Performance and accuracy of cross‐section tracking methods for hydromorphological habitat assessment in wadable rivers with sparse canopy conditions

This article investigates the performance and accuracy of continuous Real‐Time Kinematic (RTK) Global Navigation Satellite System (GNSS) position tracking for hydromorphological surveys, based on a comprehensive river restoration monitoring campaign. The aim of the research was to assess the method's suitability for efficient data collection in turbid, wadable rivers with sparse canopy conditions, and responds to the water management sector's increasing demand for efficient, low‐cost, and robust survey techniques. The methodological approach involved comparing manual, cross‐sectional water depth measurements to water depth estimations obtained by applying different interpolation methods to the continuous tracking data. The results demonstrate good agreement between both datasets (R2 = 0.77, RMSE = 0.13 m). When using a local standard deviation filter to remove noisy RTK‐GNSS measurements, estimation performance increased significantly (R2 = 0.96, RMSE = 0.06 m). The filter's influence on the hydromorphological habitat statistics mean water depth and coefficient of variation was limited but proved to be relevant for reach‐scale assessments of hydromorphological diversity. Based on a correlation analysis of >106 RTK‐GNSS position logs, we furthermore assessed the impact of tree canopy on RTK‐GNSS measurement accuracy and observed a strong influence within 6.5 m from the canopy border. Estimated accuracy deteriorated noticeably when canopy penetration exceeded 1 m, and accuracies >1 m were common beyond 4 m penetration. The study highlights the efficiency gains achieved with RTK‐GNSS tracking, and showcases its potential for hydromorphological surveys and streamgaging applications in challenging conditions, making it a promising alternative to traditional methods and remote sensing techniques.

Rivers and their alluvial zones are crucial hotspots for biodiversity, supporting a wide range of aquatic and terrestrial species (Ward et al., 1999).However, human activities such as land reclamation, channelization, and river regulation have severely impacted the morphology, habitats, and functionality of our riverscapes (Décamps et al., 1988;Elosegi et al., 2010).In response to these challenges, the restoration of degraded rivers has become increasingly important, also in a legal context (Schirmer et al., 2014), becoming an integral part of the European Water Framework Directive (WFD, 2000) in 2000, and the Swiss Waters Protection Act (WPA, 1991) after its revision in 2011.Hydromorphological habitat assessment is essential for restoration planning and monitoring to evaluate the physical characteristics governing ecological processes (Fryirs, 2015).A broad range of different assessment methods exist (Belletti et al., 2015).Field-based methods such as the Swiss Restoration Outcome Evaluation protocol (ROE; Weber et al., 2019) generally require cross-sectional water depth measurements to derive indicators of hydromorphological habitat quality, oftentimes from statistical measures such as sample mean or coefficient of variation (CV) (e.g., Gostner et al., 2013;Woolsey et al., 2007).Depending on the survey setting, it can be challenging to obtain high-precision water depth measurements within a globally efficient workflow.To overcome the limitations of traditional and state-of-the-art remote sensing techniques, this paper presents an alternative approach for cross-sectional water depth measurements and evaluates its suitability for hydromorphological habitat assessments in wadable, turbid rivers under partial canopy cover.
Manual surveys with traditional survey rods are labor-intensive and time-consuming.Recently, Real-Time Kinematic (RTK) Global Navigation Satellite System (GNSS) technology has been used to obtain instantaneous, high-precision altitude measurements at crosssectional measurement points (e.g., Bandini et al., 2023;Kasvi et al., 2019;Kinzel et al., 2021;Woodget et al., 2015), requiring each point to be approached individually, though.More sophisticated remote-sensing technologies become increasingly available and can be suitable over a wide range of scales (Carrivick & Smith, 2019;Smith & Vericat, 2014;Tomsett & Leyland, 2019).Two widely used airborne techniques are airborne lidar bathymetry (ALB) and structure-from-motion (SfM) photogrammetry (e.g., Detert et al., 2020;Dietrich, 2017;Hilldale & Raff, 2008;Kinzel et al., 2021;Shintani & Fonstad, 2017;Woodget et al., 2015).Both approaches can provide topographic data in the cm accuracy range (Eltner et al., 2021;Islam et al., 2022) but the quality is highly sensitive to flow turbidity (Mandlburger et al., 2020;Woodget et al., 2015), especially for SfM (Kasvi et al., 2019).SfM is most effective in water depths up to 15 cm whereas such shallow water depths might not be captured reliably by ALB technology (Kinzel et al., 2021).Canopy obstruction significantly reduces data availability for SfM (Javernick et al., 2014) and impacts the data accuracy of ALB surveys (Islam et al., 2022).The cost of ALB surveys can be prohibitive for small projects with <1 km reach length (Kinzel et al., 2021) whereas 100 m long reaches can be well suited for SfM (e.g., Eltner et al., 2021).Similarly to the ALB technique, ground-based through-water terrestrial laser scanning (TLS) is sensitive toward flow turbidity (Smith & Vericat, 2014).Mobile TLS systems (boat-or cart-based) have been developed to overcome spatial data collection limitations (Hohenthal et al., 2011), yet the associated cost and logistical efforts for reachscale TLS are very high (Lague, 2020).Vessel-based remote-sensing with active echo-sounders determines water depths from emitted acoustic signals, using low-cost single-beam echo-sounders for simple applications and high-resolution multibeam echo-sounders or sidescan interferometric sonars for larger scales (Arseni et al., 2019;Blondel, 2014;Halmai et al., 2020).Echo-sounding-based surveying works well in water depths >2 m (Halmai et al., 2020;Kinzel et al., 2021) but major difficulties are expected when depths <0.2 m are common (Smith & Vericat, 2014).Inverse methods based on noninvasive model calibration from terrestrial photography (e.g., Pasquale et al., 2014) require that cross-sections are integrated within the digital terrain model (Schäppi et al., 2010) and a correction map is produced.
In summary, none of the current state-of-the-art survey techniques can be considered as a low-cost, rapid, robust, and efficient alternative to manual water depth measurements when high-precision cross-sectional measurements are required in wadable, potentially turbid rivers, with partially very shallow sections in sparse canopy conditions.To improve the efficiency of such hydromorphological surveys, we optimized the manual mapping approach itself.Significant efficiency gains can be achieved for environmental surveys when GNSS positioning is coupled with mobile geographic information system (GIS) applications (Nowak et al., 2020), for example, the open-source software QField (Kuhn & Bernasocchi, 2016).For the mapping approach presented in this paper, we used QField's position tracking functionality (The QField Project/OPENGIS.ch,2023) in combination with RTK-GNSS equipment to track the three-dimensional trajectory of a surveyor who uninterruptedly wades along a set of predefined cross-sections, continuously logging the backpack-mounted receiver's position and accuracy information.The water depths at the points of interest were determined in a post-processing step based on simple interpolation techniques.The sensitivity of these depth estimations and derived hydromorphological habitat statistics toward survey technique and data processing strategies, as well as the approach's general suitability in sparse canopy conditions required dedicated analyzes.
Multipathing effects caused by signal reflections from tree canopy or the water surface can considerably reduce GNSS accuracy (Bakula et al., 2015;Johnson & Barton, 2004).The impact of canopy obstruction is well documented for different environmental conditions (e.g., Abdi et al., 2022;Andreas et al., 2019;Kaartinen et al., 2015;Morales & Tsubouchi, 2007;Sigrist et al., 1999) but the combined influence of water surface reflections and canopy proximity on estimated RTK-GNSS accuracy-as is the case for far overhanging bank vegetation-has not received much attention.This study's aim was hence threefold, namely (i) to test and evaluate the power, versatility, and robustness of the proposed continuous RTK-GNSS tracking approach in the context of a hydromorphological habitat assessment, (ii) to develop and compare data processing strategies to improve the method's water depth estimation performance, and (iii) to obtain a basic understanding of how local canopy conditions influence the suitability of the approach.Our analyses were guided by the hypotheses that (a) global estimation performance can be improved using a local standard deviation filter to detect and remove noisy tracking logs and (b) the influence range of canopy proximity (IRCP) regarding RTK-GNSS accuracy can be determined via a correlation analysis between estimated accuracy and distance to the canopy border for a series of distance-filtered data subsets.With a steadily increasing number of restoration projects in Europe (Szałkiewicz et al., 2018) and worldwide (Woolsey et al., 2007) the need for low-cost, efficient, rapid, and robust survey techniques for hydromorphological habitat assessments is more topical than ever.
Related fields of application potentially profiting from the proposed mapping approach are streamgaging operations, bathymetric surveys, or bank erosion monitoring.

