PER ALAN Methods
The PER ALAN dataset was generated while preforming field experiments assessing the impacts of artificial lighting at night on predator density and salmonid predation in the Sacramento - San Joaquin Delta (Nelson et al. 2020). Predation event Recorders (PERs) are independent drifting GPS-enabled platforms baited with a tethered hatchery-origin live Chinook Salmon smolt at 1m depth. Details of PER construction and use may be found in Demetras et al. 2016. Our PERs were slightly modified versions of those described in detail in Demetras et al. (2016). Specifically, we constructed PERs with 5.08-cm-diameter clear PVC pipe with the majority of components (GPS, timer, reed switch) contained within the PVC housing and sealed with a rubber end cap. In the field, we record the PER deployment and retrieval time on laptops. Each PER contains a GPS puck that also records GPS positions every 5 seconds. Finally if the predation timer is pulled in a given drift we record the timer time to know the exact time of predation. Upon retrieval, we also record whether a smolt is gone, or has noticeable bite marks. Furthermore, any PER deployment that gets tangled in SAV or with another PER is noted and removed prior to data analysis. During this experiment, we also deployed Hyrolabs that recorded abiotic parameters every minute.
After all PERs were retrieved, submerged aquatic vegetation (SAV), depth, and light intensity (light experiments only) within each experimental reach were surveyed. These surveys were conducted from a motorized vessel and transects were driven along the inside guide line, the middle of the reach, and outside line. During each survey, bathymetry and SAV were mapped using a Humminbird Helix 10 SI GPS and side scan sonar set to a consistent depth (0.6 m) and frequency (455/800 kHz). Light intensity (lux) was measured with an ILT2400 optometer fixed on the side of the vessel and all transects were conducted with that vessel side facing the artificial light source. Light attenuation with depth was also measured directly parallel to the light source at both the inside and outside guidelines. Starting at the surface, the optometer was lowered at 0.5 meter intervals until the bottom or 4 meters was reached and held for 1 minute at each unique depth. During this minute, an average lux value was recorded by the light meter each second and the median of these average values was used as a discrete depth value. To account for variation with distance from light source and inherent variability among nights, lux at depth was standardized by dividing the value at each depth by the surface value from each cast. This value was considered light attenuation (At) and was modeled with the following exponential decay equation,
At = e^(-(kd*depth + kt*turb))
where kd (attenuation with depth) and kt (attenuation with turbidity [turb]) were allowed to vary as a function of light type (l) and attenuation could be fit with both coefficients or only kd. The model with the lowest AICc or a less complex model with a ∆AICc <2 was selected as the most parsimonious and used to predict lux at 1 meter; the approximate depth where tethered smolts drifted during experiments.
Humminbird PC, Autochart, and GIS software were used to process SAV and depth data from the side scan sonar. The Topo to Raster tool in ArcGIS 10.5.1 was used to interpolate point depth data into a raster to create bathymetric maps of each study site. This tool was chosen because it produces hydrologically correct digital elevation models. Each PER GPS position was assigned a depth value from this raster and if any PERs drifted outside the raster bounds they received the depth value of the closet cell. To delineate where SAV was present within experimental reaches, the side scan sonar data was uploaded into SonarTRX Pro software. This software corrected for beam angle, balanced beam variation, and maximized contrast and slant range, which removed blind spots from side scan data (Chang et al. 2010). After processing, side scan data was converted into a format compatible with ArcGIS 10.5.1, and SAV was outlined and digitized to delineate the presence and extent of SAV within sites.
Lux data from light survey transects were interpolated across the experimental reach within survey bounds and assigned to each PER GPS position. Exponential ordinary kriging was implemented with the autoKrige function in R. This function generated an exponential variogram from each survey and automatically selected the values of nugget, range, and sill that resulted in the best fitting model. Using this model the value of lux was interpolated over a 500,000 cell grid for each night, which resulted in smooth fine scale lux values across the experimental reach.
