All over the world, coastal zones are in constant development and alteration because of natural processes (Neeman et al., 2015) and human-centered actions (Halpern et al., 2008). Responsible factors for changes in coastal zones may be categorized as geological including the type of sediment, distribution and resistance of sediment formations, and isostasy; geomorphological; hydrodynamic; biological including the effect of plants; climatic including rainfall and wind dynamics; and anthropogenic including settlements, industrial expansion, agricultural activities, deforestation, and conservation of coastal areas (Labuz, 2015). Coastal systems additionally have a dynamic structure and generally react in a non-linear morphological way to change (Dronkers, 2005). Also, coastal zones are an essential part of human life due to the fact that approximately 40% of the world’s population live within 100 km of the coast zones and 10% live inside coastal areas that are below 10 m above sea level (United Nations, 2017). Throughout the twentieth century, the global sea level rose at a rate of approximately 1.7 ± 0.5 mm/year, whereas the global ocean temperature has increased by 0.10°C from the water surface to the 700-m depth over the period between 1961 and 2003 (Bindoff et al., 2007). Therefore, monitoring of coastal zones can be an important task for coastal zone protection, well-being of people living inside these zones, and economic activities related to water resources. Periodic observation of coastal areas is also important for the detection of sediment transport, erosion, and accretion (Bio et al., 2015; Kavzoglu et al., 2021; Colkesen et al., 2023). The production of bathymetric maps is also an important task for monitoring marine environment. Depth information and mapping of the water body in coastal areas is critical for many fields of study such as coastal zone management and safety, navigation, oceanography, dredging planning, and construction of marine infrastructure facilities (Muzirafuti et al., 2020; Specht et al., 2020; Alevizos and Alexakis, 2022b). Various technologies and methods have been utilized over the years for bathymetry studies in coastal areas (Lubczonek et al., 2021).

In traditional methods, depth is measured with the help of a pole. With the accumulation of knowledge and developing devices over time, echosounders that form the basis of hydrographic studies have been produced. Echosounders with two different models, single-beam and multi-beam, have been a popular source for determining the depth of a body of water. Multi-beam echosounders can cover the entire seabed with their large swath widths and require a relatively short time for data acquisition (Wlodarczyk-Sielicka and Stateczny, 2016). Some studies have emphasized the disadvantages of these technologies such as cost, high labor requirements, difficulty in data processing, and narrow coverage in shallow waters. Therefore, alternative methods have been proposed for bathymetry measurements (Alvarez et al., 2018; He et al., 2022).

Airborne or space-borne remote sensing methods are relatively low-cost means of data acquisition for monitoring water areas. Airborne laser bathymetry (ALB) has developed rapidly in recent years and has become a widely used resource for bathymetry measurements (Wang and Philpot, 2007; Niemeyer and Sörgel, 2013; Mandlburger, 2019). ALB works with a green laser that can penetrate water bodies. Depth information is calculated based on the values between the water surface and the laser beam reflected from the water floor. The fact that ALB devices can be used as hybrid sensor systems together with digital cameras also increases the potential of bathymetric measurements (Mandlburger et al., 2021). However, factors such as water turbidity, wind, and seabed condition affect the working principle of this technology and cause limitations. Satellite-derived bathymetry (SDB) studies are also carried out by processing multispectral (MS) space-borne imagery. The SDB method is a much more economical resource for researchers than conventional and airborne laser technologies. The first study using satellites was carried out by NASA in 1975, and the depth of the Bahamas was calculated to approximately 20 m using the MS images of the Landsat 1 satellite. Recent advancements in satellite technology have ushered in a new era for the field of SDB, with cutting-edge satellites like WorldView-2 and -3 and QuickBird leading the way. These state-of-the-art satellites are equipped with high-resolution and MS sensors, presenting unparalleled opportunities for advancements in underwater terrain mapping and bathymetric applications (Belcore and Di Pietra, 2022). The MS sensors on these satellites enable the capture of data across different wavelengths, enhancing the ability to discriminate between various water depths and seafloor characteristics. Energy of different electromagnetic wavelengths can penetrate the water surface, and information about depths at different levels can be obtained (Philpot, 1989). Light with a long wavelength is absorbed faster than light with a short wavelength. The blue band, with its low wavelength and high energy, has a high ability to penetrate more deeply water bodies and is suitable for extracting more information for water regions. The near-infrared (NIR) band is often used for shoreline extraction because it produces low reflectance values for water areas (Jagalingam et al., 2015; Randazzo et al., 2020; Celik and Gazioglu, 2022). Pollution, water turbidity, and atmospheric effects are highly influential in the bathymetric use of satellites with passive sensors (Gao et al., 2007). SDB has a coarser resolution than ALB and does not produce information with sufficient accuracy for specific studies (Jegat et al., 2016; Cao et al., 2019).

