Optimized UAV-LiDAR Workflows for Fine-Scale Stream Network Mapping in Low-Gradient Wetlands
A Kushiro Wetland, Japan Case Study
Waruth POJSILAPACHAI (Ph.D. Student, English Engineering Education Program (e3), Hokkaido University)
Prof. Tomohito YAMADA (Supervisor, Faculty of Engineering, Hokkaido University, Japan)
Dr. Takehiko Ito (Co-Author, Faculty of Engineering, Hokkaido University, Japan)
Created: October 15, 2025 | Modified: November 12, 2025
2 Channel Network Optimization Analysis
2.1 Abstract
Accurate stream network delineation in low-gradient wetlands is essential for hydrological modeling, flood risk assessment, and ecological restoration. However, the subtle terrain features and dense vegetation in these environments present significant challenges. This study systematically evaluated 48 UAV-LiDAR processing workflows to identify the optimal approach for mapping fine-scale stream channels in the Kushiro Wetland, Japan, a Ramsar-protected site known for its ecological importance. Workflows combined three ground filtering methods Progressive Morphological Filter (PMF), Cloth Simulation Filter (CSF), Multiscale Curvature Classification (MCC), four interpolation techniques Inverse Distance Weighting (IDW), Triangulated Irregular Network (TIN), Kriging (KRG), Multilevel B-spline Approximation (MBA), two sink-filling algorithms, and two flow direction models D8, D-infinity. Performance was assessed using the Intersection over Union (IoU) metric to quantify the accuracy of channel network delineation.
The results showed that workflow configuration significantly impacts detection precision, with the optimal workflow—CSF, MBA, Planchon, and D8—achieving a high IoU of 0.85. CSF excelled at preserving complex terrain structures crucial for wetland hydrology. While KRG provided robust interpolation for general terrain representation, MBA was more effective for channel delineation within the optimal workflow. Planchon’s sink-filling algorithm substantially improved hydrological connectivity representation, outperforming Wang & Liu. Minimal differences were observed between D8 and D-infinity flow direction models, suggesting D8’s computational efficiency makes it preferable for similar environments.
These findings provide actionable recommendations for high-resolution wetland mapping and hydrological analysis. The methodological framework developed in this study supports the ongoing Kushiro Wetland Restoration Project and can be applied to other degraded wetland systems globally, contributing to conservation, restoration planning, and ecosystem management efforts.
Keywords: high-resolution mapping; multiscale curvature; classification; wetland hydrology; channelization impacts; Kushiro Wetland
Published: 06 November 2025
Conference: The 9th International Electronic Conference
on Water Sciences (ECWS 2025)
Session: Remote Sensing, Artificial Intelligence and
New Technologies in Water Sciences
Academic Editor: Nikiforos Samarinas
DOI: https://sciforum.net/paper/view/26367
2.2 INTRODUCTION
Wetlands are vital ecosystems that provide a wide range of ecosystem services, including flood mitigation, water storage, and the maintenance of diverse biological communities (Gardner & Finlayson, 2018). Accurate delineation of stream networks is essential for hydrological modeling, flood risk assessment, and effective ecosystem management. These needs are particularly evident in the Kushiro Wetland, a Ramsar site in Japan, where channelization has disrupted hydrologic connectivity and threatened biodiversity (Nakamura et al., 2014). To address these impacts, a restoration project was undertaken between 2006 and 2011 to re-establish the river’s original meandering course. For such restoration efforts, understanding the original or target hydrological regime is critical. High-resolution digital terrain models (DTMs) derived from filtered LiDAR data enable the planning of re-meandering or re-wetting interventions. The capacity to map fine-scale channels directly supports these objectives.
Global datasets such as HydroRIVERS (Lehner et al., 2013) remain valuable for macro-scale analyses (Birkel et al., 2021), including global river connectivity assessments (Grill et al., 2019). However, for site-specific restoration projects or dynamic river systems (Kondolf et al., 2014), supplementary data—such as field surveys or higher-resolution remote sensing (Jones et al., 2021)—are often required to validate and refine HydroRIVERS outputs. While high-resolution LiDAR and depression-filling techniques offer promising solutions, their application in wetland environments remains underexplored (Rapinel et al., 2011). One key limitation is the difficulty of mapping inundated channels beneath dense vegetation, where LiDAR signals are obstructed (Hooshyar et al., 2015). Ground surface accuracy depends heavily on data resolution, and advanced filtering algorithms are needed to improve ground point extraction in vegetated areas, thereby enhancing the precision of ditch network mapping (Rapinel et al., 2011).
This study addresses a critical gap in current methodological frameworks: while previous research has focused on individual algorithm comparison or used arbitrary “ground truth” datasets, we employ a dual validation approach combining comprehensive pairwise statistical analysis with independent satellite imagery validation. This methodology reveals that workflows achieving high inter-method consensus may still exhibit systematic biases when compared to observed channel features, a finding with significant implications for operational remote sensing applications in challenging environments.
2.3 METHOD
2.3.1 Study Area
The Kushiro Wetland, Japan’s largest wetland, is a critical habitat for wildlife and ecosystem services such as flood control, water purification, and carbon storage (Gardner & Finlayson, 2018). However, human activities, including agricultural development, river channelization, and sediment intrusion, have severely degraded its ecological and hydrological functions over the past century (Nakamura et al., 2014). In the 1970s-1980s, meandering rivers in the northern wetland were straightened to protect farmland, accelerating sediment/nutrient transport into the wetland and altering its hydrology (Nakamura et al., 2014). This led to drier conditions, with reed-sedge marshes replaced by alder forests, disrupting the peatland ecosystem. In response, the Japanese government launched the Kushiro Wetland Restoration Project in the 2000s to restore natural hydrology and vegetation.
Figure 2.3.1: Shaded Relief Map of Hokkaido
Figure 2.3.2: Shaded Relief Map of the Kushiro Wetland
2.3.2 Data Acquisition
UAV-LiDAR data were collected using a DJI Matrice M300 mounted with DJI L1 (LiDAR camera) from two combined areas. The master area covered 6.62 km² with a memory size of 1.7 GB, while the study area was cropped and extended to 0.91 km², with a density of 4.76 points/m² and 4.76 pulses/m².