Once field work was complete, we merged many separate data sources to get our final dataset and all data processing was completed in the statistical software R. First and foremost, we assigned each PER deployment a unique identifier for each night of sampling. Then, we assigned the start and end time of each deployment to each unique identifier. We used these start and end times to assign GPS points from the unique puck within each PER during deployment. After this step, we inspected each deployment to ensure that it was a true deployment and not a data entry mistake. Bad deployments were evident by GPS positions that did not move throughout the deployment, or had tracks that indicated they had been picked up and moved back upstream. We removed any unique deployment that was flagged during this review. For each GPS position recorded throughout PER deployments, we created unique rows in the data and a time 1 and time 2 column, to identify the start and end of each sampling interval in seconds. To assign predation events to the correct time and position, we subtracted timer time from deployment time when the timer was triggered, and a 1 was placed in a new predation column in the data. Prior to analysis, we removed any PER deployment data after predations. We assigned abiotic parameters to each PER deployment row by taking the data from the hydrolab with the closet time to the PER timestamp for a given row. To generate distance moved in a given interval we measured the distance between the current and previous GPS positions, and dived this by time to get PER speed. We calculated the distance of each PER to the shoreline with a point to line function and used a point to polygon function to obtain the distance to the closet SAV from our SAV surveys. To calculate sunset time, we used the sunset time function and time past sunset was calculated by taking the difference of a given time stamp and sunset. We obtained the water depth sampled by each PER from an extrapolated raster based on site depth survey transects and the cv of depth was the coefficient of variation for depth across the entire site. We obtained light intensity (lux) by assigning the closest lux value from our interpolated light raster and calculated lux at depth for this positon using the above equation. Finally we obtained mean large fish density between light and dark treatments on a given sampling night from our adaptive resolution imaging sonar (ARIS) and ARIS methods are described below.
Nelson, T. R., C. J. Michel, M. P. Gary, B. M. Lehman, N. J. Demetras, J. J. Hammen, and M. J. Horn. 2020. Effects of artificial lighting at night (ALAN) on predator density and salmonid predation. Transactions of the American Fisheries Society https://doi.org/10.1002/tafs.10286.
https://afspubs.onlinelibrary.wiley.com/doi/full/10.1002/tafs.10286
Demetras, N. J., D. D. Huff, C. J. Michel, J. M. Smith, G. R. Cutter, S. A. Hayes, and S. T. Lindley. 2016. Development of underwater recorders to quantify predation of juvenile Chinook salmon (Oncorhynchus tshawytscha) in a river environment. Fishery Bulletin 114(2).
https://spo.nmfs.noaa.gov/sites/default/files/demetras.pdf
Chang, Y.-C., S.-K. Hsu, and C.-H. Tsai. 2010. Sidescan sonar image processing: correcting brightness variation and patching gaps. Journal of marine science and Technology 18(6):785-789.
ARIS ALAN Methods
A pair of ARIS 3000 sonars (Sound Metrics (Bellevue, WA)) were utilized to monitor the study sites. The ARIS unit utilizes 128 beams, with each beam having an angle of 0.25 degrees in the horizontal and 14-degrees in the vertical, producing an overall 30-degree by 14-degree beam and were operated at a frequency of 1.8 Mhz.
ARIS sonars were deployed on aluminum tripods resting on the river bottom (Figure 1). Tripods helped weigh the units down in heavy current, protected the sonar from damage and facilitated deployment and removal. Each sonar head was mounted to a rotator to allow aiming once the systems were placed on the river bed. An integrated rotator was used for one ARIS, while a Remote Ocean Systems PT10 rotator was used to aim the other. Selection of the style rotator was a function what each piece of equipment was provided with, not what was optimal. Power and data cables for the units were allowed to hang slack across the river bottom, then run up the river bank to the surface control unit. ARIS units were placed in 1.5-2m of water and aimed horizontally into the water column at a depth smolts suspended from PERS would be exposed to potential predators. Range for each was set to be about 10m with the units operating in low frequency mode (1.8mhz).