| Field study
Our hydromorphological field surveys took place in the Sarine river (Fribourg, Switzerland) in August 2022 during maximum canopy cover for deciduous trees.Two 200 m long reaches were surveyed following the ROE protocol.The reaches' habitat diversity was mapped and evaluated based on a set of six indicators (Weber et al., 2019).There are no significant tributaries between the dam and the study reaches, so the water levels in both reaches were near-constant throughout the field surveys.More detailed descriptions of local environmental conditions and the river restoration context are provided in the literature (Mörtl et al., 2023;Schroff et al., 2022).

| Field data
Our mapping approach consisted of a surveyor wading along a set of predefined cross-sections while continuously tracking the position using an RTK-GNSS receiver.We used the Emlid Reach RS2 receiver, running on firmware v28.4 and configured with the application Reach-  In total, 21 cross-sections were examined, 7 in the UR and 14 in the DR.Their location was defined by single RTK-GNSS measurements at each bank.At each cross-section, we determined the water surface elevation by averaging two measurements with an estimated vertical accuracy of <20 mm.Water depth reference measurements were collected using a traditional survey rod and each position was defined by the cross-section number and distance from the right bank (366 points, 1.4 m spacing).During the continuous tracking with QField's line tracking functionality, 3D-nodes were automatically added to a geopackage (GPKG) layer while the surveyor waded twice along each cross-section: the first traverse from the right to the left bank was followed by a second traverse back to the right bank before advancing to the next cross-section.The minimum time interval between subsequent nodes was set to 1 s and the minimum distance to 0.1 m. Figure 2 shows the wading trajectory with associated nodes for the DR as well as positions of water surface elevation measurements.
The tracking data required processing before comparing it to the manual reference measurements.In the first step, the tracking line was split into partial lines, each partial line containing the two traverses of a cross-section.The 3D-nodes of all partial lines were extracted, yielding a total of 4160 nodes.For each cross-section, the corresponding nodes were orthogonally projected onto the pre-defined section.For each projected node, its relative position (distance from the right bank) was determined and water depth was calculated by subtracting the receiver altitude from the sum of water surface elevation and receiver height above ground.
To study the influence of canopy proximity on RTK-GNSS accuracy, the raw NMEA files were analyzed.The NMEA files included Then we calculated the distance to the canopy border (dist2CB) for each NMEA log as the minimal distance to the MultiPolygon border.
Negative distance values were attributed to logs under the canopy and positive values to non-covered logs, that is, outside the MultiPolygon.Figure 2b shows the NMEA logs with their color-coded dist2CB values for the DR.

