Sumatra, the sixth-largest island in the world, covering an area of 473,481 km² on the western side of Indonesia. The island shows strong east-west contrast in geomorphology and coastal dynamics. The eastern part is dominated by fluvial and alluvial deposition, delivering high sediment loads from rivers and supporting extensive peatlands and muddy coastal regions. In contrast, the western part is shaped by tectonic and volcanic activity associated with Barisan Mountains, an active volcanic range that extends along the island. This area is primarily characterized by rocky and sandy beaches and numerous coral islands. Due to these differing characteristics, two representative study areas were selected: Dumai on the eastern coast and Padang on the western coast (Table 1; Figure 2).

Dumai is designated as a Pusat Kegiatan Strategis Nasional (Center for Strategic National Activities) (Media Center Kota Dumai, 2023) and functions as a major industrial port city supporting domestic and international shipping, as well as petroleum, timber, and cement industries. The coastal zone lies along the Malacca Strait and features a semi-enclosed shoreline of approximately 125 km partly protected by Rupat Island. It is dominated by muddy environments and low-gradient topography shaped by riverine inputs, tidal currents, and anthropogenic activities (Rifardi et al., 2015). The water commonly appears brown due to peat water that consists of high organic and humic matters (Water & Wastewater Asia, 2023). Several studies have indicated that the coastal waters of Dumai exhibit varying turbidity values. Octalina et al. (2023) measured turbidity levels at seven observation sites, reporting values ranging from 5.93–29.12 Nephelometric Turbidity Units (NTU), while Sari et al. (2023) analyzed three samples from Lubuk Gaung district and found turbidity values ranged from 47.4–58.4 Formazin Turbidity Units (FTU). NTU and FTU represent comparable units for turbidity measurement, both referencing the scattering of light by suspended particles relative to a Formazin standard and are therefore broadly interchangeable.

In contrast, Padang is a coastal city that serves as the provincial capital and a major administrative and economic hub, with its economy supported by tourism and beach-related activities. Located on the western cost facing Indian Ocean, it has an 84 km shoreline strongly influenced by active tectonics near the boundary of the Eurasian and Indo-Australian plates. Coastal dynamics are dominated by high-energy waves and longshore currents (Haryani et al., 2018). Waters in Padang are generally clear and blue compared with Dumai. Turbidity levels below 4 NTU were recorded in Teluk Bungus in the southern part of Padang (Al Tanto and Kusumah, 2016), whereas higher values, averaging 73.39 NTU, were reported in the northern part, particularly in Padang Pariaman (Kamal et al., 2023). Because the urbanized nature of Padang has limited vegetation, the study area was extended northward into Padang Pariaman, which shares similar physical features but remains more agricultural, dominated by rice fields and plantations. Together, these regions encompass about 110 km of shoreline.

Both coasts have experienced substantial shoreline changes. In Dumai, the mangrove-dominated shoreline has been heavily modified by extensive land development and plantation development, with tidal currents exacerbating erosion (Rifardi et al., 2015). Landsat analyses from 1990–2020 revealed erosion averaging –2.04 m year^–1 and accretion up to 1.17 m year^–1 (Mulyadi et al., 2022). These trends are linked to peatland conversion to oil palm plantations, which alters the water table, reduce slope stability, and may lead to peat failure (Sutikno et al., 2017). Similarly, Padang has undergone extensive erosion with 77% its coastal areas have experienced erosion with a rate ranging from –0.21 to –49.4 m year^–1, while 12% of areas showed accretion of 3–9 m year^–1 (Wisha et al., 2022).

Despite these insights, notable limitations persist in recent studies of coastal changes monitoring in both sites. Previous studies have largely relied on manual interpretation of wet-dry boundaries from satellite imagery, a method that is highly sensitive to tidal conditions. The absence of tidal height data to validate shoreline position poses a significant challenge for robust coastal change monitoring. Additionally, the large spatial extent of these coastal areas makes repeated field surveys logistically challenging and expensive. These limitations underscore the urgent need for alternative methods to improve the accuracy and efficiency of coastal monitoring. The contrasting environmental characteristics of Dumai (sheltered, muddy, peat-influenced) and Padang (open, sandy, tectonically active) therefore provide an ideal testbed for validating the VedgeSat toolkit across diverse tropical environments.

