The UVP5 (Hydroptics, France) is an optical instrument developed for imaging large particles, such as aggregates and zooplankton, in aquatic environments (Picheral et al., 2010). The camera is mounted vertically, facing downward approximately 40 cm above the light source, capturing images along its path at a frequency of 20 Hz. The sampling volume spans 3.11cm in depth with a field of view of 22 cm × 18 cm. Illumination is provided by 42 red LEDs with a wavelength of 625 nm and flashes of approximately 100 µs duration, as described by the manufacturer. Images are recorded in grayscale with a pixel resolution of 88 µm. An internal processor (x86 AMD Geode GX533, 400 MHz CPU) in the camera housing automatically extracts in-focus particles, or regions of interest, to internal memory. Although the system has the option to save full-frame images, we opted to save only regions of interest to optimize memory use. The system also provides an algorithm that automatically generates size distributions; however, we applied our algorithm to generate size distributions for comparisons with other systems.

The LISST HOLO2 (Sequoia Scientific, Inc., USA) records holographic images using optical interference and diffraction to provide three-dimensional views of particles. The electronic sensor (CCD) is positioned horizontally, approximately 50 mm from the laser housing, which emits collimated laser light with a wavelength of 658 nm. As particles pass through the laser’s optical path, they generate interference patterns that are captured as holograms. These holograms are recorded at a frequency of 10 Hz, with a pixel resolution of 4.4 µm, providing a field of view of approximately 5.3 mm × 7.0 mm and a sample volume of ~1.8 cm³ per hologram. The recorded holograms require digital reconstruction to produce actual images of the particles. We used the particle image extraction software suite FastScan, developed by the University of Aberdeen, for hologram reconstruction and auto-focusing (Liu et al., 2023; Thevar et al., 2023), which extracts in-focus particles as regions of interest.

The CPICS (Coastal Ocean Vision, USA) records images in RGB color, capturing particles in true color. The camera (Prosilica GT 1380) is positioned horizontally, with white strobe lights mounted opposite the camera to provide uniform illumination. The sample volume is approximately 0.33 cm³, with a field of view measuring 1.5 cm × 1.1cm. Particles moving through the sampling volume are recorded at a frequency of 10 Hz, with a pixel resolution of 4.5 µm. An internal Jetson TX2 processor automatically detects in-focus particles, or regions of interest, and stores them in internal memory. Although full-frame savings are available, we opted to save only the regions of interest to optimize memory usage.

Olive stone granules (BioPowder.com, https://www.bio-powder.com/en/) were used as reference particles with known size ranges due to their hydrophobic properties, which help to minimize light reflection and scattering during imaging with the optical instruments. Five distinct particle size ranges (0 - 50 µm, 100 - 300 µm, 300 - 600 µm, 600 - 800 µm, and 800 - 1000 µm) were defined to represent the overall span of particle sizes considered in this study. These ranges were merged into a continuous dataset for subsequent analysis of the full size spectrum for all the instruments, enabling validation of the size spectra across the instruments.

Calibration experiments were conducted on two separate occasions (26 May 2022 and 22 April 2023) using a plastic container with dimensions of approximately 1.5 m × 1.5 m × 1 m (Figure 1b). The instrument frame, housing the three optical devices, was positioned vertically within the container to replicate its orientation during field operations. Fresh water at approximately 20 °C was used to prevent particle buoyancy or aggregation caused by salinity gradients. The water level was maintained at a consistent height of 0.5 m throughout both experiments to ensure stable conditions.

During the preliminary trials, particles were mixed with water in a small bottle and then released above the sampling volume of each instrument by gently squeezing the bottle. This method facilitated particle introduction but posed challenges in ensuring precisely controlling the volume of particles uniformly passing through the sampling volume of the instruments. Additionally, the process of squeezing the particle-water mixture out of the bottle caused excessive initial velocity to the particles, resulting in motion blur and reduced image quality. The method of particle introduction was refined to reduce particle velocity and minimize motion blur, enhancing image quality. Particles were gradually introduced using a small plastic container, held approximately 5cm from the water surface. We merged particles of five size classes (5 grams per particle size class) and released them from the container for approximately 30 seconds allowing them to become naturally moistened at the surface. Particles were then left to settle naturally through the sampling volume, minimizing turbulence and enhancing uniform distribution. Particle images were simultaneously captured by the three instruments throughout the experiments and a total of 391,828 images, 23,030 images, and 53,602 images were 244 recorded by UVP5, CPICS and LISST HOLO2, respectively.