| Data analysis
For the performance analysis of the tracking approach we calculated water depth estimates from the irregularly distributed tracking nodes at the regularly spaced positions of manual measurements.A number of appropriate interpolation methods exist (Arseni et al., 2019;Rishikeshan et al., 2014), but inverse distance weighting (IDW) and moving average (MA) are widely applied, more robust than the nearest neighbor method and less complex than kriging (Maleika, 2020).In F I G U R E 2 Field data collected for the downstream reach.The left map shows the wading trajectory of the cross-section tracking (yellow line), the associated 3Dnodes for each cross-section (green bullets) and the locations of water surface elevation measurements (yellow crosses).
The right map shows all position logs of the RTK-GNSS receiver within the survey perimeter.The logs' distance to the canopy border (dist2CB) is color-coded, ranging from green (under canopy) via white (canopy border) to red (no cover).Background © swisstopo.[Color figure can be viewed at wileyonlinelibrary.com] this study, we adopted modified versions of IDW and MA, with nonoverlapping windows.Each cross-section was subdivided into discretized intervals (or bins, windows), comparable to Bandini et al. (2023).
Bin limits were defined by the riverbanks and centers between manual measurement positions.Each tracking node was assigned to its corresponding bin.We calculated the MA water depth estimate h MA,i for each manual measurement position i as by averaging the previously determined water depths h i,j of the n i tracking nodes falling into the associated bin i.The IDW estimate h IDW,i was calculated from the depths h i,j and inverse distance weights The weights w i,j were determined considering the absolute distance d i,j of tracking node j to the manual measurement position i as well as a factor f = 0.1 m to account for the inaccuracy of manual measurement position, based on field experience.Based on a preceding analysis, the power exponent was set to p = 1, increasing the influence of more distant nodes compared to the standard exponent p = 2.
The resulting tracking evaluation dataset was comprised of 366 samples of manual depth measurements, MA estimates, and IDW estimates.Using this dataset, we calculated performance metrics to assess the tracking approach at the reach-scale and compare both interpolation methods.The coefficient of determination R2 and the root mean squared error RMSE have been used in similar settings (e.g., Arseni et al., 2019;Bandini et al., 2023;Dietrich, 2017;Islam et al., 2022;Javernick et al., 2014;Kinzel et al., 2021;Rishikeshan et al., 2014).We calculated both metrics comparing manual measurements once to MA estimates and once to IDW estimates.
To study the influence of RTK-GNSS measurement noise on performance scores we used a local standard deviation filter.High-noise bins were identified by calculating the standard deviation σ z of node water depths from the tracking nodes contained in a bin and using σ z as a filter variable.Threshold (t) specific subsets of the tracking evaluation dataset were defined where each subset k contained all bins with σ z ≤ t k (i.e., valid bins).Bins with σ z > t k were considered as outlier bins.The threshold t ranged from vertical receiver precision (t = 0.01 m) to the maximum observed σ z (t = 0.75 m).The R2 and RMSE were calculated for each subset and interpolation method.Four statistically or geomorphologically important threshold criteria were given particular attention: Criterion C1 considers the gravel bed as a continuous surface without jumps, prescribing a maximum slope angle α < 40 .Assuming a uniform distribution of tracking nodes between bin limits and considering the standard bin width of 1.4 m, the threshold value t C1 was calculated following (Walpole et al., 2017) as t C1 = 0.33 m.Criterion C2 restricts the percentage of bins identified as outliers to 5%.Criteria C3 and C4 optimize performance results, minimizing the RMSE and maximizing the R2, respectively.
Using the tracking evaluation dataset and its subsets, we furthermore analyzed the difference between hydromorphological habitat statistics calculated from manual measurements and from RTK-GNSS estimates, respectively.Reach estimates of mean water depth and depth variability were obtained by calculating the sample mean (μ) and the CV for manual water depth measurements, MA bin estimates and IDW bin estimates, respectively.For each subset, the CV was calculated from the valid bins as the ratio of sample standard deviation divided by sample mean.The CV of water depth can be used in several forms for hydromorphological assessments (e.g., Gostner et al., 2013;Weber et al., 2019).Generally, a score close to zero indicates a poor hydromorphological state while a score close to one indicates increased hydromorphological diversity (Woolsey et al., 2005).In addition to the analysis of the full tracking evaluation dataset, the same calculations were repeated for two partial sets containing only the data belonging to the UR's and the DR's crosssections, respectively.
To study the influence of canopy proximity on RTK-GNSS accuracy we performed a correlation analysis using the NMEA logs dataset.
We defined the influence range of canopy proximity (IRCP) as the width of the buffer around the canopy border, for which the correlation between estimated position accuracy and distance to the canopy border (dist2CB) of logs falling inside the buffer zone is maximized.
The degree of correlation was quantified using the Pearson correlation statistics (coefficient r and p-value) and calculated separately for the two accuracy types horizontal (acc_h) and vertical (acc_v).
To obtain a robust estimation of the IRCP, the correlation statistics were calculated for a series of subsets of the NMEA logs dataset.
Subsets were obtained by chaining two filters.Filter 1 represents the buffer width around the canopy border and uses dist2CB as the filter variable.Only logs with an absolute distance jdist2CBj ≤ F1 enter the subset.An exploratory data analysis provided the relevant filter thresholds F1 considered for IRCP determination, ranging between 1.0 m ≤ F1 ≤ 10.9 m with a 0.1 m interval.Filter 2 represents an accuracy type-dependent outlier filter.Only logs with an estimated accuracy acc ≤ F2 enter the subset.Seven meaningful threshold values F2 were considered (F2 = [0.5, 1, 2, 3, 5, 10, and 20 m]).Moreover, Filter 2 can be used for isolated analyses of a certain accuracy quality, for example, submeter accuracy logs.
The Pearson statistics were calculated for all 1400 combinations of the 100 F1 thresholds, the 7 F2 thresholds, and the 2 accuracy types.For each F2 threshold and accuracy type, the correlation-maximizing threshold value F1 = F1 opt was determined.
Given sufficient homogeneity among the F1 opt values, the IRCP was determined as Within the IRCP zone, the relationship between estimated accuracy and dist2CB was investigated in detail for submeter accuracy logs, that is, by setting F1 = IRCP and F2 = 1 m.To detect significant trends, the bias of inhomogeneous sample density along the dist2CB axis was removed.The influence range was subdivided into intervals or bins of 0.1 m.By calculating the mean accuracy of all logs contained in a bin, the bin means dataset was obtained, presenting uniform sample densities for both accuracy types.

