A Scanline-Based Sliding Window Filtering Method for UAV-Borne LiDAR Bathymetry Point Clouds

Highlights

What are the main findings?

• A novel scanline-based sliding window filtering method is proposed for the denoising of UAV-borne LiDAR bathymetric point clouds, which can effectively separate noise while completely retaining detailed features of complex terrain in shallow-water areas.
• The method achieves ≥96% noise recall and F1-score ≥ 0.9 across different terrains, with excellent filtering performance and strong adaptability, significantly improving the quality of point cloud data.

What are the implications of the main findings?

• This study innovatively applies bathymetric LiDAR scanline information to point cloud filtering, providing a new paradigm for UAV-borne LiDAR bathymetry data processing.
• The proposed method is of great reference significance for improving the data quality of UAV-borne LiDAR bathymetry and can effectively promote the application of this technology in complex shallow-water areas.

Abstract

To improve the data quality of underwater point clouds acquired by UAV-borne LiDAR bathymetry, a scanline-based sliding window filtering method is proposed based on an analysis of scanline data characteristics. Scanline data of underwater point clouds are first extracted from raw point clouds, and the radius outlier removal algorithm is employed to eliminate outliers. Taking the acquisition time of scanline points as the X-axis and elevation as the Y-axis, a 3D problem is simplified into a 2D representation, and a sliding window is constructed along the scanline. Robust least-squares fitting is applied within the window. The median absolute deviation of the fitting residuals is adopted to calculate the terrain feature values for quantifying the terrain complexity, followed by an adaptive filtering threshold determination according to terrain feature values. Fine filtering of the individual scanlines is performed using a point-by-point sliding window. Experimental results demonstrate that the proposed method is adaptable to various terrain conditions, achieving a noise recall rate ≥ 96%, an overall filtering accuracy ≥99%, and an F1-score ≥ 0.9. Particularly, the precision rate in flat-water areas reached 97.37%. Overall, the proposed filtering method effectively separates noise points while preserving detailed terrain features and supports UAV-borne LiDAR bathymetry for mapping complex shallow-water regions.

1. Introduction

Airborne LiDAR bathymetry (ALB) has been widely applied to topographic mapping in shallow-water areas, such as coastal zones and island reefs, owing to its advantages of high precision, high efficiency, strong maneuverability, and safety [1,2,3]. However, owing to factors such as water scattering, absorption, and environmental interference, ALB data often contain significant noise. The presence of noise not only affects the terrain recognition and accuracy [4] but also increases the time and complexity of data processing. With the development of unmanned aerial vehicle (UAV) technology, UAV-borne LiDAR bathymetry (UAV-ALB) systems have emerged. Compared with manned airborne LiDAR bathymetry systems, UAV-ALB systems offer advantages in flight flexibility and environmental adaptability, as well as lower cost and simpler operation, making them more suitable for fine topographic surveying of nearshore shallow-water areas [5,6,7]. However, significant variations in nearshore water quality increase the number of water-column noise points in UAV-ALB data, with complex underlying causes.

Researchers have conducted extensive studies on filtering methods for point cloud data. Existing filtering algorithms can generally be divided into four categories: mathematical morphology, density-based spatial clustering, surface-based methods, and progressive densification algorithms [8,9,10]. Mathematical morphological filtering employs structuring elements to perform opening and closing operations on a point cloud, and its filtering performance is highly sensitive to the window size (structuring element). Various researchers have improved the selection of window size [11,12,13]. For example, Hui et al. [12] proposed an improved morphological algorithm based on multi-level kriging interpolation. This method uses progressive morphological operations to initially separate ground points and constructs a reference terrain surface using multi-level kriging interpolation. Introducing a terrain slope gradient to adjust the elevation difference threshold effectively reduced omission errors for abrupt terrain features. DBSCAN (Density-Based Spatial Clustering of Applications with Noise) is a typical density-based spatial clustering algorithm. The filtering performance is strongly influenced by relevant parameters. Researchers have proposed adaptive parameter calculation methods based on various data characteristics [14,15,16,17,18,19]. Chen et al. [19] proposed a segment-adaptive photon filtering method based on the local density distribution of seabed signal photons and seabed terrain slope, achieving high-precision bathymetry in complex nearshore environments.

