Developing 3D River Channel Modeling with UAV-Based Point Cloud Data

Highlights

What are the main findings?

• K-nearest neighbor local regression (KLR) reconstructs UAV-based 3D river channels more accurately than LOWESS, with lower errors and better shape preservation across tests and field sites.
• KLR handles uneven point density and missing data well, keeping small bed features without over-smoothing.

What is the implication of the main finding?

• Cross-section delineation and hydraulic modeling for flood risk assessment can be done more accurately.
• Digital-twin river models built from UAV point clouds can be developed more reliably.

Abstract

Accurate characterization of river channel geometry is essential for hydrological and hydraulic analyses, yet the increasing use of unmanned aerial vehicle (UAV) photogrammetry introduces challenges related to uneven point density, shadow-induced data gaps, and spurious outliers. This study proposed a novel approach for reconstructing 3D river channels from UAV-derived point clouds, emphasizing K-nearest neighbor local regression (KLR), and compared it with the LOWESS model. Method performance was examined through controlled simulations of trapezoidal, triangular, and U-shaped synthetic channels, where KLR consistently preserved morphological fidelity and produced lower RMSE than LOWESS, particularly at channel bends and bed undulations, while a neighborhood selection heuristic approach demonstrated robust results across varying data densities. Synthetic channel experiments show that the proposed K-nearest-neighbor local linear regression (KLR) method achieves RMSE values below 0.06 all tested geometries. In contrast, LOWESS produces substantially larger errors, with RMSE values exceeding 0.9 across all channel shapes. Subsequent application to two South Korean field sites reinforced these findings. In the data-scarce Migok-cheon stream, KLR effectively interpolated missing surfaces while maintaining geomorphic realism, whereas LOWESS generated over-smoothed representations. Within the dense Ogsan Bridge dataset, KLR retained small-scale bed features critical for hydraulic simulations and cross-sectional delineation, while LOWESS obscured local variability. Conclusively, the results demonstrate that KLR provides a more reliable and computationally efficient framework for UAV-based 3D river channel reconstruction, with clear implications for hydraulic modeling, flood risk management, and the advancement of digital-twin systems in operational hydrology.

1. Introduction

Accurate representation of river channel geometry is a fundamental requirement in hydrological management, underpinning flood risk assessment, sediment transport studies, and the development of digital-twin systems for rivers [1,2,3]. Traditionally, in South Korea, channel surveys have relied on ground-based methods, such as total stations or RTK-GNSS, which provide high accuracy but are time-consuming, labor-intensive, and often impractical in hazardous or inaccessible environments [4,5]. The emergence of unmanned aerial vehicles (UAVs) combined with structure-from-motion (SfM) photogrammetry has transformed river surveying by enabling rapid acquisition of high-resolution point clouds over large areas with minimal ground control [6,7,8,9]. Such advances have expanded opportunities for hydrological mapping, geomorphic change detection, and hydraulic modeling [10,11].

A growing body of research has specifically investigated the applicability of UAV-SfM techniques to riverine and floodplain environments. Acharya et al. Acharya et al. [12] reviewed UAV applications in hydrology and demonstrated that UAV-based structure-from-motion (SfM) photogrammetry can rapidly acquire centimeter-scale topography across wide floodplain areas, substantially reducing survey time and human exposure in the field. Their review emphasized that UAV-SfM is particularly advantageous in dynamic river systems where frequent surveys are required. Similarly, Schumann et al. Schumann et al. [13] applied UAV-SfM to small river–floodplain environments and reported elevation accuracies comparable to airborne LiDAR, confirming that such low-cost systems can capture geomorphic detail sufficient for hydraulic and morphodynamic modeling. Duró et al. Duró et al. [14] applied UAV-SfM for monitoring bank-erosion processes along a 1.2 km reach of a mid-sized river and obtained sub-decimeter accuracy relative to RTK-GPS and laser-scanning data, showing that UAV photogrammetry can effectively capture morphological changes while reducing field effort. Despite these demonstrated advantages, previous studies consistently report that UAV-derived point clouds are affected by several systematic limitations that directly influence hydraulic applicability. UAV-derived point clouds often contain challenges such as uneven spatial density, occlusions due to vegetation or shadows, and spurious outliers over water surfaces [15,16,17]. Several studies have noted that these artifacts are not merely geometric imperfections but can significantly distort reconstructed river surfaces and extracted cross-sections. These artifacts complicate the reconstruction of hydraulically meaningful surfaces, particularly for cross-section extraction and floodplain delineation. When integrated into hydraulic models, geometric errors can significantly alter discharge estimates and stage–discharge relationships [18,19,20]. Thus, surface reconstruction methods that balance local detail preservation with robustness to noise are essential.