Figure 2.3.3: Cropped Area of Interest in green
Figure 2.3.4: Lowest Z Point clouds
2.4 METHODOLOGY
This study employed a comprehensive dual-validation comparative analysis framework to evaluate 48 distinct channel network extraction processing chains. The workflows systematically combined:
- Ground filtering methods: Progressive Morphological Filter (PMF) (K. Zhang et al., 2003), Cloth Simulation Filter (CSF) (W. Zhang et al., 2016), and Multiscale Curvature Classification (MCC) (Evans & Hudak, 2007)
- Interpolation techniques: Inverse Distance Weighting (IDW) (Shepard, 1968), Triangulated Irregular Network (TIN) (De Floriani & Magillo, 2009), Kriging (KRG) (Matheron, 1963), and Multilevel B-spline Approximation (MBA) (Lee et al., 1997)
- Sink-fill algorithms (Planchon & Darboux, 2001; Wang & Liu, 2006)
- Flow routing methods: D8 (O’Callaghan & Mark, 1984) and D-infinity (Tarboton, 1997)
Primary Analysis: Pairwise similarity analysis was conducted across all 1,128 possible combinations using multiple performance metrics including Intersection over Union (IoU), Dice coefficient, F1-score, precision, recall, and channel length ratios, with comprehensive confusion matrix calculations for each comparison. Performance differences across processing components were systematically evaluated through comparative analysis of IoU distributions, examining median values, variability measures (standard deviation, coefficient of variation), and inter-quartile ranges to identify consistent performance patterns. Component-wise comparison employed boxplot visualization and summary statistics to assess the magnitude and consistency of performance differences across workflow combinations.
Independent Validation: To address limitations inherent in purely statistical inter-method comparison, extracted channel networks were visually validated against Sentinel-2 true color composite imagery (10m resolution) acquired during a temporally proximate period to the UAV-LiDAR survey. Since the UAV survey utilized only the DJI L1 LiDAR sensor without concurrent high-resolution optical imagery acquisition, freely available satellite imagery provided the most practical independent reference for assessing channel network accuracy. This validation enabled assessment of absolute accuracy in representing observable channel features, independent of inter-workflow agreement patterns. Visual comparison focused on spatial correspondence between extracted networks and visible channel signatures in the satellite imagery, including channel centerline alignment, network connectivity, and preservation of low-order tributaries.
Advanced visualization included boxplot distributions and comprehensive summary statistics with coefficient of variation analysis to assess method stability and variability across all processing combinations, ensuring reproducible research standards through systematic logging, data validation, and complete result preservation.
2.5 RESULTS
This study presents a systematic comparative analysis of channel network extraction methodologies, evaluating 48 distinct processing chain combinations through 1,128 pairwise comparisons, with independent validation against satellite imagery.
Cluster 1: PMF
Figure 2.5.1: PMF Raster
Cluster 2: CSF
Figure 2.5.2: CSF Raster
Cluster 3: MCC
Figure 2.5.3: MCC Raster
2.6 Pixel-Based IoU Analysis
The Intersection over Union (IoU), also known as the Jaccard Index, serves as the primary quantitative metric for evaluating channel network extraction accuracy across all 1,128 pairwise workflow combinations. IoU is calculated as the ratio of the intersection (pixels classified as channels by both workflows) to the union (pixels classified as channels by either workflow), expressed mathematically as:
\[\text{IoU} = \frac{TP}{TP + FP + FN}\]
where TP represents true positives (agreement on channel presence), FP represents false positives (pixels identified as channels by workflow 1 but not workflow 2), and FN represents false negatives (pixels identified as channels by workflow 2 but not workflow 1). This metric ranges from 0 (no overlap) to 1 (perfect agreement), providing a balanced assessment that penalizes both over-extraction and under-extraction of channel networks. Unlike simpler metrics such as overall accuracy, IoU is particularly robust for imbalanced datasets where channel pixels constitute a small fraction of the total study area, making it well-suited for evaluating sparse hydrological networks in wetland environments. The complementary metrics of Precision (TP/(TP+FP)), Recall (TP/(TP+FN)), and F1-Score (harmonic mean of Precision and Recall) provide additional perspectives on workflow performance, revealing whether discrepancies arise primarily from over-prediction, under-prediction, or balanced errors across the channel network.
2.7 Statistical Analysis by Component
2.7.1 Summary Statistics Table
The statistical analysis reveals distinct performance patterns across workflow components. Among ground filtering methods, PMF (K. Zhang et al., 2003) achieved the highest median IoU (0.900), while CSF (W. Zhang et al., 2016) and MCC (Evans & Hudak, 2007) showed comparable performance (0.854-0.857). Interpolation methods (IDW Shepard, 1968; TIN De Floriani & Magillo, 2009; KRG Matheron, 1963; MBA Lee et al., 1997) exhibited remarkably similar performance (0.856-0.876), suggesting interpolation choice has minimal impact based on inter-method consensus. Sink-fill algorithms showed the largest performance gap, with Wang & Liu (Wang & Liu, 2006) (0.890) outperforming Planchon & Darboux (Planchon & Darboux, 2001) (0.846) in pairwise comparisons. Flow direction methods (D8 O’Callaghan & Mark, 1984; D-infinity Tarboton, 1997) performed nearly identically (0.860 vs 0.859).
| Component | Option | Median IoU | Mean IoU | SD | Q25 | Q75 | N |
|---|---|---|---|---|---|---|---|
| Ground Filter | CSF | 0.854 | 0.875 | 0.051 | 0.844 | 0.890 | 632 |
| Ground Filter | MCC | 0.857 | 0.878 | 0.049 | 0.845 | 0.899 | 376 |
| Ground Filter | PMF | 0.900 | 0.917 | 0.044 | 0.884 | 0.930 | 120 |
| Interpolation | IDW | 0.876 | 0.883 | 0.050 | 0.844 | 0.908 | 354 |
| Interpolation | KRG | 0.859 | 0.876 | 0.051 | 0.844 | 0.896 | 306 |
| Interpolation | MBA | 0.857 | 0.877 | 0.051 | 0.846 | 0.890 | 258 |
| Interpolation | TIN | 0.856 | 0.888 | 0.054 | 0.851 | 0.900 | 210 |
| Sink-Fill | Planchon | 0.846 | 0.860 | 0.044 | 0.837 | 0.859 | 588 |
| Sink-Fill | Wang | 0.890 | 0.903 | 0.050 | 0.873 | 0.930 | 540 |
| Flow Direction | D8 | 0.860 | 0.883 | 0.053 | 0.845 | 0.900 | 576 |
| Flow Direction | Dinf | 0.859 | 0.878 | 0.049 | 0.844 | 0.899 | 552 |
2.8 Visualization of Component Performance
This visualization presents a comprehensive comparison of channel network extraction performance across four key processing components using Intersection over Union (IoU) as the primary metric. The boxplots reveal distinct performance patterns:
Ground Filter Methods show Progressive Morphological Filter PMF, (K. Zhang et al., 2003) achieving the highest median IoU (0.900) with minimal variability, substantially outperforming Cloth Simulation Filter CSF, (W. Zhang et al., 2016) (0.854) and Multiscale Curvature Classification MCC, (Evans & Hudak, 2007) (0.857), both of which exhibit greater spread in their distributions.