Surface filtering algorithms generate a DEM by constructing an initial terrain surface and iteratively filtering noise. Representative algorithms include the cloth simulation filter (CSF) proposed by Zhang et al. [20]. Cai et al. [21] improved the algorithm by incorporating a morphological closing operation, which eliminates the necessity of manually setting the cloth rigidity parameters and adopts a terrain-adaptive elevation difference threshold to extract ground points. Hu et al. [22] proposed a filtering method based on strip curve fitting, targeting the linear scanning characteristics of airborne LiDAR. The original point cloud was divided into strips and projected from a 3D to 2D space. Ground and non-ground points were then separated through curve fitting and iterative point screening. Progressive densification algorithms include progressive triangulation filtering and moving surface fitting. Progressive TIN densification yields a favorable overall filtering performance but is sensitive to parameter settings and exhibits poor filtering performance in steep-slope areas [23,24]. The moving surface fitting method exhibits strong stability and good fitting performance at low sampling rates but is prone to oversmoothing in regions with sharp terrain features [25]. Subsequent studies proposed several improvements to enhance the adaptability of these algorithms to complex terrain [26,27,28]. Yan et al. [26] proposed a ground filtering algorithm that combines slope information with plane fitting. The algorithm used a two-level grid method to select ground seed points in a stepwise manner. In each first-level grid, plane fitting was performed using the least-squares (LS) method, and ground seed points were used to construct a ground model. The algorithm was optimized to obtain the desired filtering performance and terrain feature preservation.

The aforementioned filtering methods were primarily designed for land LiDAR point clouds. However, ALB data noise mainly arises from water-related factors, and the distribution characteristics of this noise are significantly different from those of land LiDAR point clouds, making the above filtering methods unsuitable for direct application to ALB point cloud filtering. Therefore, these methods must be improved and adapted to better satisfy the filtering requirements of ALB point clouds. To demonstrate the potential of ALB technology for mapping large underwater archeology areas, Doneus et al. [29] used a robust interpolation technique with eccentric and asymmetric weight functions to filter non-terrain points from the point cloud. Yang et al. [30] addressed the overfiltering problem in convex and concave seabed terrains by constructing a bidirectional cloth simulation correction model and optimizing the model parameters. Su et al. [31] proposed a joint ALB point cloud filtering algorithm that considers the water surface, the water column, and the bottom. First, an opposing CSF was used to remove water-surface noise. Subsequently, a statistical outlier removal filter was used to remove water-column outliers. Finally, a moving trend surface model was constructed and combined with the elevation difference threshold to achieve smooth filtering of the bottom data. The above method uses spatially complex 3D point clouds for model construction, which inevitably results in some loss of detail when dealing with highly variable and complex underwater terrain. Furthermore, it is disadvantageous in terms of its high time complexity, high parameter sensitivity, and insufficient scene robustness.

To address these limitations, inspired by the robust interpolation and hierarchical filtering ideas proposed by Kraus and Pfeiffer [32,33], this study proposes a scanline-based sliding window filtering method for UAV-ALB point clouds. The proposed method combines scanline analysis with adaptive sliding window filtering. Compared with the aforementioned methods, the method proposed in this paper can not only effectively remove noise but also better preserve the underwater topographic features. In complex terrain scenarios, it does not require repeated parameter adjustment, offering high computational efficiency, and strong applicability.

2. Methods

When using UAV-ALB systems for land–water topographic mapping, the shapes of the scanlines and point cloud densities vary with different terrains. The mainstream scanning modes of LiDAR bathymetry systems are divided into circular, elliptical, and linear [34]. Taking the oval mode as an example, as shown in Figure 1a, during UAV flight the scanning mirror rotates once, and the trajectory of the laser footprints forms an ordered scanline. A point cloud can be processed into groups by numbering the scanlines to establish unique indices. The scanlines are plotted as a time–elevation two-dimensional (2D) curve, as shown in Figure 1b, and closely resemble topographic profiles.

2.1. Workflow of the Scanline-Based Sliding Window Filtering Method

The overall workflow of the scanline-based sliding window filtering algorithm proposed in this study is presented in Figure 2.

The detailed steps are as follows:

• Read the initial UAV-ALB data according to the scanline number, extract the underwater point clouds, and remove outliers using radius outlier removal (ROR) algorithm to obtain the initial terrain points.
• Based on the results of the previous step, convert the 3D point cloud data into 2D data, define and initialize a sliding window, and perform polynomial fitting within the window using the IGG III robust least-squares (RLS) method. The residual v_i of each fitting point and the median absolute deviation M_AD are then calculated and retained.
• Determine the terrain complexity based on M_AD, obtain the feature value s, and determine the filtering threshold T. If the absolute residual $\left|v_c\right|$ of the window center point exceeds T, the center point is classified as a non-terrain point; otherwise, it is classified as a terrain point. Only the window center point is processed each time to reduce misclassification at abrupt terrain changes and prevent continuous overfiltering.
• Slide the window point by point along the current scanline and repeat the fitting process in steps 2 and 3 until all initial terrain points on the scanline have been processed and the fine filtering of the single scanline is completed. Finally, all scanlines are processed sequentially according to their scanline numbers, and the above procedure is repeated to complete global point cloud filtering, after which all classified point clouds are merged to produce the final results.