| Performance of water depth estimations
A visual comparison of manual and RTK-GNSS water depth measurements is shown in Figure 3  The results of the metric-based performance analysis are presented in Figure 4a.Performance curves are provided for both interpolation methods (MA, IDW) and performance metrics (R2, RMSE).
The curves span the entire value range considered for filter threshold t.The plot also shows the percentage of valid bins contained in the threshold-specific subsets.Starting at the lower threshold limit, the percentage of valid bins increases rapidly with increasing threshold values, exceeding 90% at t = 0.11 m.The remaining 10% of bins are F I G U R E 3 Water depth measurements at selected cross-sections in the upstream reach (UR) and the downstream reach (DR).Water depths were manually measured (black bullets) and semi-automatically tracked with an RTK-GNSS receiver (gray crosses).Water depths along cross-section UR5 were too deep for wading-based measurements, causing a data gap.A figure containing all 21 cross-sections can be accessed online as supporting information (Figure S1).
[Color figure can be viewed at wileyonlinelibrary.com] gradually included for thresholds between 0.11 m < t < 0.75 m.
Depending on the threshold value, both R2 and RMSE vary significantly.The weakest RMSE performance, that is, the highest RMSE score (RMSE MA = RMSE IDW = 0.191 m) is obtained for the most restrictive filter threshold (t = 0.01 m), for which the subset only contains 2 valid bins.The weakest R2 performance, that is, the lowest R2 score (R2 MA = 0.77, R2 IDW = 0.76) is reached when no filter is applied (t ≥ 0.75 m).The best performance for both metrics and interpolation methods is observed for a coinciding threshold t = 0.04 m with RMSE = 0.062 m and R2 = 0.96, where the subset contains 48% of bins of the full tracking evaluation dataset.For both performance metrics and across the entire range of threshold values, the MA method outperformed the IDW method or ranked equally.Therefore and for reasons of clarity, all results and analyses presented in the following are limited to the MA approach.Another important observation is the substantially different behavior of the performance metrics curves compared to the percentage curve of valid bins, allowing a good compromise between data completeness and estimation performance.
The trade-off between data completeness and estimation performance is exemplified for the threshold criteria C1-C4 in Table 1, where detailed performance results are presented along with related statistics.Criterion C1, prescribing a theoretical slope angle <40 , leads to the most conservative threshold, previously determined as t = 0.33 m.For this threshold, 12 bins were identified as outliers, leading to a 96.7% fraction of valid bins for the corresponding subset.
The RMSE minimizing criterion C3 and the R2 maximizing criterion C4 were both satisfied for t = 0.04 m.For this threshold, more than 50% of bins were filtered as outliers to obtain the optimized scores RMSE = 0.062 m and R2 = 0.96.Scatterplots comparing manual water depth measurements and its corresponding MA estimates are shown in Figure 5, distinguishing valid bins from outlier bins for threshold criteria C1-C4.The majority of scatter points are concentrated in close proximity to the identity function, indicating a good agreement between manual measurements and MA estimates.The 12 outlier bins identified by C1 correspond to scatter points that are also visually distinguishable as outliers due to their large offset from the identity function.The 6 outlier bins additionally identified by C2 include the points with the next largest offset.The performance optimizing criteria C3 and C4 identified 192 bins as outliers, of which many only have a marginal offset.