In recent years, unmanned aerial vehicle (UAV) has become an alternative technology for remote sensing community with the advantage of very high-resolution data achieved from lower flight altitudes, and low cost (Toro and Tsourdos, 2018; Casella et al., 2022). Today, UAVs are actively used in solving many problems by rapid observations about the relevant areas (Mulsow et al., 2018; Nex et al., 2022; Sefercik et al., 2022). This advantage of UAV makes it highly effective in studies on the analysis of water bodies such as water failure monitoring (Kim et al., 2023), oil spill detection (De Kerf et al., 2020), monitoring of pollution source (Cai et al., 2023), and determination of concentrations of physiochemical parameters in water (Ozdogan et al., 2021; Isgro et al., 2022). Current developments in UAV sensing technologies and data processing techniques have made it possible to imaging and modeling of the seafloor. The processing of high-resolution imagery with the help of the structure from motion (SfM) approach, the development of high-accuracy point clouds, and surface models produced from these data have paved the way for the use of UAVs in underwater areas such as bathymetric mapping (Starek and Giessel, 2017). The presence of rocks at specific elevations along the coastline poses a significant hindrance to smooth ship movement, thereby underscoring the practical utility of employing UAVs for bathymetric studies. Bathymetric mapping using UAVs emerges as a notably cost-effective alternative source, particularly in the case of clear waters below 10 m in depth, when contrasted with ship-borne echo sounding or underwater photogrammetric methods (Agrafiotis et al., 2018; Alevizos et al., 2022b). UAVs offer numerous advantages, such as swift data acquisition across expansive areas, cost-effective alternatives to traditional methods, and minimal labor requirements for conducting bathymetric measurements (Rossi et al., 2020; Slocum et al., 2020; Alevizos et al., 2022a). In addition to providing depth data based on image-based bathymetric mapping, it also provides simultaneous acquisition of many data with the help of spectral information contained in the image and indices derived from them. In light of the aforementioned advantages, the UAV-derived bathymetry (UDB) method (Rossi et al., 2020), which builds upon the foundations of traditional SDB, is increasingly being adopted in the shallow bathymetry mapping. Recently, learning-based methods, including deep convolutional neural network (Alevizos et al., 2022a), gene-expression programming (Lee et al., 2022), and support vector regression (Specht et al., 2023), have been employed in the field of UDB studies. Nevertheless, empirical SDB methods are still employed in UDB studies, offering reasonable results in shallow bathymetry mapping (Rossi et al., 2020; Specht et al., 2022; Alevizos and Alexakis, 2022a; Apicella et al., 2023).