Coastal classification was an essential first step to identify the various coastal types and their land cover that were used for VE extraction. The classification was based on environmental characteristics and vegetation or land cover types. To define the classification, the 1:50,000 scale Rupabumi Map (Badan Informasi Geospasial (BIG), 2019) for each site was used, provided by the National Geospatial Agency of Indonesia (Badan Informasi Geospasial). Due to scale limitations, additional supporting data were necessary for more detailed classification. To achieve this, we used the ESRI World Imagery basemap, sourced from Maxar technologies, which provided high-resolution satellite imagery across various acquisition dates. Although imagery dates varied by location, most scenes in our study area were captured between 2020-2024. This basemap was used to visually support the coastal classification process and was complemented by ground-truth surveys. The final classification of coastal types was determined based on the predominant vegetation cover and sediment (mud, sand, and gravel), derived from the Rupabumi map, visual interpretation of satellite imagery, and field expertise. This integrated approach ensured that the classification reflected the actual environmental conditions at each site.

Dumai was primarily characterized by muddy environments, while Padang had sandy environments. However, certain parts of Dumai’s coastal area, such as in the Sungai Sembilan Sub-district, exhibited sandy environments, as confirmed through the ESRI basemap and ground-truth surveys. Heavily built-up urban environments were excluded from further analysis due to the lack of clearly distinguishable vegetation boundaries in these areas. The presence of artificial surfaces (roads, buildings) introduced spectral noise that complicated edge detection. Dumai was classified into three vegetation types: mangroves, mixed vegetation, and anthropogenic vegetated areas (such as oil palm plantations). Padang, on the other hand, was classified into mixed vegetation, grassland, and anthropogenic vegetated areas (such as managed vegetation).

The VedgeSat toolkit detected VEs by analyzing pixel-level NDVI values, which were influenced not only by vegetation but also by adjacent coastal waters. The spectral qualities of these waters affected the contrast between vegetated and non-vegetated areas, thereby influencing edge detection accuracy. In addition to classifying land cover, coastal waters were also categorized based on their visual appearance in satellite imagery. Despite turbidity measurements showing in Section 2.1, Dumai had relatively low turbidity values, it was still classified as having turbid water due to its consistently brown appearance in satellite imagery. In contrast, Padang’s water appeared clear and blue, which aligned with a clear water classification. This appearance-based distinction reflects the spectral conditions relevant for remote sensing analyses and ensures the classification better supports the evaluation of VedgeSat performance under differing spectral environments. Detailed examples of these settings, including representative satellite and ground-level imagery, are provided in Supplementary Table S1.

The VedgeSat toolkit interfaces with the GEE API, providing access to a variety of satellite image datasets, including Top-of-Atmosphere reflectance images from Landsat 5, Landsat 7, Landsat 8, Sentinel-2, and Landsat 9 (Muir et al., 2024). For this study, Sentinel-2 imagery was employed to extract VE positions. It offers an advantageous combination of high spatial (10 m) and temporal resolution (5 days revisit time), along with multispectral coverage. Furthermore, previous studies have demonstrated that Sentinel-2 imagery provides high positional accuracy. Cabezas-Rabadán et al. (2019) reported an RMSE of 3.01 m, while Muir et al. (2024) reported of 9.3 m for vegetation edge accuracy.

VedgeSat is a toolkit built on the CoastSat infrastructure, designed to extract the vegetation edge information through four main steps: (i) identifying the user requirements; (ii) satellite image pre-processing; (iii) supervised satellite image segmentation and classification; (iv) thresholding and contouring (Muir et al., 2024). Here we summarize the workflow with specific reference to this study and refer readers back to Muir et al. (2024) for further detail (Figure 3).

Referring to 2.1, each study site was classified based on its vegetation types and environmental conditions, providing a comprehensive framework for determining sample numbers and locations. This classification considered as key factors such as vegetation type, muddy or sandy environments, and water turbidity, ensuring that the selected samples represented the diverse environmental context of Dumai and Padang. A total of 15 samples were analyzed for this study, with eight samples from Dumai and seven samples from Padang (Figure 4).

