**1. Introduction**

Peatlands represent a diverse array of wetlands that accumulate partially decomposed organic material. Whilst they may only cover a small proportion (~3%) of the Earth's land surface, these ecosystems are highly important in terms of functional and ecological values. Indeed, undisturbed, global peatland systems act as net atmospheric carbon sinks, storing approximately a third of the world's soil organic carbon [1], the vast majority of which (450–547 GtC (Gigatons of Carbon)) is held in northern peatlands (those above 45◦N [2]). From an ecological perspective, these environments also provide important habitats for a number of rare plant and animal species [3].

Traditionally, species discrimination for floristic mapping needs intensive field work, including taxonomical information and the visual estimation of percentage cover for each species which are costly and time-consuming and sometimes inapplicable due to their poor accessibility [4]. Remote sensing is a technique that gathers data regularly about the earth's features. The main advantages that make remote sensing preferable to field-based methods in land cover classification, are that it has repeat coverage potential, allowing continuous monitoring, and its digital data can be easily integrated into a geographic information system (GIS) for more analysis which is less costly and less time-consuming [5,6].

Historically, aerial photography was the first remote sensing method to be employed for mapping wetland vegetation [7]. Currently, a variety of remotely sensed images are available for mapping wetland vegetation thanks to of airborne and space-borne vectors with multi-spectral sensors or hyperspectral sensors which operate within the different optical spectra [8].

Mapping and monitoring wetlands' (and even though peatland) floristic diversity is really challenging. Indeed, both temporal and spatial resolutions of remotely sensed imageries and in situ plant diversity and mixing contribute to the limitation of such techniques. Wetland plants are not as easily detectable as terrestrial plants since herbaceous wetland vegetations exhibit high spectral and spatial variabilities because of its steep environmental gradients [5,8]. Besides, the reflectance spectra of wetland vegetation canopies are often very similar and can be combined with reflectance spectra of the underlying soil, hydrologic regime and atmospheric vapour [9,10].

However, plant species have been successfully classified in estuarine [11], palustrine [12] and riparian habitats [13], as well in saltmarsh [5], in mangrove [14,15], in swamp [16] but not in peatlands, to our knowledge. Peatland mapping faces two grea<sup>t</sup> challenges at local and global scales due to their high environmental function (biodiversity hotspot, greenhouse gas fluxes, etc.): characterizing their internal diversity [8] and delineating their extent [17]. This study focuses on the first challenge for which only high-spectral or spatial-resolution imageries appear appropriate (see for instance [18–20]).

Plant species classification can benefit from several existing and recent techniques commonly used in remote sensing. Two main methods are applied for vegetation discrimination: the similarity measurement techniques and the supervised classification methods with sometimes application of a preliminary spectral band reduction technique. On one hand, similarity measures enable us to discriminate between similar classes from a set of spectra, extracted from images or acquired on the field. Some spectral measures, such as the Spectral Angle Mapper (SAM) are related to the difference of the spectral shape (e.g., Yagoub, H. et al. [21] identified forests of the Liege oaks from other forests, grain crops and steppes using the multispectral Advanced Very High Resolution Radiometer (AVHRR) with five bands from 580 nm to 1250 nm, 1 km spatial resolution (Overall Accuracy (OA) = 94.10%, *κ* = 0.93); Bahri, E.M. et al. [22] discriminated between tree species using the multispectral Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) sensor with 9 spectral bands from 520 nm to 2430 nm and a spatial resolution of 15 m or 30 m (*κ* = 0.66)). Other spectral measures, such as the Spectral Information Divergence (SID) are related to probabilistic behaviour (e.g., Sobhan, I. [23] classified different tree species at leaf and vegetation cover scales using the hyperspectral HyMap sensor: 126 spectral bands from 436 nm to 2485 nm and a spatial resolution of 4 m (OA = 91.10%, *κ* = 0.87)). On the other hand, the supervised classification methods may contribute as well to discriminate between (group of) spectral signatures for plant species discrimination. The Linear Discriminant Analysis (LDA) is a method assuming that independent variables are normally distributed and which attempts to look for linear combination of variables to model the difference between the classes of the data (e.g., Clark, M.L. et al. [24] succeeded in classifying different tree species at leaf and vegetation cover scales using the HYperspectral Digital Imagery Collection Experiment (HYDICE) sensor with 210 spectral bands from 400 nm to 2500 nm, 1.6 m spatial resolution (OA = 86% using an object-based approach)). The Random Forest is an ensemble learning method based on the construction of multiple decision trees (e.g., Lawrence, R.L. et al. [25] succeeded in mapping invasive plants using the hyperspectral Probe-1 sensor: 128 bands from 450 nm to 2507 nm, 5 m spatial resolution (OA = 86% for the leafy spurge classification)). The Support Vector Machines (SVM) is a classifier that looks for the best separating hyperplane (e.g., Dalponte, M. [26] succeeded in classifying different tree species in boreal forest using HySpex VNIR-1600-instrument: 160 spectral bands ranging from 410 nm to 990 nm , with a spatial resolution of 0.4 m (OA = 79.2%); Vyas, D. et al. [27] classified successfully tropical vegetation using the Hyperion (EO-1) sensor (OA = 80%)). The Regularized Logistic Regression (RLR) is the combination of a linear model (logistic regression) and a regularization term. It is usually used for feature selection (e.g., Pant, P. et al. [28] applied it to reduce the 64 spectral bands from the hyperspectral AisaEAGLE II sensor to classify tree species in boreal forest using SVM; Pal, M. [29] applied it for reducing the 79 bands from the hyperspectral Digital Airborne Imaging Spectrometer (DAIS) sensor and the 220 bands from the hyperspectral Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) sensor to classify different land covers using SVM) is investigated in this paper as a classifier.