The accuracy of bathymetry data is significant for a large variety of marine-related applications such as monitoring of sand mobility, coastal erosion, and hydrodynamic structure in the coastal zone, and maritime navigation (Muzirafuti et al., 2020; Doukari et al., 2021; Alevizos and Alexakis, 2022a; Lange et al., 2023). Like any data acquisition technique, the UAV bathymetry has limitations and it should be taken into consideration that bathymetric mapping performed with UAV cannot produce results as consistent as in-situ measurements (Santos et al., 2023). In this context, it is necessary to consider factors affecting the accuracy of bathymetry data. In shallow bathymetric mapping, a key limitation is the influence of refraction caused by bending of electromagnetic energy or light as it passes through the water surface, causing the apparent depth to be shallower than the actual depth (Tewinkel, 1963; Harris and Umbach, 1972; Dietrich, 2017; Cao et al., 2019). To overcome this problem, various refraction correction approaches have been developed in recent years for SfM point clouds and UDB models (Woodget et al., 2015; Dietrich, 2017; Agrafiotis et al., 2019, 2020; Lambert and Parrish, 2023; Lingua et al., 2023). A straightforward refractive correction method (Westaway et al., 2000) was employed in Woodget et al. (2015) by multiplying apparent depths, calculated using an underlying digital elevation model (DEM) and estimated water surface elevations, by the refractive index of clear water, thus obtaining corrected water depths with an error reduction from 16 mm–89 mm to 8 mm–53 mm. Dietrich (2017) proposed an iterative approach for off-nadir multiview stereo photogrammetry, based on calculating a set of equations for each point/camera combination in the point cloud. Compared with the approach applied in Woodget et al. (2015), this iterative approach is also applicable to off-nadir photos, which in turn provides higher accuracy and precision, as well as better camera calibration results (James and Robson, 2014; Carbonneau and Dietrich, 2016). The iterative approach proposed in Dietrich, (2017) was also applied in a shallow stream bathymetry by Lingua et al. (2023) through the usage of raster data instead of point cloud. The results were satisfactory, with accuracies and precisions of ~0.019% and ~0.07%, respectively, of the flying height. Among these approaches is a learning-based refraction correction approach, DepthLearn, which employs an SVR model developed on the basis of known depth values to correct the systematically underestimated depth values present in the SfM point cloud (Agrafiotis et al., 2019, 2020). Moreover, the SVR model trained on synthetic UAV data with varying flight heights, depths, and seabed anaglyphs displayed notable performance in the correction of estimated depth values while reducing the root mean square error (RMSE) in the z-direction in the range between 7 cm and 24 cm (Agrafiotis et al., 2021). In a recent study, Lambert and Parrish (2023) modified the well-known Stumpf algorithm to test the effect of refraction correction on UAV data. The results demonstrated that the modified algorithm significantly improved accuracy, reducing the RMSE by 3 cm–11 cm.

In this paper, the effect of various UAV imaging bands and different DBM generation methods on the bathymetric 3D modeling quality was comprehensively analyzed through visual and statistical model-based comparison approaches using reference bathymetry data obtained via a single-beam echosounder. Accordingly, four different DBMs were generated and evaluated, two of which were generated using dense point clouds acquired from single-band red–green–blue (RGB) and MS five-band aerial photos, and the other two were generated utilizing Stumpf and Lyzenga empirical SDB models adapted to UAV data. Moreover, the most suitable two-by-two band pair for the Stumpf log-band ratio was determined by means of optimal band ratio analysis (OBRA) and refraction correction was applied to generate SfM point clouds.

Tavşan Island is located at the 2-km South of Büyükada in the Sea of Marmara, Istanbul, Turkey. Marmara is an 11,000-km² inland sea, and seven cities bordering it have vital importance for the country’s economy with industrial, trade, tourism, and agricultural activities. The area of Tavşan Island is 3.6 ha, and the elevation reaches up to 30 m from the sea level. The coastal length of the island is approx. 1 km, and the average depth is around 5 m. The Island is home to a diverse range of marine life, including fish, dolphins, sea turtles, and a rare type of soft coral known as yellow gorgon (Topaloglu, 2016). The study area is placed at the east side of the Island, and the size is 18 m × 90 m. In the study area, the underwater topographic elevations are between −1 m and −5.5 m according to the reference bathymetric model. Figure 1 shows the Istanbul in the Turkey map, location of Tavşan Island in Istanbul, and the study area on RGB orthomosaic of the Island.

The aerial photos of the Tavşan Island were captured with an average ground sampling distance (GSD; i.e., pixel size) of 3.4 cm by using DJI Phantom IV MS UAV. The UAV includes a FC6360 MS camera which have Red, Green, Blue, Red Edge, and NIR monocrome sensors and a traditional RGB sensor for visible light imaging. In addition, the UAV has a real-time kinematic (RTK) global navigation satellite system (GNSS) receiver for precise positioning. Although the UAV has an RTK receiver, ground control points (GCPs) have been established in the Island against RTK signal interruptions that frequently occur when flying over water surfaces. The GCPs were measured by utilizing a CHC i80 GNSS receiver in fixed positioning mode for the highest positioning accuracy (>3 cm). Since the MS UAV flights took 3 h–4 h, the aerial photos were affected by different lighting due to atmospheric and solar conditions; that is why radiometric calibration, converting the digital number into a unit of scene reflectance, was required (Guo et al., 2019). For radiometric calibration, MAPIR Camera Reflectance Calibration Ground Target Package V2 was employed. Table 1 shows the specifications of the used UAV, GNSS receiver, and radiometric calibration ground target.

