RESULTS

The training and validation of the CNN for surface pattern classification achieved high accuracies (Figure 3), both for the two-class and the three-class classification. Training losses were similar to validation losses, suggesting a reasonable model fit. There was no evidence of overfitting, as validation loss continued to decline with epoch. The change in training loss tended to decline with each successive epoch, but was still gently declining by epoch 15 for the three-class surface classification system, suggesting slight underfitting of the model. However, the CNN performed well when predicting on new datasets, whether based on two classes (Sections 3.1) or three classes (Section 3.2).

Hydromorphological unit classification

Overall, the HMU classification system predicted HMUs that were consistent with those that had been identified manually.
Alta. The Alta stretch, showing a sequence of alternating slow flowing, mild-smooth HMUs and fast flowing, mild-turbulentHMUs (Figure 4; upper panel), was classified into a similar sequence by the automated HMU classification system (Figure 4; middle panel). The gentle gradient of the watercourse (<4%) meant thatsteep-smooth HMUs (associated with runs) were absent. Only one cell was classified as being a steep-turbulent HMU, where the surface gradient locally exceeded 4%: its small size suggests that it is not a functionally “realistic” classification of flow conditions. The automated classification showed a finer level of detail than the manual classification, reflecting the surface patterns identified by the CNN (Figure 4; lower panel). For example, standing waves extended around the outer bank of a meander into an area that had been manually classified into a mild-smooth HMU.
Nidelva. The automated HMU classification for the Nidelva stretch alternated between mainly mild-smooth and mild- orsteep-turbulent HMUs (Figure 5). Steep-smooth HMUs were rare and found as isolated cells. The HMUs with more turbulent flows were found in the mid-part of the imaged stretch, where higher gradients occurred (Supplementary figure 1). The spatial configuration of mesohabitat classes largely concurred with those identified by Borsányi (2006) (Supplementary figure 1). HMU classifications depended on flow conditions, and the middle parts of the imaged stretch showed an increase in the prevalence of turbulent HMUs at high discharge (see inlet panels in Figure 5).
Orkla. Automated HMU classification of the Orkla stretch was consistent with the manual classification in the mid- to downstream part but diverged from the manual classification in the upstream part (Figure 6). The middle part, mainly downstream of an island, consisted of a smooth water surface in a mild channel gradient (mainly classified as the mild-smooth HMU); the surface became more turbulent further downstream (mainly classified as the mild-turbulent HMU). The upstream part of the Orkla stretch consisted of smooth surfaces, separated by short regions of white water associated with rapids or cascades. While the automated HMU classification system was successful in classifying the mild-turbulent and steep-turbulent areas successfully, it did not, however, always correctly identify areas free from standing waves. For example, some areas were misclassified asmild-turbulent HMUs due to the two-class CNN assigning then astanding wave classification, when manual examination of the imagery showed the presence of air bubbles/foam rather than standing waves. These misclassifications typically occurred in mild gradient areas, downstream of rapids or cascades where there was white water at the surface from the downstream advective diffusion of air bubbles/foam.

Refined surface pattern classification

The Refined surface pattern classification system (three classes:smooth or rippled , standing waves and diffusing foam ) classified surface patterns in the Orkla stretch that were consistent with visual inspection of the imagery (Figure 7). In particular, it identified the presence of advective diffusion of air bubbles/foam, immediately downstream of rapids and mainly in pool mesohabitats. Here, white water was present on the surface but was not in the form of standing waves generated by interaction between the flow and the riverbed immediately beneath the white water. Smooth or rippled surface patterns were more prevalent at the downstream side of pools, where surface foam had disappeared due to diffusion.