2.2. Terrain and Noise Characteristics

UAV-borne LiDAR bathymetry used for nearshore topographic surveying covers land–water interface zones, flat underwater areas, undulating underwater areas, and deep-water areas. Using the measured data, this section analyzes the scanline characteristics and point cloud distribution patterns of typical terrains in these areas.

(1)

Land–water interface

The land–water interface can be divided into two types: continuous and discontinuous, as shown in Figure 3. Figure 3a shows the continuous type, where the elevation of data points from different classes at the interface changes gradually; additionally the underwater data are continuous—without notable data truncation, and with few near-terrain noise points. Figure 3b shows the discontinuous type, where islands and reefs exist in the area, interrupting the continuity of the underwater data. In addition, steep terrain leads to the loss of data points on the vertical topographic surfaces, and the elevation of data points changes abruptly at the interface. The data contain numerous outliers, clustered noise points, and near-terrain noise points.

Figure 2. Workflow of the proposed method.

Figure 2. Workflow of the proposed method.

(2)

Flat underwater areas

Low-turbidity water exhibits reduced absorption and scattering of laser pulses. On flat terrain, the laser can penetrate the water column uniformly and reflect stably from the bottom. The point cloud distribution is shown in Figure 4. The elevation of the terrain points varies slightly, and the points are continuously distributed—without notable local data gaps; however, numerous clustered noise points with elevated protrusions, outliers, and individual near-terrain noise points are present.

Figure 3. Land–water interface areas. (a) Continuous; (b) discontinuous.

Figure 3. Land–water interface areas. (a) Continuous; (b) discontinuous.

Figure 4. Flat underwater areas.

Figure 4. Flat underwater areas.

(3)

Undulating underwater areas

Some shallow-water areas feature complex terrain, such as undulations, potholes, and reefs, as shown in Figure 5. In Figure 5a, the terrain exhibits slight undulations, the terrain points are continuously distributed, and the noise points are sparse, with only a few clustered noise points accompanied by elevated protrusions. In Figure 5b, the point cloud distribution at the bottom of the pothole is unevenly distributed and contains data gaps with dense noise points. In Figure 5c,d, areas of abrupt terrain change exhibit sparse point clouds, local data discontinuities, pronounced data gaps along scanlines, and numerous clustered noise points.

(4)

Deep-water areas

In deep-water areas, the absorption and scattering of laser pulses by the water column become more pronounced, causing some laser signals to fail to reach the bottom, as shown in Figure 6. As a result, the available topographic data are limited, and numerous noise points are present.

By analyzing the terrain data using scanlines, the terrain complexity can be divided into three types: (1) flat: the terrain is flat with almost no undulations, and the data exhibit similar depth values; (2) slight undulations: the terrain changes gently with minor convexities and concavities, and the data show relatively consistent depth values; and (3) complex: this terrain is characterized by abrupt terrain changes, including reefs, islands, and potholes, where the point cloud is sparse and exhibits local gaps. Noise in underwater data can be classified into three categories: (1) numerous outliers suspended in the water column, whose dispersion degree is related to the water quality; (2) small-scale clustered noise points far from the terrain, often accompanied by elevation protrusions originating from fish schools and secondary echoes from complex seabeds; and (3) a small number of near-terrain noise points, which are easily confused with terrain points resulting from coral reef scattering errors and edge effects caused by scanning angles.

To address the filtering problem in UAV-ALB point clouds, a coarse-to-fine joint filtering method based on scanlines and sliding windows is proposed: (1) an ROR method is used for preprocessing to remove outliers; and (2) RLS polynomial fitting is performed within sliding windows. Based on the fitting results, the terrain types are classified, and an adaptive filtering threshold is designed to simultaneously separate far- and near-terrain noise points, thereby achieving flexible filtering for different terrains.

2.3. Radius Outlier Removal (ROR)

The ROR method is commonly used to remove outliers from point cloud data. By setting a search radius, the type of point is determined by checking whether the number of neighboring points within the radius meets a predefined threshold. If the number of neighboring points within the specified radius is less than the threshold, the point is identified as an outlier. The filtering principle is illustrated in Figure 7. Assuming that the search radius is set to d and the point-count threshold is k = 1, the yellow point is identified as an outlier. If k = 2, both the yellow and green points are identified as outliers, whereas the blue point is considered a valid point.

The basic steps for applying ROR filtering to 3D scanline data are as follows:

• Set the search radius r and neighbor point-count threshold k for ROR filtering.
• For the underwater point cloud data, read scanlines individually according to their scanline numbers, extract the 3D (X_i, Y_i, Z_i) coordinates of the underwater points, and perform the filtering process.

  a.

  Topology construction: A k-d tree method was used to construct a spatial topological structure for the scanline point cloud.

  b.

  Noise point identification: Each underwater point on the scanline was traversed, and a k-nearest neighbor search was performed in a 3D space. Points with fewer than k neighbors within the search radius r were identified as outliers and removed; otherwise, they were retained as terrain points.

