**1. Introduction**

Monitoring coastal areas is essential to the preservation of the land-water continuum's habitats and the services they provide, particularly in a context of local and global changes [1]. Seagrasses, salt marshes, mangroves, macroalgae, sandy dunes, or beaches are examples of such habitats that continually interact with the tide levels. They can be found along the temperate shorelines and play key roles in the ecological equilibrium of these ecotones. Seagrasses ensure water quality and are a significant carbon sink, along with salt marshes and mangroves [1,2]. Coastal habitats also provide protection from marine hazards to coastal communities and infrastructures and supply many recreational activities such as snorkeling, fishing, swimming, and land sailing [1–3]. Finally, they support a wide range of endemic species by offering them nurseries, food, and oxygen [1,2]. However, coastal (including estuarine) habitats are exposed to a plethora of natural and anthropic threats that may be amplified by global changes. Thorough observation of coastal processes is necessary to identify the trends of evolution of these fragile environments. It requires regular data acquisition along the shoreline with spatial resolution and time spacing both adapted to the task. However, the surveying complexities inherent to land-water continuum areas hinder their monitoring at a time scale relevant to their fast evolution, and over large, representative extents. Remote sensing can adequately address this issue.

**Citation:** Letard, M.; Collin, A.; Corpetti, T.; Lague, D.; Pastol, Y.; Ekelund, A. Classification of Land-Water Continuum Habitats Using Exclusively Airborne Topobathymetric Lidar Green Waveforms and Infrared Intensity Point Clouds. *Remote Sens.* **2022**, *14*, 341. https://doi.org/10.3390/ rs14020341

Academic Editor: Junshi Xia

Received: 15 November 2021 Accepted: 7 January 2022 Published: 12 January 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

Due to the presence of water, coastal surveys are conventionally split between topographic and bathymetric campaigns, both constrained to the tide and the field accessibility. Subtidal areas can be surveyed with waterborne acoustic techniques, while supratidal domains are documented with passive or active imagery using satellite, airborne, or unmanned aerial vehicles (UAV) [4–6]. Boats being unable to reach very shallow areas and imagery being limited by the water surface, intertidal, and shallow water areas are harder to accurately monitor [7]. Distinct terrestrial and marine surveying campaigns can also be difficult to merge, considering they might rely on different reference systems, and their thin overlapping area can be challenging to sample thoroughly with ground control points. Existing seamless surveying techniques over land-water interface areas are summarized in the following section.

Multispectral or superspectral imagery can be used for coastal habitat mapping. In clear and shallow water, traditional image classification techniques can be applied [7]. A more accurate approach consists in suppressing the effects of water on light refraction and diffusion by using inversion models on superspectral imagery. Using such models, it is possible to obtain satellite-derived bathymetry [8] or satellite-derived topobathymetry [9,10], which are proven to improve the classification of coastal covers obtained [11]. Bathymetry can also be extracted using multispectral imagery, as demonstrated in [12]. Multispectral imagery has the advantage of being accessible with different platforms: UAVs, planes, or satellites nowadays all benefit from multiband sensors. The cost of acquisition can therefore be lowered depending on the chosen source, and the revisit time can allow high temporal resolution monitoring.

Hyperspectral imagery is the last passive imagery-based method to map the landwater continuum [13,14]. The key principle of methods using hyperspectral imagery to study submerged areas is to model the interactions of light with the water column and correct them to obtain imagery unaffected by these processes. By inversing a radiative transfer model of the water column, it is possible to derive the seafloor reflectance and estimate the bathymetry [15]. These products are adapted to the characterization of sea bottom types and can be used for benthic classification tasks.

Although satellite passive imagery overcomes the issue of accessibility and temporal resolution, its spatial resolution is sometimes too coarse to spot specific changes (depending on the sensor and the quality of pan-sharpening), and it only penetrates water in shallow, clear areas [16]. The main issue with passive imagery remains the depth range in which it is usable. Due to optical phenomena, past a certain depth threshold that varies with water clarity, passive imagery can no longer give information on what lies beneath the water surface. Bathymetry extraction via active imaging then becomes the only way to gather information on these areas. Furthermore, even in shallow waters, bathymetry derived from active sensors gives access to the seabed covers' elevation but also to the seabed's elevation itself, providing 3D information on these covers, which enables biomass estimation or other structural assessments [17–20]. The bathymetry measurements obtained with active sensors (airborne bathymetric lidar or waterborne acoustic soundings) also leverage a higher vertical precision, useful for ecological structural assessments [12].

Airborne topobathymetric lidar is a reliable alternative: it ensures information continuity between land and water, covers vast areas quickly, penetrates a depth of up to dozens of meters and has a higher spatial resolution than satellite imagery [21,22]. Current approaches to map coastal interfaces using airborne lidar mostly make use of the digital terrain models or digital surface models derived from the lidar point clouds (PCs), including those obtained for the water bottoms after removing points corresponding to the water bodies [17]. Fewer studies rely on the 3D PCs of the lidar surveys to generate coastal or riverside habitats maps. Directly processing the PCs and avoiding rasterization has the advantage of preserving the dense spatial sampling provided by lidar sensors. It also opens possibilities for 3D rendering of the results, and structural analysis thanks to the rich spatial information contained in PCs. Indeed, the vertical repartition of the points offers useful information on scene architecture, providing relevant features to determine their origin, namely for vegetation or building identification. Analysis of this geometrical context is the most frequently used method to produce maps of land and water covers [23–25]. Research works conducted on PCs processing mostly rely on the computation of geometrical features using spherical neighborhoods [23] and, more recently, on deep neural networks [26].