Image filtering to separate particles and artefactual features (i.e. extreme motion blur, bubbles and interference patterns, Supplementary Figure S1) was performed by trained convolutional neural network (AlexNet, F-1 score = 0.9, detailed method in Supplementary Materials). Although 20 - 40% of images were excluded during filtering, most removals reflected optical artifacts or poorly resolved features rather than true particle absence. Including such detections would distort size spectra due to motion blur, focus errors, or scattering artifacts. We therefore retained only particles with reliably measurable morphology, following standard plankton imaging practice. All filtered images were then converted into grayscale and binarized to facilitate size analysis. We used the regionprops function in MATLAB (R2023b, Mathworks Inc.) to approximate each particle as an ellipse and extract geometric properties. Two metrics were used to characterize particle sizes; major axis length (MajAL, in cm) corresponding to the length of the major axis of the best-fit ellipse, and equivalent spherical diameter (ESD, in cm), calculated as the diameter of a two-dimensional circle with the same area as the detected particle region. For all instruments, a minimum particle size threshold of 5 pixels was set to ensure reliable particle recognition. Since each imaging system produces different types of images, we applied four binarizing algorithms to determine the most effective method to detect particles - Otsu, Canny, Sobel, and Prewitt. Each algorithm offers a distinct approach to separating particles from the background and identifying particle features and they are commonly compared when selecting robust edge detection method (Canny, 1986; Liu et al., 2018; Onyedinma Ebele et al., 2025; Otsu, 1979). We visually inspected the subset of binarized images based on three criteria: (1) edge clarity, the degree to which particle are clearly separating particles from the background, with continuous boundaries and minimal noise; (2) fragmentation, evaluating whether the full shape of each particle is retained without breakage; (3) completeness, assessing whether all visible particles are fully captured. To confirm that our simplified approach with a default setting provided robust results, a 10% subset of images was also processed with additional preprocessing (background correction, denoising, and contrast normalization) and with optimized Canny thresholds (the most parameter-sensitive method) following Ohman et al. (2019). Results differed by less than 10% from the defaults and the changes in particle outlines were negligible; the optimized thresholds and the number of edge pixels remained consistent. These checks confirm that the default settings provided robust and reproducible edge detection across instruments.

We constructed number size spectra using logarithmically increasing size bins (following Petrik et al., 2013), a powerful method to describe size distribution of marine particles, where large particles are typically less common. Number of particles, N, in each logarithmically increasing size bin (MajAL and ESD; 0.01 cm - 0.56 cm with 100.065 cm bin width) was calculated. Nlog was divided by the bin width and the sample volume (V_im) to obtain normalized particle abundance (n = N/(bin width)/V_im), and log₁₀(n) was plotted against the logarithm of the average bin size to construct a number size spectrum. Any bin with the insufficient number of particles N < 10 was discarded before computing the spectrum. Normalized particle abundance - defined as (raw particle count divided by total sample volume and bin width, n, cm⁴) was plotted as a function of particle size. For each instrument, size spectra were plotted together, and raw particle counts in each bin were indicated in color (Figure 2), highlighting the most abundant size bin for each instrument. This comparison allowed us to confirm that the normalization procedure did not alter the relative shape of the size distributions. We also fitted linear regression lines from the peak of number.

For high particle-background contrast images, four thresholding methods (Otsu, Sobel, Canny and Prewitt) showed similar outcomes of binarized images (Figure 2) and passed our visual inspection criteria (edge clarity, fragmentation and completeness). In contrast, for images with irregular patterns, the method choice became more critical for the binarization quality (Figure 3), resulting in failing the visual inspection and inaccurate particle size measurements. For example, when UVP5 captured multiple particles within the regions of interest and there was a contrast between particles (with one darker than the other), Otsu thresholding failed to accurately capture the full shape of the lighter-colored particle (i.e. failing completeness) (Figure 3b). It incorrectly detected the dark spot within the light-colored particle and erased the lighter portion, leading to underestimation of particle abundance and sizes- up to 80% smaller when the particle-background contrast was weak (contrast ratio, minimum intensity/maximum intensity > 0.5). Similarly, Otsu thresholding was less effective when CPICS captured particles with highly reflective spots and weak contrast between the particle and background (i.e. failing edge clarity and fragmentation) (Figure 3g), resulting in underestimating particle size - up to 65% smaller when the contrast ratio exceeded 0.8. In both cases, Canny thresholding provided more accurate particle descriptions (Figures 4d, i). In contrast, Otsu thresholding accurately described particles for LISST HOLO2 even when the interference pattern around the particles was prominent (Figure 3k), while Canny thresholding tended to overestimate particle sizes in this instance - up to approximately four times larger - by detecting the noise around the particles (i.e. failing the edge clarity) (Figure 4n).

