*4.1. Similarity Measures*

Considering all transformed spectral signatures, spectral ranges and similarity measures, only the Canberra distance on [350 nm to 2500 nm] gives an overall accuracy higher than 50 % whatever the spectral reference database (Table 7). Indeed, the Canberra distance gives the higher overall accuracy because it is sensitive to a small change when both coordinates are closed to zero [140,141].

Because of the high variability of some vegetation types (Appendix B), spectral reference database built from median spectra, that are *real* spectra, gave worse results than spectral reference database built from median and mean spectra, that are *theoretical* spectra not representative of a in situ measured vegetation type (Table 7). There is a need to collect more spectral signatures to build a consistent spectral database.

As spectral signatures can be considered as high dimensional vectors, a specific distance is needed to compare them. It is well known that Euclidean distance is not good when comparing high dimension data [142]. Table 8 shows that the Canberra distance always outperforms other distances, including SAM, which is commonly used in remote sensing, when considering the whole spectral range (1823 wavelengths).


**Table 7.** Overall accuracy (%) for Canberra distance on [350–2500 nm].

**Table 8.** Overall accuracy (%) for different distances on [350–2500 nm] considering Median reflectances as spectral reference database.


Using the Canberra distance, best results (overall accuracy higher than 60 %) are given with the second derivative, first derivative and CRDR (Table 7), that are closely related to absorption features rather than reflectance magnitude [38]. Indeed, it is possible to discriminate between vegetation types thanks to their biophysical components which will be discussed in details in Section 4.2.1. Furthermore, Table 9 shows that the whole spectral range gives the best results. Although spectral ranges are related to specific biophysical components (Table 6), the whole spectral range is needed to discriminate between the 13 vegetation types because some of them are sharing same plant species (Table A1) and the spectral signatures are mixed. Worse results are obtained in [1940–2400 nm] whatever the transformed spectral signature. Table 9 show that worse results are obtained by the spectral signature whatever the spectral range. Indeed those transformations are related to absorption features as explained above, which confirms that transformed spectral signatures are more suitable to discriminate between vegetation types than spectral signatures.


**Table 9.** Overall classification accuracy (%) for different spectral ranges considering Median reflectances as spectral reference database and Canberra distance.

Considering classification accuracy for each vegetation type, Table 10 shows that best F1-score is obtained by *Sphagnum* sp. (SPHA) (98%), *Juniperus communis* (JUCO) (97%), Aquatic type b (AQ\_B) (93%) and *Salix* sp. (SALI) (92%). Except for JUCO, all of these vegetation types are well classified and their user's accuracy is higher than 85%. Indeed these vegetation types are less mixed than others: Table A1 shows that SPHA is mainly dominated by different kinds of *sphagnum*; AQ\_B is dominated by *Utricularia* sp; JUCO is dominated by *Juniperus communis* and SALI is dominated by *Salix*. Only three other vegetation types have user's accuracy equal to 100%: *Rhododendron ferrugineum* (RHFR), *Calluna vulgaris* (CAVU) and Aquatic type a (AQ\_A). However, only around 57% of spectral signatures are well identified for CAVU and AQ\_A. This can be explained by the high variability of these sample plots. Contrary to SPHA, JUCO, AQ\_B and SALI, there is not a single dominated plant species neither for CAVU nor for AQ\_A (Table A1). Worse F1-score is obtained by *Pinguicula* sp. (PING) (54%) which is not dominated by only one plant species: this vegetation type is mainly dominated by *Eleocharis quinqueflora* (ELQU) (40%), bare ground (15%), *Molinia caerulea* ssp *caerulae* (10%) and *Tomenthypnum nitens* (10%). It can explain the difficulty to identify this vegetation type in particular rather than the low number of spectra: PING has eight spectra whereas AQ\_B has seven spectra.

**Table 10.** Confusion matrix of the classification based on Second derivative, Canberra Distance on [350–2500 nm] with Median reflectance as reference spectral database. The producer's and user's accuracies, the overall accuracy and the F1-score are also shown.