Compared with the global ROR filtering of point clouds, the dispersion of noise points in a single scanline is higher, resulting in improved filtering performance. In addition, the volume of topological data is smaller, leading to reduced computational complexity.

2.4. Robust Polynomial Fitting Based on Sliding Windows

2.4.1. Sliding Window

This study uses scanlines and performs local terrain fitting using sliding windows to improve the ability of the algorithm to preserve detailed terrain features. The sliding window method is illustrated in Figure 8. Starting from the beginning of the scanline, the window slides point by point towards the end with a fixed length, and terrain fitting is performed on the time-elevation data within the window. Each underwater point on the scanline is processed once as the center of the window, ensuring the continuity and completeness of the terrain fitting results. During the sliding process, if discontinuities occur in the local data (i.e., when the time interval between adjacent underwater points exceeds a predefined threshold t), the data are segmented to avoid excessive window spans that could lead to fitting failure.

2.4.2. Robust Least Squares Fitting

A polynomial model is used to fit the 2D data within the window. The observed data are (x_i, y_i) (i = 1, 2, …, n, where n is the number of underwater points within the window), and x_i and y_i represent the time and elevation values, respectively, of the underwater points. The optimal polynomial function is constructed based on the LS criterion, as shown in Equation (1).

$$\widehat{y}=l_0+l_{1}x+l_2{x}^2+\cdots +l_m{x}^m(m\le n)$$

(1)

In Equation (1), $\hat{\mathit{y}}$ represents the fitted elevation, m is the polynomial degree, and l is the parameter to be determined. The fitting error equation can be expressed as

$$\mathit{V}=\mathit{A}\hat{\mathit{X}}-\mathit{L}=\left(\begin{array}{c}{a}_1\\ a_2\\ ⋮\\ a_n\end{array}\right)\hat{\mathit{X}}-\left(\begin{array}{c}{L}_1\\ L_2\\ ⋮\\ L_n\end{array}\right)$$

(2)

where V is the residual vector between the fitted and observed values; A is the coefficient matrix; a_i is the i-th row of A; $\hat{\mathit{X}}$ is the parameter to be determined; and L is the observation vector.

According to robust estimation theory, the objective function to be minimized is expressed as

$${\Phi }_{\min }={\displaystyle \sum _{i=1}^n{P}_i{v}_i^2}$$

(3)

where P_i and v_i denote the weight and residual of the i-th observation, respectively. Based on the principle of M-estimation, the extreme value function is constructed as follows:

$$V_{\min }={\displaystyle {\displaystyle \sum _{i=1}^n}{P}_i\rho (v_i)}={\displaystyle {\displaystyle \sum _{i=1}^n}{P}_i\rho (a_i\hat{\mathit{X}}-L_i)}$$

(4)

where ρ is a convex function of the residual v. Based on equivalent weight theory, the robust M-estimate of the parameter vector can be obtained as

$${\widehat{X}}^k={\left(A^{\mathrm{T}}{\overline{P}}^{k}A\right)}^{-1}{A}^{\mathrm{T}}{\overline{P}}^{k}L$$

(5)

where ${\overline{\mathit{P}}}^{\mathit{k}}$ is the equivalent weight matrix after k iterations and can be determined using the IGG III weighting function [35,36]:

$${\overline{P}}_i^k=\left\{\begin{array}{ll}1.0,\hfill & \left|{\tilde{v}}_i^k\right|<k_0\hfill \\ {\displaystyle \frac{k_0}{\left|{\tilde{v}}_i^k\right|}}{({\displaystyle \frac{k_1-\left|{\tilde{v}}_i^k\right|}{k_1-k_0}})}^2,\hfill & k_0\le \left|{\tilde{v}}_i^k\right|\le k_1\hfill \\ 0,\hfill & \left|{\tilde{v}}_i^k\right|>k_1\hfill \end{array}\right.$$

(6)

where ${\tilde{\mathit{v}}}_{\mathit{i}}^{\mathit{k}}$ is the standardized residual of the i-th observation after k iterations; ${\tilde{\mathit{v}}}_{\mathit{i}}^{\mathit{k}}={\mathit{v}}_{\mathit{i}}^{\mathit{k}}/{\mathit{\sigma }}_{{\mathit{v}}_{\mathit{i}}}$ and ${\mathit{\sigma }}_{{\mathit{v}}_{\mathit{i}}}$ are the standard deviations of v_i. Usually, k₀ is chosen within the range 1.0–2.5 (signal region), and k₁ is chosen within the range 3.0–8.0 (gross error region).