Interpolation Methods demonstrate remarkably similar performance across all four techniques IDW (Shepard, 1968), TIN (De Floriani & Magillo, 2009), Kriging (Matheron, 1963), MBA (Lee et al., 1997), with median IoU values clustering tightly between 0.856-0.876, suggesting interpolation choice has minimal impact on final channel network accuracy.
Sink-Fill Algorithms reveal a stark contrast, with the Wang & Liu algorithm (Wang & Liu, 2006) (median IoU 0.890) showing consistent high performance and tight distribution, while the Planchon & Darboux algorithm (Planchon & Darboux, 2001) produces notably lower accuracy (0.846) with increased variability.
Flow Direction Methods indicate D8 (O’Callaghan & Mark, 1984) and D-infinity (Tarboton, 1997) perform nearly identically (medians 0.860 and 0.859 respectively), with comparable distributions, suggesting flow routing algorithm selection has negligible effect on extraction outcomes.
The red diamonds marking median values and the relatively narrow interquartile ranges across most components indicate stable, reproducible performance, with ground filtering (Evans & Hudak, 2007; K. Zhang et al., 2003; W. Zhang et al., 2016) and depression handling (Planchon & Darboux, 2001; Wang & Liu, 2006) emerging as the most influential processing steps in determining overall channel network quality.
2.9 Optimal Workflow Identification
| Workflow | Median IoU | Mean IoU | SD | Min IoU | Max IoU | N Comparisons |
|---|---|---|---|---|---|---|
| PMF_TIN_Planchon_Dinf | 1.0000 | 1.0000 | 0.0000 | 1.0000 | 1.0000 | 2 |
| PMF_TIN_Wang_D8 | 1.0000 | 1.0000 | NA | 1.0000 | 1.0000 | 1 |
| MCC_TIN_Wang_D8 | 0.9299 | 0.9335 | 0.0464 | 0.8894 | 1.0000 | 17 |
| MCC_TIN_Wang_Dinf | 0.9144 | 0.9294 | 0.0445 | 0.8894 | 0.9993 | 16 |
| CSF_IDW_Wang_D8 | 0.9050 | 0.9025 | 0.0522 | 0.8368 | 1.0000 | 45 |
| PMF_IDW_Planchon_D8 | 0.9050 | 0.9249 | 0.0425 | 0.8839 | 1.0000 | 15 |
| PMF_IDW_Planchon_Dinf | 0.9050 | 0.9195 | 0.0385 | 0.8839 | 1.0000 | 14 |
| PMF_IDW_Wang_D8 | 0.9050 | 0.9134 | 0.0320 | 0.8839 | 1.0000 | 13 |
| PMF_IDW_Wang_Dinf | 0.9050 | 0.9061 | 0.0195 | 0.8839 | 0.9295 | 12 |
| MCC_IDW_Wang_D8 | 0.9039 | 0.9102 | 0.0500 | 0.8383 | 1.0000 | 29 |
This table presents the top 10 channel extraction workflows ranked by their median Intersection over Union (IoU) scores based on pairwise statistical comparisons across all 1,128 workflow combinations. The results reveal that workflows combining PMF ground filtering with TIN interpolation achieved perfect agreement (IoU = 1.0000, SD = 0.0000) when paired with either Planchon or Wang sink-filling algorithms, indicating these workflows produce nearly identical outputs that represent the statistical consensus across methods. However, the low number of comparisons (N = 1-2) for these top-ranked workflows suggests they have limited diversity in pairwise evaluations. In contrast, workflows like CSF_IDW_Wang_D8 (median IoU = 0.9050, N = 45 comparisons) and MCC_IDW_Wang_D8 (median IoU = 0.9039, N = 29 comparisons) show more robust performance across a broader range of methodological comparisons, though with slightly lower agreement scores and higher variability (SD = 0.0522 and 0.0500, respectively). Notably, while these rankings identify workflows with highest inter-method consensus, they do not necessarily indicate the most accurate representation of actual channel features—as demonstrated in Section 5.11 where the validated workflow (CSF_MBA_Planchon_D8) provides superior correspondence with satellite imagery despite not appearing in this top 10 statistical ranking, underscoring the critical distinction between methodological agreement and ground-truth accuracy.
2.10 Best Individual Workflow
The statistical analysis identified the workflow combining PMF (K. Zhang et al., 2003), Kriging interpolation (Matheron, 1963), Wang & Liu sink-filling (Wang & Liu, 2006), and D8 flow direction (O’Callaghan & Mark, 1984) as achieving the highest inter-method consensus. However, as detailed in Section 2.17, independent validation against satellite imagery revealed that this statistical optimum does not necessarily represent the most accurate workflow for representing actual channel features in low-gradient wetland environments.
| Selection Criterion | Workflow | Mean IoU | SD IoU |
|---|---|---|---|
| Highest Mean IoU | PMF-TIN-Planchon-Dinf | 1 | 0 |
| Most Robust (Lowest SD) | PMF-TIN-Planchon-Dinf | 1 | 0 |
The perfect inter-method agreement (IoU = 1.0000, SD = 0.0000) indicates that this workflow produces channel networks that are most representative of the central tendency across all 48 methodological variations. This high consensus suggests methodological robustness—the PMF-TIN-Planchon-Dinf combination generates results that are minimally sensitive to the specific algorithmic choices made by other workflows.
However, statistical consensus does not guarantee accuracy. A workflow could achieve high agreement by consistently over-extracting or under-extracting channels relative to ground truth. The distinction between inter-method reliability and ground-truth accuracy becomes particularly critical in low-relief wetland environments where subtle topographic variations challenge all automated extraction methods. Section 2.17 presents independent validation using high-resolution satellite imagery, revealing important discrepancies between statistical optimality and ecological reality, thereby highlighting the necessity of multi-source validation in wetland hydrological analysis.