For the unit utilizing the PT10 rotator, cabling was run up the bank to a large plastic job box. Within the job box were placed the sonar control unit, sonar power supply, rotator controller, and laptop computer utilizing Windows 10 and ARIScope software ver 2.7.3. An uninterruptable power supply (UPS), was also located in the job box and provide uninterrupted power during short power fluctuations which occurred quite often. The unit with integrated ARIS rotator control was placed with a laptop and UPS in a small plastic storage box in a similar fashion. A Yamaha 2000 watt generator was placed midway between the two units and extension cords run to the surface control hardware for each unit. Once started, the sonar units collected data continuously throughout the study period. Software parameters were set so that in the event of power loss to the sonar data collection would resume automatically on power up. Data were stored as 30 minute files to external hard drives attached to each computer.
Data processing
Although the ARIS sonars do have an apparent vertical component to the beam, the data obtained with the vertical component are not available for multibeam analysis. For this reason, all data output is 2D (range and left-right positioning in the beam). It is possible to account the volume sampled within different ranges from the units to calculate average target density over time.
Acoustic processing ARIS data included the following series of steps, resulting in a final output and was all completed using Echoview version 10.2: Steps 1-5 were done in one processing file with step 6 completed by creating a separate processing file in the interest of analysis efficiency and step 7 done outside of echoview.
(Step 1: Raw data examination) Raw data examination determined where obvious potential processing problems arose and helped determine which techniques might prove most useful for analysis. Based on previous experience, files had to be deleted from analysis when partial shading of the image occurred because of movement of debris such as submerged aquatic vegetation (SAV), incorrect aiming of the sonar, or when excessive background interference limited the utility of the data (Figure 2).
(Step 2: Excess noise and background data removal) Excess noise was not true noise, but instead refers to portions of the acoustic signal outside the realm of targets of interest. For this study targets of interest were fish, or fish sized targets. Noise removal for the most part thus consisted of removing the portions of the sonar image that were not relevant to the analysis, typically the fixed bottom or structure images. The multi-beam background removal operator available in Echoview was selected for this study over the use of CSOT files generated using the ARIS software as it performed significantly better when removing unwanted targets and provided more options for optimizing the data. This technique did not necessarily assume a static background, but allowed for some slow movement of the background. For this study a mean image of a surrounding number of specified pings was developed then subtracted from the current ping. Thus, slow-moving objects (e.g., slowly moving vegetation), or stationary objects would not be detected as motion because the previous pings were only fractions of a second apart. We could also specify a signal to noise ratio to further refine what was excluded from the echogram. This technique was effective, and was used directly in Echoview though processing time was slow because of the large file sizes, and the need to calculate a data subtraction for each ping. Other patterns of unwanted noise were introduced by plumes of fine sediment passing through the sonar beam could not be removed using a background filter. Sediment and other small false targets were minimized in a later step via target size filtering.
(Step 3: Smoothing) After the background was removed from the multibeam echogram, an additional step was employed in Echoview to improve target detection. This involved using a median operator, which helped smooth the image without imparting any serious changes to future analyses. Adjusting this smoothing variable allowed for greater or less target definition. For smaller targets, proper operation of target recognition can be difficult and smoothing significantly helped increase target resolution and detection.