We further investigated how the different binarization methods impact on the number size spectra by comparing the spectra generated from the sizes measured using Otsu thresholding and Canny thresholding methods (Figures 4a-d-f). The underestimation of weakly contrasted particle sizes by CPICS and UVP5 under Otsu thresholding resulted in aparent high abundance of small particles (MajAL < 0.01 cm for and CPICS, MajAL < 0.1 cm for UVP5) (Figure 4a) which was not observed in the size spectra generated using the Canny thresholding method (Figure 4c). This trend is reflected in the fitted linear regression slopes from the peak of each size spectrum, which were steeper for Otsu (-3.44 for UVP5, -3.91 for CPICS, -5.86 for LISST HOLO2) than for Canny (-2.98, -4.30, and -5.63, respectively), indicating that Otsu detects proportionally fewer large particles. In contrast, Canny thresholding tends to overestimate the size of particles in LISST HOLO2 images when interference patterns remain visible around the particles (Figure 3n). This led to an apparent excess of larger particles (MajAL > 0.1 cm) and absence of small particles (MajAL < 0.01 cm) in the resulting size spectra (Figure 4c).

The influence of the thresholding algorithm on particle size measurements and consequently size spectra, as observed in our results, is particularly surprising given that we are using fairly homogeneous and ‘simple-shaped’ particles with clearly defined edges. This finding suggests the substantial impact of the thresholding method selection on particle size measurements and subsequent interpretations with in-situ biological particles, which typically possess more complex shapes and less defined edges (Giering et al., 2020b). Indeed, the limitations of edge detection, and resulting inaccurate measurements of particle size and abundance, have been reported for measurements of complex-shaped natural particles (i.e. filament, spines), which can have critical consequences when size estimates are converted to ecological parameters such as biovolume (Huang and Bochdansky, 2025; Kraft, 2021; Sosik and Olson, 2007). Although, we selected Canny for UVP5 and CPICS, and Otsu for LISST HOLO2 for the further analysis because these methods performed the most accurate edge detection, these earlier studies on natural particles highlight that routine visual inspection remains crucial to identify the most effective method and reduce misinterpretation in future in-situ work. One effective approach would be to generate a collage of images (raw + binarized, such as Figures 2, 3) for, at least, 5 randomly selected images within each size bin. This visual inspection will help researchers assess whether the chosen thresholding algorithm is appropriate to describe the particle size spectra. Such careful inspection and selection of suitable binarization methods will be even more crucial when measuring the natural particles, such as marine snow, as these exhibit more complex shapes and edges (Giering et al., 2020b; Liu et al., 2023) compared to the uniform particles with well-defined edges used in this study.

The observed particle size across the three instruments ranged from 100 µm to 1 cm, which exceeded the maximum size of the source particles (1 mm) used in the experiment. Visual inspection of the images revealed that these unexpected large particles were due to particle flocculation. Since flocculation is a natural process that can be seen in any aquatic system (e.g. marine snow), we did not exclude these large particles as noise. The size spectra for each instrument (Figures 4, 5) shows distinct peaks corresponding to the most frequently detected particle sizes. The highest raw counts of particles per bin were observed at 0.02 cm, 0.03 cm, 0.09 cm for LISST HOLO2, CPICS, and UVP5, respectively. These bins provide more statistically stable (i.e., less noisy) estimates of relative abundance because they contain more observations, although this does not necessarily imply greater sizing accuracy. While normalized particle numbers typically decrease continuously as bin sizes increase, fluctuations along the size spectra (peaks and troughs) can occur naturally [e.g. through cohorts or high abundance of a particular species (Zhou et al., 2009; Giering et al., 2019)] or because the bin size approached the limit of detection for a given instrument. The latter is often associated with a ‘dropping off’ in the slope towards the extremes of the instrument-specific spectra. Bins containing very few detected particles carry higher sampling uncertainty in their estimated concentrations and should therefore be interpreted with caution, particularly when combining spectra from different instruments. We suggest including the raw particle counts in the number size spectra (as indicated in color in Figures 4, 5) as a potential guideline to validate the reliability of the number size spectra. In our case, this examination method revealed that the UVP5 and LISST-HOLO2 size spectra align best when two different thresholding algorithms are used (Figure 4e).