| Sensitivity of hydromorphological habitat statistics
The mean water depth and CV values calculated from the tracking evaluation dataset as well as from the UR and DR partial sets are presented in Table 1a-c.When comparing the statistics derived from manual measurements and MA estimates we observe a generally good agreement.The mean depth of all manual measurements (μ Manual = 0.51 m) is close to the mean depth calculated from the corresponding MA estimates (μ GNSS-MA = 0.48 m), differing by 3 cm or 6%.A similar agreement is observed for the UR partial set (μ Manual = 0.66 m, μ GNSS-MA = 0.62 m) and the DR partial set (μ Manual = 0.44 m, μ GNSS-MA = 0.42 m).The CV of all manual measurements (CV Manual = 0.54) is only slightly overestimated by the MA estimates (CV GNSS-MA = 0.56).For the UR partial set identical CV F I G U R E 4 Performance metrics and hydromorphological habitat assessment statistics for the threshold-filtered subsets of the tracking evaluation dataset.(a) The coefficient of determination R2 and the root mean squared error RMSE are plotted against the subset-specific threshold values, both for the moving average (MA) and the inverse distance weighting (IDW) method.(b) Mean water depth and (c) coefficient of variation CV were calculated from the manual measurements and MA estimates of each subset.
T A B L E 1 Joint (a) and reach-specific (b, c) results of water depth estimation performance and hydromorphological habitat assessment statistics for the investigated filter criteria C1-C4.

| Influence of canopy on measurement accuracy
The correlation analysis performed to determine the IRCP is visualized in Figure 6.Two subplots show the results of all threshold combinations F1 and F2, separately for the vertical and the horizontal accuracy dataset.Depending on the accuracy type and applied thresholds, the correlation coefficient r varies significantly.
However, the correlation-maximizing threshold values F1 opt , which were determined for each combination of F2 threshold and accuracy type, vary little and seem to be insensitive to F2 or accuracy type.By averaging the F1 opt values, the IRCP was calculated as IRCP = 6.5 m.
Within the IRCP zone, significant correlations are found between estimated accuracy and distance to the canopy border.Figure 7

| Lessons learned
Our findings indicate that water depth estimates obtained from continuous RTK-GNSS tracking perform well when compared to manual depth measurements.The two interpolation methods applied to calculate depth estimates-IDW and MA with non-overlapping windows-obtained similar results for the performance metrics R2 and RMSE.
However, the MA method performed slightly better in our case and is more computationally efficient.
Submerged boulders and rocks, dead wood accumulations, and low-hanging branches resulted in locally increased RTK-GNSS measurement noise.By using a standard deviation filter to remove faulty estimations stemming from high measurement noise, the overall estimation performance was significantly improved for both performance metrics.In certain cases, manual filtering might be preferable, for example, to distinguish between riverbed measurements and wood surface measurements of sparse dead wood piles.The four filtering criteria that were given particular attention can be described as either being informed or non-informed, depending on whether knowledge about the "true" manually measured depths is required.For surveys focusing on (reach mean) water depth estimations in environmental conditions similar to our study, Criterion C2 can be recommended (maximum percentage of outliers = 5%).Its advantages include being non-informed, simple, rather conservative, and efficient.For C2, a nearly optimal RMSE score was obtained (RMSE = 0.07 m).Interestingly, even without noise filtering, the RMSE scores fall within the range of the standard deviation that we obtained for static measurements at geodetic points in immediate canopy proximity.This suggests no significant accuracy loss when tracking continuously while wading, further supporting the RTK-GNSS cross-section tracking approach.
Although noise filtering can significantly improve global estimation performance, its influence on the prediction quality of hydro- Note: n_valid(F1,F2): number of samples in the filtered data subset; n_outliers(F2): number of outliers removed from the subset by Filter 2; Pct_valid(F2): percentage of samples left in the dataset by Filter 2; Pct_valid(F1,F2): percentage of n_valid(F1,F2) compared to the unfiltered NMEA logs dataset.p-values <5eÀ324 were smaller than the minimum positive float64 value on the computer system used.
single cross-section of manual depth and flowmeter measurements or $15 min for depth measurements only.Without considering tape measure preparations, a total of 5 h 15 min can be attributed to the manual depth measurement of all 21 cross-sections.Hence, RTK-GNSS tracking provided significant efficiency gains in the field, reducing the time effort by at least 80%.Time savings might be less extreme when comparing the tracking approach to individual RTK-GNSS measurements at the regularly spaced measurement points.
Further time savings could be achieved when tracking only one traverse per cross-section.The potential time savings of such a single traverse approach need to be weighed against the presumably better robustness and interpolation quality of the double traverse approach.
The postprocessing tasks associated with RTK-GNSS tracking can be standardized and mainly automated, making it a comprehensive approach for hydromorphological surveys and instream habitat mapping.

