A set of 1000 images produced in two sites with different suspended sediment concentrations was selected to test the noise-reduction algorithms. The images were obtained on November 27, 2015, two days after the collapse of the dam located in Mariana (MG), when the waste with a high concentration of suspended sediment reached the outfall of the Doce River. During the operations of the R/V H39 “Vital de Oliveira”, from the Brazilian Navy, images were recorded in situ through a Light Frame On-Sight Key Species Investigation (LOKI, Isitech, Bremerhaven, Germany). The hauls were vertical from near the bottom with a mesh of 200 μm and an acquisition rate of 20 frames per second. The hauls were carried out at two distinct sites along the plume of sediments in the coastal zone, throughout a gradient of turbidity. The first site (1) was located as close as possible to the outfall of the Doce River, under high turbidity and shallow copepod assemblage (~20 meters deep). The second site (2) was located far from the Doce River opening with less turbid waters and deeper copepod assemblage (~100 meters).

The vertical profile of turbidity (NTU) and suspended particle concentration (CPS, particles·L⁻¹) in the two sites were simultaneously addressed using a turbidity sensor attached to a Multiparameter Datalogger (Horiba) and a Moving Vessel Profiler (MVP) equipped with a Laser Optical Plankton Counter (LOPC). The turbidity sensor was down from the deck up to 20 meters on both sites 1 and 2, and the water turbidity was annotated. The MVP + LOPC was deployed from site 1 to site 2 with seven sequential launches to address the suspended particle concentration in the plume.

The LOKI is a system with 5 main parts: 1) A concentration net, with a mouth opening of 0.28 m² and a mesh aperture of 200 μm, 2) a computer with a solid-state driver to store the images, 3) a CTD with temperature, pressure, dissolved oxygen, and fluorescence sensors, 4) a Prosilica GC 1380H (AVT-Allied Vision Technologies, Canada) with the Pentax 2514-M lens and 5) a set of battery. The raw image has 1360 1024 pixels with a final resolution of 23 μm·pixel⁻¹. It has a high-power LED unit, synchronized with the camera’s exposure-shooting signal, which allows a fast shut-off time (55 μs) avoiding motion blurring that causes image distortion. In combination, it has an image channel 4 mm high (length = 31.3 mm, width = 20.75 mm, volume = 2.6 cm³), causing all the particles of the image to stay in focus. All images are stored in the LOKI solid-state drive, where they can be accessed for further analysis. The recording starts instantly after the power is on, resulting in many waste images gathered either before or after the actual haul. From all images gathered since the equipment started functioning on the ship deck, only those generated during the haul from near the bottom to the surface represent a valid set of images of organisms in the water column and, therefore, were selected for characterization and quantification of copepod populations. The distinction and separation of these images of interest were performed automatically according to the actual depth measured by the pressure sensor. Hence, the onset of the haul was defined as the instant of maximum pressure ( Figure 1 ). The end of the haul, in contrast, includes safety procedures to bring the LOKI back to the ship deck. These procedures produce an oscillation in the equipment due to the waves at the surface that are captured by the pressure sensors and might bias the density estimation near the surface ( Figure 1 ). All images produced during this oscillation, therefore, were removed to avoid counting the same particle twice.

After sorting the valid set of raw images, we performed 1) automatic recognition of the region of interest (ROI) and cropping, 2) background removal, and 3) counting and measurement of organisms ( Figure 2 ). A vignette is a smaller image cut from the original raw image and containing a prominent particle. To check if the counting and measurement steps are less accurate in places with high turbidity, thus biasing the estimation of copepod size and abundance, simultaneous manual and semi-automated procedures were compared. To perform the automated measurements of organisms surrounded by distinct amounts of suspended sediments, a subset of 1000 vignettes were randomly selected, after which all vignettes with a single copepod (singlets) were manually counted and sorted into a different folder. This subset of copepods was processed in the 16 different binarization algorithms available in the ImageJ software (2.3.0) [9] . The accuracy of each algorithm was calculated by the comparison of the counting made manually by the taxonomist and automatically by the ImageJ contour detection (function “Analyze particle”, ImageJ). The same approach was adopted for morphometric measurements of the copepod’s body (e.g., area, perimeter, and fitted ellipses). The abundance was calculated by the percentage of singlets correctly identified by the algorithms.

To improve the accuracy of copepod enumeration and measurement, three additional algorithms based on convolution matrices (herein named algorithms 2 to 4) were elaborated with different sequential steps of noise reduction done before the binarization. To evaluate the constancy and linearity of the results, the algorithms were tested with 3 subsets of 1000 randomly selected images from each site. Algorithm 2 includes an increase in contrast before the binarization step to accentuate details in the target ROI by applying a kernel to replace each pixel with a weighted average of the 3 × 3 neighborhood:

$\begin{array}{ccc}-1& -1& -1\\ -1& 12& -1\\ -1& -1& -1\end{array}$

The third algorithm, in contrast, reduced the exposition of the vignettes to check if noise could be removed before the binarization step to let only those pixels from a target ROI be left in the vignette. The reduction of exposition was achieved by reducing each pixel value by 25 (10% of the full 255 grayscale). Pixel values between 0 and 24 were set to zero. The fourth algorithm applies the “Statistical Region Merging” algorithm (Nock and Nielsen 2004) to highlight the copepod from the surrounding suspended particles (“noise”).