2.11 Consensus Analysis
| Workflow Component | Best Option (Statistical) | Median IoU | Improvement vs. Worst |
|---|---|---|---|
| Ground Filter | PMF | 0.9000 | 5.3% |
| Interpolation | IDW | 0.8764 | 2.3% |
| Sink-Fill | Wang | 0.8897 | 5.2% |
| Flow Direction | D8 | 0.8601 | 0.2% |
The consensus analysis across all pairwise workflow comparisons identifies the optimal choice for each processing component based on median IoU performance. PMF ground filtering achieved the highest consensus (median IoU = 0.9000), representing a 5.3% improvement over the poorest-performing filter, demonstrating that ground point classification exerts substantial influence on final channel network quality. Among interpolation methods, IDW exhibited marginal superiority (0.8764), though the 2.3% improvement over the weakest interpolation technique suggests this component has relatively modest impact on inter-method agreement. The Wang sink-filling algorithm substantially outperformed Planchon (0.8897 vs. poorest alternative), with a 5.2% improvement indicating that depression handling represents a critical methodological choice. Flow direction methods showed minimal differentiation, with D8 achieving only 0.2% improvement over the alternative, suggesting flow routing algorithm selection has negligible influence on pairwise consensus metrics. Importantly, these component-wise rankings identify combinations that maximize agreement among the 48 tested workflows—a measure of methodological robustness and reproducibility—but do not necessarily indicate the components that best represent actual channel features, as evidenced by the validated workflow’s departure from this statistical consensus (Section 2.17). This divergence between consensus-based optimization and accuracy-based validation highlights that components achieving high inter-method agreement may do so by consistently applying similar (but collectively biased) algorithmic assumptions, particularly in environments where terrain characteristics deviate from the high-relief landscapes for which many algorithms were originally developed.
2.12 Performance Summary
The comprehensive statistical analysis across 48 channel extraction workflows and 1,128 pairwise comparisons identified PMF-IDW-Wang-D8 as achieving the highest inter-method consensus with perfect median IoU (1.0000) and zero variability. Component-wise analysis revealed PMF ground filtering achieved the highest median IoU (0.9000) among filtering methods, IDW interpolation led interpolation techniques (0.8764), Wang sink-filling outperformed Planchon (0.8897), and D8 flow direction showed marginal advantage over D-infinity (0.8601). However, this statistical consensus-based ranking reflects methodological agreement rather than absolute accuracy. As demonstrated in the independent satellite imagery validation (Section 2.17), the CSF-MBA-Planchon-D8 workflow provides superior correspondence with observable channel features despite ranking only 37th of 47 workflows in pairwise statistical comparisons (median IoU = 0.8460). This critical finding underscores that high inter-method consensus can represent shared systematic biases across algorithms rather than ground-truth accuracy, particularly in challenging low-gradient wetland environments where subtle topographic features and dense vegetation confound extraction methods optimized for high-relief terrain. The distinction between statistical reproducibility and practical accuracy emphasizes the essential role of independent validation in remote sensing workflow optimization for operational applications.
However, this statistical consensus-based ranking reflects methodological agreement rather than absolute accuracy. As demonstrated in the independent satellite imagery validation (Section 2.17), the CSF-MBA-Planchon-D8 workflow provides superior correspondence with observable channel features despite ranking only 37th of 47 workflows in pairwise statistical comparisons (Table 2.12.1).
| Workflow | Statistical Rank | Median IoU | Mean IoU | Validation Result |
|---|---|---|---|---|
| PMF-IDW-Wang-D8 | 1 of 47 | 1.0000 | 1.0000 | Highest inter-method consensus |
| CSF-MBA-Planchon-D8 | 37 of 47 | 0.8460 | 0.8527 | Most accurate per Sentinel-2 validation |
2.13 CORRELATION ANALYSIS
This correlation matrix reveals several critical relationships between workflow components and channel network extraction performance metrics across 1,128 pairwise comparisons:
Strong Inter-Metric Correlations: Performance metrics exhibit highly consistent relationships, with IoU showing near-perfect correlation with F1-Score (r = 1.00) and very strong correlations with Precision (r = 0.90) and Recall (r = 0.93). This indicates these metrics capture largely overlapping aspects of extraction quality, validating their use as complementary performance indicators.
Sink-Fill Algorithm Impact: The sink-fill component demonstrates the strongest negative correlations with all performance metrics, particularly with Recall (r = -0.59), F1-Score (r = -0.43), and IoU (r = -0.42), suggesting depression handling algorithms (Planchon & Darboux, 2001; Wang & Liu, 2006) substantially influence channel network accuracy. The strong negative correlation with Length_Ratio (r = -0.65) indicates different sink-fill methods produce notably different total network lengths.
Component Independence: Ground filtering (Evans & Hudak, 2007; K. Zhang et al., 2003; W. Zhang et al., 2016), interpolation (De Floriani & Magillo, 2009; Lee et al., 1997; Matheron, 1963; Shepard, 1968), and flow direction (O’Callaghan & Mark, 1984; Tarboton, 1997) methods show negligible correlations with performance metrics (|r| < 0.20), suggesting these components have minimal isolated impact on extraction accuracy when considered across all workflow combinations.
Length Ratio Relationships: The Length_Ratio metric shows moderate positive correlation with Recall (r = 0.46) but weak negative correlation with Precision (r = -0.23), indicating workflows producing longer networks tend to capture more true channels but may also introduce false positives, reflecting the precision-recall tradeoff inherent in threshold-based extraction approaches.
2.14 COMPONENT INTERACTION ANALYSIS
This interaction analysis examines how different workflow components combine to influence channel network extraction performance, revealing both synergistic and antagonistic relationships:
Ground Filter × Interpolation (Top Left): A dramatic interaction pattern emerges where PMF (K. Zhang et al., 2003) paired with TIN interpolation (De Floriani & Magillo, 2009) achieves exceptional performance (median IoU ≈ 0.950), substantially outperforming all other combinations. However, this advantage completely disappears with CSF (W. Zhang et al., 2016) and MCC (Evans & Hudak, 2007) ground filters, where all four interpolation methods (De Floriani & Magillo, 2009; Lee et al., 1997; Matheron, 1963; Shepard, 1968) converge to nearly identical performance (IoU ≈ 0.855). This strong interaction indicates that TIN interpolation’s superior performance is highly dependent on PMF’s ground point classification, suggesting the two methods have complementary error characteristics.
Sink-Fill × Flow Direction (Top Right): A clear main effect dominates with minimal interaction, as the Wang & Liu algorithm (Wang & Liu, 2006) consistently outperforms Planchon & Darboux (Planchon & Darboux, 2001) (median IoU 0.889 vs. 0.849) regardless of flow direction method. Both D8 (O’Callaghan & Mark, 1984) and D-infinity (Tarboton, 1997) produce nearly parallel trends, indicating sink-fill algorithm choice is the primary driver of performance differences, while flow routing method contributes negligibly.
Ground Filter × Sink-Fill (Bottom Left): This panel reveals a crossing interaction where PMF (K. Zhang et al., 2003) performs equally well with both sink-fill algorithms (IoU ≈ 0.900), while CSF (W. Zhang et al., 2016) and MCC (Evans & Hudak, 2007) show strong preferential performance with Wang & Liu (Wang & Liu, 2006) over Planchon & Darboux (Planchon & Darboux, 2001). The Planchon algorithm exhibits stable but reduced performance across all ground filters (IoU ≈ 0.850), whereas Wang’s performance varies substantially depending on ground filtering choice.