*4.2. Supervised Classification Based on Feature Selection of Spectral Vegetation Indices*

## 4.2.1. Feature Selection

The Kruskal-Wallis method (Section 3.4.2, p. 13) does not show any significant index (frequency discrimination > 75 %) that allow discrimination between vegetation types (Figure 5, only the first 69 indices are drawn). The best vegetation index (NDWI [860, 2130]) only allows us to discriminate between 49 pairs of vegetation types, that may be explained by the plant species mixing within several vegetation types. The proposed method reduced the number of selected indices from 129 to 26 (Table 11). More precisely, on the first step of the method, only 17 single indices amongs<sup>t</sup> 26 are needed to discriminate between 59 pairs of vegetation types amongs<sup>t</sup> 78. On the second step, these single indices must be completed by 7 additional spectral vegetation indices to discriminate between 17 more pairs of vegetation types (Table 12; ∅ means either a pair of vegetation type can not be discriminated thanks to a pair of spectral vegetation indices built from single ones selected on the first step, either more than two vegetation indices are needed to discriminate between a pair of vegetation types). On the last step, a single index is added to discriminate between two vegetation types whereas a combination of previous selected indices allows us to discriminate between another pair of vegetation type (Table 11). Finally several different—single or pair or triplet—vegetation indices allow us to discriminate between pairs of vegetation types. However, none single spectral index allows us to discriminate between all pairs of vegetation types nor the majority: e.g., the most discriminating single spectral index, the Water Index (WI), only discriminates between around 42% pairs of vegetation types (Table 11).

Table 13 shows that one single biophysical component can discriminate between most of the vegetation types except for *Carex* sp. homogeneous vegetation (CA\_HV). More precisely, three kinds of vegetation types (*Sphagnum* sp. (SPHA), Aquatic type b (AQ\_B) and Aquatic type c (AQ\_C)) are separated thanks to a single biophysical component. However, some biophysical components are more discriminant than others according to vegetation types: e.g., the chlorophyll is more discriminant than the water content for AQ\_C whereas the water content is the only discriminant biophysical component for AQ\_B; the water content, the chlorophyll and water, cellulose, starch, lignin (w., c., s., l.) equally discriminate between SPHA and all other vegetation types.

Only two indices related to water content are needed to separate AQ\_B from all other vegetation types: WI and NDWI[860,1240] (Table 13) because AQ\_B vegetation type is mainly composed of *Utricularia* sp. and water (Table A1). The AQ\_B spectral signatures are lower than the spectral reflectance values of the other vegetation types and the water absorption band at 900 nm and 970 nm are highlighted (Figure 6).

**Figure 5.** Frequency distribution of the Kruskal-Wallis test for the 129 spectral indices for paired species across the 13 vegetation types. The horizontal red line stands for 75 % of all 78 possible combinations of the 13 vegetation types.


#### **Table 11.** Single selected indices from the Hellinger distance and their occurrences.

\* index selected on first step. \*\* index selected on second step. \*\*\* index selected on third step.

**Table 12.** Single spectral index or pairs of spectral indices retained to discriminate between vegetation types pairs.


∅ means either it is not possible to discriminate between a pair of vegetation typess thanks to a pair of spectral vegetation indices built from single ones selected on the first step, or more than two vegetation indices are needed to discriminate between a pair of vegetation types.

**Table 13.** Single main discriminating biophysical components for each vegetation type and their occurrences (%).


The chlorophyll is the main biophysical component (86.33 %) able to discriminate between AQ\_C and all other vegetation types, except with Aquatic type a (AQ\_A) and AQ\_B differentiated by considering additional water indices (MSI and NDWI [860,1240]). Indeed, dry matter can be seen on spectral signatures (Figure 7): AQ\_B has the lowest slope on the spectral range [705–730 nm] whereas other vegetation types (except AQ\_A and AQ\_B) have higher values because they still contain chlorophyll. However, as AQ\_B and AQ\_C have low values of Boochs2 index, it is possible to discriminate between them thanks to a water index (right side of Figure 8 shows that those vegetation types can be clearly separated; indeed, those vegetation types have different shapes and values that characterize each type).

**Figure 6.** Mean spectral reflectance of the 13 vegetation types. Dashed lines represent the wavelengths used by Water Index (WI).

