Artificial reef effectiveness changes among types as revealed by underwater hyperspectral imagery

Artificial reefs (ARs) are designed to mimic natural habitats and promote marine life. Their effectiveness is however debatable and can depend on factors such as structural complexity and construction material. Old artificial reefs (OARs) were made of concrete mold of simple geometric shapes, limiting their ability to mimic the complexity of natural reefs. Recent advancements in three‐dimentional (3D)‐printing technology have enabled the creation of 3D‐printed artificial reefs (3DRs) with biocompatible material and complex structures that can better simulate the natural habitats. We employed underwater hyperspectral technology to estimate the performance of these reefs and compare the benthic photosynthetic signal of natural reefs (NATs) with those of ARs (OARs and 3DRs) in coastal area of the north‐western Mediterranean (France and Monaco Principality). We expected differences in reflectance signals between OARs and NATs, and signals closer to NATs in 3DRs than OARs. Underwater hyperspectral technology was able to detect higher chlorophyll‐a derived signals on NATs than OARs. Moreover, the magnitude of differences between 3DRs and NATs was smaller than that between OARs and NATs. Although ARs were not capable of mimicking natural reefs, the use of 3D‐printed ARs might ameliorate their effectiveness for coastal reconciliation.


Introduction
Artificial reefs (ARs), which are placed on the seabed to mimic some characteristics of natural reefs (NATs) (Pickering et al. 1998), are nowadays used to ameliorate the ecological conditions of key habitats (Colsoul et al. 2020).They should thus be constructed based on the ecological knowledge of the natural system, respecting natural habitat complexity and using substrates made of inert materials (UNEP 2009).So far, when used as "natural surrogates," old-generation of ARs, mainly built of concrete casting of simple geometric shape, have shown to support less diversity than adjacent natural habitats (Ferrario et al. 2016;Hill et al. 2021).The three-dimentional (3D)-printing technology can contribute to improve this use of ARs, by enhancing design complexity and supporting the use of biocompatible materials, as recently observed for coral reefs (Levy et al. 2022).Such a structure may be more effective in facilitating the settlement of communities that better mimic the characteristics of natural assemblages than the old generation of ARs.
The effectiveness of ARs in enhancing marine life has been mainly evaluated using fish monitoring (Paxton et al. 2020).Monitoring benthos is, however, fundamental since ARs are "fouling objects" and their benthic assemblages are expected, over time, to converge toward natural communities (Svane & Petersen 2001).Benthic communities are sampled by scraping or using photoquadrat, which are destructive and/or time-consuming methods.Recent advances in underwater hyperspectral (HP) imaging could be valuable for greatly improving fast data collection nondestructively.HP imaging allows integrating spatial information of an object with its spectral characteristics, by combining imaging and spectrometry (Liu et al. 2020).It has been widely used for monitoring terrestrial and shallow coastal habitats since the early 2000s (Khan et al. 2018) and, more recently, underwater, as an emerging tool for coral reefs (Chennu et al. 2017) or deep-sea benthic megafauna (Dumke et al. 2018).
This case study investigates how photosynthetic benthic assemblages differ between natural and artificial reefs types using underwater HP imagery to detect these putative differences.We expected that underwater HP imagery would detect the spectral signature of photosynthetic communities and discriminate between natural and artificial reefs and between AR types (old-generation and 3D-printed reefs).We also expect that the magnitude of differences between natural reefs and artificial reefs will be reduced using 3D-printed reefs.

HP Camera
We used the DiveRay camera, a dive operator prototype version of a commercial hardware solution provided by planblue GmbH (https://www.planblue.com/)to scan the seafloor.It includes red-green-blue (RGB, resolution: 2448 Â 2048 pixels) and HP (resolution: 680 pixels across track resolution, Field of view of approximately 30 , spectral resolution: approximately 2.9 nm) cameras, equipped with two modular lights (intensity: 4215 lm each).Distance to the seabed is derived from stereo camera measurements.Depth, position, and velocity are derived from a combination of sensors, including a Doppler velocity log measurement, a magnetometer, and an inertial measurement unit.This information is available on the display screen of the DiveRay in addition to the view of the scanned habitat (in HP or RGB format as required).The DiveRay is neutrally buoyant underwater, making it extremely flexible to handle, allowing the camera to be oriented in any direction the diver wishes.A reference board (a gray-level plastic board that reflects 55% of the light) must be placed at the beginning and end of each transect for data processing afterwards.When recorded at the beginning and end of the transect, and at similar distances under consistent artificial lighting conditions, it serves to compensate for variations in natural light penetration into the water column for each transect scanned.After each dive, data are transferred from the DiveRay to a computer.The data are composed of RGB, HP, and stereo (3D) video files, and geo-positional data.