All the validation lines in Dumai were obtained through manual digitation of PlanetScope images. To ensure the accuracy of these images, orthorectification was performed using ArcGIS with ESRI Basemap as a reference. In contrast, the ground-truth survey in Padang were collected through a ground-truth survey, where vegetation edge was tracked using an GPS EOS Arrow 100 Sub-meter GNSS receiver connected to ArcGIS FieldMaps. This setup ensured high positional accuracy by receiving signals from multiple global navigation satellite systems (GLONASS, Galileo, and BeiDou). The EOS Arrow 100 is known for submeter accuracy, achieving a positional accuracy of approximately 20 cm, as demonstrated by (Muir et al., 2024), and even 15 cm, as reported in Dynamic Coast project (Rennie et al., 2021). These precision levels make it a reliable tool for validating vegetation edge positions in diverse environments. The validation process in these environments uses dense vegetation to define the edge. Additionally, this study seeks to refine the detection of sparse vegetation within transition zones, addressing the challenges associated with mapping these areas.

The difference in validation approaches between Dumai and Padang was due to field conditions. In Dumai, ground-truth surveys were not feasible due to the hazardous environmental conditions, characterized by uncertain mud depths, unstable ground, and dangerous fauna. As a result, manual digitization from higher-resolution (3 m) PlanetScope imagery was used as an alternative. Although PlanetScope-based digitization involves visual interpretation, it was applied systematically as part of an RMSE-based validation process. To reduce potential spatial uncertainty, the closest available PlanetScope imagery was prioritized to reduce the date gap between Validation Lines (VLs) and extracted VEs. However, image availability and cloud cover sometimes restricted the ability to obtain an exact temporal match, potentially introducing minor positional discrepancies. Despite these limitations, both validation methods provide meaningful assessments of VedgeSat’s performance across different environments, ensuring that vegetation edge detection is evaluated in both field-accessible and less accessible coastal settings.

To validate the results of VedgeSat, the positions of extracted vegetation edges were compared to the validation lines along shore-normal transects spaced at the default interval of 10 m (Muir et al., 2024). VedgeSat toolkit offers the validation function to quantify the error or biases between the extracted Vegetation Edges (VE) and Validation Line (VL). This function provides key statistical metrics, including Root Mean Square Error (RMSE) and Mean Absolute Error (MAE). The RMSE value (Equation 2) and MAE (Equation 3) are widely used statistical metric for evaluating model performance (Chai and Draxler, 2014). VedgeSat utilized both metrics to assess error margins for each weighting combination and satellite image platform, based on cross-shore distances. In these equations, X and x in the equations refer to the cross-shore positions of validation line and the extracted vegetation edges, respectively. This study focuses on the RMSE value as the primary validation parameter. RMSE reflects the variability in model performance by penalizing larger errors through squaring, making it sensitive to variations in errors (Chai and Draxler, 2014). This sensitivity is particularly suitable for datasets in vegetation edge detection, where capturing variabilities in positional accuracy is crucial.

In addition to RMSE and MAE value, the VedgeSat toolkit also provides statistical measures including MPE and the Coefficient of Determination (R²). The median positional error (Equation 4) was a statistical measure that represents by the median of all positional errors of extracted vegetation edge and validation lines used as references. This measure was useful for identifying systematic positional bias, as it indicated whether the extracted edges consistently deviated landward (negative values) or seaward (positive values) from the reference line. Meanwhile, the Coefficient of Determination (R²) shown in Equation 5 was used to evaluate the performances of VedgeSat by assessing the alignment between the extracted vegetation edge compared to the validation line.

In Equations 2–5, n denotes the number of cross-shore transects used for validation. Xi and xi represent the cross-shore positions (in meters) of the xtracted vegetation edge (VE) and corresponding validation line (VL) at transect I, with positive values indicating seaward bias and negative values indicating landward bias. In Equation 5, yireference and yidetected refer to cross-shore positions of validation line and the extracted vegetation edge, respectively. All positional metrics are expressed in meters.

Figure 5 summarizes VedgeSat performance across 15 validation samples representing a variety of coastal environments in Dumai and Padang. Although the VEs are generally considered more stable indicators for coastal change than dynamic wet-dry indicators, the time difference (date gap) between VLs and extracted VEs may still influence positional accuracy and model performance. In this study, the average date gap was 8–9 days, with the longest gap observed for sample S6-Mixed vegetation muddy of Dumai, with a 17-day difference. Meanwhile, the smallest date gap of one day was noted for sample S1-Mangroves-muddy in Dumai, as well as for sample S12-Grass sandy and sample S14-Managed vegetation sandy in Padang.