**Figure 7.** Mean first derivative spectral signatures of the 13 vegetation types on [695–730 nm]. The green dashed line represents the wavelength used by the Boochs2 index.

In some case, there is no single biophysical component allowing us to discriminate between vegetation types: e.g., both water content (33.33%), chlorophyll (33.33%) and w., c., s., l. (33.33%) are needed to distinguish SPHA from all other vegetation types (Table 13). More precisely, biophysical components related to water (WI, MSI) are discriminating SPHA from CA\_HV, *Pinguicula* sp. (PING), *Pinguicula* sp. combined vegetation (PI\_CV) and AQ\_B ; biophysical components related to chlorophyll (CCCI, OSAVI [800,670]) are differentiating SPHA from AQ\_A, AQ\_C, *Eleocharis quinqueflora* (ELQU) and *Menyanthes trifoliata* (METR) ; biophysical components related to w., c., s., l. (F\_WP) are separating SPHA from *Calluna vulgaris* (CAVU), *Rhododendron ferrugineum* (RHFR), *Salix* sp. (SALI) and *Juniperus communis* (JUCO) (Table 13). Unlike an index related to water content (Figure 9), an index related to the chlorophyll will discriminate between SPHA and AQ\_A. Indeed, the right side of Figure 9 shows that some AQ\_A plant species can not be distinguished from SPHA because it is a dry moss and the left side of Figure 9 shows that SPHA and non discerned AQ\_A have the same spectral signature shape. The right side of Figure 10 shows that these two vegetation species can clearly be separated despite the class variability of AQ\_A. A complex biophysical component such as F\_WP will differentiate SPHA from CAVU (left side of Figure 11) shows that different spectral shapes between those vegetation types can be exploited on the [1220–1280 nm] domain. The right side of Figure 10 shows that the wavelengths corresponding to the maximum of the first derivatives can clearly discern these two vegetation types even if these vegetation types can be mixed.

**Figure 8.** (**Left**) spectral signatures of AQ\_B (blue) and AQ\_C (dark slate gray). Red dashed lines are the wavelengths used by the Normalized Difference Water Index (NDWI) [860,1240] index; (**Right**) NDWI [860,1240] values for each vegetation type, H is the Hellinger distance.

**Figure 9.** (**Left**) spectral signatures of Sphagnum sp. **(SPHA)!** (black) and AQ\_A (green). Red dashed lines are WI wavelengths; (**Right**) WI values for each vegetation type, H is the Hellinger distance.

**Figure 10.** (**Left**) spectral signatures of SPHA (black) and AQ\_A (green). Red dashed lines are Optimised Soil-Adjust Vegetation Index (OSAVI) [800,670] wavelengths; (**Right**) OSAVI [800,670] values for each vegetation type, H is the Hellinger distance.

**Figure 11.** (**Left**) spectral signatures of SPHA (black) and Calluna vulgaris (CAVU) (gray); (**Right**) F\_WP values for each vegetation type, H is the Hellinger distance.

In most cases, a single biophysical component is sufficient to class a vegetation type from the others (except for CA\_HV), but a pair of biophysical components is needed to discriminate more specifically between some vegetation types (Table 12), apart from some particular cases where a pair of biophysical components is needed CA\_HV (Figure 12). Indeed, CAVU and SALI are differentiated with the stress index (CARTER [695, 420]) and the water index (NDII).

Among the 78 combinations of pair of vegetation types, only two require three indices to be separated: CA\_HV vs. PING and AQ\_A vs. METR. Indeed, because of its within class variability (Table A1), only 33.33% of a single biophysical component can discriminate between CA\_HV and all other vegetation types (Table 13). Besides, as mentioned in Section 4.1, none of the main plant species of PING represents more than 50% of this vegetation type. The advent of a third index only improves significantly their discrimination (Figure 13).

**Figure 12.** (**Left**) spectral signatures of CAVU (gray) and Salix sp. (SALI) (cyan); (**Right**) map of CARTER[695,420] and Normalized Difference Infrared Index (NDII) values for each vegetation type, H is the Hellinger distance.