| Limitations and future work
The study acknowledges certain limitations and suggests potential avenues for future research.First, the use of RTK-GNSS tracking requires a stable mobile network connection.Future studies could explore alternative approaches, such as base-rover configurations or satellite-based correction services (e.g., Trimble's CenterPoint ® RTX).
With these approaches RTK-GNSS tracking might be possible in rugged or remote areas with limited network coverage while also minimizing the dependence on costly, mobile network-based correction services.Second, the commonly used approximation of a constant water surface elevation across a single cross-section introduces a source of inaccuracy that has not been accounted for in the presented analyses.To address this limitation, which is of particular relevance when superelevation occurs in channel bends, water surface measurements could be taken at each bankside, improving the accuracy of depth estimations.Third, we observed a systematic offset between manual depth measurements and RTK-GNSS estimates.Depending on the reach considered and the filtering threshold applied, RTK-GNSS estimates underestimated the mean water depth by $1-4 cm compared to the manual measurements.This phenomenon might be explained with the survey techniques' individual penetration depth into the bed substrate.The surveyor's boots penetrated the top layer only in patches covered by sand or fine sediment, whereas the survey rod penetrated the top layer also in zones presenting a coarser granulometry.Future studies could develop technique-specific normalization rules to improve the agreement between manual depth measurements and RTK-GNSS estimates.We also observed measurement discrepancies in immediate bank line proximity at steep banks.
In particular, for cross-sectional area surveys in narrow channels, steep banks will require special attention and the assumption of fixed receiver height above ground might have to be revised.Fourth, the noise filter's categorical distinction into valid and outlier bins might be revised to improve reach-scale CV estimations.First tests showed that estimation performance can be improved by assigning penalization factors to high-noise bins instead of removing them entirely.
Finally, further investigation is required to understand the sensitivity of the IRCP not only to accuracy type and accuracy outliers but also to other criteria.The IRCP could be determined separately within and outside the canopy perimeter, as well as separately for each study reach.To describe the non-linear relationship between estimated accuracy and dist2CB more accurately, pre-analysis data scaling could be explored.Our study concentrated on a pre-alpine, medium-sized river.By expanding the study to different river widths and alluvial vegetation settings, the suitability and limitations of RTK-GNSS cross-section tracking could be tested over a whole range of geographical and environmental conditions.

| CONCLUSIONS
This article introduced the method of RTK-GNSS cross-section tracking and showed it to be an efficient approach for hydromorphological habitat assessments.The water depth estimates calculated from tracking data using inverse distance weighting and moving average interpolation methods performed well when compared to manual depth measurements (R2 = 0.77, RMSE = 0.13 m).The moving average method performed slightly better and was computationally more efficient than inverse distance weighting.Noise filtering significantly improved estimation performance (R2 = 0.96, RMSE = 0.06 m).However, excessive filtering can lead to significant underestimations of hydromorphological habitat diversity.Special attention is also required

ACKNOWLEDGMENTS
We would like to thank Nico Bätz for helping shape our article's thematic focus and we thank the anonymous reviewer for the valuable and concise comments.Thanks are also due to our colleagues involved in fieldwork.We highly appreciate the good cooperation with the adjacent land owner and cantonal authorities for facilitating access to the study sites and thank OPENGIS.ch for technical support.
Open access funding provided by Ecole Polytechnique Federale de Lausanne.

Figure 1
Figure1shows both reaches, the upstream reach (UR) and the downstream reach (DR), as well as their relative location.The mean width of the active channel was 24.4 m for the UR and 27.5 m for the DR.Both reaches presented similarly sparse canopy cover conditions.Based on drone imagery collected prior to this study, alluvial