(Step 4: Multibeam Target Detection) The next phase was to detect multibeam targets and to filter targets according to an initial length criteria. Again, a series of target criteria were tried until it visually appeared the software was correctly identifying most fish targets. In Echoview a target overlay image was created to visually examine how well the multibeam operator was tracking targets. (Figure 3) This was only a minor consideration for larger targets; changes in parameters primarily affected small targets close to the lower limits of the detection parameters. To filter targets at this stage, the minimum acceptable fish length was identified, and all targets below that threshold were excluded. To determine the best settings for false target removal, files were randomly selected and then manually replayed and notes made on times when fish were present. To parameterize Echoview to detect small fish without introducing undue levels of background noise, files were identified that had small fish, as well as those having forms of noise present. Adjusting the initial data threshold was useful for removing most very small targets, after which parameters were adjusted in the multibeam and single targets algorithm until visual analysis of the echogram indicated performance was acceptable. Small debris, which when present, could very quickly overwhelm small fish and reduce the utility of automation. In this case 5 cm was selected as the smallest target we could expect to reliably detect given the sonar settings. At sizes smaller than 5cm a variety of false targets, particularly bubbles, sediment plumes and small debris started appearing and quickly swamped the echogram. Once final filter parameters were set they were not changed across all study dates and sites. Any change in the filter does impact the number of targets observed and could produce miss-leading results in the final analysis if the same set of parameters were not used uniformly across the data set.
(Step 5: Target conversion and export) Target conversion is a one step process in Echoview which takes a multibeam target generated from the sonar image and condenses the information into a single summary target. This gives the appearance of a typical echogram one might see on a split-beam or boat mounted echosounder (Figure 2). To allow enumeration of target numbers and sizes, and to allow Echoview to identify fish traces and provide data suitable for analyses, multibeam data then were converted to a single-target echogram. The actual length of the fish as detected by the multibeam target selection criteria was used as a substitute for target strength at this point. The single target echogram, consisting of target length and other multibeam descriptors was then exported as a comma separated variable file (CSV).
(Step 6: Fish track detection and track spreadsheet summary creation) To develop potential fish tracks the exported CSV file of targets was imported into a new file set in Echoview and further thresholding and fish tracking was completed on this new file set. Tracking could be done in the original Echoview file set, but took orders of magnitude more time to process, as all ping calculations were redone for each step. All fish tracking with ARIS data occurred without the minor axis component (used to calculate depth of fish in the water column) as multibeam data is only two dimensional (range and left-right), thus reducing one source of positioning error. Additionally, because multiple beams were used to reference the target, the result was accurate positioning within the beam, and the default parameters in Echoview yielded acceptable tracking results. With better positioning data, fish tracks also were likely to be better defined because the position accuracy allowed the tracking algorithm to perform better. When developing size criteria as mentioned earlier initially a 5cm filter was applied to all targets, however, visual examination of the data indicated small debris generally overwhelmed the dataset thus more restrictive size filters were to reduce inclusion of unwanted targets (Figure 5). As shown in figure five, increasing the filter size reduces the number of targets for a potential fish track. Removing smaller targets would increase the average size of a track, however, we used only the maximum value for each track as a measure of size so removing small values did not impact final size estimates.
(Step 7: Separating fish from debris) Twenty cm was selected as a target filter to exclude any fish too small to be a likely predator on smolts. Fish, or target tracks, were only considered if there were four or more targets that could be added to the trace using a 20cm target filter. Once detected, fish tracks and pings were exported as time-stamped variables for further manual review to ensure that fish tracks were indeed fish and not debris. We manually reviewed each exported track to separate fish from non-fish (debris) tracks. When true fish were identified each ping of these tracks was retained to calculate large fish (predator) density.
Dataset
The ARIS data is recorded in 30 minute time bins so our final fish density data has this time resolution. To analyze ARIS footage a dataset was created that included the ARIS unit, the sight and night of sampling, whether the ARIS was in light or dark treatments and associated data from the above Echoview processing. Fish density was calculated in R by taking the total number of true fish pings derived from step 7 above, divided by the Echoview exported beam volume sampled for a given time period. However, it is important to keep in mind that fish density is relative given the 2D nature of the ARIS image; true volume cannot be calculated. To standardize for changing sunset times on each given night, each 30 minute ARIS sample was assigned a time past sunset bin from 1 - 8, representing time past sunset in 30 minute increments. Time past sunset was calculated using the same R function as PER data. The mean temperature from hydrolabs from a given night at a given site was also added to this dataset.