To further evaluate the effective detection ranges of each instrument, we quantified their relative contributions to total particle counts across five defined size classes (Figures 4b, d, f). Under the Otsu thresholding, LISST-HOLO2 dominated detections in the small size classes (0 - 50 µm and 100 - 300 µm; 84%), whereas UVP5 accounted for nearly all detections in the largest ranges (> 300 µm; >98.6%). CPICS contributed primarily to the intermediate range (100 - 300 µm; 56%), reflecting its optimal detection window. Under the Canny thresholding, the overall pattern was similar: LISST-HOLO2 and CPICS detected the majority of particles below 600 µm (up to 96.1% and 90.7%, respectively), whereas UVP5 was dominant above 600 µm (up to 99.9%). These results confirm that each instrument performed best within its designed operational size range - LISST-HOLO2 for small particles, CPICS for intermediate sizes, and UVP5 for larger particles. Although both thresholding methods produced consistent overall trends, our comparison showed least misinterpretation of particle features during binarization (Figure 3), and the size spectra aligned most closely when the Otsu method is applied to the LISST-HOLO2 and the Canny method to the CPICS and UVP5 (Figure 4e). This indicates that applying the Otsu thresholding to LISST-HOLO2 and the Canny thresholding to CPICS and UVP5 provides the most effective approach in our setup.

We also constructed size spectra using ESD to evaluate how different size approximations influence the size spectra (Figure 5). Since particles are transformed into spherical shapes to obtain ESD, ESD naturally appears smaller than MajAL for elongated particles. The degree of difference between the two can vary with particle shape, and this sensitivity becomes important when analyzing marine particles that have large variations in the shape. While our number size spectra did not show significant differences between MajAL and ESD - likely because our particles were relatively uniform in shape - we observed a slightly higher abundance of small sizes with ESD. The choice between MajAL and ESD should therefore depend on the aim of analysis with natural marine particles. For example, ESD may be more useful when studying volume-based or mass-related processes, while MajAL may be more important when studying the dynamics of particles. In particular, turbulence can cause fragmentation, and the Kolmogorov scale (the smallest eddy scale) can be a key factor in fragmentation (Takeuchi et al., 2019).

Overall, the three number size spectra were well aligned despite variations in the recording system between the instruments (Figures 4e, 5e), demonstrating that multiple optical imaging systems can be effectively combined to generate a continuous size spectrum across broad size ranges. Comparable comprehensive size spectra have also been produced by combining sequentially collected in-situ plankton images (Jackson et al., 1997; Stemmann et al., 2008; Zhang et al., 2023; Dugenne et al., 2024; Lombard et al., 2019; Greer et al., 2020b). A key distinction of our study is that all instruments are mounted on the same frame capturing images simultaneously and ensuring consistent measurement conditions across instruments. Since environmental conditions can alter particle number size spectra, simultaneous deployment is crucial when using multiple instruments to construct a comprehensive size spectrum in dynamic systems. Time scale matters as if the interval between sequential deployments is longer than the time processes that alter size spectra, the resulting conclusions may be misleading. Turbulence, for example, can make change rapidly (on the order of seconds to minutes) and may also drive a significant change during extreme event such as storms. Such change in turbulence can influence the particle dynamics through dispersion and avoidance behavior of plankton (Haury et al., 1990; Incze et al., 2001) (Basterretxea et al., 2020; Durham, 2013) or aggregation and fragmentation of marine snow (Rühl and Möller, 2024). For the longer time scale (several hours to days), biological processes may have a pronounced influence on particle dynamics; such as growth of phytoplankton (Prairie, 2019), zooplankton grazing (Giering, 2014) or zooplankton diurnal vertical migration (Ohman and Romagnan, 2016). Our setup which allows deploying the entire system on a single frame is therefore an effective setup to minimize spatiotemporal variability and produce more directly comparable size spectra from each instrument. On a practical level, the operation of all instruments on one frame makes deployment easier and ensures that the sampling conditions, such as lowering speed and potential ship-based influences (i.e. waves and ship motions) - all of which can affect size spectra through e.g. disaggregation, avoidance and attraction of zooplankton - are the same.