Regression-based techniques have been widely adopted for smoothing and interpolating riverbed elevations. Multiple linear regression (MLR) is simple and computationally efficient but assumes globally linear relationships, making it poorly suited to nonlinear and spatially heterogeneous morphologies [21,22]. Several river-related studies have shown that MLR tends to oversimplify channel geometry, particularly in braided or compound sections where spatial variability is pronounced. Nonparametric approaches such as locally weighted scatterplot smoothing (LOWESS/LOESS) have become popular because they flexibly adapt to local structure using kernel weighting [23,24]. LOWESS has been successfully applied in hydrology for smoothing hydroclimatic time series and environmental data. In river surface reconstruction, LOWESS has been reported to improve surface continuity and reduce noise; however, multiple studies indicate that its smoothing behavior may obscure fine-scale geomorphic features in dense UAV point clouds. However, its tendency to oversmooth in dense data environments may obscure small-scale geomorphic features such as bed undulations and bank breaks, which are critical for hydraulic resistance and flow simulation [1,20].

K-nearest neighbor (KNN) local linear regression (KLR) provided by Lee et al. [25] represents a promising alternative. By fitting local linear models to the k nearest points around each location, KLR preserves geomorphic variability while limiting computation to localized neighborhoods [26,27]. Its theoretical basis aligns with the nearest-neighbor bootstrap methods long used for hydrological resampling and stochastic simulation of hydroclimatic time series [28,29,30]. More recently, KLR-type approaches have been extended to drought forecasting, precipitation downscaling, and flood frequency analysis, highlighting their adaptability to nonlinear and data-limited contexts [31,32,33]. However, existing applications of KLR in river studies have been largely confined to one-dimensional analyses, primarily focusing on cross-sectional smoothing rather than full three-dimensional surface reconstruction. Such regression-based surface reconstruction directly affects hydraulic analyses, as the accuracy of modeled river geometry determines flow simulation and flood prediction performance. The hydraulic implications of accurate surface reconstruction are well documented. Cross-sectional spacing and fidelity strongly influence model stability and flood risk predictions [3,34]. Errors in bathymetric representation propagate into 1D and 2D models, affecting flow conveyance and inundation extents [35,36]. In operational flood forecasting, where real-time UAV deployment is increasingly feasible, efficient and reliable surface reconstruction methods can directly enhance warning systems and emergency planning [37,38].

Although regression techniques such as MLR and LOWESS have been widely applied for riverbed interpolation, most existing studies emphasize one- or two-dimensional representations and provide limited qualitative comparison of their suitability for complex three-dimensional channel morphology. These methods tend to oversmooth local geomorphic variations that are important for hydraulic interpretation and are sensitive to uneven point density and noise in UAV-derived datasets. Consequently, their applicability to complex and spatially heterogeneous channel morphology remains limited, indicating the need for more adaptive surface-reconstruction techniques. This study aims to evaluate and compare KLR and LOWESS for reconstructing river channels from UAV-derived point clouds since the method with 1D was applied in Lee et al. Lee et al. [39,40,41] only for cross-section. The proposed method for the 3D river channel was tested with a simulation case and a field study. The objective of the current study is to focus on providing the development of the KLR for 3D river channels and comparing it with the LOWESS across synthetic and real-world channels, quantifying their respective strengths and weaknesses. The proposed 3D KLR method was validated through both synthetic simulations and field experiments to provide quantitative evidence supporting its performance and applicability. It demonstrates practical implications for hydraulic modeling, cross-section delineation, and digital-twin development, highlighting how method choice affects flood modeling and early-warning systems. By addressing these objectives, this research advances both methodological development and applied hydrology, providing guidance for practitioners seeking robust surface-reconstruction techniques for UAV-based river surveys.

2. Mathematical Description

The point cloud data can be obtained from the UAV photogrammetry. The river channel can be developed with regression-based methods employing the x and y location data as the predictors and the height z as the explanatory variable. Here, a nonparametric regression approach is mainly adopted in the current study, called K-nearest neighbor local regression (KLR). Further description of the KLR model, as well as the base model and the multiple linear regression model, is provided in the following.

2.1. Multiple Linear Regression

A multiple regression model can be used when the relationship between multiple predictors (e.g., x and y) and an explanatory variable (z) is linear. The multiple regression model with two predictor variables can be expressed as

$$z={\beta }_0+{\beta }_{1}x+{\beta }_{2}y+ϵ=\mathbf{x}\mathsf{\beta }+ϵ$$

(1)

where $ϵ$ is considered to be random noise with zero mean and x indicates a variable vector as x = [1, x, y] and $\mathbf{\beta }=[{\beta }_{\mathbf{0}},{\beta }_{\mathbf{1}},{\beta }_{\mathbf{2}}]$. Note that x and y are represented as the location of cloud points, and $z$ is the smoothing elevation. In estimating the parameter vector of β, the least squares method has been mostly applied as

$$\widehat{\mathsf{\beta }}={\left({\mathbf{X}}^{\mathbf{T}}\mathbf{X}\right)}^{-1}{\mathbf{X}}^{\mathit{T}}\mathbf{z}$$

(2)

2.2. KNN-Based Local Linear Regression (KLR)

The current predictor variables of the location data are assumed to be ($X_t$, $Y_t$) with the observed data ($x_i$, $y_i$) for i = 1,…,n that are provided from the selected cloud points. Furthermore, the explanatory variable of the height is denoted as $Z_t$ for the current condition and the observed data $z_i$ for i = 1,…,n. Note that the number of neighbors (k) is also assumed to be known. The predictor of the height $Z_t$ is estimated according to the following steps:

(a) Estimate the distances between the current and observed predictors (i.e., location x and y) for all n observations, as follows:

$$D_j=\sqrt{{\left(x_j-X_t\right)}^2+{\left(y_j-Y_t\right)}^2} j = 1,\dots ,n$$

(3)

(b) Save the location indices for the k smallest distances.