Across all samples, RMSE ranged from 2.96 to 53.08 m and R² values varied from 0.01 to 0.89, indicating differences in performance depending on vegetation type and sediment. In general, VedgeSat showed strong performance in dense vegetation settings with RMSE values consistently below 7 m. For example, S2-Mangroves muddy and S6-Mixed vegetation muddy in Dumai achieved R² values of 0.89 and 0.86, respectively, with RMSE values below 6 m. The lowest RMSE of 2.96 m was recorded in sample S3-Mangroves muddy in Dumai. Conversely, accuracy was lower in sparse vegetation areas, especially in sandy environments. S12-Grass sandy exhibited the highest RMSE (53.08 m) and the lowest R² (0.12), while both S4-Mangroves sandy and S5-Mangroves sandy in Dumai returned poor R² values (0.08 and 0.01) and high RMSE values of 29.22 and 36.94, respectively.

MPE, which reflects the typical direction and magnitude of offset between extracted VEs and VLs, ranged from –38.3 to +6.4 m. Negative values indicate that the VE was placed more landward than the corresponding VL, while positive values reflect a seaward offset. Most samples in Dumai showed negative MPEs, consistent with muddy environments except for S3-Mangroves Muddy, which had a slightly positive offset of 0.2 m. In Padang, most samples exhibited positive MPE values, with the exception of S12-Grass sandy which had a substantial landward error with MPE = –38.3 m. Overall, dense vegetation on muddy sediments showed the highest accuracy, while sparse sandy environments exhibited the largest errors. The following section provides a more detailed breakdown of performance by vegetation type and environmental conditions.

The selection of an appropriate NDVI threshold was a key factor in determining the accuracy of vegetation edge detection using the VedgeSat toolkit. Applying a single automated threshold across diverse environments may lead to inconsistent or suboptimal outcomes. Figure 14 summarizes the results of threshold tuning process across 15 samples, which tested NDVI values from -0.05 to 0.30 in increments of 0.025, and compared their performance with that of the threshold automatically calculated using the Weighted Peaks threshold. Automated thresholds were frequently outperformed, with RMSE values ranging from 4.1 m (S13-Grass gravel) to 88.9 m (S8-Oil palm plantation muddy), while tuning reduced RMSE to as low as 2.67 m (S14-Managed vegetation sandy). Optimal thresholds varied by vegetation type and density. Dense samples typically performed best between 0.10 and 0.20, whereas sparse vegetation required lower thresholds from –0.025 to 0.10. Mixed vegetation in muddy environments and managed vegetation in sandy settings showed consistent optimal thresholds around 0.125 and 0.10, respectively, whereas other environments exhibited varying values.

The sensitivity of VE detection to thresholds adjustments varied across vegetation types and environmental settings. Sparse vegetation, such as mangroves and grass on sandy sediments, showed narrow optimal ranges where small NDVI changes produced large positional shifts. For example, in sample S12-Grass sandy, RMSE increased from ~10 m at 0.025 to nearly 40 m at 0.05, reflecting a narrow NDVI threshold window. By contrast, mixed vegetation on sandy sediments and grass on gravel showed broader tolerance with relatively stable RMSE values across thresholds. Overall, sparse settings were highly sensitive while denser vegetation types were more robust, with Padang samples (sandy and gravel) generally showing wider optimal ranges than Dumai’s muddy environments. These results suggest that vegetation density, sediment, and spectral contrast together determine the robustness of VE detection to NDVI threshold variation.

To further explore how vegetation type and sediment influence edge detection, the 15 samples were grouped into four categories: (1) mangroves on muddy vs sandy environments, (2) mixed vegetation in muddy vs sandy environments, (3) grass in sandy vs gravel environments, and (4) managed vegetation vs oil palm plantation (Figure 15). Mangroves on muddy settings showed stable RMSE around 0.10, while those on sandy sediments were highly sensitive with sharp increases beyond 0.025.

Mixed vegetation in both muddy and sandy environments exhibited well-defined RMSE minimum near 0.125, although sandy sites were slightly more variable. S13-Grass gravel displayed a stronger contrast with low sensitivity and steady RMSE across a wide threshold range. Conversely, S12-Grass sandy was highly sensitive, as RMSE rose dramatically from 10.08 m to over 30 m with only small change in threshold. Lastly, the comparison of anthropogenic vegetated landscape between managed vegetation in sandy environment and oil palm plantation in muddy environment revealed contrasting patterns. Managed vegetation on the sandy coasts produced smooth and stable results with an optimum near 0.10, while oil palm plantations on muddy coasts showed narrow optimal ranges around 0.15 and greater variability.

