Masfjorden is a 20 km long fjord at the west coast of Norway (∼60°50’ N, ∼005°30’ E). It has a maximum depth of 494 m and a sill depth of 75 m bordering to Fensfjorden, which is connected to the coastal ocean (details in Kaartvedt et al., 1988). In Masfjorden, scattering layers shallower than 200 m are almost entirely formed by M. muelleri with juvenile fish in a shallow scattering layer (<100 m at night and ∼100 m during the day) and the adults in a deeper scattering layer (around 200 m) during autumn and winter (Giske et al., 1990; Staby et al., 2011). The two layers merge in spring (Staby et al., 2011). The glacier lanternfish Benthosema glaciale prevails below ∼200 m, with limited diel vertical migration in winter (Giske et al., 1990; Kaartvedt et al., 2009).

Three upward facing stationary split-beam echosounders (EK60, Simrad) were deployed in Masfjorden between October 5–7, 2010 and recovered on August 17, 2011. Echosounders were moored at ∼370 m (38 kHz), ∼250 m (120 kHz), and ∼90 m depth (200 kHz) to enable high resolution of acoustic signals throughout the water column. The deepest echosounder was located at the bottom, the two shallower mounted in floating, anchored rigs in the same part of the fjord. The echosounders were cabled to shore and, with the exception of short periods of power failures, continuous recordings are available from all three frequencies for the entire study period. For further details about the setup of the echosounders see Prihartato et al. (2015). Here, we use data from the 200 kHz echosounder for the analysis of scattering layer depth and individual behavior. In addition, we consulted data from the deeper located 120 kHz echosounder for an overall assessment of the population behavior. This allowed to monitor the full diel migrating cycle including when juveniles migrated to below the 200 kHz echosounder (Figure 1), and also the largely non-migrating adults.

We analyzed data from December 2010 to March 2011, which enabled detailed analysis of individual swimming behavior of juvenile M. muelleri without the inclusion of deeper-living targets that only to a limited degree migrated vertically during this period (see e.g., Prihartato et al., 2015). Echosounder data with a temporal resolution of ∼2 s^–1 were available from 98 days. Surface light levels (2 m above water) were measured continuously after the end of December 2010, yet the sensitivity of the light meter was too low for winter night-time measurements (see Prihartato et al., 2015).

For information on approximate nocturnal light levels during the study period, we obtained cloud cover and precipitation data from the nearby (∼20 km) weather station Takle from https://seklima.met.no/observations/ (download on November 12, 2019) and moon phases from https://www.timeanddate.com/moon/phases/ (accessed on November 12, 2019). Night time cloud cover and precipitation were calculated by averaging data collected within 0.3 days around midnight on each day. Time of sunrise and sunset were calculated for each day using the function sunRiseSet¹. All dates and times are presented in UTC (local time -1 h; maximum deviation from apparent solar time about 33 min on December 1, 2010).

One aim of the study was to relate individual behavior to that of the M. muelleri population. Therefore, the location and migration velocity of the scattering layer formed by juvenile M. muelleri were determined. The raw EK60 data were reorganized into a three-dimensional grid, with a daytime (resolution of 90 s), depth (resolution of 0.3 m) and date dimension, in order to improve computation times and data accessibility. Each grid cell contained the average (calculated in the linear domain) volume backscattering strength (S_v, dB re 1 m^–1 (MacLennan, 2002), in the following referred to as backscatter), calculated from the raw data of the given depth and time interval. The gridded backscatter data were used in all of the following population analyses. We calculated monthly median backscatter from the 120 kHz data by computing the median backscatter for each depth and daytime interval over all dates in the respective month.

The depth of the scattering layer was determined for each day of the study period using night time data (between sunset and sunrise) between 2 and 84 m depth. Data closer to the transducer could not be used due to noise (ping interference) at about 86 m depth. Between 0.1 day (144 min) after sunset and before sunrise, all values with depths < 19 m were excluded to reduce the inclusion of night time surface signals (Supplementary Figure S1). Backscatter values > −53 dB re 1 m^–1 usually occurred due to extended surface signals or strong individual targets (potentially larger fish). These high backscatter values were thus not representative for the M. muelleri population and therefore excluded from the analysis. The thresholds for these data exclusions were determined empirically from the echograms. The remaining backscatter data were linearized to s_v (m^–1), the volume backscattering coefficient (MacLennan, 2002). For each time point, the cumulative s_v over depth was calculated and the depth where the cumulative s_v reached 50% of its maximum value was defined as the midpoint of the scattering layer. In some cases, the M. muelleri scattering layer could not be clearly identified due to low fish densities, which we defined as a cumulative s_v < 5 × 10^–6 m^–1. Periods where the cumulative s_v fell below this threshold were excluded from the scattering layer analysis. We calculated the vertical velocity of the scattering layer by applying a moving slope approach² with a time window of 0.05 decimal days (72 min) on the scattering layer depth.