(c) Fit the local weighted linear regression to the observed dataset of the selected location indices [x_(p), y_(p)] and the height z_(p) for p = 1,…,k, where (p) indicates the pth decreasing ordered location index relative to the distance measure in step (a).

(c-1) Calculate the weight matrix using the weight as follows:

$${\mathit{W}}_{KLR}=diag\left[{\displaystyle \frac{1}{\delta }},{\displaystyle \frac{1/2}{\delta }},\dots ,{\displaystyle \frac{1/k}{\delta }}\right]$$

(4)

where $\delta ={\sum }_{p=1}^{k}1/p$.

(c-2) Calculate the feature matrix ${\stackrel{⃡}{\mathbf{X}}}_t$ with the difference of the location predictors as

$${\stackrel{⃡}{\mathbf{X}}}_t=\left(\begin{array}{c}\begin{array}{cc}1& X_t-x_{\left(1\right)} Y_t-y_{\left(1\right)}\\ 1& X_t-x_{\left(2\right)} Y_t-y_{\left(2\right)}\end{array}\\ \begin{array}{cc}& ⋮\\ 1& X_t-x_{\left(k\right)} Y_t-y_{\left(k\right)}\end{array}\end{array}\right)$$

(5)

(c-3) Derive the parameter vector ${\widehat{\mathit{\beta }}}_t^{KLR}$ from the weight matrix W_KLR in Equation (3) with the weighted least square estimator as

$${\widehat{\mathit{\beta }}}_t^{KLR}=({{\stackrel{⃡}{\mathbf{X}}}_t}^T{\mathit{W}}_{KLR}{\stackrel{⃡}{\mathbf{X}}}_t{)}^{-1}{{\stackrel{⃡}{\mathbf{X}}}_t}^T{\mathit{W}}_{KLR}{\mathbf{z}}_{KLR}$$

(6)

where ${\mathbf{z}}_{KLR}$ is the predicted value for the ordered observations $[z_{\left(1\right)},z_{\left(2\right)},\dots ,z_{\left(k\right)}{]}^T$.

(d) Calculate the current predictor as follows:

$$Z_t={\overrightarrow{\mathbf{X}}}_t^T{\widehat{\mathit{\beta }}}_t^{KLR}$$

(7)

where ${\overrightarrow{\mathbf{X}}}_t=(1 X_t Y_t)$.

(e) Repeat steps (a)–(d) until all the required data are simulated.

A heuristic approach for estimating k has been used by $k=\sqrt{n}$ for the selection of k [26,27,28]. In the current study, we tested the multiplier with the base heuristic estimation method:

$$k=a\sqrt{n}$$

(8)

where a is a positive value and is denoted as a multiplier.

2.3. Locally Weighted Scatterplot Smoothing (LOWESS)

LOWESS was devised by Cleveland [23] as a nonparametric and nonlinear regression. The LOWESS, in the current study, consists of two explanatory variables (x_t and y_t, the distance from the base location for x- and y-coordinates) and one predictor variable (z_t, the elevation from the base bottom), which can be described as

$$Z_t=m\left(\right[X_t,Y_t\left]\right)+{\epsilon }_t$$

(9)

where the regression curve m(·) is the conditional expectation $m\left(\right[X_t,Y_t\left]\right)=E\left(Z\right|X=X_t,Y=Y_t)$. The LOWESS estimate can be denoted as

$${\widehat{m}}_{LOWESS}([X_t,Y_t])={\stackrel{⃡}{\mathit{X}}}_t^T{\widehat{\mathit{\beta }}}_t^{LOWESS}$$

(10)

where

$${\widehat{\mathit{\beta }}}_t^{LOWESS}=({{\stackrel{⃡}{\mathbf{X}}}_t}^T{\mathit{W}}_t{\stackrel{⃡}{\mathbf{X}}}_t{)}^{-1}{{\stackrel{⃡}{\mathbf{X}}}_t}^T{\mathit{W}}_t\mathbf{z}$$

(11)

with

$${\stackrel{⃡}{\mathbf{X}}}_t=\left(\begin{array}{c}\begin{array}{cc}1& X_t-x_1 Y_t-y_1\\ 1& X_t-x_2 Y_t-y_2\end{array}\\ \begin{array}{cc}⋮& ⋮\\ 1& X_t-x_n Y_t-y_n\end{array}\end{array}\right)$$

(12)

and

$${\mathit{W}}_t={\mathit{H}}^{-1}diag\left[K_d\left({\mathit{H}}^{-1}\right({\mathit{X}}_t-{\mathit{x}}_1\left)\right),\cdots ,K\left({\mathit{H}}^{-1}\right({\mathit{X}}_t-{\mathit{x}}_n\left)\right)\right]$$

(13)

with the H bandwidth matrix. For the kernel function K, the following is generally employed as

$$K_d(z)=\left\{\begin{array}{c}(1-|z{|}^3{)}^3 \left|z\right|<1\hfill \\ 0 \mathrm{o}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{w}\mathrm{i}\mathrm{s}\mathrm{e}\end{array}\right.$$

(14)

3. Simulation Study with Synthetic Channels

The performance of the model comparison in fitting the point cloud data for river channel smoothing was verified with the simulated point cloud data. The river channels were classified into three shapes: trapezoidal, triangular, and U-shaped. The workflow of the current study is shown in Figure 1.