The variation in VedgeSat performance across vegetation types and environments highlights how ecological structure and local spectral context influence the robustness of automated NDVI-based detection. Dense vegetation produced reliable results, with sub-pixel positional accuracy and strong correlations to validation data. These findings indicate that canopy structure, coverage and biomass enhance NDVI separation from adjacent bare sediments, creating well-defined gradients that support consistent edge detection. Previous studies (Gao and Zhang, 2006; Olmo et al., 2024) similarly demonstrated that vegetation type and canopy cover strongly influences NDVI-based indices, supporting the idea that structural complexity drives spectral stability in vegetation edge detection. In contrast, sparse or patchy vegetation caused substantial positional uncertainty. Limited canopy cover and fragmented patches such as pioneer mangroves and grasses produced weak NDVI signals that often fall below classification thresholds and reduce detection accuracy. In line with these findings, Muir et al. (2024) noted that sparse vegetation within diffuse and dynamic vegetation edge presents significant challenges for edge detection, often leading to larger positional uncertainties.

Spectral mixing emerged as the primary factor driving bias patterns across sites. In Dumai’s brown waters, most samples (except S3-Mangrove muddy) consistently exhibited negative MPE values, indicating that vegetation edges were positioned landward relative to validation lines. Transitional pixels in this environment often contained mixed signals of brown water, muddy sediment and vegetation signals, producing intermediate NDVI values that fell below the vegetation threshold and shifted the edge inland. In contrast, sandy and gravel environments in Padang consistently produced positive MPE values. The transitional pixels combined signals from vegetation, and sandy or gravel sediments, generating intermediate NDVI values and pushing the VE further seaward. Despite these opposing patterns, MPE values for dense vegetation across both sites remained below 7 m, confirming that VedgeSat maintained robust performance under varied water and sediment conditions.

In addition to environmental conditions, several technical factors also affected VedgeSat’s performance and positional accuracy. The 10 m spatial resolution of Sentinel-2 imagery limited the ability to capture narrow or highly dynamic vegetation edges, particularly the sparse or patchy vegetation. Additionally, temporal mismatch between satellite image acquisition and validation surveys also introduced uncertainty, as vegetation growth, erosion and/or accretion, or other environmental change could occur between dates, increasing RMSE values. However, perfectly matching dates for VLs and extracted VEs are often unachievable due to limitations in satellite revisit frequencies, cloud cover, and data availability. In this study, these temporal effects were present but generally less pronounced than the environmental factors such as vegetation density, sediment, and water conditions.

Overall, despite challenges in sparse and dynamic environments, VedgeSat demonstrated strong and consistent performance in detecting vegetation edges across the two contrasting tropical coastal settings examined in this study, particularly in areas with dense vegetation. While the accuracy values reported reflect site-specific environmental conditions and validation approaches, several methodological insights are broadly transferable, including the sensitivity of vegetation edge detection to vegetation density and NDVI threshold selection. These findings provide an initial benchmark of VedgeSat’s performance in tropical coastal environments and highlight its potential to support long-term coastal change assessments, provided that site-specific calibration and validation undertaken.

The performance of VedgeSat’s automated thresholding, which is based on the Weighted Peaks algorithm, reflected both its potential and limitations when applied to complex tropical coastal environments. Designed to enhance detection in diffuse vegetation edges by assigning greater weight to non-vegetation pixels (0.8) than to vegetation (0.2), the algorithm aimed to better capture transition zones. However, its performance across 15 samples in Dumai and Padang was inconsistent. While the method performed well in some cases (e.g., RMSE of 4.1 m in sample S13-Grass gravel), in most samples, especially sparse vegetation settings, it produced high errors, with RMSE values exceeding 30 m. Overall, these results indicate that automated NDVI thresholding does not consistently deliver optimal performance across the tropical environments examined in this study.