**Figure 13.** (**Left**) spectral signatures of CA\_HV (pink) and PI\_CV (magenta); (**Right**) map of Optimised Soil-Adjust Vegetation Index (OSAVI) [800,670] and GITELSON values for each vegetation type, H is the Hellinger distance value.

## 4.2.2. Supervised Classification

The 26 indices selected with the Hellinger distance enables overall classification accuracy scores ranging from 72.90% to 85.20% depending on the training size, whereas when considering all indices overall accuracy, scores range from 66.70% to 82.80% (Table 14). Moreover, these selected indices are robust because no significant difference between classifiers score (except for RF) regardless of the training size is noted (Figure 14). As expected, the worst results are given by the Kruskal-Wallis method (to compare performance of the two features selection methods, 26 first indices given by Kruskal-Wallis method have been selected).


**Table 14.** Vegetation types identification (overall accuracy (±standard deviation) in %) with indices.

RLR gives better results than SVM and RF (Table 14, Figure 14) except when the size of the training set equals 50% for the Hellinger distance. That may be explained by the possible confusion between some vegetation types due to their plant species composition. Indeed, SVM aims to find the best hyperplane that can separate data, whereas RLR aims to find a probability (according to a logistic function) to separate them.

Considering RLR--2 some vegetation types are not easily discriminated whatever the indices. Tables 15 and 16 show that PING has the lowest F1-score (20.99 % and 33.13 % respectively) which can be explained by the mixed composition of this habitat (Appendix B) and not the low number of spectra. Indeed, AQ\_B has about the same number of spectra: 7 spectra whereas 8 spectral measurements have been collected for PING. Yet it has a F1-score = 91.95 % considering all indices and F1-score = 91.66 % considering indices selected by the Hellinger distance that can be explained by its composition dominated by *Utricularia* sp.

**Figure 14.** Vegetation types identification accuracies (overall accuracy) with indices.

**Table 15.** Confusion matrix of the classification based on Regularized Logistic Regression (RLR)--2 with all indices and training size = 25%. The producer's and user's accuracies and the overall accuracy average (OAA) are also shown.


Focusing on shrubs, JUCO has the best performances (F1-score = 94.83%) whereas SALI and RHFR are often confounded. Table 16 shows that on average 2.53 spectra of RHFR (20.02%) are classified as SALI and on average 2.30 spectra of SALI (19.15%) are classified as RHFR. Indeed, as JUCO has a higher foliage density, the overall spatial signature is less sensitive to the ground influence and as a result JUCO spectral reflectance is close to a pure endmember (Appendix B). In the latter case, the spectral measurements are composed of soil and more affected by mixed signatures. Another pair of vegetation types is hardly discriminated: PI\_CV and CA\_HV. Table 16 shows that on average 4.93 spectra of CA\_HV (25%) are classified as PI\_CV which may be explained by the plant species they have in common: *Carex* (50–100% depending on the location) and *Molinia caerulea* ssp. *caerulae* (40–70%) (Appendix B).


**Table 16.** Confusion matrix of the classification based on RLR--2 with indices selected by the Hellinger distance and training size = 25%. The producer's and user's accuracies and the overall accuracy average (OAA) are also shown.

#### *4.3. Supervised Classification According to the Spectral Ranges*

Only the best results are presented, obtained with the four spectral ranges ([350–750 nm], [750–1350 nm], [350–1350 nm], [350–2500 nm]) and the spectral signature as reference and the three transformed spectral signatures (second derivative, first derivative, Continuum Removed Derivative Reflectance).

Tables 17–20 show the best results obtained with RLR--2 on [350–1350 nm] whatever the transformed spectral signatures.




**Table 18.** Vegetation types identification accuracies (overall accuracy (±standard deviation) in %) on [750–1350 nm].

Considering wavelengths used by selected indices (Section 4.2.1), most of them use spectral bands located on [350–1350 nm] either: 50% are located in visible range and 32.35% in near-infrared range. Indeed, in this spectral range all the biophysical components discriminating the peatland vegetation types can be taken into account. That is confirmed by Figure 15 which shows that the best results are given by [350–1350 nm] considering the training size = 25% regardless the transformed spectral signatures and the the classifier, except for RF applied on the spectral signature. In this case, considering the whole spectral range improves the result by 1% compared with [350–1350 nm].