PER contact Methods
Other contact points within the Sacramento-San Joaquin Delta such as diversions, bridges, pilings, docks, and/or submerged aquatic vegetation (SAV) were sampled following similar methods to the artificial light experiments to determine if proximity to these contact points affected relative predation risk. These experiments were conducted in the afternoon at five different sites and unlike the light study took advantage of existing structures. All PER contact data was collected in the same way as described above for the PER ALAN experiments. However, distance to pilings were calculated by taking the point to point distance from PER positions to the closet piling in a site, if present. Distance to other contacts was the closet distance to docks, bridges, and diversions. Distance to all contacts was defined as the distance of each PER GPS position to the closet contact point that was not SAV. Lux data was not collected for this dataset given that this work was conducted in the daylight.
Sundial Bridge ALAN Experimental Design (sampling metadata and descriptions)
The Sundial Bridge is an iconic illuminated structure in the City of Redding located within the little remaining spawning habitat of endangered winter run Chinook Salmon. To assess the impacts of Sundial Bridge artificial lighting at night (ALAN), we developed a field experiment to test the relationships of predator density and relative predation rate across four different ALAN treatments (0%, 25%, 50%, and 100% illumination intensity). Similar to previous work, we deployed ARIS cameras to quantify the response of relative predator density to ALAN treatments, as well as the temporal relationship of predator density during the transition form day to night. To quantify relative predation risk, we developed micro predation event recorded (mPERS). These were based on the design in Demetras et al. (2016), but we small enough to be casted from fishing poles. This data set is our sampling data for this experiment, including sampling dates, times, and field crews.
ARIS Data (RED ARIS data Sundial ALAN experiment, Yellow ARIS data Sundial ALAN experiment, and ARIS Sundial Continuous Deployment)
During all but the last sampling week, we deployed ARIS (Adaptive Resolution Imaging Sonar, Sound Metrics Corp.) cameras from 1 hour prior to sunset until 4 hours post sunset on each river bank to quantify the relative density of presumably piscivorous Rainbow Trout. On the last week of sampling we deployed an ARIS continually on River Right to quantify the full diel patterns of Rainbow Trout density underneath the Sundial Bridge. To remove background noise, we processed ARIS footage using contiguous samples over threshold (CSOT) in ARIS Fish software and imported this into Echoview software to automatically identify and filter fish. To generate fish tracks, we converted multibeam data to a single-target echogram, removed targets < 200 mm TL, and tracked fish pings to generate individual fish tracks. For each fish track, we exported the number of pings within 10-minute time bins and summed these pings to generate total fish pings per 10 minutes for each ARIS throughout sampling periods. For every 10-minute bin within each ARIS and sampling period, we divided the total number of fish pings by relative beam volume to generate relative fish density. Fish density is a relative measure given that ARIS footage is a 2D representation of a 3D space.
mPER Sundial ALAN
The mPERS are a miniaturization of the original PER design that worked well in the swift flowing Upper Sacramento River. Every second of each deployment, the mPERS recorded the date, time and whether a tethered juvenile salmonid had been predated, and this data was all stored on an internal micro SD card. We could not use endangered winter run fry for this experiment so we tethered rainbow trout fry to our mPERS as a surrogate species. Given that Rainbow Trout are piscivorous and cannibalistic this should not have biased our results. We tethered each fry to the mPER using 2m of 8lb fluorocarbon fishing line and attached each mPER to a fishing pole. During each deployment, we would cast the mPER above the influence of Sundial ALAN, let the PER drift past the bridge under the lights, and then retrieve the mPER once it was downstream of ALAN influence. When a fry was predated, a read switch was triggered and it logged a predation event on the internal micro SD card. We quantified the ALAN intensity of each treatment from a boat and related these lux levels to PER locations and predation events. Unfortunately, we only received 4 events throughout the study so we did not quantitatively analyze this data. However, we did test these mPERS in the Sacramento San-Joaquin Delta and had 60 events out of 520 deployments, ensuring that the sampling method worked well.