Discriminating between and classifying plant species can be done. Firstly, using different techniques hyperspectral measurements can be made thanks to a portable spectroradiometer (FieldSpec Pro FR, Analytical Spectral Devices—ASD) which ranges on the reflective domain ([350–2500 nm] with a spectral resolution of 3 nm in Visible and Near InfraRed (VNIR) and approximatively 10 nm in the ShortWave InfraRed (SWIR)) either on laboratory [14] or immediately after the leaf was cut using the leaf clip accessory [16]. This can be an indicator of the ability of discriminating plant species using specific wavelengths or evaluating the performance of a classifier. Then, the wetlands heterogeneity mixing vegetation types can be catched still using a portable spectroradiometer: orbick, N. et al. [12] used the ASD spectroradiometer, Ground Field of View (GFOV) = 0.43 m; Schmidt, K. et al. [5] used the GER 3700 (Geophysical and Environmental Research Corporation) which ranges from 350 nm to 2509 nm) with a spectral resolution of 2 nm below 1000 nm and from 6 to 10 nm beyond 1000 nm, GFOV = 0.13 m. Secondly, with airborne imageries, hyperspectral sensors (SOC-700: 120 spectral bands between 394 and 890 nm with a 4 nm bandwidth and a spatial resolution of 0.5 m and a spatial resolution of 3 m [13]; HyMap: 128 bands in the visible and near infrared (VNIR: 0.45–1.50 μm with a 10 nm bandwidth) through the shortwave infrared (SWIR: 1.50–2.50 μm with a 15–20 nm bandwidth [11]). Thirdly, with spaceborne imageries using hyperspectral sensors (Hyperion: 242 spectral bands from 357 to 2756 nm with a spectral interval of 10 nm and a spatial resolution of 30 m [15]) or multispectral sensors (SPOT-5: 4 bands with 10 m resolution [15]) can be used to map wetlands.

This study aims at inventorying and evaluating the performance of discrimination techniques for peatland habitats based on in situ spectra. These habitats are characterized by more or less homogeneous vegetation mixing and have been chosen because of their ecological values (i.e., biodiversity). As defined by [30], mapping these habitats is therefore important to identify potential and/or effective areas with (at least) a floristic biodiversity function. For instance, we do not aim at detecting *Drosera rotundifolia* but at mapping the habitat favorable to this species (*Sphagnum* ...). Similarity measures and classifiers were applied on spectral signatures and some of their transformations (first and second derivatives, continuum removal, first derivative of continuum removal, normalized spectral signatures, log transformation). These transformations have been chosen because they enhance biophysical components which may help to distinguish plant species. These techniques were applied on different spectral ranges that either characterize specific biophysical

components [31]. Classifiers were applied on spectral vegetation indices, characterizing specific biophysical components such as chlorophyll, pigments, nitrogen, cellulose, water.

This paper is organized as follows. After presenting the study site located in the Pyrenees (France) and associated data collection in Section 2 the methodology is detailed in Section 3. Then Section 4 presents and discussed the results of the different classifications that are suitable for distinguishing vegetation types. Finally, in Section 5, the conclusion summarizes the main results and some perspectives that have arisen in applying these techniques to hyperspectral imageries.
