DISCUSSION

This study has shown that an automated HMU classification system, based on a CNN classification of surface patterns evident in aerial orthophotos, and a rule-based classification of surface gradient based on topographic LiDAR-derived DSMs, can be used to successfully classify river stretches into sections of distinct hydromorphology. For the three river stretches examined, classifications broadly coincided with those previously identified (Borsányi 2006, Hindar et al. 2019). However, further refinement of the CNN with respect to classification of surface patterns may lead to a better characterization of hydromorphological conditions, particularly with regard to distinguishing between turbulence generated locally (standing waves) and upstream (advective diffusion of air bubbles/foam). In the following sections, we discuss the HMU classifications for our study rivers, discuss the sensitivity of the classification to the data used for training and validating the CNN and for making subsequent predictions, and examine how the approach outlined here can be further developed.

Hydromorphological unit classifications in the study rivers

HMU classifications produced by the automated classification system were generally consistent with previously made manual classifications. The automated classification was derived from the best currently available aerial orthophotos for comparison with previously made manual classifications. Differences between automated and manual classifications may partly be attributed to differences in flow conditions at the time of surveying, which are known to affect mesohabitat classifications (Hauer et al. 2009).
No large contiguous area of steep-smooth HMU, associated with the run mesohabitat, was found. Conditions of a smooth or rippled surface pattern alongside a steep gradient are rare in Norway because steep gradients are often associated with standing wave surface patterns; in 16 Norwegian rivers for which manual mesohabitat classifications are available (Alta, Aurland, Eidselva, Enningdalselvam, Flomselvi, Halselva, Imsa, Jolstra, Laerdal, Laerdal, Laukhelle, Nausta, Orkla, Stjorda, Stryn, Suldalslågen; see Hindar et al. 2019), no locations have been classified as corresponding to steep-smooth conditions. Cases where isolated cells were classified as steep-smooth could be argued to be misclassifications, suggesting that it is necessary to set a minimum size for an HMU and/or that the classification should take into account the HMU classifications of the neighboring cells when defining an HMU in any given cell.
The HMU classification was dependent on the surface gradient of the river. We used the 4% threshold of Borsányi (2006), but expert opinion has suggested that this threshold may be too high for a river reach close to the estuary, such as the Nidelva (Borsányi 2006). The HMU classification did not differentiate between turbulent conditions that were generated locally (standing waves) and the advective diffusion of air bubbles/foam generated upstream. This resulted in inaccurate predictions of hydromorphology in the Orkla in locations downstream of cascades or rapids. These were typically pool habitats, a mesohabitat noted for low turbulence (Stone & Hotchkiss 2007), which were incorrectly assigned a mild-turbulent HMU. A refined surface pattern classification based on three classes correctly classified these areas as diffusing foam , suggesting that the HMU classification needs further refinement.

Training and validation of the CNN

Key to ensuring an effective hydromorphological classification is using a suitable dataset for training and validating the CNN. The HMU classification system used two very dissimilar classes – smooth or rippled versus standing waves – so it was easy to train the CNN such that it could distinguish between these patterns for predicting on new datasets. Additionally, it was easy to create training classes to differentiate between standing waves and diffusing foam . Differentiating between further unsmoothed conditions (e.g. between broken and unbroken standing waves) requires careful selection of sample classes. This may be difficult because the optical properties of the phenomena may overlap, depending on both local flow conditions and how the imagery is acquired. Training and validation therefore require selection of class examples over a sufficiently broad range of conditions for effective prediction in other river stretches.
The CNN-based approach for identifying flow features requires high-resolution aerial photographs of the water surface that are not adversely influenced by sub-optimal light conditions, both for training and validating the CNN, and predicting on new datasets. Firstly, the approach may be less effective for identifying mesohabitats when shadows fall across the water surface, making surface patterns less visually identifiable (darker, “noisier” imagery). The approach will therefore be less effective for narrow watercourse headwaters in steep, tree-lined valleys where the entire channel may be obscured by shadow (see Hedger et al. 2022). Secondly, the optical properties of areas in direct sunlight depend on the interaction between the water surface and the solar position (Mount 2005, Zeng et al. 2017). Surface undulations may be more apparent under direct rather than diffuse sunlight, and specular reflection from unbroken stranding waves in imagery taken under direct sunlight could appear more like the white water inherent in broken standing waves. Therefore, the relative frequency of the classification type will depend on the light environment. Thirdly, in shallow areas, it is possible that the texture on the riverbed from coarse substrates could be confused with surface patterns.

Further development

The method proposed here provides a system by which river surface pattern recognition may be applied to remotely sensed imagery within a decision system to obtain an automated map of river hydromorphology. Improved application of the remote sensing data may provide a more refined classification: specifically, better characterization of flow features via a refined surface pattern classification, and the use of additional hydromorphological information to refine the habitat classification.
Refining the surface pattern classification. The HMU classification system only categorized surface patterns into two classes: (1) smooth or rippled and (2) standing waves . We have shown here that there is potential to distinguish between standing waves and the advective diffusion of air bubbles/foam. There is, however, also potential to differentiate between unbroken and broken standing waves. Unbroken standing waves might have alternating bright pixels, associated with reflections on the side of the wave facing the sun, and darker pixels facing away from the sun. Broken standing waves will have more white water distributed randomly in the cell. The channel characteristics influence hydraulics, so more detailed information on the forms of surface flows might be used in further defining the hydromorphology. For example, unbroken standing waves are more associated with riffles whereas broken standing waves are more associated with rapids and cascades (Newson & Newson 2000).
Refining HMU classes. The full classification system on which the current work is based (Borsányi et al. 2004) uses depth and velocity to further refine the mesohabitat class: for example, cascades are defined by shallower depths than rapids. Our HMU classification system used class divisions that were only based on surface patterns and gradient, so only provided four broad HMU types. However, extraction of depth via remote sensing is possible, using bathymetric LiDAR (Hugue et al. 2016, Puig-Mengual et al. 2021, Sundt et al. 2022), or analysis of image spectra (Legleiter & Harrison 2019, Sundt et al. 2021), or Structure from Motion applied to UAV imagery (Dietrich 2017). Velocity may be determined using large-scale particle image velocimetry, applied to UAV imagery (Detert & Weitbrecht 2015). There may also be potential for extraction of depth and velocity using a CNN approach applied to the imagery of the type used in this analysis, based on direct properties within the imagery (e.g. the darkness of an area indicating depth, or foam patterns indicating flow speed). Refining the HMU classes with information obtained directly from the imagery will give a more detailed classification of hydromorphology.