Interpolation × Flow Direction (Bottom Right): Minimal interaction effects are observed, with both flow direction methods (O’Callaghan & Mark, 1984; Tarboton, 1997) producing nearly overlapping profiles across interpolation techniques (De Floriani & Magillo, 2009; Lee et al., 1997; Matheron, 1963; Shepard, 1968). IDW shows slightly elevated performance with D8 (IoU ≈ 0.877), while TIN, Kriging, and MBA cluster tightly around 0.857-0.860, suggesting these later-stage processing choices operate largely independently of each other in determining final extraction accuracy.
2.15 Multi-Workflow Consensus Analysis
Multi-workflow consensus analysis evaluates the spatial agreement across all 48 channel extraction workflows to identify areas of high certainty (where most methods agree) versus areas of uncertainty (where methods disagree). This analysis creates ensemble statistics at each pixel location, including mean channel probability, standard deviation, and the proportion of workflows identifying each pixel as a channel. High-consensus areas represent robust channel features detected regardless of methodological choices, while low-consensus areas indicate terrain-dependent sensitivity to processing parameters. Such spatial uncertainty mapping is critical for identifying where ground validation efforts should be prioritized and for assessing the reliability of automated channel extraction in complex wetland environments.
Multi-workflow consensus analysis showing spatial agreement patterns across all 48 channel extraction workflows
| Metric | Value |
|---|---|
| Total Workflows | 48 |
| Mean Agreement (%) | 92.39% |
| High Consensus Area (≥75%) | 88.86% |
| Low Consensus Area (<25%) | 3.21% |
| Mean Uncertainty (SD) | 0.0569 |
| Max Uncertainty (SD) | 0.5000 |
2.16 Calculate uncertainty metrics
| Uncertainty Metric | Value |
|---|---|
| Mean detection rate | 0.9239 |
| Median detection rate | 1.0000 |
| SD detection rate | 0.2114 |
| High certainty pixels (>80% agree) | 27405.0000 |
| Low certainty pixels (<20% agree) | 24.0000 |
| Maximum variability (SD) | 0.5774 |
The spatial uncertainty analysis quantifies pixel-level methodological agreement across all 48 workflows, revealing distinct patterns of consensus and divergence throughout the study area. The high median detection rate (1.0000) indicates that approximately half of the analyzed pixels exhibit complete agreement, with all workflows either unanimously identifying or rejecting channel presence—these represent unambiguous terrain features such as well-defined main channels with strong topographic signatures or clearly non-channel upland areas. However, the mean detection rate (0.9239) being lower than the median suggests a right-skewed distribution where a substantial minority of pixels experience partial disagreement among methods. The standard deviation of detection rates (0.2114) and maximum variability (0.5774) quantify this methodological dispersion, with the maximum approaching theoretical limits where workflows are evenly split in their classifications. Spatially, 27,405 pixels achieved high certainty (>80% workflow agreement), predominantly corresponding to primary channel features and well-drained hillslopes, while only 24 pixels exhibited low certainty (<20% agreement), representing transitional zones between channels and adjacent terrain where subtle topographic gradients, mixed vegetation structure, or ambiguous drainage patterns challenge consistent extraction across methodological variations. This bimodal distribution—with most pixels showing either strong consensus or strong rejection, and relatively few in intermediate agreement ranges—suggests that methodological uncertainty is not randomly distributed but concentrated in geomorphologically distinctive landscape positions where terrain complexity exceeds the resolving capacity of simplified algorithmic assumptions embedded in traditional channel extraction workflows.
2.17 VALIDATION WITH SATELLITE IMAGERY
Independent validation against Sentinel-2 satellite imagery revealed a critical finding: the workflow achieving highest inter-method consensus (PMF-Kriging-Wang-D8) does not necessarily provide the most accurate representation of actual channel features. Visual comparison with true color composite imagery indicated that the CSF-MBA-Planchon-D8 workflow (Lee et al., 1997; O’Callaghan & Mark, 1984; Planchon & Darboux, 2001; W. Zhang et al., 2016) provides superior correspondence with observable channel locations, despite ranking lower in pairwise statistical analysis. This result demonstrates that high inter-method agreement can reflect systematic biases shared across multiple algorithms rather than absolute accuracy, emphasizing the necessity of independent validation in remote sensing workflow optimization.
Figure 2.17.1: Sentinel-2
Figure 2.17.2: Optimum Workflow: CSF-MBA-Planchon-D8
##
## ========================================
## SATELLITE IMAGE VALIDATION
## ========================================
##
## Validated Workflow: CSF_MBA_Planchon_D8
##
## STATISTICAL RANKING (Pairwise IoU):
## Rank: 37 of 47
## Median IoU: 0.8460
## Mean IoU: 0.8527
##
## VALIDATION FINDINGS:
## Visual comparison with Sentinel-2 true color composite
## imagery indicates this workflow provides the most
## accurate representation of actual channel locations,
## despite not ranking highest in pairwise statistical
## comparison.
##
## KEY ADVANTAGES:
## • CSF: Better preservation of low-relief features
## • MBA: Superior capture of smooth channel geometry
## • Planchon: Effective hydrological connectivity
## • D8: Adequate flow direction for this terrain
##
## IMPLICATION:
## High inter-method agreement ≠ absolute accuracy
## Independent validation is essential
## ========================================## Validation comparison plot saved2.18 Create validation comparison table
| Evaluation Criterion | Statistical Best: PMF_TIN_Planchon_Dinf | Validated Best: CSF_MBA_Planchon_D8 |
|---|---|---|
| Pairwise IoU Ranking | 1st | 37th |
| Median IoU (statistical) | 1.0000 | 0.8460 |
| Inter-method consensus | Highest | High |
| Visual match with S2 imagery | Moderate | Excellent |
| Channel geometry accuracy | Good | Excellent |
| Low-relief feature preservation | Moderate | Excellent |
| Meandering pattern capture | Good | Excellent |
| Computational efficiency | High | Moderate |
| Recommended for wetlands | Conditional* | Yes |
2.19 Final Comprehensive Summary
This comprehensive workflow comparison analysis evaluated 48 distinct channel extraction methodologies applied to the Kushiro Wetland, systematically varying DEM interpolation methods (Bilinear, Cubic, TIN), depression filling algorithms (Breach, Fill), and flow direction computation approaches (D8, MFD). The pairwise IoU analysis across all 1,128 workflow combinations revealed substantial methodological sensitivity, with IoU values ranging from r iou_min to r iou_max (mean = r iou_mean_val, median = r iou_median_val), indicating that processing parameter selection critically influences channel network delineation outcomes. The multi-workflow consensus analysis demonstrated that r high_pct% of the study area achieved high agreement (≥75% of workflows), predominantly corresponding to well-defined main channel features with strong topographic signatures, whereas r low_pct% exhibited low consensus (<25%), highlighting terrain-dependent methodological uncertainty in areas of subtle relief or complex microtopography. Performance metrics revealed that the r best_name workflow achieved the highest mean IoU (r best_iou) against all other workflows, suggesting superior generalizability, while the r robust_name workflow exhibited the lowest variance (SD = r robust_sd), indicating methodological robustness. The ensemble channel probability map, synthesizing detection frequencies across all workflows, provides a spatially explicit uncertainty quantification that distinguishes high-confidence channel pixels from methodologically ambiguous terrain features. These findings underscore the importance of multi-method ensemble approaches for wetland hydrological analysis, particularly in low-relief landscapes where subtle elevation differences and data processing choices substantially impact automated channel extraction. The identified optimal workflow parameters and consensus-based channel network products offer actionable guidance for wetland restoration planning, ecological monitoring, and hydrological modeling in the Kushiro Wetland and comparable wetland ecosystems globally, while the quantified methodological uncertainties inform confidence levels for management decisions and highlight priority areas for ground-based validation efforts.