2.5. Adaptive Threshold Setting

After the RLS fitting of the data within the window, the fitting residuals v_i for each point are calculated. The median absolute deviation of the residuals, denoted as M_AD, is calculated as shown in Equation (7).

$$M_{\mathrm{AD}}=\mathrm{median}(\left|v_i-\mathrm{median}\left(v\right)\right|)$$

(7)

Considering the bathymetric accuracy σ of UAV-ALB (the Mapper-20kU system used in this study has a manufacturer-specified bathymetric accuracy of 0.2 m; therefore, σ = 0.2 m), a terrain feature value s is introduced to define the terrain-adaptive threshold, as shown in Equation (8).

$$T=s\times \sigma$$

(8)

M_AD is used to classify terrain complexity and determine the parameter s. A smaller M_AD indicates a good fit over flat terrain, requiring a strict threshold with a small s. A larger M_AD indicates a poor fit over complex terrain, requiring a relaxed threshold with a larger s to mitigate overfiltering and preserve fine features. The specific rules are defined in Equation (9).

$$\left\{\begin{array}{ll}{M}_{\mathrm{AD}}<0.05,\hfill & s=1\hfill \\ 0.05<M_{\mathrm{AD}}<0.1,\hfill & s=2\hfill \\ M_{\mathrm{AD}}>0.1,\hfill & s=3\hfill \end{array}\right.$$

(9)

where the M_AD segmentation threshold is determined based on the distribution characteristics of M_AD values in the measured data, and the value of s is selected from the experimentally optimized results.

This method dynamically adjusts the filtering intensity, reducing the overfiltering and underfiltering caused by fixed thresholds, and improving the robustness of filtering in complex shallow-water environments.

3. Experimental Data

To evaluate the filtering performance of the algorithm, experiments were conducted using data from Wuhua Lake in Jiuzhaigou County, Sichuan Province, and Feicui Lake in Suzhou, Jiangsu Province. The data were acquired using the UAV-ALB system Mapper20KU (parameter indicators are shown in

Table 1. Design parameters of the Mapper-20KU system.

Parameters	Indicators
Laser pulse frequency	≥20 kHz
Laser wavelength	532 nm & 1064 nm
Scanning mode	Elliptical scanning
Scanning speed	1200 rpm
Field of view (FOV)	40°
Data acquisition mode	Full-Waveform sampling
Maximum detection depth	2 × SD (at flight height 75 m)
Bathymetric accuracy	0.2 m
Weight	≤4.2 kg

), developed by the Shanghai Institute of Optics and Fine Mechanics, Chinese Academy of Sciences. The data acquisition in this study was carried out using the DJI FlyCart 30 multi-rotor UAV (DJI, Shenzhen, China). The flight altitude was 80 m and the flight speed was 10 m/s. The acquired underwater point cloud density was 28 pts/m², with an average point distance of approximately 0.19 m. The average distance between adjacent flight paths was 25 m, and the average overlap ratio of adjacent strips was 50%.

The experimental datasets are shown in Figure 9. Figure 9a,b show the flight trajectories over Wuhua Lake and Feicui Lake, respectively. Figure 9c,d show the digital surface model (DSM) generated from the raw point clouds of the two lakes. Affected by water quality conditions, the Wuhua Lake dataset contains numerous far-terrain clustered noise points and a small number of near-terrain noise points. Consequently, the underwater terrain in the directly generated DSM is almost completely obscured. Feicui Lake is a mine-pit lake with complex terrain and large undulations. Its dataset contains relatively few noise points, among which some are near-terrain noise points.

4. Results

The algorithm proposed in this study was applied to filter data from the two survey areas. First, ROR was performed on the underwater point cloud. To remove outliers in the scanline while fully preserving valid terrain points, the optimal parameters were set to r = 1 and k = 1. Subsequently, fine filtering was conducted. Considering the data characteristics, the key parameters were set to n = 17 and m = 5. These parameter settings are elaborated in Section 5. Finally, the filtering results are evaluated both qualitatively and quantitatively, as presented below. The DEM grid resolution in Figure 9 and Figure 10 is 0.2 m, which is determined according to the measured underwater point cloud density (28 pts/m²) and an average point distance (0.19 m). It can effectively balance the trade-off between the interpolation insufficiency and excessive smoothing of terrain details, thereby ensuring DEM accuracy.

4.1. Qualitative Evaluation

As shown in Figure 10, compared with the raw data, the filtering algorithm proposed in this study achieves a significant filtering effect: it effectively removes numerous noise points while preserving detailed terrain features, and accurately restoring the underwater terrain. Only a small number of noise points remain in data edges, undulating areas, and sparse point cloud areas, with a minor loss of sparse terrain points. Figure 11 and Figure 12 show detailed local views.

There are numerous submerged logs in Wuhua Lake, some lying on the lake bottom and others near the water surface. Part of the point cloud is sparse and exhibits elevated protrusions. Consequently, these points were inevitably misclassified as noise by the algorithm, as shown in Figure 11g,h. The misclassified points can subsequently be corrected manually. If the research objective is to obtain only basic terrain data, no further processing of the filtering results is required. In summary, from a qualitative perspective, the algorithm proposed in this study demonstrates strong filtering performance, and the DEM generated from the filtered terrain point cloud is of reliable quality.