Considering RLR--2 in [350–1350 nm], Table 21 shows that the best overall accuracies are given by first derivative, second derivative and CRDR. First and second derivatives overall accuracies are very close (difference lower than 1%). However, those transformations are sensitive to noise. However, CRDR delivered better results than spectral signatures and similar performances to the first and second derivatives (difference is lower than 4%). As mentioned in Section 4.1, those transformations are closely related to absorption features rather than reflectance magnitude [38], and are helpful to discriminate between peatland vegetation types which are clearly characterized by different biophysical components as mentioned in Section 4.2.1.



**Figure 15.** Vegetation type identification accuracies with the training size = 25%.


**Table 20.** Vegetation types identification accuracies (overall accuracy (±standard deviation) in %) on [350–2500 nm].



Considering RLR, -1 regularization, which controls the selection or the removal of variables, always underperforms -2-regularization, which handles collinear variables [16]. Because of mixed plant species, it is difficult to remove variables that are not involved in the classification of all the vegetation types. Although SVM and RF are popular classifiers in remote sensing community, they are outclassed by RLR in [350 nm to 1350 nm] which is the spectral range where results are the best (Figure 16). Results given by SVM RBF are lower than those obtained with RLR and can be explained by the difficulty to find adapted parameters considering this high dimensionality problem. However, it is interesting to note that results from SVM linear are close to RLR ones considering first derivative, second derivative and CRDR. Further investigations should be conducted to better understand the link between those classifiers and improve the choice of the parameters. Figure 16 shows that PLS-DA is the least sensitive classifier to training size regardless transformed spectral signatures in [350–1350 nm].

**Figure 16.** Vegetation type identification accuracies on [350–1350 nm].

Table 22 shows that *Pinguicula* sp. (PING) has the lowest F1-score (66.67% and 56.00% respectively) as well as for the spectral vegetation indices (Section 4.2.2). Besides, this vegetation type can hardly be discriminated from the other ones (Producer's accuracy (PA) = 53.33%) and some *Pinguicula* sp. combined vegetation (PI\_CV) spectra are classified as PING). However, it should be kept in mind that PING has a small number of spectra. Considering Aquatic type b (AQ\_B) which has about the same number of spectra (7 spectra against 8 for PING), User's Accuracy (UA) = 60.98% and some Aquatic type a (AQ\_A) spectra are predicted as AQ\_B ones. These poor UA results compared to one obtained by spectral vegetation indices can not be explained by the spectral domain. Indeed, the best spectra vegetation index (NDWI[860,1240]) that discriminate between AQ\_A and AQ\_B has both wavelengths in [350–1350 nm]. However, this result may be qualified by PA. Indeed, on [350–1350 nm] domain, UA = 100.00% whereas UA = 84.60% for spectral vegetation indices. Nevertheless, using a continuous spectral domain can lead to worse results for other vegetation types such as *Sphagnum* sp. (SPHA), *Calluna vulgaris* (CAVU), AQ\_A: F1-score is always better considering the same classifier (RLR--2) applied on spectral vegetation indices selected by the Hellinger distance (SPHA: 91.12% vs. 82.80%; CAVU: 77.62% vs. 71.43%; AQ\_A: 86.80% vs. 82.81 %). Considering SPHA, if PA = 90.59 % for spectral vegetation indices or for [350–1350 nm], the latter predicts more SPHA than observed (UA = 76.24%) and is more confused with CAVU. This can be explained by plot 7 which is mainly composed of *Calluna vulgaris* (20%), *Carex rostrata* (25%), *Molinia caerulea* ssp. *caerulae* (20%) and *Sphagnum palustre* (20%) (Appendix B).