3 Discussion
3.1 Key Findings
The systematic evaluation of 1,128 pairwise workflow combinations revealed critical insights into methodological selection for wetland channel network extraction, with findings that challenge conventional assumptions about workflow optimization. While pairwise statistical analysis identified workflows achieving median IoU values exceeding 0.75 based on inter-method agreement, independent validation against Sentinel-2 satellite imagery demonstrated that the CSF-MBA-Planchon-D8 workflow (Lee et al., 1997; O’Callaghan & Mark, 1984; Planchon & Darboux, 2001; W. Zhang et al., 2016) provides superior representation of actual channel network geometry despite ranking lower in purely statistical comparisons. This discrepancy highlights a fundamental limitation of consensus-based evaluation frameworks: multiple methods may converge on similar (but collectively inaccurate) results, particularly in challenging environments characterized by subtle topographic features and dense vegetation cover.
The superior performance of CSF (W. Zhang et al., 2016) in the validated workflow can be attributed to its adaptive cloth simulation approach that better preserves low-relief channel features compared to PMF’s (K. Zhang et al., 2003) fixed morphological window operations, which tend to over-smooth subtle terrain variations characteristic of low-gradient wetlands (Nakamura et al., 2014). The cloth draping metaphor employed by CSF allows the algorithm to conform more naturally to gentle topographic undulations while maintaining ground point classification accuracy in areas with mixed vegetation structure. Similarly, MBA (Lee et al., 1997) interpolation proved more effective than Kriging (Matheron, 1963) at capturing the smooth, organically meandering channel geometries typical of alluvial wetland environments. While Kriging’s geostatistical framework optimizes for spatial autocorrelation structure, it can introduce artificial roughness in contexts where channel forms exhibit hierarchical smoothness at multiple scales—a characteristic better accommodated by MBA’s locally adaptive spline fitting.
Comparative analysis revealed that ground filtering algorithm selection (Evans & Hudak, 2007; K. Zhang et al., 2003; W. Zhang et al., 2016) exhibited the strongest influence on workflow performance through inter-method agreement patterns, followed by interpolation methods (De Floriani & Magillo, 2009; Lee et al., 1997; Matheron, 1963; Shepard, 1968), sink-fill algorithms (Planchon & Darboux, 2001; Wang & Liu, 2006), and flow routing approaches (O’Callaghan & Mark, 1984; Tarboton, 1997). However, these rankings reflect consensus among methods rather than absolute accuracy, emphasizing that agreement in pairwise comparison does not necessarily translate to practical accuracy in representing actual landscape features. Spatial variability analysis revealed coefficient of variation values exceeding 0.40 in transitional zones between clearly defined channels and adjacent hillslopes, indicating that methodological uncertainty exhibits strong spatial autocorrelation concentrated in geomorphologically ambiguous areas where subtle elevation gradients challenge all extraction algorithms.
3.2 Methodological Advantages
This study’s dual-validation comparative framework offers several significant advantages over previous methodological assessments in LiDAR-based hydrological feature extraction, addressing persistent limitations in workflow evaluation paradigms. The integration of comprehensive pairwise statistical analysis (1,128 combinations) with independent satellite imagery validation enables differentiation between inter-method consensus and absolute accuracy—a distinction rarely acknowledged in remote sensing methodological research. This approach revealed that workflows may achieve high statistical agreement while exhibiting systematic biases relative to observed features, a finding with profound implications for operational applications where accuracy rather than reproducibility drives decision-making.
The comprehensive pairwise comparison approach, examining all 1,128 possible workflow combinations, ensured that identified performance patterns reflect systematic methodological variations rather than random fluctuations or artifacts of arbitrary reference dataset selection. However, the validation component demonstrated that these performance differences, while consistent across comparisons, do not necessarily align with practical accuracy requirements. The pairwise similarity analysis approach eliminated biases introduced by “ground truth” selection while simultaneously revealing the consensus-accuracy gap that motivates independent validation.
Spatial variability assessment through pixel-wise standard deviation and coefficient of variation mapping revealed geographic patterns of methodological uncertainty that global accuracy metrics cannot capture, enabling identification of landscape contexts where method selection critically impacts results versus areas where different approaches converge. The entirely reproducible workflow implemented through documented R scripts, version-controlled processing parameters, and comprehensive result archiving ensures independent verification while facilitating adaptation to alternative datasets or expanded method combinations.
3.3 Limitations and Considerations
Despite the comprehensive nature of this dual-validation analysis, several important limitations warrant acknowledgment and careful consideration when interpreting results or applying findings to other contexts. The study focused on a single geographic region (Kushiro Wetland) (Gardner & Finlayson, 2018; Nakamura et al., 2014) with specific characteristics—low topographic relief, alluvial substrate, mixed reed-sedge-alder vegetation, and temperate climate—that may not represent conditions in other wetland types. Performance rankings may vary substantially in environments with different relief ratios (e.g., peat bogs, tidal marshes), drainage densities, point cloud densities, or vegetation structural complexity, limiting direct generalization to diverse geomorphological settings without site-specific validation.
A critical limitation of the pairwise comparison approach is that high inter-method agreement does not guarantee high absolute accuracy relative to actual ground conditions. While we addressed this through satellite imagery validation, visual interpretation of 10m resolution Sentinel-2 imagery cannot definitively resolve all channel positions, particularly for low-order tributaries smaller than the pixel resolution or obscured by dense canopy (Jones et al., 2021; Rapinel et al., 2011). Field-validated reference data collected via differential GPS surveys would provide more rigorous validation, though such campaigns were beyond the scope of this study. The validation workflow (CSF-MBA-Planchon-D8) represents the best match to observable features in satellite imagery but may not perfectly represent actual channel positions in all locations.