Individual targets could not be resolved when in surface waters at dusk and dawn, and we here focus on the nocturnal behavior in-between these timepoints. The midnight sinking period was marked by the population starting to descend from the surface to deeper layers in the evening (from now on referred to as dusk descent) and the ascent to the surface in the morning (from now on dawn ascent). We used the location of the scattering layer to determine the start of the dusk descent and the end of the dawn ascent of the population (Figure 2). We chose the first time-point when the center of the scattering layer was at its shallowest depth in the afternoon and morning, respectively, as criteria. We also calculated the time spent in near-surface waters < 20 m depth at dusk and dawn for each day.

The proximity of the M. muelleri scattering layer to the 200 kHz transducer at night enabled the identification of night-time individual swimming behavior. Densities of fish in the core of the scattering layer were too high for separating individuals during some parts of the night and at long distance from the transducer (i.e., shallow depth), while individuals outside the core and during periods of lower density could be distinguished successfully. Types of swimming patterns were first visually identified from echograms in the Sonar5-Pro software (Balk and Lindem, 2017) on separate days and thereafter quantified using target tracking.

Successive echoes of individual targets can be combined and thus the path of individuals over depth, time and in along and athwart direction can be followed (target tracking; Brede et al., 1990). We used target tracking to assess the individual vertical swimming of M. muelleri over the entire study period. Target tracking was performed in the software Sonar5-Pro (Version 6.0.4, Balk and Lindem, 2000, 2017). Prior to the tracking, the files were pre-processed by cross-filter detection (Balk and Lindem, 2001, 2017), a procedure which improves the detection of individual tracks in single echo detection mode. Default settings were selected for the cross-filter detection. The thus modified single echo detection echograms were then analyzed by automatic target tracking. We optimized the parameters for the target tracking (Table 1) beforehand by comparing manual and automatic tracking at different time points of the study period. We adjusted the tracking parameters depending on distance from the transducer (range), which reduces the uncertainties at longer ranges compared to when using a single set of parameters. Individuals were tracked between 5 and 85 m range (i.e., ∼9–89 m depth). The use of cross-filter detection introduces uncertainties. While enabling tracking at increased ranges and at higher densities (Balk and Lindem, 2002), it also increases the risk of multiple detections, increases ping gaps and introduces noise into the target strength (MacLennan, 2002) of tracks. Yet, we decided to accept these uncertainties to be able to analyze individual movements throughout most of the water column and with longer track durations that enabled the differentiation between behaviors. Only few individuals (∼3%) were tracked at ranges > 60 m (shallower than 34 m depth; Supplementary Figure S2). While these few individuals have limited influence on the overall results, they contribute information about processes and behavior in the upper water column.

High population densities hamper individual tracking detection (Handegard, 2007). Correspondingly, individuals close to the transducer are tracked more effectively due to the higher resolution of targets. Differences in population densities over time, e.g., lower densities during the dusk descent, and distances to the transducer, e.g., shorter distances in the morning when many individuals already started descending, thus led to a higher number of descending tracks. Nevertheless, a visual comparison of the tracks and echograms indicated that the tracks’ position information and thus behavior of the individuals were represented well. This also makes us confident that we are not assessing the behavior of outliers with deviating behavior and that the large-scale picture of the behavioral repertoire observed here is representative.

The workflow during track processing and filter criteria as outlined below are summarized in the Supplementary Figure S3. The data were processed and analyzed in Matlab (Mathworks; R2017b). We first applied a filter to exclude potential other targets than M. muelleri. This identity filter included a threshold in mean TS (Figure 3) and excluded daytime tracks from after sunrise and before sunset. We determined the TS thresholds by manually tracking M. muelleri individuals on one randomly selected day of each of the 4 months (about 800 tracks per day) and then identifying the TS range that included about 90% of the tracked individuals on these 4 days. Tracks with a mean TS outside the defined range were excluded from further analysis. This procedure allowed identification of M. muelleri with a reduced influence of tilt angles (Miyashita et al., 1996) as fish in all different angles were included during the manual tracking. A minimum track duration of 30 s was used. The first time-point of each track determined whether the tracks were within the defined night-time hours.