Study Location and Sampling Design
The sampling took place during the last 2 weeks of June 2020 along the northern-western Mediterranean coastline.We sampled within urban marine protected area (MPA) of the Principality of Monaco (MON, in Fig. 1) managed by the Monegasque Association for Nature Protection (Debernardi & Allemand 1993), and within the MPA of Roquebrune Cap Martin (RCM, in Fig. 1) managed by "Service de l'ingénierie environnementale du Département des Alpes Maritimes."ARs have been immersed for several years in these MPA for reconciliation measures and fishing support (Allemand et al. 2000;Tessier et al. 2015).We sampled natural reefs (NATs) as control, old artificial reefs (OARs), and the 3D-printed artificial reefs (3DRs).The latter was present only in MON area and therefore was sampled (Fig. 1).The ARs were at comparable depths (30 m), and orientations, in similar conditions with regard to northern Mediterranean current as the natural reefs, and at maximum 2 km apart (see details on sampling effort and reef details in Fig. 1).
Scuba divers recorded the vertical surfaces of the ARs because they were oriented similarly to the natural reefs.Measurement were taken with the DiveRay camera at a constant distance, maintained by the diver, about 80 cm from the reef scanned.Recorded transects were approximately 42-cm wide and 6-20 m long as a function of the circumference of the reef unit, and we ran 6 m linear transects on the natural reefs (Fig. 1).

HP Data Processing
To generate HP data's true-and false-color representations, we used a program delivered by planblue.We then annotated regions of interest (ROIs) for each scanned transect: including the two reference boards recorded at the beginning and end of each transect, and ROIs of benthic community.We were careful to avoid void spaces of the structure.All annotations are stored in a database collecting spectral, spatial, and environmental data.These recorded spectra of each ROI are standardized to unit range and divided by the average spectrum of the reference board recorded at the beginning and end of each transect.That way, the effects of light conditions on different recordings can be accounted for.Among the 302 annotated ROIs (all types of reef), we randomly selected 30 ROIs for each reef type to standardize our sampling effort.We standardized their size by 12 pixels in width and 444 pixels in length, corresponding to the reference board size (25 Â 25 cm).

Wavelength Selection
In our study, we used the second derivative (R 00 ) of the reflectance signal to assess the presence of photosynthetic pigments.We employed two methods: (1) converting positive and negative R 00 values into a binary presence/absence as a cover proxy and (2) using the positive values of R 00 as a semiquantitative proxy for pigment intensity.The second derivative spectrum was computed from the reflectance spectrum for each pixel in each ROI with the Savitzky-Golay algorithm (Savitzky & Golay 1964).
The different steps of the HP data process were done on Python (3.6.15) and are summarized in a diagram (Fig. S1).

False-Positive Photosynthetic Signal
To ensure the accuracy of data processing and WL identification, we visually inspected all HP images in conjunction with the RGB images from ROIs.We also checked if the WLs we associate with pigments of photoautotrophs could be linked to other possible signals in the spectral data.The 567-nm WL that was associated with the presence of PhycoE pigment occured not only in red algae but also in the sponge Crambe crambe (Fig. S2).Moreover, both PhycoE and PhycoC can be present in microbial cyanobacteria biofilm that can cover every organic or nonorganic surface (Callow & Callow 2006;Salta et al. 2013).Therefore, we were cautious with our interpretation regarding these WLs and used ratio of Chl-a/PhycoE and Chl-c/PhycoE to cross check out results.An example of this cross-check evaluation is available in Figure S2.