In our case, reducing feature space by selecting most discriminant wavelengths (using PCA or MNF) has not been implemented, whereas it can be an interesting track to explore to see if it improves results for RLR--2. *Juniperus communis* (JUCO), *Eleocharis quinqueflora* (ELQU) and Aquatic type c (AQ\_C) have about the same F1-score considering spectral vegetation indices or [350–1350 nm]: less than 2% difference. However, they have better PA on the continuous spectral range (PA = 100.00% for JUCO; 95.56% for AQ\_C) which means that this spectral range contains discriminant wavelengths able to catch characteristics of those vegetation types.

*Rhododendron ferrugineum* (RHFR), *Carex* sp. homogeneous vegetation (CA\_HV), *Salix* sp. (SALI) and *Menyanthes trifoliata* (METR) have better results considering [350–1350 nm]. This can be explained by the fact that the spectral vegetation indices used have not been built for that kind of vegetation type. Further investigations can be undertaken to find specific indices that can discriminate between those vegetation types and other ones.


**Table 22.** Confusion matrix of the RLR--2 classification using Continuum Removed Derivative Reflectance (CRDR) on [350–1350 nm] (training size = 25%). The producer's and user's accuracies, the overall accuracy and the F1-score are also shown.

#### **5. Conclusions and Perspectives**

This study aimed at inventorying and evaluating the performance of discrimination techniques for peatland habitats based on in situ hyperspectral measurements with a high spectral resolution and high signal-to-noise ratio. To evaluate the potential of hyperspectral data to separate and classify those habitats, three classes of methods were investigated and compared:


This study demonstrated that is it possible to discriminate between peatland vegetation types by using the Canberra distance on the whole spectral range [350–2500 nm]. This distance is sensitive to a small change when both coordinates approach zero which is the case of reflectance especially in the visible ranges and in the SWIR (Figure 2). Further investigations should be conducted to see if combinations of spectral range can improve overall accuracy or if the lack of spectral signatures in the reference database (which is a weakness of this method) may explain why the whole spectral range is needed to compare spectra in that case. Besides, it is of importance to collect more spectral signatures from peatland vegetation types to build a spectral reference database of peatland vegetation types that can catch more spectral variability.

Although there are no spectral vegetation indices built to discriminate between peatland vegetation types, this study showed that some indices could be selected using the Hellinger distance. Although those indices have not been built to discriminate between peatland vegetation types, they were able to classify them because they focus on biochemical properties such as chlorophyll, nitrogen, water stress, etc. Further investigations have to be done to see the impact of spectral bandwidth around the wavelength of selected indices instead of working with one particular wavelength. For instance, there are lots of indices that catch the same biochemical property but wavelengths of interest change because they focus on specific plant species (e.g., for the chlorophyll, SR [700,670] is built for field corn, whereas SR [675,700] is built for soy beans leaves; contrary to SR [675,700], SR [700,670] has been selected with the Hellinger distance).

Contrary to similarity measures which had the best results considering the whole spectral range, supervised classification on specific spectral range as defined by [31] achieved the best overall accuracy considering [350–1350 nm] domain. This is in agreemen<sup>t</sup> with the spectral vegetation indices: only 4 indices (NDWI [860, 1240], NDWI [860, 2130], NDWI [1110, 1450], MSI) over the 26 selected have a discriminant wavelength which is not in this spectral range. More precisely, the discriminant wavelength is located in the SWIR and all concerned vegetation indices are linked to the water status. Further investigations should be conducted on the extraction or the reduction of features of this spectral range to understand why this domain sometimes gave worse results than spectral vegetation indices depending on the vegetation type.

Among the three methods, the best results are obtained considering a specific spectral domain [350–1350 nm] with RLR regardless of the transformed spectral signatures and the size of the training size (overall accuracy ranges from 81.47% to 96.36%). However, it should be of interest to apply feature reduction methods usually applied on remote sensing (such as PCA or MNF) to see it results are improved or specific spectral wavelength can be selected.

To our knowledge, although not popular in remote sensing for classifying (but already used for feature selection), the RLR classifier achieves the best overall classification accuracy when applied to the spectral vegetation indices selected by the Hellinger distance (77.21%) on the [350–1350 nm] domain (83.84%) considering training size = 25%.

Furthermore, this study showed that CRDR gave encouraging results event if it is slightly below those obtained by the first derivative and the second derivative considering RLR classifier.