4.2. Quantitative Evaluation

(1)

Evaluation metrics

The filtering results were quantitatively evaluated using statistical metrics. Before the experiment, the data were manually classified in detail, with noise and terrain points manually labeled as the ground truth. The adaptability and reliability of the algorithm were evaluated using four metrics: noise recall (R), precision (P), F1-score (F1) and overall accuracy (OA). R reflects the ability of the algorithm to identify noise points, P measures the reliability of noise detection, F1 provides a balanced evaluation by integrating recall and precision, avoiding the bias caused by a single metric, and OA reflects the overall classification accuracy. The calculation formulae for each quality evaluation metric are listed in

Table 2. Calculation of quality evaluation metrics. TP and TN represent the number of points correctly identified as noise points and terrain points, respectively, and FP and FN represent the number of points incorrectly identified as noise points and terrain points, respectively.

Category	Algorithm Recognition		R	P	F1	OA
	Noise Points	Terrain Points
Noise points	TP	FN	TP/(TP + FN)	TP/(TP + FP)	2PR/(P + R)	(TP + TN)/(TP + FN + FP + TN)
Terrain points	FP	TN

.

(2)

Quality evaluation

Two datasets with uniformly distributed point clouds and different terrain complexities were extracted from the Wuhua Lake and Feicui Lake datasets, denoted as Data 1 and Data 2, respectively. Data 1 contains several small, gently undulating pits with a uniform terrain point cloud density and a large number of noise points, most of which are far-terrain noise points. Data 2 contains continuous raised terrain with a relatively dense terrain point cloud and fewer noise points. Some noise points are confused with terrain points, increasing the difficulty of filtering. The joint filtering algorithm proposed in this study was applied to the datasets, and the noise points were effectively separated (Figure 13). The two-stage filtering results were statistically analyzed, and the filtering quality was evaluated, with the results shown in

Table 3. Quality evaluation results.

Area	Stage	TP	FN	FP	TN	R/%	P/%	F1	OA/%
Data 1	1	1771	6345	3	165,347	21.82	99.83	0.358	96.34
	2	8072	44	218	165,132	99.46	97.37	0.984	99.85
Data 2	1	1315	1997	59	167,058	39.70	95.71	0.561	98.79
	2	3206	106	510	166,607	96.80	86.28	0.912	99.64

.

As shown in Table 3, the preprocessing stage effectively removed outliers while preserving terrain points. The filtering performance is mainly related to the type and dispersion of noise points. After the fine-filtering stage, the noise recall and F1-score improve significantly. In Data 1, both the noise recall and overall accuracy exceed 99%, the precision rate is 97.37%, and the F1-score reaches 0.984, indicating a high filtering performance. In Data 2, affected by the terrain complexity, all performance metrics are slightly lower than those in Data 1. Nevertheless, the noise recall and F1-score still reached 96.8% and 0.912, respectively. Notably, the precision rate is only 86.28%. This is mainly because, in complex terrain areas with an uneven point cloud density, it is difficult for the algorithm to accurately distinguish between noise and terrain points. A small number of terrain points were lost during filtering, which did not significantly affect the overall quality of DEM construction.

4.3. Comparison with Another Method

To further demonstrate the superiority of the proposed method, we continued to use Data 1 and Data 2 as samples for further evaluation. Under the same processor, the proposed filtering algorithm was compared with the SVB joint-filtering algorithm proposed by Su et al. [31]. The filtering results are shown in

Table 4. Performance comparison between our algorithm and the SVB joint-filtering algorithm.

Data	Filtering Method	R/%	P/%	F1	OA/%	Time
Data 1	SVB	97.04	91.1	0.940	99.42	4 min 12 s
	our	99.46	97.37	0.984	99.85	7.355 s
Data 2	SVB	87.02	74.14	0.801	99.16	4 min 56 s
	our	96.80	86.28	0.912	99.64	7.196 s

.

As shown in Table 4, the proposed method outperforms the SVB algorithm in all metrics across both scenarios, demonstrating excellent filtering performance. Specifically, in Data 1—which features a flat terrain and substantial water-column noise—the SVB algorithm can achieve satisfactory filtering results. However, its performance significantly degrades in Data 2, characterized by an undulating terrain and abundant near-terrain noise, resulting in persistent near-terrain noise and the loss of valid terrain points. This is primarily because the SVB algorithm performs filtering based on surface fitting after constructing a grid, resulting in some loss of the terrain detail and efficiency. It is only suitable for conventional 3D point cloud processing. Instead, our method processes data in a scanline-by-scanline manner and offers advantages in terrain detail preservation and computational efficiency. However, this requires access to the original scanline information of the raw point cloud.