F
I G U R E 1 Location of the study reaches in Switzerland (upper left) and in the Sarine river (lower left).On the right, the reach perimeters are shown for both the upstream reach (UR) and the downstream reach (DR).Blue arrows indicate flow direction.Background © swisstopo.[Color figure can be viewed at wileyonlinelibrary.com] for the continuous tracking on the surveyor's backpack.For project preparation and postprocessing, QGIS Desktop v3.24.3 was used with the QField Sync plugin.Data analysis tasks were performed with the Python v3.8.11 suite and packages pandas v1.3.2 and geopandas v0.9.0.
>10 6 position logs, resulting from $60 h of net RTK-GNSS mapping time.Using a custom Python script inspired by the open source nmea2qgis plugin (MaciekO & Mackoo, 2014), all NMEA messages were parsed and each position log's position and accuracy information (estimated vertical and horizontal accuracy) saved as feature attributes in a GPKG point layer.To reduce dataset size, only one log per second was retained, resulting in $200,000 points.Next, the point layer was clipped to the actual mapping area, defined as the active channel surface of both study reaches extended by a 5 m lateral buffer.This reduced the data set to 134,495 point features (NMEA logs dataset).In a subsequent step, we delineated the alluvial forest canopy along the riverbanks using high-resolution drone imagery, saving the entire set of canopy patches as a single MultiPolygon feature.
for selected cross-sections in the UR and DR.A figure containing all 21 cross-sections can be accessed online as supporting information (Figure S1).The agreement between manual and RTK-GNSS measurements was generally good, both in shallow and deeper waters, as seen in cross-sections UR2 and DR14.Stronger deviations and increased RTK-GNSS measurement noise were locally constrained and often concentrated in bank proximity.For the example of DR3, overhanging branches at a distance of 1-3.5 m from the right bank obliged the surveyor to duck, resulting in increased noise.Deadwood accumulation partially obstructed the riverbed at DR5 (0-6 m from the right bank) and submerged boulders caused deviations at DR8 (3-7.5 m from the right bank).
The performance metrics root mean squared error (RMSE) and coefficient of determination (R2) were calculated by comparing the manual water depth measurements to the RTK-GNSS moving average (MA) estimates of valid bins (vb).The coefficient of variation of water depth is abbreviated as CV.values were obtained (CV = 0.39).For the DR partial set, the difference is moderate with CV Manual = 0.57 and CV GNSS-MA = 0.62.For both statistics, that is, mean water depth and CV, the difference between reaches outranges the difference between survey techniques.Table1a-c also shows the influence of filter criteria C1-C4 on the habitat statistics.For the tracking evaluation dataset, the curves plotted in Figure 4b,c additionally show the statistics' evolution over the entire range of threshold values.The mean depth and CV values obtained for low thresholds t (from subsets with <50% valid bins) differ significantly from the values presented in the above paragraph for the unfiltered dataset.For subsets containing <90% of valid bins, the values derived from manual measurements and MA estimates generally show a similar behavior.However, the values derived from MA estimates systematically underestimate the values derived from manual measurements, significantly for CV (Δ = 0.05) and marginally for mean water depth (Δ = 0.01 m).Both for mean depth and CV, the values calculated from manual measurements are stable when the percentage of valid bins exceeds 90%, while the values calculated from the MA estimates exhibit an almost linearly decreasing (mean depth) or increasing (CV) behavior from the 90% mark to the upper threshold limit.Generally speaking, noise filtering led to decreased CV values derived from MA estimates, aligning optimally with manual measurements' CV values for thresholds between 0.67 m ≤ t ≤ 0.71 m.Moderate noise filtering, that is, retaining >90% samples, improved the quality of reach-scale mean water depth derived from MA estimates.
contains two scatterplots showing the accuracy and distance to the canopy border of all submeter accuracy logs within the IRCP zone (filter setting: F1 = 6.5 m and F2 = 1.0 m).The subdivision of the influence range into 0.1 m wide intervals yielded a set of 131 bins.The mean accuracy of submeter logs within each bin (bin means) and the percentage of samples removed by Filter 2 (F2 outliers) are also represented in Figure 7.When considering the submeter accuracy logs, stronger correlations are observed for vertical than horizontal accuracy (vertical: r = À0.31 and horizontal: r = À0.25).For the bin means, the correlation is significantly more pronounced, with similar correlations for both accuracy types (vertical: r = À0.78 and horizontal: r = À0.77).The curve described by the bin means is non-linear and indicates a rapid deterioration of RTK-GNSS accuracy with increasing penetration into canopy-covered areas, starting approximately at dist2CB = À1 m.Logs with an estimated accuracy >1 m (F2 outliers) are concentrated in the deep canopy, with a significant and sudden increase at dist2CB = À4 m.Table 2 provides an overview of correlation, accuracy and filtering statistics of the bin means dataset, the NMEA logs dataset, as well as three selected F1-F2-filtered subsets.The generally very low p-values indicate statistical significance.The application of Filter 1, limiting the NMEA logs dataset to the IRCP zone, increases the correlation coefficient by more than 100% (acc_v/acc_h: from r = À0.04/À0.03 to À0.09/À0.08).The isolated application of Filter 2 (discarding outliers with acc >1 m) removes only 0.1% of samples from the NMEA logs dataset while leading to a major increase for the correlation coefficient (acc_v/acc_h: r = À0.27/À0.20).As also seen in Figure 7, combining both filters further increases the correlation (acc_v/acc_h: r = À0.31/À0.25).The highest correlation coefficients are obtained for the bin means dataset with uniform sample density, exhibiting F I G U R E 5 Comparison of water depths obtained from manual measurements and RTK-GNSS moving average (MA) estimates for the filtering criteria C1-C4.similar correlations for both accuracy types (acc_v/acc_h: r = À0.78/À0.77).