The methodology of the study consists of three main stages: i) UAV data acquisition and photogrammetric processing, ii) generation of UAV DBMs by multiple methods, and iii) accuracy assessment of UAV DBMs. These stages were presented in detail as subsections of Methodology.

As aforementioned, four different DBMs as RGB, MS, Stumpf, and Lyzenga were generated and comprehensively evaluated by visual and statistical approaches utilizing a reference DBM produced from single-beam echosounder data. Figure 10 shows the generated reference DBM and four evaluated DBMs of the study area. According to the reference DBM, the elevation of deep topography is maximum and minimum in the Southwestern and Northeastern sides of the study area, respectively and around −1 m to −5.5 m. In the first view, the bathymetric surface descriptions of RGB and MS DBMs are closer to the reference than Stumpf and Lyzenga DBMs. In fact, the height scale of the MS DBM produced from five-band aerial photos is almost in the same range as the reference. However, the minimum elevation of RGB DBM is around −4.8 m whereas the reference is around −5.5 m. This shows that the limitations of RGB DBM increase as water depth increases. A similar situation exists for Stumpf DBM, which could not identify topographic details lower than ~−4.8 m.

The model-based absolute vertical accuracy assessment results of the generated DBMs are presented in Table 3 . The systematic bias values were calculated and eliminated before the assessment, and SZ and NMAD values do not include bias. The systematic biases were determined as −0.09 m and 0.07 m for RGB and MS DBMs, respectively. As can be seen in Table 3 , while the accuracies of RGB and MS DBMs based on SZ and NMAD are around ±0.23–0.27 m, they are around ±0.40–0.55 m for Stumpf and Lyzenga DBMs. This result demonstrates that generation of DBMs directly from 3D dense point clouds of the aerial photos is more reasonable than generation of DBMs using the reflectance values of the MS UAV orthomosaic. In parallel with R 2 values, the accuracy of Lyzenga DBM is higher than Stumpf DBM as both SZ and NMAD. This result also reveals that in-depth estimation via UDB adopting more than spectral bands can improve the performance in comparison with the band ratio approach.

The frequency distribution, SZ, and NMAD of height differences between generated DBMs and the reference DBM is given by distribution graphics in Figure 11 . In the graphic, while the peaks of frequency distribution of height differences (FDZ) of 3D dense point cloud-based RGB and MS UAV DBMs are around 30K pixels, it is around 15K pixels for UDB DBMs. In SZ and NMAD sides, symmetric distribution, which reveals a normal error propagation, is available for all DBMs. The NMAD distributions of RGB and MS DBMs are very similar to each other whereas SZ is slightly different. SZ and NMAD distributions prove that Lyzenga DBM performs better compared with Stumpf DBM.

Figure 12 shows the contour maps which enable to interpret relative vertical accuracies between neighboring pixels on generated DBMs. While the contour lines are clearly detectable in RGB and MS DBMs, small contour islands stand out in UDB DBMs with a discontinuous structure. This points out irregular morphologic description of the seabed. The RSZ values in Figure 13 can assist to interpret this situation. While the RSZ values of RGB and MS DBMs begin around ±0.03 m, they are around ±0.11 m for Stumpf and Lyzenga DBMs. Particularly in Stumpf DBM, the RSZ trend is sharper than expected and the values reach up to ±0.3 m. The main reason of this result may be very high resolution (0.1 m) of the UAV dataset against space-borne datasets that are commonly utilized for Stumpf and Lyzenga methods.

The DIFFDBMs, differential DBMs of evaluated DBMs with the reference DBM, are presented in Figure 14 . For better interpretation, the 3D dense point cloud-based RGB and MS DIFFDBMs and Stumpf and Lyzenga UDB DIFFDBMs were validated separately and the height differences between these validated DBMs and the reference DBMs were classified as between ±1 m, ± 0.5 m, and ±0.1 m. The dark-blue color points out that the height difference interval is exceeded. As mentioned before, water deepens from Southwest to Northeast of the study area and the first result of DIFFDBMs is that the accuracy of all analyzed DBMs decreases with increasing water depth. According to the reference DBM, the depth of the water in the study area is around −1 m and −5.5 m and considering the Secchi disc value (7.4 m), turbidity does not exist. From this point of view, it can be discussed that there is a quality limitation in all four UAV-based DBMs, although there is no turbidity in the study area. In general view, the 3D dense point cloud-based RGB and MS DBMs are more coherent with the reference and in the large part of the ±1-m DIFFDBMs, continuous structure is available without exceeding the height difference scale. There are gaps at the scales of ±0.5 m and ±0.1 m, as expected under underwater conditions.