In several cases, the derived thresholds unexpectedly aligned more closely with NDVI values of vegetated areas, despite the higher weighting assigned to non-vegetation. This counterintuitive result likely arises from differences in the statistical properties of pixel distribution. First, vegetated pixels often produce sharper and more distinct NDVI peaks which can disproportionately influence the threshold calculation despite their smaller assigned weight. Second, the NDVI values of non-vegetated areas such as sand or clear water often exhibit broader or flatter distributions, leading to poorly defined peaks that reduce the influence of non-vegetated class in determining threshold. Third, the default weighting approach may not be fully appropriate for tropical coastal environments, where spectral conditions are more complex. These results indicate that the default weighting scheme is less suited to tropical conditions, where vegetation, sediment and water mixtures create complex spectral signatures that challenge automated threshold estimation. Further adjustment of the weighted values or local tuning may be necessary to achieve optimal automated results in complex tropical coastal environments.

The key motivation behind the automated thresholding method is to reduce manual effort, enabling faster and more scalable VE detection across large datasets or long time series. However, these results underscore a trade-off between automation and accuracy. While automation allows users to avoid labor-intensive threshold selection for each image, it may compromise edge precision in heterogeneous environments. This highlights the importance of improving the automated approach to balance efficiency with positional accuracy, especially for long-term change assessments where manual inspection of each image is impractical.

The threshold sensitivity analysis further highlighted how environmental and vegetation structure influenced VedgeSat’s response to NDVI variation. Dense vegetation exhibited relatively low sensitivity to threshold adjustments compared to sparse vegetation samples. For example, in sample S12-Grass sandy, a small threshold change from 0.025 to 0.05 increased RMSE from 10.08 m to 38.53 m, reflecting instability in transitional zones. In contrast, S15-Managed vegetation sandy showed consistently low sensitivity, with RMSE increasing by only 0.8 m between thresholds of 0.20 and 0.225. Furthermore, muddy environments tended to be more sensitive than sandy environments, indicating that vegetation edges in turbid and dynamic settings are more responsive to small spectral changes. These contrasting responses demonstrate strong site-specific sensitivity of NDVI thresholds, with S12-Grass sandy exhibiting a narrow optimal threshold window, in contrast to other environments where a broader range of thresholds produced comparable RMSE values. Furthermore, these findings are further illustrated in Figure 16, which compares positional accuracy before and after NDVI threshold tuning. The plot shows a marked improvement in RMSE, R², and MPE after tuning for most samples, highlighting the effectiveness of threshold adjustments in enhancing vegetation edge detection accuracy in complex tropical coastal settings. The substantial RMSE gap observed between automated NDVI thresholds, and the best-performing manually selected thresholds provides direct empirical support for the need for regional or site-specific threshold calibration when applying VedgeSat in tropical coastal environments.

Overall, these findings demonstrate that while the automated threshold may not consistently deliver the best results, VedgeSat’s threshold flexibility allows users to adapt the tool to local environmental contexts. Through systematic threshold tuning significant improvements in positional accuracy can be achieved, especially in complex or transitional coastal environments. This adaptability underscores VedgeSat’s potential as a flexible and scalable framework for monitoring vegetation edges in diverse tropical settings.

This study highlights several key strengths of applying VedgeSat for vegetation edge detection in tropical regions. First, the toolkit proved to be robust and adaptable across diverse vegetation types and environmental settings, even without retraining of the model classifier. Despite being originally trained in temperate settings, the default configuration performed well for a range of dense vegetation types in tropical coasts, underscoring its transferability and scalability for data-limited regions. Second, the systematic NDVI threshold tuning revealed that simple adjustments can lead to notable improvements in edge accuracy when validation data are available. This adaptability strengthens its potential for broader use across varied coastal settings. Finally, the toolkit’s flexible input structure and transparent configuration also allow for customization, making it accessible to a broader range of coastal researchers and practitioners. These strengths position VedgeSat as a promising proxy for monitoring coastal change, particularly in challenging settings such as turbid and muddy environments, as demonstrated by López et al. (2025) who applied it successfully in estuarine settings in Argentina.

However, VedgeSat is built upon several methodological assumptions, including the presence of a discernible vegetation boundary separating vegetated and non-vegetated zones, sufficient vegetation density to produce a clear NDVI signal, and sensor characteristics that allow consistent detection across image acquisitions. The method further assumes that vegetation edges represent a more temporally stable shoreline proxy than wet-dry boundaries and are therefore less sensitive to short-term tidal variability. As a result, VedgeSat is particularly well suited for long-term and seasonal analyses of coastal change, where reducing tidal-induced positional noise is critical.