F
I G U R E 6 Determination of the influence range of canopy proximity (IRCP) on RTK-GNSS accuracy.The IRCP is obtained by determining Filter 1 thresholds that optimize the correlation between estimated accuracy and distance to the canopy border (dist2CB) of NMEA logs.The correlation maximizing F1 thresholds (F1 opt ) are plotted with enlarged symbols and averaged to obtain accuracy type-specific IRCP estimates.Subplot (a): vertical accuracy, acc_v and subplot (b): horizontal accuracy, acc_h.F I G U R E 7 Relationship between estimated RTK-GNSS accuracy and distance to the canopy border (dist2CB) of submeter accuracy logs within the influence range of canopy proximity.The main, left ordinate shows the estimated accuracy of NMEA logs resp.their bin means.The secondary, right ordinate represents the percentage of logs removed from each bin by Filter 2. The correlation coefficient r was calculated for the filtered NMEA logs dataset as well as for the bin means dataset.Filter setting F1: À6.5 m ≤ dist2CB ≤6.5 m; F2: acc ≤1.0 m; bin size = 0.1 m.Subplot (a): vertical accuracy, acc_v and subplot (b): horizontal accuracy, acc_h.
morphological habitat statistics (CV and mean water depth) is limited.As long as enough valid bins remained in the dataset (>50%), filtering had a moderate effect on mean water depth estimations calculated from MA estimates at the reach scale.The quality of mean depth estimations consistently improved for conservative filter thresholds (retaining at least 90% valid bins).However, filtering strongly impacts CV estimations and by penalizing zones with increased structural diversity, excessive noise filtering has the potential to distort diversity-oriented, hydromorphological habitat assessments.The study's findings confirmed our hypothesis that the influence range of canopy proximity (IRCP) can be determined from correlation analysis between the estimated RTK-GNSS accuracy of NMEA logs and their distance to the canopy border (dist2CB).We determined the IRCP in our study area as the correlationmaximizing dist2CB value, consistently evaluated as IRCP = 6.5 m across different outlier filters and accuracy types.Within the IRCP zone, there were significant correlations between estimated accuracy and dist2CB.Removing the bias of inhomogeneous log sample density further increased the degree of correlation.The study identified noticeable accuracy deterioration when penetrating >1 m into a canopy-covered area, with a higher risk of accuracies >1 m when penetration exceeds 4 m.This finding highlights the importance of considering canopy proximity when estimating water depths with RTK-GNSS technology.4.2 | Wider significanceOur results provide fundamental insights to evaluate the performance and accuracy of RTK-GNSS tracking under different canopy conditions.The study thus has significant implications for ROE campaigns and similar hydromorphological habitat surveys.By combining crosssection tracking with instream habitat mapping, the possibility of multi-tasking was demonstrated.The time dedicated to the tracking of all 21 cross-sections was 4160 nodes =4160 s =1 h 10 min.In comparison, on average it took 30 min for two people to complete a T A B L E 2 Correlation, accuracy, and filtering statistics of the canopy influence analysis for both accuracy types of the bin means dataset, the NMEA logs dataset, and selected filter subsets.
tracking provides high-resolution cross-sectional point data and detects subtle riverbed features not captured by individual wading measurements (c.f.Kinzel et al., 2021).Moreover, only basic technical skills are required for RTK-GNSS handling and potential challenges are less complex compared to the deployment of unoccupied aerial vehicles (UAVs) and point cloud postprocessing.RTK-GNSS tracking can be applied in highly turbid waters without compromising data accuracy.These reasons make RTK-GNSS tracking an interesting alternative to traditional methods and state-of-the-art technologies for a whole range of fields of application, such as streamgaging and bank erosion monitoring.Depending on the survey scale and site conditions, RTK-GNSS tracking might offer the best compromise between resolution, cost, effort, and suitability.For sediment budget analyses and hydromorphological numerical models, RTK-GNSS tracking can provide high-resolution bathymetric input data, especially when also tracking longitudinal profiles.Such grid-based trajectories could also be used as high-resolution input data for bathymetric interpolation algorithms (c.f.Schäppi et al., 2010) and increase the reliability of model results (c.f.Frank et al., 2007).To the best of authors' knowledge, this study quantified for the first time the impact of combined multipathing effects from tree canopy and water surface reflections on estimated RTK-GNSS accuracy.In sparse canopy conditions, RTK-GNSS tracking can deliver reliable results, whereas ALB surveys can be considerably affected by vegetation reflection of the laser signal during the vegetation period and data gaps can be expected for SfM photogrammetry.The advantages of the RTK-GNSS tracking approach come along with observed versus predicted reach-scale R2 values of R2 = 0.77 and RMSE = 0.13 m for the unfiltered dataset and up to R2 = 0.93 and RMSE = 0.07 m for moderate noise filtering but without systematic data correction.In comparison, ALB surveys can obtain R2 values ranging from 0.6 to 0.97, depending on site conditions such as turbidity(Kinzel et al., 2021).Mandlburger et al. (2020) reported deviations >0.06 m for underwater reference targets at water depths >1 m in moderately turbid conditions (Secchi depth ≈1.1 m).SfM photogrammetry can provide a standard deviation of 0.06 m in quasi-clear water conditions(Eltner et al., 2021).
regarding the influence range of alluvial canopy on RTK-GNSS accuracy.The influence range might stretch over several meters around the canopy border and noticeable accuracy deterioration can develop when canopy penetration exceeds 1 m.Cross-section tracking offers substantial efficiency gains in the field, reducing the mapping time by more than 80% compared to manual measurements.RTK-GNSS tracking allows to reliably capture small-scale riverbed features of O (10 À1 m) and provides highresolution point data at cross-sections comparable to remote sensing technology, but without the challenges and effort associated with UAV deployment.For hydromorphological surveys in deep rivers or for large-scale surveys in clear waters without significant canopy cover, vessel or UAV-based remote-sensing will still be more suitable.Beyond hydromorphological habitat assessments, cross-section tracking can also constitute a powerful, robust survey technique for streamgaging applications, bank erosion monitoring, and sediment budget analyses, offering a whole series of advantages in terms of efficiency, cost, resolution, and suitability for challenging environmental conditions.