Each vertical haul lasts for nearly two minutes, starting approximately five minutes after the LOKI is powered on ( Figure 3 , red dotted rectangle). In site 1 ( Figure 3A ), the equipment was towed over nearly 20 meters (pressure ~260 mbar) for 2 minutes and a half, with an oscillating minute in the surface (8' - 9'). In contrast, the net in site 2 was quickly towed for 100 meters (5× longer) and only 2 minutes, with no oscillation near the surface. The tow speed of the first haul was estimated as 0.12 meter·sec⁻¹, while the second haul reached 0.9 meter·sec⁻¹.

As expected, the water turbidity was extremely different between sites, mainly near the bottom where the plume of suspended sediments was more intense ( Table 1 ). Therefore, the number of particles in suspension near site 1 (>5000) was 5 times higher than that at site 2 (<1000 CPS). In total, 40,375 vignettes were produced under high turbidity conditions, contrasting to only 21,664 under low turbidity conditions. By considering the different depths and operating times at each point, more vignettes were obtained in shallower and highly turbid waters (2018 vignettes·m⁻¹) than in deeper ones (217 vignettes·m⁻¹).

The performance of the LOKI system to automatically recognize an ROI in a raw image and crop the corresponding vignettes was significantly affected by the turbidity and concentration of suspended particles ( Table 2 ). From 1000 randomly selected images obtained under high turbid waters ( Figures 4A-J ), the majority (96%) were either badly segmented or non-segmented ROI ( Figure 4B ) that needed additional processing steps. In contrast, under less turbid waters (site 2), nearly 99% of the stored images were correctly segmented singlets ( Figures 4K-R ).

The size of the ROI cropped reflected the smaller particles dominating site 1 with high turbidity. Nearly 60% of the random set represents vignettes with 1 to 10 kb ( Figure 5A ), while those from site 2 predominate in the 10 - 50 kb ( Figure 5B ). Full images resulting from failed segmentation have 1.2 Mb and were more frequent under high turbidity.

The presence of suspended sediments can be seen in vignettes from site 1 as white dots around the copepods ( Figure 6A ). In contrast, copepods from site 2 with less suspended sediments have less “noise” and the resulting histogram of grayscale values is less skewed to the right ( Figure 6B ). Nevertheless, the presence of “noise” in one vignette is revealed by the differences between modes, rather than the mean pixel value ( Table 3 ). Both pictures have very similar means (50 and 56), while the vignette with more suspended particles exhibits a surprisingly darker mode (17). In both cases, the standard deviation was higher than the mean.

The resulting number of copepods according to the four automatic approaches revealed percentages of accuracy ranging between 90% and 507% ( Table 4 ). A total of 678 and 1038 copepods were registered from a set of 1000 randomly selected images, respectively from high and low turbid waters. The abundance was either highly overestimated (>300%) or underestimated in turbid waters due to the presence of sediments when either the contrast is increased, or the adjacent mean pixel value (Statistical Region Merging) is used as pre-treatment. Despite underestimation, the most efficient method for counting organisms under highly turbid water was just binarization using “the first higher average” algorithm (or Default in ImageJ) without pre-treatment or by performing a slight reduction of exposition to remove some noise. The abundance under high turbidity was 5% to 10% underestimated (90% to 95%), while under low turbidity it was 13% overestimated (113%). Such results are observed in ( Table 4 ) and the binarization evident in ( Figure 7 ).

The use of distinct binarization steps without pre-treatment to reduce the noise in turbid waters also resulted in different estimates of abundance according to the algorithm used. The total abundance was again underestimated in turbid waters while under less turbid ones it was occasionally overestimated ( Table 5 ). The three best algorithms for high turbid waters (Huang, Intermodes, and Li) estimated nearly 88% of the actual abundance. As expected, higher efficiency (>95%) occurs when addressing the abundance of copepods from low turbid waters. There was no coincidence in algorithms with the best performance under low and high turbid waters, except for the “Intermodes”.

The surrounding particles that might be miscounted as a copepod were evidenced in vignettes with increasing contrast ( Figures 8E-H and Figures 8M-P ) but smoothed in vignettes with high degree of erosion ( Figures 8A-D and Figures 8I-K ). An apparent good contour of copepod, with full antennae preserved and proportionally less surrounding particles was achieved after the “Huang” binarization algorithm ( Figure 8B ).

From a dorsal view, the average area of a copepod (prosome + urosome + antennae) as manually estimated by visual analysis was equal to 312,639 μm² (mean = 591 pixels² ± 369 pixels²). Most of the automatic algorithms tested herein resulted in a good approximation of the actual size, as demonstrated by the high linear correlation (R² > 80%) ( Table 6 ). The “Renyi Entropy” algorithm had the highest correlation (R² = 0.895, Figure 9M ).