Several limitations and challenges were also identified. VedgeSat’s performance decreased in environments characterized by sparse vegetation, such as sparse mangroves and grasslands with limited spatial extent, where NDVI values tend to be low and fall near the classification threshold. This method-related limitation was observed to be further amplified under tropical coastal conditions characterized by heterogeneous vegetation composition, high turbidity, and mixed sedimentary environments. Furthermore, the 10-m spatial resolution of Sentinel-2 imagery may not be sufficient for detecting narrow or fragmented vegetation zones, especially when these patches are smaller than a single pixel. In addition, the automated thresholding using the Weighted Peaks algorithm, though conceptually promising, showed mixed results in tropical settings. While it worked well in gravel environments, it tended to underperform in other environments, reinforcing the need for localized threshold adjustments in such settings.

Lastly, the validation process in Dumai, which relied on manual digitization of high-resolution PlanetScope imagery to define VLs, introduces an element of subjectivity and potential positional uncertainty. Differences in spatial resolution between Sentinel-2 (10 m) and PlanetScope (3 m) may also affect alignment and contribute to discrepancies between extracted VE and VL. These resolution mismatches and manual interpretation challenges should be considered when evaluating the accuracy of the results. In contrast, validation in Padang was conducted using a GPS EOS Arrow 100 Sub-meter GNSS receiver integrated with ArcGIS FieldMaps surveys, providing higher positional accuracy and a more direct representation of the true vegetation boundary. While this approach reduces subjectivity and image interpretation uncertainty, GNSS-based surveys remain subject to errors related to satellite geometry, atmospheric conditions and receiver performance. Accordingly, the use of different validation approaches between Dumai and Padang introduces method-dependent uncertainty, which may contribute to differences in reported accuracy metrics. Accuracy results are therefore interpreted primarily within each study, rather than as a direct quantitative comparison between sites.

Future research should address the challenges of detecting edges in sparse or fragmented vegetation environments. While VedgeSat currently incorporates the Soil Adjusted Vegetation Index (SAVI) using a default L value of 0.5, there is potential to further refine its performance in areas with sparse or dynamic vegetation by adjusting the L values based on the environmental characteristics. Additionally, integrating alternative indices such as Enhanced Vegetation Index (EVI), which is more responsive to canopy structure and sediment variability, could help resolve spectral ambiguity of sparse vegetation. These indices may improve edge detection where the NDVI contrast is weak.

Beyond vegetation indices, the integration of spectral unmixing techniques could further enhance edge detection by decomposing mixed spectral signals into their constituent components. This approach would allow for the identification of sparse vegetation within pixels that contain a combination of vegetation, sand, and water. For instance, spectral unmixing has been successfully applied by Ettritch et al. (2018) to monitor the coastal dune system in Wales and by de Vries et al. (2021) to characterize diffuse boundaries of subtidal mud banks. Applying similar techniques to tropical coastal environments could improve the detection of vegetation edge.

Furthermore, while the current automated NDVI thresholding approach using Weighted Peaks showed promising results in specific cases such as gravel environment, its performance was inconsistent in muddy and sandy settings, particularly in areas characterized by sparse vegetation. To enhance the performance of VedgeSat in diverse tropical environments, future work should focus on refining the weighted peaks values. While the current Weighted Peaks algorithm provides a computationally efficient solution, it may not adequately capture the spectral complexity found in tropical coastal settings. One promising direction is to evaluate alternative weighting combinations tailored to specific vegetation and environmental types. This would not require the development of new algorithms, but rather a systematic calibration of the existing approach to improve its adaptability and accuracy.

Ultimately, these refinements guided by the tuning result can reduce the need for manual threshold adjustments and enable more consistent and scalable automated VE detection in coastal monitoring. Despite the challenges outlined, this study demonstrates the promising capabilities of VedgeSat in characterizing vegetation edges across complex tropical coastal environments. Its ability to detect long-term vegetation changes using freely available satellite imagery makes it a scalable and replicable tool for supporting coastal monitoring in data-limited regions. By establishing the baseline accuracy and environmental sensitivity of the toolkit, this study provides a critical foundation for expanding its application to broader coastal change assessments. These insights will be further explored in a follow-up study focused on evaluating the long-term spatiotemporal dynamics of coastal change in both study sites.