Statistical Analysis
We ran all analyses using R v3.4.1.We fitted generalized linear models (GLMs) on the mean signal intensity of pixels for each selected WL in each selected ROI.To quantify the differences between reef types (old artificial and natural reefs; hereafter OARs and NATs), we combined reef types and areas (MON and RCM) in a new factor, which we named "Habitat".We did this because we observed large variability between these areas, which could mask differences between AR types (see results in Figs. 2 & 3).The factor "transect" was nested in "Habitat."We expected significant differences between OAR and NAT situated both in MON and RCM.To estimate the Table 1.Characterization of green, brown, red algae, and cyanobacteria according to their pigment content related to their associated wavelength in our hyperspectral system (+ means the presence of the pigment, while À means its absence).Green algae differences between the 3D-printed reefs (3DRs), the OARs and the NATs on MON, we used the factor "Reef type" (NATs, OARs, and 3DRs) and the factor "transect" nested in "Reef type."The factor transect was random, except for multivariate data as it was not implemented in the package.Multivariate data (mean R 00 for each WL) were analyzed after square-root transformation using the function manylm from the "mvabund" package (Wang et al. 2020).These data were represented by an nMDS to visualize the dissimilarity of the HP signal using Bray-Curtis distance with the function vegdist from the "vegan" package (Oksasen et al. 2022).Percentage dissimilarity between factors is also shown in the Results section.
For the univariate data, we used the glmmTMB function from the "glmmTMB" package (Brooks 2022).The family distribution was selected after visually examining the scaled residual using the simulateResiduals function from the package   S1 and S3.

Wavelengths Selection and Global Signal
Analysis on each ROI showed that the selected WLs typical of photosynthetic pigments had in average positive signals, except for 648 nm related to Chl-b that showed negative signal on the  majority of ROIs (Fig. 2).On all reef type scanned the signal of 666 nm for Chl-a was the highest, followed by 567 nm for Phy-coE, 635 nm for Chl-c, 620 nm for PhycoC, 546 nm for Fuco and 648 nm for Chl-b (Fig. 2).
Based on the presence/absence signals for each pixel of the selected ROIs, both natural and artificial reefs were almost entirely covered of Chl-a (90-100%), whereas the other pigments showed a variable cover.PhycoE covered between 63 and 95% of the surface, Chl-c 57 and 88%, PhycoC 57 and 86%, Fuco 70 and 81%, and Chl-b 4 and 41% (Table 2).
These differences were also found in univariate analyses of the HP signal.The GLM(M) models showed significantly different signals among habitats, except for the Chl-b signal (Table 3).Pairwise a posteriori tests showed the lowest signal of Chl-a, Chl-c, Fuco, PhycoC, and PhycoE on the OAR immersed nearby the MON coastline (Table S2; Fig. 4).In addition, although this OAR showed the lowest signal associated to Chl-b (666 nm), the 567 nm signal was comparatively high and visual observation showed numerous sponges of the species Crambe crambe.The NATs in MON had higher signal of PhycoE but lower signal of Chl-c than the NATs situated in RCM area, where OARs showed lower PhycoE signal than the respective NATs (Table S2; Fig. 4).
Each response variable differed significantly among reefs, except Chl-b (Table 4).Pairwise a posteriori tests showed that the OARs had less Chl-c and Fuco signals than both NATs and 3DRs, whose values were similar (Table S4; Fig. 4).Finally, the 3DRs showed intermediate values between the OARs and NATs of Chl-a, PhycoC, and PhycoE signals (Table S4; Fig. 4).

Discussion
The photosynthetic assemblages detected using underwater HP imaging showed a signal of red and brown macroalgae, followed by green algae, as typical in coralligenous habitats (Ballesteros 2006).Visual cross-identification was difficult due to low light conditions and the small size of some of the organisms.Improved camera resolution and light input would allow automatic species identification at a finer scale using morphological features observed in the RGB images associated with each HP image as it has been done in other study using artificial intelligence for identification (Chennu et al. 2017;Dumke et al. 2018).In addition, further calibration using ground truth sampling would distinguish signal spectra typical of nonphotosynthetic organisms.For these reasons, we acknowledge the limitations of our method in revealing spectral signatures to quantify and identify the photosynthetic community in detail.Nevertheless, our results highlighted relative differences in the signal of the photosynthetic community on different habitats and reduce human and time effort required for underwater observations.In our study, one diver could sample 24 transects on five different sites at 30-m depth over 3 days.We believe that our pioneering study will encourage further development of this approach.
The results of our comparative study, based on HP signals and their relative differences, distinguished different uni-and multivariate signals between natural and artificial reefs (NAT vs. OAR).Within each study area, the dissimilarities among OARs were greater than those among NATs, suggesting greater variability among ARs.In particular, the old artificial reef in Monaco Principality showed a high signal at 567 nm, which could not be solely attributed to the presence of red algae, as it exhibited the lowest signal associated with Chl-a at 666 nm.Visual observations of this AR revealed significant coverage by the sponge Crambe crambe, which shows a signal at 567 nm.In contrast, in the RCM sampling area, the OAR had signals typical of coralligenous reefs and showed some similarities to the signals measured on the natural reef.However, the OARs in RCM had a weaker signal for red algae compared with the natural reef.Thus, our analysis not only confirmed the differences between natural and artificial reefs, in line with previous research (Perkol-Finkel & Benayahu 2005;Carvalho et al. 2013;Ferrario et al. 2016), but also highlighted that the  S3) for each WL for second derivative of the reflectance signal (R 00 ) of photosynthetic pigments data using Gaussian family distribution.magnitude of these differences varied depending on the shoreline.