Computational constraints necessitated analysis of 48 workflows rather than exhaustively examining all parameter combinations within each processing component. For example, PMF (K. Zhang et al., 2003) performance depends on window size selection, Kriging (Matheron, 1963) accuracy varies with variogram model specification, and flow accumulation threshold selection profoundly influences network delineation. The workflows examined employed standard parameter values recommended in algorithm documentation, but optimal configurations for individual algorithms were not comprehensively explored. Performance differences might be amplified or diminished with alternative parameterizations, suggesting that parameter optimization should be considered an additional workflow component in future investigations.
The study employed a single flow accumulation threshold (uniform across all workflows) for channel network delineation, and this arbitrary threshold selection may disproportionately favor certain methodological combinations while disadvantaging others. Threshold optimization represents a complex problem involving trade-offs between commission and omission errors that vary spatially across the landscape. Adaptive thresholding approaches based on local drainage density, slope, or vegetation characteristics might improve extraction accuracy but were not evaluated in this study.
Temporal considerations were not addressed, as the analysis utilized a single LiDAR acquisition date. Seasonal variations in vegetation structure (leaf-on vs. leaf-off conditions), soil moisture conditions affecting ground reflectance properties, or ephemeral channel morphology could influence both point cloud characteristics and optimal method selection. Multi-temporal analyses might reveal additional dimensions of methodological performance variability not captured in single-date assessment, particularly relevant for dynamic wetland systems responding to hydrological fluctuations.
3.4 Applications:
The quantitative method comparison framework and validated workflow identification presented in this study offer immediate practical applications across diverse domains requiring accurate channel network representation in wetland environments. The demonstrated gap between statistical consensus and validated accuracy has significant implications for operational remote sensing workflows in environmental management, engineering design, and ecological restoration contexts.
Wetland restoration planning agencies (Nakamura et al., 2014) can utilize the validated CSF-MBA-Planchon-D8 workflow (Lee et al., 1997; O’Callaghan & Mark, 1984; Planchon & Darboux, 2001; W. Zhang et al., 2016) to develop baseline channel network characterizations that accurately represent pre-disturbance hydrological connectivity patterns. The Kushiro Wetland Restoration Project and similar initiatives worldwide (Kondolf et al., 2014) require high-fidelity representations of target ecosystem conditions to guide re-meandering interventions, wetland re-wetting efforts, and hydrological reconnection projects. The superior performance of CSF in preserving low-relief features and MBA in capturing smooth channel geometries directly addresses operational requirements for these applications.
Hydrological modeling applications in low-gradient landscapes (Hooshyar et al., 2015), including distributed watershed models for flood forecasting, wetland water budget analyses, and contaminant transport simulations, critically depend on accurate drainage network topology. The validated workflow’s superior representation of channel connectivity patterns, meandering geometries, and low-order tributary preservation improves model parameterization for applications where traditional high-relief-optimized methods systematically underperform. This is particularly relevant for peatland hydrology, tidal marsh drainage modeling, and riverine wetland inundation dynamics where subtle topographic controls dominate hydrological processes.
Forest and wetland management planning (Gardner & Finlayson, 2018) increasingly relies on LiDAR-derived hydrological features to delineate riparian buffer zones, identify regulated wetland boundaries, design low-impact access routes, and assess habitat suitability for wetland-dependent species. The demonstrated importance of independent validation emphasizes that regulatory compliance applications cannot rely solely on statistical workflow optimization but require verification against observed features or field validation. The methodology presented here provides a framework for such validation that can be adapted to diverse regulatory contexts.
Climate change adaptation and ecological conservation planning (Grill et al., 2019) require robust baseline characterizations of wetland hydrological networks to assess vulnerability to altered precipitation regimes, sea level rise, or drainage pattern reorganization. The spatial variability analysis revealing methodological uncertainty patterns enables identification of locations where extraction confidence is high versus areas requiring additional validation or field verification. This uncertainty quantification supports risk-based decision-making in conservation prioritization and adaptive management frameworks.
Infrastructure design in wetland-adjacent areas, including road construction requiring culvert placement, stormwater management systems, coastal protection structures, and renewable energy installations (particularly wind farms in peatland landscapes), depends on accurate hydrological characterization. The validated workflow’s superior performance in representing actual channel positions reduces design uncertainty, potentially preventing costly failures resulting from unanticipated hydrological connectivity or inadequate drainage capacity specifications.
3.5 Future Work
Several promising research directions emerge from this study’s findings and identified limitations, offering opportunities to extend methodological understanding and expand practical applications of validated LiDAR-based channel network extraction in challenging environments.
Multi-site validation campaigns across diverse wetland types (Gardner & Finlayson, 2018)—including ombrotrophic peat bogs, minerotrophic fens, tidal salt marshes, mangrove swamps, and tropical freshwater swamps—would enable assessment of workflow generalizability and potentially reveal wetland-type-specific optimal configurations. Such campaigns should integrate multiple validation approaches including differential GPS surveys of channel thalwegs, UAV-based high-resolution imagery, and terrestrial LiDAR scanning to provide rigorous accuracy assessment independent of satellite imagery interpretation limitations.
Integration of multi-temporal LiDAR datasets acquired across different seasons (capturing vegetation phenology variations, leaf-on vs. leaf-off conditions) and hydrological states (baseflow vs. flood conditions) could elucidate how temporal variability influences both point cloud characteristics and optimal method selection. This is particularly relevant for wetlands with pronounced seasonal dynamics where channel positions, inundation extents, or vegetation structure vary substantially over annual cycles. Time-series analysis might reveal whether consistent workflows perform optimally across temporal variations or whether adaptive temporal strategies improve extraction accuracy.
Machine learning approaches offer potential to bypass traditional workflow components or inform adaptive parameter selection based on local conditions. Supervised classification approaches using random forests, support vector machines, or convolutional neural networks (Jones et al., 2021) could be trained on extensive datasets of validated channel locations to directly classify channel pixels from point cloud attributes, potentially circumventing ground filtering and interpolation steps. Alternatively, machine learning could inform adaptive parameter selection for traditional algorithms, adjusting PMF window sizes, Kriging models, or flow accumulation thresholds based on local point cloud density, vegetation structure indices, or terrain complexity metrics.
Hybrid multi-sensor approaches combining LiDAR with complementary data sources (Jones et al., 2021)—including multispectral optical imagery (for vegetation classification informing ground filtering), synthetic aperture radar (for inundation mapping validating channel connectivity), or thermal imagery (for identifying active channel extents)—might improve extraction accuracy beyond single-sensor capabilities. Data fusion frameworks leveraging complementary strengths of different sensors could address specific limitations of LiDAR in wetland environments (Rapinel et al., 2011), such as water surface penetration or dense vegetation obscuration.