3.1. Shape of Synthetic Channels

3.1.1. Trapezoidal Channel

A manmade river channel is generally trapezoidal to economically discharge floods and for easy construction [21,22]. In the current study, a trapezoidal channel was designed with a 10 m base width and a 10 m top at both sides as well as a 1:3 side slope with a 10 m height, for a total of 70 m wide. The length of the channel (y-coordinate) was set as 100 m, and its slope along the river course was set as 0.01, as illustrated in the top panel of Figure 2a. The channel base points were placed at 1/4 m intervals for the x and y coordinates, for a total of 281 × 401 points.

3.1.2. Triangular Channel

To further validate the model performance, a triangular channel was synthesized as shown in the middle panel of Figure 2b. Triangular channels are generally located in upper-reach regions with high gradients and strong downward erosion, and they often naturally develop when a riverside slope is cut down and attacked by weathering [42]. Sometimes, this shape of the channel is built to provide an accurate solution for open channel flow measurement [23]. As the trapezoidal channel, there is 10 m of a levee crown part at the top of both sides of the channel. The riverside slope is set as 2:5 (10 m height and 25 m length), and its total length is 70 m, the same as the trapezoidal channel. Also, the channel has a slope of 0.01 along the river. The channel was also separated with a 1/4 m interval for the x and y coordinates, with 281 × 401 points, for a total of 112,681 points.

3.1.3. U-Shaped Channel

U-shaped channels that are close to a natural river were also tested, as shown in the bottom left panel of Figure 2. U-shaped channels are developed in the middle reaches where the stream valleys broaden in a graded condition [42]. In the literature, the U-shaped channels have been modeled with a power function from Neal et al. [43] as

$$w_i=w_F{\left({\displaystyle \frac{h_i}{h_F}}\right)}^{{\displaystyle \frac{1}{s}}}$$

(15)

$$h_i=h_F{\left({\displaystyle \frac{w_i}{w_F}}\right)}^m$$

(16)

where $w_i$ and $h_i$ represent the flow width, while $w_F$ and $h_F$ are the width and height in a bank-full condition, respectively. Also, m is the shape parameter in this channel function. Here, m = 5 was tested in the current study, as used in Neal et al. [43] as a basic value.

The U-shaped bank in the current test was designed as h_F = 5 m and w_F = 20 m, and the levee width was set to 4 m on both sides, and the length along the river was set as 100 m. The base points for this U-shaped channel are located at 0.1 m and 1 m intervals for the x and y coordinates, for a total of 26,462 (262 × 101) points.

3.2. Simulation Methodology

The base cloud point dataset was simulated over each division point, as follows:

$$Z=Y+\epsilon$$

(17)

where Y is the divided points at each channel shape, and $\epsilon ~N(0,{\mathsf{\sigma }}_{\epsilon }^2)$, i.e., normally distributed error with the error variance ${\mathsf{\sigma }}_{\epsilon }^2$. Here, ${\sigma }_{\epsilon }^2$ = 0.2 was presented in the current study following the variability of the case study after testing several values. Further simulations with different error variances were made, and not much different results can be observed from the presented result.

In the right panels of Figure 2, the simulated data points with Equation (17) were illustrated for trapezoidal, triangular, and U-shaped channels at the top, middle, and bottom panels, respectively. The cloud points of the simulated data are scattered well over the synthetic shape channels while illustrating the variability from the location of the base plane, as shown in the left panels of Figure 2. Particularly, 281 × 401 points (=112,681) were simulated for the trapezoidal and triangular synthetic channels, while 26,462 (262 × 101) were simulated for the U-shaped channel, as discussed.

3.3. Simulation Results

3.3.1. Selecting the Multiplier

In order to apply the KLR method, the multiplier ‘a’ must be known as in Equation (8). An experiment to select the appropriate multiplier was performed with different numbers of simulated data. The root mean square error (RMSE) result was calculated with the estimates from the KLR and the base shape data as

$$RMSE=\sqrt{{\displaystyle \frac{1}{\mathrm{N}}}{\sum }_{t=1}^N{\left({\widehat{z}}_j^{KLR}-z_j\right)}^2}$$

(18)

where ${\widehat{z}}_j^{KLR}$ is the KLR estimate from Equation (9) while $Z_z$ shows the base points for the synthetic channel with N, the number of data points, as shown in Figure 2.

At first, different numbers of point cloud data were simulated with Equation (17) as N = 1000, 5000, 10,000, and 50,000 to investigate the performance of the KLR model and to select the multiplier according to the number of simulated points. Figure 3, the KLR estimated surfaces of the trapezoidal shapes with the number of simulated points as N = 1000 and 10,000 are presented.