Variable
It is known that the complexity of the structure and the texture of the material used to construct ARs can influence colonization patterns (Bohnsack 1991;Baine 2001;Riera et al. 2018), and that ARs are considered mature after 5 years of immersion (D'Anna et al. 2000).The studied old artificial reefs were similar in design and material, and were both immersed for over 30 years.During the field campaign, we carefully sampled habitats that were no more than 4 km apart, at comparable depths (30 m), and with similar orientations to the light and the Mediterranean Northern Current.However, the coastal area of the MON has undergone extensive development, with only 11% of the original natural coastline remaining (Bottin et al. 2022).Waters close to urban areas with elevated coastal development receive high inputs of organic matter (Pooters 2013), which have been shown to promote a shift from the dominance of macroalgae to that of filamentous or other ephemeral algae, and filter feeders (Airoldi & Beck 2007).The pattern observed in our study could therefore indicate an effect of increased particulate organic matter, on macroalgal community, related to coastal development and urbanization.However, the lack of replicated coastline conditions prevents us from inferring general patterns, and further studies replicating coastlines with comparable ARs are needed.
Regarding the effect of reef type, the 3DRs were not able to fully mimic the signal of the photosynthetic benthic assemblages of the natural adjacent rocky reefs (NATs).However, the HP signal of 3DRs was closer to that of NATs than OARs.It showed higher signal of photosynthetic community related to brown algae (666, 635, and 546 nm) compared to the OARs.However, there was still a low signal of red algae (567 nm) compared to NATs which can be explained by a slow growth of coralligenous algae (Ballesteros 2006).Research on the use of 3D-printed ARs for enhancing marine life is still in its infancy, but our results suggested a slight positive increase in ARs' effectiveness compared to old artificial reefs.Monitoring over time is however necessary to determine whether the benthic community they support will be able to converge toward the natural ones.However, our results are promising and in agreement with the studies on fish community (Pérez-Pag an & Mercado-Molina 2018; Ruhl & Dixson 2019) or on substrate comparison for the settlement of oyster, coral, or primary biofilm (Riera et al. 2018;Ly et al. 2021;Matus et al. 2021), suggesting that this approach may help to facilitate the colonization of benthic communities typical of natural reefs on ARs, to improve the success of reconciliation measures.
In conclusion, this pioneering study in the use of HP technology on underwater photosynthetic communities of natural and artificial reefs shows that, after decades of immersion, ARs are still not able to mimic adjacent natural habitats.In addition, the use of reefs (such as 3D-printed ones) that can mimic the complexity of natural reefs and made out of biocompatible materials appears promising to improve the global success of reconciliation measures.Nevertheless, this study is an important step toward developing HP imaging as a useful tool for mapping underwater photosynthetic communities and reducing the human and time resources required for such efforts.
(Fox & Weisberg 2022) likelihood ratio tests to test the most parsimonious model; e.g.ANOVA function analyses with chi-squared test from package "Car"(Fox & Weisberg 2022)and tested for significant factors for the most parsimonious model.Post hoc analyses were done using pairwise t tests from the R "stats" packages.Information on family choice and model parameters are available in Tables

Table 3 .
Summary of results of the analyses of deviance (ANOVA type II Wald χ 2 tests) on the fixed factor "Habitat" for the selected GLM(M) models (see details in TableS1) for each WL for second derivative of the reflectance signal (R 00 ) of photosynthetic pigments data using Gaussian family distribution.

Table 4 .
Summary of results of the analyses of deviance (ANOVA type II Wald χ 2 tests) on the fixed factors "Reef type" for the selected GLM(M) models (see details in Table