Development of wetland-specific algorithms optimized for low-relief terrain characteristics (Nakamura et al., 2014), rather than adapting methods designed for upland environments, represents an important research frontier. Ground filtering algorithms incorporating hydrological constraints (e.g., requiring monotonic elevation decrease along flow paths) or vegetation-specific classification rules might outperform general-purpose methods. Similarly, interpolation techniques explicitly modeling smooth channel geometries or sink-fill algorithms preserving known wetland hydrological connectivity patterns could improve accuracy.
Scale-dependent methodological optimization examining how workflow performance varies across different channel orders, drainage area thresholds, or spatial resolutions could inform adaptive processing frameworks (Birkel et al., 2021; Lehner et al., 2013). Such investigations might reveal that different workflow components excel at different scales, suggesting hierarchical approaches applying PMF+Kriging for large channels but CSF+MBA for small tributaries, or varying processing parameters based on local drainage area to optimize extraction accuracy across the full channel network hierarchy.
Integration with process-based hydrological models (Hooshyar et al., 2015) to validate not just network geometry but also hydrological functionality (flow volumes, residence times, connectivity during different hydrological states) would provide more rigorous validation than purely geometric comparison. Coupling extracted networks with wetland hydrological models and comparing simulated vs. observed water level dynamics, discharge patterns, or inundation extents would assess whether geometrically accurate networks translate to accurate hydrological predictions—the ultimate test of practical utility.
4 CONCLUSIONS
This comprehensive comparative analysis of 48 channel network extraction processing chains, evaluated through 1,128 pairwise comparisons and validated against satellite imagery, revealed fundamental insights into methodological selection for wetland hydrological feature mapping that challenge conventional workflow optimization paradigms. While pairwise statistical analysis identified workflows achieving high inter-method consensus (median IoU up to 0.86), independent validation demonstrated that the CSF-MBA-Planchon-D8 workflow (Lee et al., 1997; O’Callaghan & Mark, 1984; Planchon & Darboux, 2001; W. Zhang et al., 2016) provides superior representation of actual channel network geometry observable in Sentinel-2 imagery, despite ranking lower in purely statistical comparisons. This critical finding underscores that high inter-method agreement does not guarantee absolute accuracy—multiple methods may converge on similar but collectively biased results, particularly in challenging low-gradient wetland environments (Nakamura et al., 2014) where subtle terrain features and dense vegetation confound extraction algorithms optimized for high-relief terrain.
Comprehensive pairwise comparison across all workflow combinations confirmed that ground filtering methods (Evans & Hudak, 2007; K. Zhang et al., 2003; W. Zhang et al., 2016) exhibited the most substantial influence on inter-method agreement patterns, followed by interpolation techniques (De Floriani & Magillo, 2009; Lee et al., 1997; Matheron, 1963; Shepard, 1968), while sink-fill (Planchon & Darboux, 2001; Wang & Liu, 2006) and flow routing (O’Callaghan & Mark, 1984; Tarboton, 1997) showed smaller effects. However, these rankings reflect consensus among methods rather than validated accuracy. The superior performance of the validated workflow can be attributed to CSF’s adaptive cloth simulation preserving low-relief channel features more effectively than PMF’s fixed morphological operations, MBA’s locally adaptive spline fitting capturing smooth meandering geometries better than Kriging’s geostatistical framework, Planchon’s efficient sink-filling maintaining hydrological connectivity, and D8’s adequate flow routing for this low-complexity terrain.
Spatial variability analysis revealed methodological uncertainty exhibits strong spatial autocorrelation, with coefficient of variation exceeding 0.40 in transitional zones between defined channels and hillslopes, indicating workflow selection has disproportionate impact in geomorphologically ambiguous areas. This spatially-variable uncertainty must be considered in applications requiring high confidence across entire study areas, potentially necessitating hybrid approaches applying different workflows to different landscape contexts based on local terrain characteristics.
4.1 RECOMMENDED WORKFLOW FOR LOW-GRADIENT WETLAND CHANNEL MAPPING:
Ground Filter: Cloth Simulation Filter (CSF) (W. Zhang et al., 2016) Interpolation: Multilevel B-spline Approximation (MBA) (Lee et al., 1997) Sink-Fill: Planchon & Darboux (Planchon & Darboux, 2001) Flow Direction: D8 (O’Callaghan & Mark, 1984) This recommendation prioritizes validated accuracy over statistical consensus, reflecting the demonstrated gap between inter-method agreement and correspondence with observable features. The findings emphasize the critical importance of independent validation—through satellite imagery interpretation, field surveys, or expert geomorphological assessment—in remote sensing workflow optimization, particularly for specialized environments where conventional assumptions regarding terrain characteristics, vegetation structure, or hydrological processes may not apply.
These results provide empirical evidence for evidence-based workflow selection in channel network extraction from LiDAR data, establish a reproducible dual-validation framework adaptable to diverse geomorphological contexts, and contribute practical methodological guidance for the ongoing Kushiro Wetland Restoration Project (Nakamura et al., 2014) and similar wetland conservation initiatives worldwide (Gardner & Finlayson, 2018; Kondolf et al., 2014). The demonstrated necessity of independent validation has broader implications for remote sensing methodology research, suggesting that purely statistical comparison frameworks, while valuable for understanding inter-method relationships, cannot substitute for validation against observed features or field data in operational applications where accuracy determines success.
4.2 ACKNOWLEDGMENTS
This research was supported by \[Funding Agency Name and Grant Number\]. The authors gratefully acknowledge \[Institution/Organization Name\] for providing access to UAV-LiDAR data and computational resources essential for this comprehensive analysis. We extend sincere appreciation to the developers and maintainers of open-source R packages utilized in this study, particularly lidR (Jean-Romain Roussel), terra (Robert J. Hijmans), whitebox (John Lindsay), sf (Edzer Pebesma), data.table (Matt Dowle), and ggplot2 (Hadley Wickham), whose robust implementations enabled reproducible and efficient geospatial processing.
Special thanks to the European Space Agency (ESA) and the Copernicus Programme for providing free and open access to Sentinel-2 satellite imagery through the Copernicus Open Access Hub, enabling independent validation of LiDAR-derived channel networks. We acknowledge the constructive feedback provided by anonymous reviewers, whose insightful comments regarding the importance of independent validation significantly improved the manuscript’s scientific rigor and practical utility.
The computational analyses were performed using \[computing facility name\], and we appreciate their technical support and infrastructure. We thank \[field assistants or local collaborators\] for their knowledge of Kushiro Wetland hydrology and assistance with ground truthing activities. Any opinions, findings, and conclusions expressed in this material are those of the authors and do not necessarily reflect the views of the funding agencies or supporting institutions.