N = 1000 is tested for the case in which the cloud points from the UAV-based photogrammetry are rather scarce. Due to photographic conditions such as varying lighting conditions, shadows, and background complexity [44], a relatively small number of cloud points can be generated from the photogrammetry. The 1000 points are presented in panel (c) of Figure 3, overlapped with the smoothed panel from the KLR estimate. Due to the scarcity of points, the smoothed surface from the KLR estimated is curved, while the original surface is trapezoidal, as shown in panel (a) of Figure 3. Therefore, some deviations of the KLR surface from the base surface (see the panels (a) and (c)) can be found at the bent area at the bottom and the top, as seen in X = 30 and 40 m with Z = 0 m and X = 10 and 60 m with Z = 0 m. In Figure 3b,d, the larger number of points N = 10,000 (i.e., 10 times larger) was simulated, and its surface is presented. The KLR estimate with N = 10,000 is more accurate than the case of N = 1000, especially at the bent locations.

In Figure 4, the RMSE along with the multiplier ‘a’ in Equation (8) for the number of neighbors (k) is presented at the trapezoidal case. As discussed, four different numbers of simulated points were tested: N = 1000, 5000, 10,000, and 50,000. The RMSE of a lower number of points is higher than that of a higher number of points since the KLR estimate can be more accurate with a higher number of points, as discussed. As shown in Figure 4, the multiplier that results from the smallest RMSE is about 0.5 for all the numbers of simulated points. Therefore, the multiplier a = 0.5 was employed in Equation (8) as $k={\displaystyle \frac{1}{2}}\sqrt{N}$ in the current study. The same multiplier value was considered in the KNNR-based stochastic simulation of streamflow time series [45]. Note that the multiplier can play a role as a tuning parameter to control the smoothness of the surface, and it is adjustable according to the objective of each study. However, it has been noticed that this multiplier is not very sensitive to most cases, meaning that there is no need to estimate the value at each application [26].

3.3.2. Results of the Synthetic Channels

As mentioned, three different synthetic shapes were tested to validate the model performance of the KLR model. The data points employed as the simulated point cloud data are the same as shown in Figure 2 (i.e., N = 112,681, 112,681, and 26,462 for trapezoidal, triangular, and U-shaped, respectively). Furthermore, the LOWESS model was also applied to all three synthetic channels for comparison. Those results are presented in Figure 5, Figure 6 and Figure 7 for the trapezoidal, triangular, and U-shaped channels, respectively.

The synthetic trapezoidal channel shown in panel (a) of Figure 5, with a blue line, was estimated from the simulated cloud points from Equation (9) as in panel (c) of Figure 5. The KLR estimate reproduces the base synthetic trapezoidal shape well. In contrast, the LOWESS estimate over smooths the trapezoidal surface as the bottom part is higher than the base channel, as shown in panel (b) of Figure 5. Subsequently, the points for the bottom part in the 30–50 m of X-coordinate cannot be seen, as in panel (d) of Figure 5. Also, the slope part over 5 m height shows that the LOWESS estimate is lower than the base.

Similar behaviors of the KLR and LOWESS estimates can be observed at the triangular and U-shaped channels in Figure 6 and Figure 7. The bottom part of the LOWESS estimate for the synthetic triangular and U-shaped channels is lower than the base, as in panel (b) of both figures, while the slope portion of the LOWESS estimate is at a higher elevation than the base. Meantime, the KLR estimate of the triangular channel preserves the bent area as well as the straight slope and horizontal top as in panel (a) of Figure 6, except for the slight oversmoothness at the vertex point of the triangular shape (see X = 35 m and Z = 0 m). Furthermore, the KLR estimate of the synthetic U-shaped channel reproduces the U-shaped curvature, as shown in panel (a) of Figure 7. Slight oversmoothness can be observed at the connection part between the U-shape and the crown of the channel, but with a much lower magnitude than the LOWESS estimate, as in panel (b).

Overall,

Table 1. Performance measurements of the KLR and LOWESS methods.

	RMSE		MAE		R2
	KLR	LOWESS	KLR	LOWESS	KLR	LOWESS
Trapezoid	0.02766	1.23205	0.00810	1.01219	0.99998	0.96580
Triangle	0.02710	1.12136	0.00648	0.84962	0.99998	0.97416
U-shaped	0.05118	0.93679	0.02786	0.80662	0.99966	0.88507

summarizes the performance of the KLR and LOWESS methods for three synthetic channel geometries: trapezoidal, triangular, and U-shaped. Across all cases, KLR consistently outperforms LOWESS, exhibiting substantially lower RMSE and MAE values and near-perfect coefficients of determination (R² ≈ 1.0). For the trapezoidal and triangular channels, KLR achieves RMSE values of approximately 0.03, whereas LOWESS produces much larger errors exceeding 0.9. Even for the U-shaped geometry, which represents a more nonlinear and curved channel form, KLR maintains high accuracy (R² = 0.99966), while LOWESS shows noticeably reduced performance (R² = 0.88507). These results indicate that KLR more effectively preserves local geometric features and adapts to nonlinear channel morphology, leading to more accurate surface reconstruction compared to LOWESS.