Another possibility with airborne lidar is to exploit the spectral details contained in the backscattered signals. These can be recorded under the form of time series of intensities received by the sensor: lidar waveforms. Each object of the surveyed environment illuminated by the sensor's laser reflects light in a specific way, generating a characteristic signature in the signal. Waveforms consequently provide additional information on the structure and physical attributes of the targets. The shape, width, and amplitude of their spectral signature—a peak—are information that can be used for land and water covers mapping [27–29]. Waveforms are therefore a useful indicator of the diversity of coastal areas. Though many methods have been proposed to process airborne topographic lidar full-waveforms, airborne bathymetric lidar full-waveforms are, to the best of our knowledge, much less explored. They are even less employed for classification tasks, and often only analyzed to retrieve bathymetry. There are currently three main approaches to waveform processing. The first consists in decomposing the waveforms to isolate each element of the train of echoes in the signal [30,31]. The second consists in reconstructing the signal by fitting statistical models to the waveforms [32]. Knowing how to approximate the sensed response allows to extract the position and the attributes of each component. The last approach is to analyze waveforms straightforwardly as any 1D signal to detect their peaks [27], using first derivative thresholding for example. Identifying waveform components is useful in order to localize the objects populating the scene, but also to extract features to describe them and prepare their automatic classification [27,33,34].

Classification of land or water covers using lidar data has been well explored recently. Even when using waveform data, most of the published research is based on 2D data classification [17,25,27,29,33] while fewer articles exploit PCs [24,34,35]. Many studies researching ways to classify lidar data used machine learning algorithms such as support vector machine (SVM), maximum likelihood (ML), or random forests. The maximum likelihood is mostly used for 2D lidar data, while SVM and random forests have been proofed on PCs. SVM and random forest seem to have similar classification performances on 3D lidar data [36]. However, with these algorithms, the spatial context around each point is not considered and does not impact the prediction [36]. Research papers show that conditional random field (CRF) and Markov random field (MRF) classifications produce better results in that way [36,37]. However, these require heavier computation and are more difficult to apply to large datasets. Currently, there is a consensus on the efficiency of random forest on that aspect [36]. Contrary to SVM, CRF, or MRF, it is easy to apply to large datasets. Random forest is, furthermore, robust to overfitting issues and offers the possibility to retrieve predictors contribution easily. In this article, we therefore wish to implement a hand-crafted features' random forest classification to map coastal habitats. Although machine learning classification of lidar waveform features has been explored previously, we have found no point-based studies dedicated to mapping a large number of habitats both marine and terrestrial. Previously cited studies such as [24,34,35] classified either only marine or only terrestrial habitats from PCs and [27] processed 2D data to produce their map of coastal habitats.

The present study aims at mapping an array of 21 habitats of the 3D land-water continuum seamlessly using exclusively airborne topobathymetric bispectral lidar. Our objective is to bridge the gap between marine and terrestrial surveys, demonstrate that efficient methods can be developed to automatically map the land-water interface, and show that an integrated vision of coastal zones is feasible and advised. Our contributions consist in (1) developing a point-based approach to exploit full-waveform data acquired during topobathymetric surveys for subtidal, intertidal, and supratidal habitats mapping, (2) quantifying the contribution of green waveform features, infrared (IR) intensities, and relief information to the classification accuracy based on a random forest machine learner. We improve an experimental method presented in a previous work [29] and test it on a wider area including both emerged and submerged domains to determine the suitability of full-waveform lidar data for coastal zone mapping. UAV data, inexpensive to implement, is used to estimate the accuracy of the resulting very high spatial resolution maps, which are produced under the form of PCs, paving the way for 3D classification of land and sea covers using solely topobathymetric lidar data.

#### **2. Materials**

#### *2.1. Study Area*

The study location was chosen along the northern coasts of Brittany, France, for its ecological diversity and due to the availability of full-waveform lidar data acquired by the French Hydrographic Office (Shom) as part of the Litto3D® project [38]. The study area, presented in Figure 1, features typical coastal habitats such as fine sand or pebble beaches, a sandy dune, rocky areas provided with macroalgae, seagrass meadows, and salt marshes at the estuary of a local river. It also hosts a small resort town, Sables d'Or les Pins (48◦38 27N: 2◦24 24W). Buildings, tar or concrete-covered paths, boats in mooring, vehicles in parking lots, wooded areas populated with evergreen pine trees or deciduous species, and crop fields are also present in the selected zone. All these habitats are home to a rich variety of species: shellfish; dune plant vegetation; green, red, and brown seaweed; eelgrasses; evergreen and deciduous trees; crops; and salt marsh plants such as glasswort, common soda, sea purslane, or sea lavender.