At the ±0.1-m scale, a considerable part of MS DBM was still coherent with the reference in the deepest areas whereas RGB not. No doubt, this result demonstrates the advantage of multiband usage in UAV-based bathymetric modeling. In the UDB DBM side, it is clearly seen that Lyzenga DBM’s compatibility with the reference is much higher than Stumpf DBM in every height difference range. Additionally, the ±1-m scale DIFFDBMs of RGB-MS and Stumpf-Lyzenga DBMs are also included in Figure 14 . While the consistency of RGB and MS DBMs is high at minimum and medium depths, differences of close to 1 m occur at maximum depths. Stumpf and Lyzenga DIFFDBM have an irregular distribution of height differences, and considerable differences are noticeable even at minimum depths in addition to deepest areas.

The use of UAVs in DBM generation in shallow waters is increasing permanently due to its potential to provide rapid, periodic, and very high-resolution data. However, several parameters related with sensor characteristics, imaging geometry, and calibration requirements limit the UAV-derived DBM quality. In this paper, four DBMs, two from single-band RGB and five-band MS 3D dense point clouds and Stumpf and Lyzenga UDB methods, were generated and comprehensively analyzed by visual and statistical model-based comparison approaches. Before the analyses, radiometric and geometric calibrations of UAV aerial photos were carefully completed. While 13 GCPs and 16 additional TPs were used for absolute orientation of RGB aerial photos, 13 GCPs and 4 TPs were used for MS. In both models, five similar ICPs selected from GCPs were employed and approx. ± 0.05-m general orientation accuracy were achieved. In optimal MS band combination determination to properly apply Stumpf and Lyzenga UDB models using very high-resolution MS orthomosaic, OBRA were applied and blue–red band combination was preferred accordingly. To avoid misleading depth description, refraction correction was applied in the generation of all evaluated DBMs.

While DBMs and the DIFFDBMs were generated for visual accuracy analyses, SZ and NMAD of the height differences between the DBMs and the reference model were utilized in statistical analyses. In addition to absolute vertical accuracies, the relative accuracies of the DBMs were also exposed by RSZ calculation for neighboring pixels by distance grouping and morphologic analyses by creating contour-lines. The results of visual and statistical analyses demonstrated that i) the accuracy of all analyzed DBMs decreases as water depth increases considering the DIFFDBMs and ii) the accuracy of RGB and MS DBMs produced from 3D point clouds of aerial photos is higher than DBMs produced from Stumpf and Lyzenga UDB methods. While SZ and NMAD of RGB and MS DBMs are around ±0.23 m–0.27 m, this is around ±0.40 m–0.55 m for UDB DBMs. In detailed view, the representation capability of MS DBM is higher than that of RGB DBM and that of Lyzenga DBM is higher than that of Stumpf as the water depth increases. In contrast, it has been observed that the situation reverses in minimum depths. In addition, the systematic bias determined in RGB DBM between reference DBM was determined as −0.09 m, whereas in MS, it was determined as 0.07 m. iii) Relative accuracies of RGB and MS DBMs produced from 3D point clouds are also higher than UDB DBMs. While RSZ values are between ±0.03 m and 0.15 m for RGB and MS DBMs, they are ±0.11–0.30 m for Stumpf and Lyzenga DBMs. In addition, generated contour lines revealed the noisy structure of Stumpf and Lyzenga DBMs in comparison with RGB and MS DBMs that cause small contour islands. The reason of this result may be the very high resolution (0.1 m) of the UAV dataset against space-borne datasets that are commonly utilized for Stumpf and Lyzenga methods.

Overall, the accuracy of DBMs generated with different techniques has been documented, thus allowing their use for different bathymetric purposes to be evaluated. The advantages and disadvantages of generated DBMs compared with each other are discussed. The common conclusion is that the results given by the examined DBMs can be evaluated for shallow waters up to 5-m depth. It has been documented that at depths greater than 5 m, height differences (i.e., height error) between tested and reference DBMs are partially greater than ±1 m.