P12 was identical) datasets led to emphasize gains of both healthy
P. rus (55.8%) and unhealthy
P. rus (4.3%), as well as a slight loss of
A. pulchra (2.2%) (
Figure 8(A)). Although the diversity index-related gain was not huge, the SVM produced the best CA using the
E35, the
E45 and the
E123. Retrieving the changes from the CM, we could infer that the index based upon (1) the green and the red reflectance refined distinction among healthy
P. lobata (10%), unhealthy
P. lobata (11.1%) and healthy
P. rus (7%), (2) the yellow and the red reflectance augmented the CA of healthy
P. lobata (10%), unhealthy
P. lobata (11.1%), unhealthy
P. rus (4.3%) at the expense of the healthy
P. rus (2.3%), (3) the “coastal”, the blue and the green bands increased the CA of healthy
A. pulchra (2.2%), healthy
P. lobata (10%), healthy
P. rus (4.6%) at the expense of unhealthy
P. lobata (11.1%) and unhealthy
P. rus (4.4%) (
Figure 8(B)). CA of
P. lobata derived from the SVM
WV2 CM approximated those computed from an airborne 1 m spectrographic imager [
42]. However, when boosted by
E35,
E45,
E123, the CA equaled 100%, surpassing the airborne survey by around 10%. Compared to CA dedicated to remote sensing of (1) acroporid assemblages by a spaceborne multispectral 4 m sensor [
43] and (2) an Acropora dataset by an
in situ hyperspectral spectrometer [
16], a gain of 10% occurred with and without
E contribution. This finding proved the greater power of classification of the SVM algorithm
vis à vis traditional classifiers used in those studies. We furthermore mapped the
E123 index compounded with the
E123-enhanced SVM classification (
Figure 8(B)) of the first subregion (
Figure 9) to demonstrate that our proof-of-concept was able to furnish meaningful information intelligible to marine spatial planners and lagoon practitioners.
The two series of spectral indices,
i.e.,
E and
P, differentially modified CA in respect to the classifiers. While the contribution’s differences between
E and
P negated 16 times (out of 26 possible) for the RF classifier, only five cases were reported for the SVM. Conversely, the maximum contribution difference topped at 0.21 and 0.02 for RF and SVM, respectively. The
P index, computed on the basis of Shannon-Weaver entropy [
29], weighed each band according to the attendant fraction of the whole measured light, that is to say, the frequency. On the other hand, the
E index, based upon the Simpson index [
29], attached much importance to the most dominant band since it involved the sum of the squares of the frequencies (note that the square of a small frequency was a very small number). We could deduce that (1) the
E index tended to be lower than the
P index when a band dominated, and (2) the greater the dominance in the spectrum, the greater the difference between these two indices. Across the four classifiers, the four worst and least bad contributions were found for the
P series, while the
E series contained the seven worst and four least bad contributions. In the light of previous theoretical findings, the prevalence of
E in terms of highest absolute values of contribution indicated that the most significant spectral indices were more sensitive to spectral dominance. The corollary of this stronger sensitivity indicated that infrequent bands hardly contributed to the
E indices. One way around this dilemma would be to use Hill’s diversity index, merging both previous diversity indices, allowing for a sharper view of the spectral diversity. While providing the measurement of the dominant bands, it would have the ability to incorporate the infrequent bands.
Comparing the spectral indices involved in the best contributions with the spectra of pigments including coral tissues highlighted spectral windows as valuable to discern between healthy and unhealthy corals (
Table 3). The diversity index
E12 provided the highest contribution to the RF in specifically increasing recognition of both health states of
P. rus. This result suitably matched a spectroscopic study showing that the spectral region comprised between 430 and 490 nm displayed the highest differentiation among various health states of corals [
16]. The absorption of the first peaks of chlorophyll-a and -c, diadinoxanthin and β-caroten were encompassed by the oint “coastal” and blue bands, ranging from 400 to 510 nm. We could therefore hypothesize that the four latter pigments were found to be relevant predictors of
P. rus health state. Assuming that the absorption magnitude of these pigments responded to the zooxanthella density and that the zooxanthella density differed remarkably among various
P.rus health states, the intensity of the remotely sensed reflectance between 400 and 510 nm would accurately measure the zooxanthella density, thus reflect the gradient of the
P. rus health state. Beyond the binary classification of the health state (healthy and unhealthy), intermediate states would be inferred, offering a time flexibility, favorable to managing coral reefs. The diversity index
E123 improved the SVM classification in gradually increasing healthy
A. pulchra,
P. rus and
P. lobata at the expense of unhealthy
P. rus and unhealthy
P. lobata. Adding the green band to the previous diversity index pushed detection boundaries to 580 nm, thus integrating the peridinin into pigments detection (the four latter ones). While adding the peridinin detection led to slight misclassifications towards unhealthy corals, it augmented the detection of the healthy state of the three coral-dominated assemblages. These findings concurred with those of Clark
et al. [
44], showing a good discrimination of healthy and unhealthy coral in the 515–596 nm range. The slight misclassifications akin to unhealthy
P. rus and unhealthy
P. lobata might be explained by the presence of turf algae on colonies (e.g.,
Figure 3(B)). The within-pixel mixing of living and bleached coral tissue with algae cover is very susceptible to alter the spectral signature of a pure dead or bleached coral (deprived of pigment-related absorption
per se) by diminishing the reflectance due to the chlorophyll-related absorption. The diversity index
E35, involving the green and red bands, enabled the two health states of
P. lobata and the healthy
P. rus to be clearly separated. The monitoring of the second half of the peridinin spectrum and the last peak (the third) of chlorophyll-a could be deemed as meaningful for assessing
P. lobata health state. The diversity index
E45, bringing into play yellow and red bands, led to a better distinction among the two health states of
P. lobata and the unhealthy
P. rus. The adjacency of the two bands ensured the two last peaks (the second and third) of chlorophyll-a and -c to be fully covered. The knowledge of the red band, associated with either the green or yellow band, therefore suggested an enhancement in distinctness of
P. lobata state health. Whereas the red band unequivocally fitted with last peak of chlorophyll-a, strong attenuation of the inherent radiation by water hindered its use for discriminating coral health state below a few meters depth (>4–5 m). In addition, when combined with green, red highlighted the healthy
P. rus, while it emphasized the unhealthy
P. rus with yellow. The presence of peridinin appeared decisive for the assessment of healthy
P. rus, while the sensing of the two last peaks of cholorophyll-c suggested
P. rus with poor health. This latter conclusion corroborated results derived from
in situ spectroscopy analysis, specifying that spectral signature (reflectance) of worldwide-averaged “brown” corals was clearly characterized by a “triple-peaked pattern” ranging from 560 to 660 nm [
15]. It is not possible to speculate on the spectral partitioning of the diversity index in the present study, but work is in progress to validate or refute previous inferences based upon assemblage of pigment spectra.