Considering the habitats, some vegetation types were more easily separated. For instance, JUCO had the best F1-score with the spectral vegetation indices selected by the Hellinger distance (94.83%) or on the [350–1350 nm] (95.24%) with RLR and the training size = 25%. In some cases, this specific spectral domain gave better results (F1-score = 92.21% whereas with spectral vegetation indices F1-score = 64.72% for SALI) while in other case, the spectral vegetation indices gave better results (F1-score = 91.12% whereas F1-score = 82.80% for SPHA). As mentioned earlier, reducing feature space needs to be investigated to see if a particular feature space exists that can discriminate between and classify all vegetation types or if we need to consider either spectral vegetation indices or a specific spectral domain depending on the vegetation type to classify.

Although all the results strongly depended on the current dataset, this study illustrated promising methods for classifying peatland vegetation types using in situ hyperspectral measurements. The next step concerns the application or adaptation of those methods to airborne hyperspectral imageries with high spatial resolution acquired on September 2014 (simultaneously with in situ measurements). With the objective of evaluating the benefits of airborne or spaceborne sensors with a lower spectral resolution a lower signal-to-noise ratio, these conclusions may change. For that purpose, some indices (involving wavelengths lower than 480 nm) will not be used because of the camera spectral range sensitivity and some transformed spectral signatures such as second derivative will also not be used because of signal-to-noise ratio. Similarly, the first derivative transformation is very sensitive to the noise coming from the instrument but also from the atmosphere correction and this can degrade its performance.

Additional imageries acquired in October 2012 and July 2013 would allow us to test these methods with spectral signatures extracted from the ancillary dataset. Multi-temporal analysis could also be conducted to discriminate between vegetation types thanks to the phenological changes. This step would be of interest to evaluate the robustness of spectral measurements, spectral vegetation indices and classifiers selected previously from in situ hyperspectral measurements to airborne data.

**Acknowledgments:** The authors would like to thank Rosa Oltra-Carrió and Olivier Vaudelin for their help with field measurements and acknowledge the LabEx DRIIHM and the Observatoire Hommes-Milieux (OHM-CNRS) Haut-Vicdessos for funding and supporting the study.

**Author Contributions:** Thierry Erudel conducted the analyses and wrote most of the manuscript. Florence Mazier helped with floristic survey data. Thomas Houet, Sophie Fabre and Xavier Briottet helped with the field measurements and contributed as supervisors. All authors contributed to the preparation of the manuscript.