The simulation study presents that the KLR estimate can be a good alternative to produce a smoothing surface of the channels for different shapes, compared to the LOWESS model. Also, the multiplier of 0.5 for the KLR estimate can be applicable for different numbers of cloud points.

4. Case Study of Field Sites

4.1. Data Acquisition

4.1.1. Specification of Employed UAV

Aerial imagery of the Migok-cheon stream was acquired using the Autel EVO II Dual 640T RTK V3 unmanned aerial vehicle (UAV). The unmanned aerial vehicle employed in this study was the Autel EVO II Dual 640T RTK V3, which integrates an RTK positioning system and a visible-light imaging sensor. The visible camera is equipped with a fixed lens corresponding to a 35 mm-equivalent focal length of approximately 35 mm, enabling stable and geometrically consistent image acquisition for mapping and analysis purposes. This platform is specifically designed for high-precision surveying and is equipped with an advanced real-time kinematic (RTK) positioning module, which applies real-time corrections to satellite signals and achieves centimeter-level spatial accuracy. The flight altitude was set to 75 m, as suggested in previous studies for river modeling [46]. Such positional accuracy is critical in hydrological applications, where even minor horizontal or vertical discrepancies in georeferencing may lead to substantial errors in delineating floodplain boundaries, estimating channel cross-sections, and interpreting stage–discharge relationships. In contrast to conventional GPS-based methods, the RTK system markedly reduces systematic positioning errors and minimizes reliance on extensive ground control networks, thereby improving both efficiency and reliability in data acquisition. Detailed information about the available RTK network sites can be found on the website https://www.gnssdata.or.kr/cors/getCorsView.do (accessed on 22 December 2025).

The UAV also incorporates a 48-megapixel optical sensor capable of capturing images at a resolution of 8000 × 6000 pixels, ensuring fine-scale representation of surface features within the study area. Field operations were executed using the Autel Explorer mission-planning software, which enabled pre-programmed flight paths with controlled altitude, image overlap, and coverage consistency, facilitating accurate photogrammetric processing. The integration of RTK-enhanced navigation and ultra-high-resolution optical imagery produced a robust dataset for spatial analysis of the Migok-cheon watershed. This dataset not only enhanced the precision of hydrological mapping but also established a reliable foundation for subsequent applications in flood early-warning system design and regional flood risk assessment.

4.1.2. Ground Survey and Post-Processing

RTK positioning can reduce the ground control points (GCPs) to validate the simulated cloud points. Only a few of the GCPs were measured over the study area on the ground with GPS surveying. In the current study, EMLID Reach (https://emlid.com/reachrs2/ (accessed on 22 December 2025)), which is a multiband RTK GNSS receiver, was adopted.

The aerial photographs were post-processed to construct a dense point cloud dataset using Pix4Dmapper (Pix4D S.A., Lausanne, Switzerland). Pix4D is a widely applied commercial photogrammetric software that converts overlapping UAV images into georeferenced spatial products, including dense point clouds, digital surface models (DSMs), and orthomosaics. Its advanced structure-from-motion (SfM) and multi-view stereo (MVS) algorithms provide high accuracy in three-dimensional reconstruction, making it well-suited for hydrological and geomorphological applications. In this study, the software was utilized to generate high-resolution datasets necessary for channel cross-section extraction, floodplain characterization, and subsequent hydrological analysis.

4.2. Scarce Data Area in Migok-Cheon

4.2.1. Study Area

The research site is situated along the Migok-cheon, a tributary running through Jinju-si in South Korea, as illustrated in Figure 8. Migok-cheon extends about 8.8 km in length and drains a catchment of roughly 13.9 km². The stream gradient varies between 1/50 and 1/400, while the specific study reach has an average slope of 1/350. At its downstream end, Migok-cheon merges with the Hwanggang River, which subsequently flows into the Nakdong River—one of the nation’s four major rivers. Consequently, the hydrology of Migok-cheon is strongly influenced by the stage levels of both the Hwanggang and Nakdong Rivers.

The Hapcheon Dam is situated midstream on the Hwanggang River, serving both hydropower generation and water supply functions. The upper reaches of the Hapcheon River are surrounded by mountainous terrain with steep gradients, which results in rapid runoff, short concentration times, and sudden flood events. A notable case occurred in August 2020, when large discharges from the Hapcheon Dam significantly raised water levels in the Hwanggang River. This elevated stage caused several tributaries, including Migok-cheon, to overflow [46]. To mitigate flood impacts in the Migok-cheon basin, the implementation of a flood early-warning system has been proposed. Developing such a system requires detailed channel cross-section data to determine the threshold water levels that should trigger alerts.

4.2.2. Result

The study site, Migok-cheon Stream, is illustrated in Figure 8. This area experienced severe flooding in 2020, primarily due to backwater effects from the Hwanggang River, whose flow regime is influenced by the Hapcheon Dam, shown in the middle-left panel of Figure 8 and the according point cloud dataset generated though Pix4D is shown in Figure 9a. The overlay figure of the 3D point cloud data and river channel imagery is shown in Figure 10. During this event, the residential zone on the right bank was placed at considerable risk, as not all residents were able to evacuate safely. In response, the local government-initiated plans to enhance flood resilience in the area. A detailed field investigation was subsequently conducted to provide essential baseline data for flood risk management.