5. Discussion

5.1. Analysis of Fitting Results

Based on the scanlines, the complex UAV-ALB 3D point cloud data were converted into 2D data, simplifying their spatial distribution characteristics. The RLS algorithm was then used to perform the fitting, forming the core component of the filtering algorithm proposed in this study. To more clearly demonstrate the fitting accuracy of the RLS algorithm under different terrains conditions, Figure 14 shows the 2D scanline fitting results for several terrains after ROR filtering. All fitting results adopt the default parameters: n = 17, m = 5. The curves were formed by sequentially connecting the fitted values of the window-center points obtained during the sliding fitting process, where yellow rectangles indicate areas with large deviations in the fitting curve.

As shown in Figure 14, when the terrain points are continuously distributed, the fitting curve accurately reproduces the shape of the 2D scanline in both flat and undulating terrain areas, demonstrating robust fitting performance. Slight smoothing occurs only in areas with elevated protrusions, sharp terrain features, and sparse point clouds, as shown in Figure 14c,d. The fitting results demonstrate that the proposed method achieves robust fitting under diverse terrain conditions. It can effectively suppress noise interference and preserve topographic details, exhibiting excellent terrain adaptability.

Figure 15 shows the absolute residuals of window-center points for local terrain in Data 1 after fitting using the traditional LS method and the proposed method. The red line represents the absolute residual after fitting using the LS method, and the blue line represents the residual obtained using the proposed method. As shown in Figure 15b,c, when no noise points are present, the fitting results of the LS method and the proposed method are similar, with residual values less than 0.05 m in flat areas and greater than 0.2 m in complex areas. Figure 15a shows distant terrain noise points. In this case, the LS fitting method is affected by noise points, and the terrain-point residuals fluctuate abnormally, failing to reflect the local terrain. In contrast, the proposed fitting method effectively suppresses noise point interference, and the terrain-point residuals remain stable within a reasonable range, further verifying the reliability of the proposed fitting method.

5.2. Influence of Filtering Parameters

5.2.1. ROR Parameters

The parameter settings for ROR filtering directly affect the effectiveness of outlier removal, and the filtering results are sensitive to these parameters. Improper parameter settings can easily lead to insufficient filtering or overfiltering. To analyze the impact of parameter settings on the filtering performance, a strip point cloud from the Wuhua Lake dataset was selected as the test sample. By adjusting the parameters, we compared the filtering accuracy (P) and F1-score (F1) under different parameter combinations. The results are shown in Figure 16.

As shown in Figure 16a, as r increases from 0.6 m to 1.5 m, the precision for all values of k rises and stabilizes when r exceeds 1 m. For a given value of r, the precision rate is highest when k = 1, which eliminates isolated outliers while effectively preserving sparse terrain points. As shown in Figure 16b, the F1-score declines for all k values as r increases. Although the F1-score is high when k = 2 or 3, the precision is the lowest, resulting in the loss of a significant number of sparse terrain points. By integrating the two evaluation metrics and considering the characteristics of low local point cloud density and high parameter sensitivity of the dataset, the optimal parameter combination is determined as k = 1 and r = 1. Filtering under these parameters reduces the interference of noise on the accuracy of subsequent polynomial fitting while avoiding a decrease in the overall filtering accuracy due to the loss of terrain points.

5.2.2. Polynomial Fitting Parameters

Polynomial fitting is a core steps in subsequent filtering processing, and its fitting performance is primarily determined by the window size n and polynomial order m. To analyze parameter sensitivity and determine the optimal configuration, comparative experiments were conducted using Data 1 as the sample dataset. The filtering results under various combinations of n and m are listed in

Table 5. Performance comparison under different parameter combinations.

m	n	TP	FN	FP	TN	R/%	P/%	F1	OA/%	Time/s
3	13	8068	48	201	165,149	99.41	99.88	0.985	99.86	5.574
	17	8072	44	218	165,132	99.46	99.87	0.984	99.85	5.751
	21	8073	43	267	165,083	99.47	99.84	0.981	99.82	6.004
5	13	8068	48	201	165,149	99.41	99.88	0.985	99.86	6.582
	17	8072	44	218	165,132	99.46	99.87	0.984	99.85	7.073
	21	8073	43	267	165,083	99.47	99.84	0.981	99.82	7.381
7	13	8068	48	201	165,149	99.41	99.88	0.985	99.86	8.099
	17	8072	44	218	165,132	99.46	99.87	0.984	99.85	8.533
	21	8073	43	267	165,083	99.47	99.84	0.981	99.82	8.937

.

As shown in Table 5, all accuracy metrics achieve satisfactory performance across the different parameter combinations. Within a reasonable range, these metrics show minimal variation as n and m change. Specifically, when m is fixed, increasing n from 13 to 21 results in a marginal improvement in recall, but a gradual decline in other accuracy metrics, accompanied by a moderate rise in computational time. When n is fixed, varying m from 3 to 7 produces no notable changes in any of the metrics, with only the computational time rising.