**Conflicts of Interest:** The authors declare no conflict of interests.





**Table A1.** *Cont*.

**Table A1.** *Cont*.



**Table A1.** *Cont*.

*Remote Sens.* **2017**, *9*, 748

#### **Appendix B. Data from Vegetation Types**

*Appendix B.1. Sphagnum sp. (SPHA)*

**Figure A1.** Location of the in situ spectroradiometer measurements for the plots of *Sphagnum* sp. (SPHA).

**Figure A2.** Mean reflectance (*μ*) and standard deviation (*σ*) of *Sphagnum* sp. (SPHA).


**Table A2.** Pictures, plots, geographic coordinates and number of spectra of *Sphagnum* sp. (SPHA).

*Appendix B.2. Calluna vulgaris (CAVU)*

**Figure A3.** Location of the in situ spectroradiometer measurements for the plots of *Calluna vulgaris* (CAVU).

**Figure A4.** Mean reflectance (*μ*) and standard deviation (*σ*) of *Calluna vulgaris* (CAVU).


**Table A3.** Pictures, plots, geographic coordinates and number of spectra of *Calluna vulgaris* (CAVU).

**Table A4.** Pictures, plots, geographic coordinates and number of spectra of *Eleocharis quinqueflora* (ELQU).


**Figure A5.** Location of the in situ spectroradiometer measurements for the plots of *Eleocharis quinqueflora* (ELQU).

**Figure A6.** Mean reflectance (*μ*) and standard deviation (*σ*) of *Eleocharis quinqueflora* (ELQU).

*Appendix B.4. Pinguicula sp. (PING)*

**Figure A7.** Location of the in situ spectroradiometer measurements for the plots of *Pinguicula* sp. (PING).


**Figure A8.** Mean reflectance (*μ*) and standard deviation (*σ*) of *Pinguicula* sp. (PING).

*Appendix B.5. Menyanthes trifoliata (METR)*

**Figure A9.** Location of the in situ spectroradiometer measurements for the plots of *Menyanthes trifoliata* (METR).

**Figure A10.** Mean reflectance (*μ*) and standard deviation (*σ*) of *Menyanthes trifoliata* (METR).


**Table A6.** Pictures, plots, geographic coordinates and number of spectra of *Menyanthes trifoliata* (METR).

*Appendix B.6. Juniperus communis (JUCO)*

**Figure A11.** Location of the in situ spectroradiometer measurements for the plots of *Juniperus communis* (JUCO).



**Figure A12.** Mean reflectance (*μ*) and standard deviation (*σ*) of *Juniperus communis* (JUCO).

*Appendix B.7. Rhododendron ferrugineum (RHFR)*

**Figure A13.** Location of the in situ spectroradiometer measurements for the plots of *Rhododendron ferrugineum* (RHFR).

**Figure A14.** Mean reflectance (*μ*) and standard deviation (*σ*) of *Rhododendron ferrugineum* (RHFR).


**Table A8.** Pictures, plots, geographic coordinates and number of spectra of *Rhododendron ferrugineum* (RHFR).

*Appendix B.8. Salix sp. (SALI)*

**Figure A15.** Location of the in situ spectroradiometer measurements for the plots of *Salix* sp. (SALI).


**Table A9.** Pictures, plots, geographic coordinates and number of spectra of *Salix* sp. (SALI).

**Figure A16.** Mean reflectance (*μ*) and standard deviation (*σ*) of *Salix* sp. (SALI).

*Appendix B.9. Aquatic Type a (AQ\_A)*

**Figure A17.** Location of the in situ spectroradiometer measurements for the plots of Aquatic type a (AQ\_A).

**Figure A18.** Mean reflectance (*μ*) and standard deviation (*σ*) of Aquatic type a (AQ\_A).


**Table A10.** Pictures, plots, geographic coordinates and number of spectra of Aquatic type a (AQ\_A).

*Appendix B.10. Aquatic Type b (AQ\_B)*

**Table A11.** Pictures, plots, geographic coordinates and number of spectra of Aquatic type b (AQ\_B).


**Figure A19.** Location of the in situ spectroradiometer measurements for the plots of Aquatic type b (AQ\_B).

**Figure A20.** Mean reflectance (*μ*) and standard deviation (*σ*) of Aquatic type b (AQ\_B).

*Appendix B.11. Aquatic Type c (AQ\_C)*

**Figure A21.** Location of the in situ spectroradiometer measurements for the plots of Aquatic type c (AQ\_C).

**Figure A22.** Mean reflectance (*μ*) and standard deviation (*σ*) of Aquatic type c (AQ\_C).


**Table A12.** Pictures, plots, geographic coordinates and number of spectra of Aquatic type c (AQ\_C).

*Appendix B.12. Carex sp. Homogeneous Vegetation (CA\_HV)*

**Figure A23.** Location of the in situ spectroradiometer measurements for the plots of *Carex* sp. homogeneous vegetation (CA\_HV).

**Figure A24.** Mean reflectance (*μ*) and standard deviation (*σ*) of *Carex* sp. homogeneous vegetation (CA\_HV).

**Table A13.** Pictures, plots, geographic coordinates and number of spectra of *Carex* sp. homogeneous vegetation (CA\_HV).


*Appendix B.13. Pinguicula sp. Combined Vegetation (PI\_CV)*

**Figure A25.** Location of the in situ spectroradiometer measurements for the plots of *Pinguicula* sp. combined vegetation (PI\_CV).

**Figure A26.** Mean reflectance (*μ*) and standard deviation (*σ*) of *Pinguicula* sp. combined vegetation (PI\_CV).

**Table A14.** Pictures, plots, geographic coordinates and number of spectra of *Pinguicula* sp. combined vegetation (PI\_CV).