Certain regions were missing in the point cloud, likely due to unfavorable UAV camera angles or shadowing effects. These data gaps were effectively reconstructed using the proposed KLR method, which produced a continuous, smoothed surface. The frontal view Figure 9d reveals some spurious points on the lower right side, located at approximately 40–50 m along the x-axis and 16–17 m in height. These erroneous points generated localized peaks of 19–20 m; however, their influence was spatially limited by surrounding data. Aside from this anomaly, the KLR model successfully reproduced the river channel with high fidelity.

For comparison, Figure 11 presents the test site surface modeled with LOWESS. The resulting surfaces in Figure 11b,c appear excessively smoothed, obscuring local undulations and geomorphic features. As illustrated in Figure 11d, this smoothing effect led to both underestimation and overestimation in critical areas. Nonetheless, in the data-deficient right-bank section, the LOWESS model provided a more stable representation than the KLR model, as it avoided abrupt discontinuities.

The comparison between Figure 8 and Figure 9 shows that KLR more accurately represents the local topography of the Migok-cheon reach. In Figure 8, KLR preserves small-scale bed variations and realistic gradients, reconstructing the channel surface even in areas where data gaps occurred. Although a few spurious points appear at 40–50 m along the x-axis and 16–17 m in height, their spatial influence is minimal. By contrast, Figure 9 indicates that LOWESS produces smoother and more visually continuous surfaces but obscures real undulations and geomorphic transitions near the banks. This excessive smoothing causes under- and over-estimation in certain areas, reducing the representation of micro-topographic variability that affects flow conveyance. Therefore, while LOWESS provides slightly greater continuity in the right-bank region with sparse data, KLR offers a more faithful reproduction of the actual riverbed form, which is critical for hydraulic interpretation of the site.

4.3. Dense Data Site in Ogsan Bridge

4.3.1. Study Area

The study site is located along the Yeongcheon River, a tributary within the Nam River basin in Jinju, South Korea, as depicted in Figure 12. The Yeongcheon River flows through predominantly agricultural and peri-urban landscapes before discharging into the Nam River, which ultimately contributes to the Nakdong River, one of the country’s principal river systems. The selected reach is situated in proximity to the Ogsan Bridge, marked in the figure, and represents a section where flood risk and sediment transport dynamics are of particular concern. The river section extends approximately 9.5 km, with a contributing catchment area of about 15.6 km². Channel gradients in the Yeongcheon River vary from 1/60 to 1/500, with the investigated test area exhibiting an average slope of nearly 1/400.

The Nam River is regulated by the Namgang Dam, which plays a critical role in flood control, municipal water supply, and irrigation for the Jinju region. During heavy rainfall events, controlled discharges from the Namgang Dam can significantly alter the stage levels of the Nam River, thereby influencing backwater effects in tributaries such as the Yeongcheon River. Historical flood records indicate that local overbank flooding has repeatedly occurred near Ogsan Bridge, particularly during typhoon-induced rainfall events. The hydrological behavior of this tributary is thus highly sensitive to both direct precipitation and regulated discharges from upstream reservoirs.

In August 2020, a series of extreme precipitation events in the Nakdong basin resulted in elevated flows in the Nam River. Subsequent backwater effects propagated into the Yeongcheon River, inundating agricultural fields and infrastructure adjacent to the test site. To reduce the vulnerability of this basin, improved hydrological monitoring and early-warning systems have been recommended. Developing such systems requires high-resolution cross-sectional data of the channel, as shown in the selected test area, to identify threshold water levels for flood warnings and to enhance hydraulic modeling accuracy.

4.3.2. Result of Ogsan Bridge

The overlay figure of the 3D point cloud data and river channel imagery is shown in Figure 13 while the point cloud data captured for the Ogsan Bridge are presented in Figure 14a, and The dataset consists of densely distributed points, accompanied by some erratic outliers both above and below the river channel. Owing to the high density of the point cloud, the surface reconstructed using the KLR model is well covered by these points, as illustrated in the panels of Figure 14c,d. The outliers, located at elevations of approximately 28–32 m above the riverbanks and 18–20 m below the riverbed, exert only a minimal influence on the modeled surface. Consequently, the KLR model generates a realistic river channel characterized by small bumps and peaks, particularly along the riverbed. Such fine-scale reproduction enhances the accuracy of cross-sectional delineation, which is advantageous for physical flow simulations and digital-twin modeling of a river.

Figure 15 illustrates the river channel reconstructed with the LOWESS model, shown prominently in panel (b). Unlike the KLR model, the LOWESS-based surface remains discernible beneath the overlapping point cloud, revealing its tendency to over-smooth the river channel. While the LOWESS approach offers the benefit of smoothly interpolating across gaps shown in the Migok stream case, this advantage is less relevant in the context of dense point cloud data in this Ogsan Bridge case. In contrast, the KLR model is better suited to such cases, as it preserves fine topographic features while maintaining computational efficiency by utilizing only the K nearest neighboring points at each calculation step.

