INTRODUCTION

Categorizing fluvial habitat is an essential step for supporting river habitat management and conservation programs. Indeed, the physical habitat composition in a river or stream and the corresponding hydraulic parameters are considered to be basic elements to river health assessment (Maddock 1999). A range of approaches with associated terminology – river landform , morphological unit ,mesohabitat type , hydromorphological unit (HMU) ,physical/hydraulic biotope , ecotope, channel geomorphic unit (CGU) (see Belletti et al. 2017) – have been developed for categorizing fluvial habitat, reflecting differences in research focus, rationales, and the scale at which the research is conducted. The term “mesohabitat” (e.g. Tickner et al. 2000) applies to a locally contiguous area (typically less than several 100 m in length) consisting of similar hydromorphological conditions, such as flow velocity, turbulence and depth (see Wegscheider et al. 2020). Pardo and Armitage (1997) characterize mesohabitats as being “visually distinct units of habitat within the stream, recognizable from the bank and with an apparent physical uniformity”. Mesohabitats are often associated with particular depth-velocity conditions (Kemp et al. 1999) and it is these interactions between flow and physical habitat characteristics that create the variety of mesohabitats. In particular, surface flow type is considered a major descriptor of physical habitats. The composition and particular assemblage of mesohabitats in a river is an indicator of conditions for fish, macroinvertebrates, and other river biota. Fausch et al. (2002) advocated for the pertinence of mesohabitat characterization for the study of fish ecology, arguing that features relevant to fish movement and behavior, as well as barriers and obstacles, were best assessed at this scale. Mesohabitats are usually defined as distinct classes, such as riffle, glide, pool or run, to cite a few (see for example Parasiewicz 2007). There is some inconsistency in how mesohabitat classes are defined (Newson & Newson 2000), but typically, these classes compartmentalize a continuum of overall hydromorphological conditions, from slower and more laminar flows in low gradient areas (e.g. glides), to faster and more turbulent flows in higher gradient areas (e.g. rapids). The term hydromorphological unit is often used as a synonym for mesohabitat (Alcaraz-Hernandez et al. 2011, Suska & Parasiewicz 2020), but in the current study we define this as a broad class, based on the flow features and gradient of the water surface.
A range of methods for classifying fluvial hydromorphology and mesohabitats exist (see Harby et al. 2004 for a first summary and overview), ranging from simple field-based qualitative assessment [e.g. the River Habitat Survey method (Newson et al. 1998); theNorwegian Mesohabitat Classification method, (Borsányi et al. 2004)] to numerical modelling approaches [e.g. MesoCASiMir(Eisner et al. 2005, 2007), MesoHabsim (Parasiewicz 2007)]. Field-based mesohabitat classification may simply involve observing features such as surface flow type from the riverbank and inferring the habitat types from these. For example, in Norway, Borsányi (2004) developed a system based on surface flow features or types, surface gradient, flow velocity and depth, all of which can be visually assessed in the field. Such a system requires no specialized instrumentation and can be done using bankside observations, possibly supplemented with qualitative interpretation of aerial photographs. The diversity of methods for identification of mesohabitats in the field has four major difficulties in common: 1. They require some experience to provide consistent and robust results; 2. Researcher variability may lead to the same mesohabitat being characterized differently depending on the surveyor (e.g. riffle vs shallow run); 3. The same mesohabitat type may be identified differently depending on the method used, and similar terms are used by different methods to identify different features; 4. They can be time-consuming depending on the method used and length of river to be surveyed. Numerical modelling approaches, for instance hydraulic modeling, may remove some of the subjectivity but can be very time-consuming, both in terms of obtaining data for model validation and for setting-up the model.
Remote sensing, typically based on airborne or satellite true color or multispectral imagery and/or LiDAR data, offers the advantage of providing synoptic coverage of the river at a range of spatial scales pertinent to the mesohabitat and over larger distances. Remote sensing is particularly useful as a source of empirical data for numerical models. For instance, channel bathymetry may be derived from a range of methods (Sundt et al. 2022), and such data can then be used to derive flow properties and other habitat metrics, either by empirical-based hydraulic rules (Hugue et al. 2016) or by hydraulic models (Hauer et al. 2009, Sundt et al. 2022). For example, Hauer et al. (2009) used a rule-based system to classify a watercourse into distinct mesohabitat types (riffles, fast runs, runs, pools, shallows, backwaters) based on predictions from a hydraulic model calibrated using LiDAR data. Drawbacks and limitations to this approach are that establishing hydraulic properties is time-consuming and requires specific expertise. It may also not always be possible to use remote sensing to obtain depth estimates, for instance if the river bottom is not visible in aerial images and/or LiDAR data are unavailable.
Remote sensing may be used to directly determine hydromorphology and mesohabitat because properties of the river’s water surface may provide information on its hydromorphology. For example, turbulent flows, shown by broken or unbroken standing waves on the surface, may indicate the presence of steep channels with coarse bed material (associated with rapids or cascade mesohabitats), whereas a smooth or rippled water surface is more indicative of laminar flows (associated with glide or run mesohabitats). While surface water patterns have been incorporated into mesohabitat classification schemes (e.g. Borsányi et al. 2004), they have not been fully exploited in remote sensing-based mesohabitat mapping. The advantage of deriving these properties directly from remote sensing is that it is possible to retain the primary advantage of remote sensing (synoptic coverage) without the requirement to estimate bathymetry and/or introduce subsequent modelling (e.g. hydrodynamic modelling). Ideally, the remote sensing approach should be automated to reduce subjectivity and effort.
Automated extraction of information on river habitats from remote sensing imagery is difficult due to a large number of confounding factors (Hedger et al. 2022). Artificial intelligence techniques offer the potential to deal with the complexity existing in imagery of river habitats, and have been used to classify surface cover type (Carbonneau et al. 2020), river sediment sizes (Takechi et al. 2021), hydromorphological features (Casado et al. 2015), and salmon redds (Harrison et al. 2020). There is potential for such an approach to be used to classify features of the water surface (e.g. identifying whether the surface is smooth or whether the surface has standing waves). Flow features identified in such a manner can then be combined with information on surface gradient (readily obtainable from digital terrain model, DTM, or digital surface model, DSM, data) for a refined hydromorphological classification. Convolutional neural networks (CNNs) are particularly appropriate for this type of work. These are deep learning algorithms that can be trained on images to assign importance (learnable weights and biases) to features of the image within a network linking the images to a predefined classification. Such a network can then be used to predict class types on new images.
Here, we evaluate the potential for automated river classification into hydromorphological units (HMUs) based on airborne remote sensing data. We use a classification system that requires information on only (1) surface patterns, identified by applying a CNN to aerial photographs, and (2) surface gradient, identified from NIR (topographic) LiDAR-derived DTMs, to classify areas into one of four broad HMU classes. Following Borsányi’s rule system, we begin with a simple surface pattern classification (smooth or rippled versusstanding waves ), but we also investigate a more detailed surface pattern classification that allows the distinguishment between standing waves (generated locally) and air bubbles/foam (generated upstream and being advected downstream).