Balancing the filtering accuracy and computational efficiency, we selected n = 17 and m = 5 as the default optimal parameters for the algorithm. Under these parameters, the filtering precision reached 99.87%, and the F1-score was as high as 0.984. This setting not only maintains a high filtering performance and avoids the computational redundancy associated with high polynomial orders and excessively large windows, but also accommodates locally complex terrain and effectively prevents underfitting. Furthermore, combined with the fitting results from Section 5.1—where scanlines over varied terrains were processed using identical parameters—it can be concluded that the proposed method exhibits low parameter sensitivity during the fitting stage. When processing terrain with varying complexity, there is no need for repeated parameter adjustments, demonstrating excellent applicability and robustness.

5.3. Influence of Flight Parameters and Water Conditions

The density and quality of the UAV-ALB point clouds are mainly governed by flight parameters and water conditions, which directly affect filtering performance.

A lower flight altitude and higher flight speed lead to a higher point cloud density, improving the fitting accuracy and terrain detail preservation. Nevertheless, excessive data redundancy will reduce the computational efficiency. Conversely, a higher altitude and lower speed lead to sparse point clouds, degrading the fitting reliability and filtering effect.

The water conditions are a critical factor. Turbid water intensifies laser absorption and scattering, reduces the terrain points density, and generates substantial complex noise, thereby increasing the filtering difficulty. Additionally, increasing the water depth limits the number of acquired point clouds and may cause local data gaps, impairing the filtering accuracy. By contrast, nearshore and clear water provide uniformly distributed point clouds with distinct noise characteristics, allowing the proposed algorithm to maintain high accuracy and stability.

5.4. Limitations

Compared with multi-beam and single-beam measurement systems, UAV-ALB systems can achieve a very high measurement efficiency in shallow-water areas (water depth within 10 m). However, the terrain in such shallow-water areas is complex, and a single filtering method can result in the loss of a large number of terrain feature points. Therefore, the filtering process requires a comprehensive consideration of both terrain complexity and detail preservation. Although the proposed algorithm performs well in most scenarios, it still exhibits certain limitations in specific local areas, such as abrupt terrain areas, areas with non-terrain objects, and areas with high-density noise. These cases are discussed in detail below.

5.4.1. Abrupt Terrain Areas

In areas with abrupt terrain changes, the algorithm is prone to misclassifying abrupt points during filtering, leading to the loss of a small number of terrain point clouds and residual near-terrain noise, as shown in Figure 17. The more drastic the terrain change, the more easily point clouds are lost, potentially resulting in missing local terrain features during DEM construction and a loss of the computational efficiency. Conversely, if the underwater terrain changes are gradual, the algorithm performs better and is faster.

5.4.2. Non-Terrain Objects

When non-terrain objects such as trees and aquatic plants that protrude significantly from the underwater terrain exist, their point clouds are sparse and exhibit similar statistical characteristics to noise, making them prone to being misclassified as noise by the algorithm. However, such point clouds are not terrain data in nature, and their loss does not constitute a defect in algorithm performance, as shown in Figure 11. If it is necessary to retain information on these objects for research purposes, they can be manually restored in the corresponding regions.

5.4.3. High-Density Noise

Additionally, if the density of locally clustered noise in a single scanline is excessively high, it will interfere with the fitting process, making it difficult to remove them effectively, which may result in the loss of terrain points, as illustrated in Figure 18. However, such noise rarely occurs in real data and has little impact on the overall quality of the DEM, hardly affecting practical applications.

Overall, the above limitations mainly stem from the complexity of underwater terrain rather than fundamental defects in the algorithm design. In future work, effective solutions will be explored to address these limitations and further improve the robustness of the algorithm in complex scenarios.

6. Conclusions

To address the complex spatial distribution of noise points in UAV-ALB underwater data, lack of dedicated filtering methods, and unsatisfactory filtering performance in shallow water with complex terrain, a point cloud filtering method based on scanlines and sliding windows is proposed to effectively separate various types of noise points from underwater point clouds. By combining scanline-based ROR filtering with adaptive-threshold sliding fitting, the method effectively filters noise while preserving detailed terrain features. The results show that the proposed method achieves a noise recall rate ≥ 96% and an F1-score ≥ 0.9 in two water terrains with different levels of complexity. Compared with the SVB joint-filtering algorithm, our method achieves better filtering performance as well as higher processing speed, exhibiting advantages of high precision and high efficiency simultaneously. Furthermore, the fitting performance remains stable across different terrain types, and the residual distribution before and after fitting are reasonable, further demonstrating the applicability and robustness of the method for fitting various terrain data. Thus, the method effectively separates noise points from complex UAV-ALB underwater data while preserving detailed terrain features.