5. Discussion

The comparative evaluation of KLR and LOWESS through both simulation and field applications highlights several key findings regarding their applicability to UAV-derived river surface reconstruction. In the synthetic channel experiments with trapezoidal, triangular, and U-shaped geometries, KLR consistently maintained morphological fidelity and achieved lower RMSE than LOWESS. The advantage of KLR was particularly evident at channel bends and bed undulations, where preserving local variability is essential for hydraulic interpretation. These results confirm that the localized fitting strategy of KLR more effectively captures spatial heterogeneity than the globally weighted smoothing employed by LOWESS.

Field validations further reinforced the performance consistency of KLR under contrasting data conditions. In the data-scarce Migok-cheon basin, where photogrammetric limitations produced gaps and irregular sampling, KLR successfully reconstructed missing areas while maintaining geomorphological realism. Although LOWESS yielded smoother continuity in such regions, it simultaneously suppressed critical small-scale variability. Conversely, in the dense Ogsan Bridge dataset, KLR effectively retained fine-scale features, such as subtle bed bumps and peaks that are hydraulically significant for modeling flow resistance and delineating cross-sections. These findings demonstrate that KLR adapts to varying point densities without overfitting or excessive smoothing.

From an applied hydrological perspective, the outcomes indicate that reliable surface reconstruction directly enhances flood modeling, river monitoring, and digital-twin development. Realistic representation of cross-sectional variability improves the accuracy of hydraulic simulations and flood risk assessments, which depend heavily on detailed bed geometry. Furthermore, the local-neighborhood computation framework of KLR offers higher efficiency than global regression approaches, making it suitable for large-scale UAV datasets that require rapid and robust processing.

In relation to previous studies, the present work extends the 1D cross-section application of KLR proposed by Lee et al. [39] to full three-dimensional surface reconstruction. This expansion enables continuous spatial modeling across the entire channel domain rather than isolated transects, addressing a key limitation of earlier regression-based methods. The comparative analysis with LOWESS also provides quantitative insight into the trade-offs between local detail preservation and smoothness, which had not been explicitly evaluated in prior research.

Overall, the discussion affirms that the 3D KLR framework effectively balances detail preservation, computational efficiency, and robustness to uneven data distribution. These characteristics make it a practical tool for operational hydrology and geomorphic modeling. Future work may focus on optimizing neighborhood-size selection, integrating multi-temporal UAV datasets for dynamic surface monitoring, and extending the method toward automated processing pipelines for digital-twin systems.

6. Summary and Conclusions

In the current study, regression-based approaches for reconstructing river channels from UAV-derived point cloud data were developed and evaluated, focusing on KLR and LOWESS. Through both controlled simulation experiments and real-world case studies, the research demonstrated the advantages and trade-offs of each method under varying data density and geomorphic conditions. By systematically comparing synthetic trapezoidal, triangular, and U-shaped channels, the KLR method was found to produce surface reconstructions that closely adhered to the underlying channel morphology, while the LOWESS method consistently introduced oversmoothing effects that diminished critical features.

The identification of an effective heuristic for selecting the number of neighbors (k) was found from the current study. Across multiple simulations with varying point densities, a multiplier of 0.5 applied to the heuristic approach yielded consistently low RMSE values and stable reconstructions. This parameter tuning has practical importance, as it enables hydrologists to achieve further accurate results without excessive trial and error. Additionally, the results also demonstrated that KLR adapts flexibly to both data-scarce and data-dense environments, making it suitable for diverse UAV-based surveying conditions.

Field applications provided further validation of the proposed methodology. In the Migok-cheon basin, where photogrammetric limitations produced data gaps, KLR effectively reconstructed missing surfaces while maintaining geomorphological realism. Although LOWESS offered smoother continuity in these scarce-data areas, it did so at the expense of local variability, which is crucial for accurate hydraulic modeling. In contrast, the dense Ogsan Bridge dataset highlighted the strengths of KLR in capturing fine-scale features, such as bumps and peaks in the riverbed, that directly influence flow resistance and cross-sectional delineation. Here, LOWESS obscured important details, reinforcing the conclusion that its utility is limited when high-resolution point clouds are available.

From an applied perspective, the findings of this study have significant implications for flood modeling, river monitoring, and digital-twin development. The ability of KLR to reproduce cross-sectional variability with high fidelity enhances the accuracy of hydraulic simulations, which rely heavily on realistic channel geometry. Furthermore, the computational efficiency of KLR, achieved through local neighborhood selection rather than global fitting, makes it well-suited for large-scale applications where high-resolution UAV data must be processed more quickly and reliably. These advantages consider KLR as a valuable tool for operational hydrology, including early-warning systems, infrastructure risk assessment, and adaptive flood management strategies.

In conclusion, this research establishes KLR as a superior regression-based approach for reconstructing river channels from UAV-derived point cloud data. By preserving local detail, maintaining computational efficiency, and adapting across varying data densities, KLR addresses key challenges in hydrological modeling and geomorphological analysis. The outcomes of this work not only advance methodological understanding but also provide practical guidelines for implementing UAV-based surface reconstruction in both research and applied water-resource management. Future studies should aim to refine parameter selection strategies, extend applications to dynamic monitoring, and explore integration with machine learning frameworks to further enhance the scalability and versatility of this approach.