After applying the identity filter, 613003 tracks remained and were further processed. Depth outliers of each individual track

were removed by first calculating a running median as well as a running median absolute deviation (MAD) (each with a window of 10 echoes) of the track’s depths and then replacing depth values that deviated from the running median by more than the MAD × 10 by the running median. In the next step, a running mean (window size of 10 echoes) was applied to the depth values of each such treated track. We calculated net vertical displacement (m), track duration (s) and vertical velocity (vertical displacement divided by track duration; cm s^–1) of each track. Each track was also examined for patterns of alternations between segments of active vertical ascent or descent, and pauses in which fish maintained a constant depth, although they might have been active in the horizontal plane (Supplementary Figure S4). We determined and subsequently counted the pauses in each track using thresholds. We defined parts of the tracks with depth changes < 0.01 m between at least four subsequent echoes, as pauses and those with larger depth changes as active phases. Single values of pauses or active phases were filled by the surrounding values. For example, when in the vector x = [0,0,0,1,0,0,1,1,1,1,0,0], the digit 1 stands for “pause” and 0 for “active phase,” then the resulting index vector of which echoes should be regarded as pause would look like this: y = [0,0,0,0,0,0,1,1,1,1,0,0]. The first and last segments (i.e., the first and last ascent, descent or pause) of each accepted track with more than two pauses were excluded in order to reduce errors from the target tracking, stemming from potentially overlapping tracks. Tracks with < 60 echoes were considered fragments and were therefore excluded from further analysis. The remaining 272290 tracks were used for behavioral analysis.

Based on net vertical displacement, number of pauses and track duration, the tracks were assigned to either of three main swimming patterns (Supplementary Figure S3):

We then calculated the average active swimming speed (swimming speed, cm s^–1), i.e., the vertical velocity during active segments of each step-wise track. In stationary and direct tracks, the swimming speed equals the vertical velocity. Furthermore, we determined the proportion of the three different swimming patterns in relation to time from the start of the dusk descent and time to the end of the dawn ascent, in order to account for seasonal changes. We also calculated the proportion of ascending and descending step-wise tracks, as well as their average vertical velocities in relation to daytime and depth using data binned by depth (3 m intervals) and time (72 min). The proportion of ascending tracks for each bin was determined when more than five tracks were found in the respective bin.

The vertical distribution of the scattering layer ascribed to juvenile M. muelleri differed distinctly between day and night and was characterized by short-term migrations to the surface at dusk and dawn (Figure 1). After the dusk ascent, the population returned to intermediate depths, deepening from an average of 60 m in December to 70 to 75 m in February and March. From mid-January, the vertical distribution changed continuously throughout the night, with the scattering layer moving deeper before midnight and reversing direction after midnight (Figure 4). The daytime distribution was generally deeper and below 100 m. The deeper scattering layer ascribed to the adults mostly stayed below 120 m depth throughout the diel cycle, but a small proportion of this adult population sometimes migrated to shallower depths at night (Figure 1).

The timing of the dusk descent and dawn ascent of the juveniles changed over the study period in accordance with seasonal changes in sunrise and sunset (Figure 4A). The dusk descent from near-surface waters started about 20–70 min after sunset after the fish had spent around 20–40 min in waters shallower than 20 m. The dawn ascent ended around 40–70 min before sunrise on most days and the fish stayed in near-surface waters (<20 m) for about 40–60 min. The duration between these migrations and sunset/sunrise shortened over the study period (Figure 4A). On darker, foul weather days (e.g., March 20–25, 2011; Figure 4A), dusk descent started earlier and dawn ascent ended later. We did not find a clear relationship of population movement with moon phases (see Supplementary Figure S5 for a combination of Figure 4 with cloud cover and moon phase data). Interrupted upward migrations in the afternoon, where the majority of the population started descending again before reaching the surface, were observed on 39 out of 98 analyzed diel echograms.

Population descent velocities were highest ∼30 min after the start of the dusk descent, while ascent velocities were highest 30–40 min before reaching the surface layers in the morning (Figure 4B). The maximum velocity of the scattering layer during the dusk descent increased from ∼0.5 cm s^–1 in December to ∼0.8 cm s^–1 in March. The maximum velocity during the dawn ascent increased from ∼1 cm s^–1 in December to ∼1.5 cm s^–1 in March. On most days, the dawn ascent was faster than the dusk descent.

Of the 272290 tracks, 142547 were classified as step-wise swimming, with 56744 tracks ascending and 85803 descending (Table 2). In total 127221 tracks displayed no or little vertical movement and were classified as stationary. An additional 2522 tracks were directly ascending/descending without steps. Duration and net vertical displacement of individuals depended on the respective swimming pattern (Table 2 and Figure 5).

Step-wise swimming was especially prominent during the dusk and dawn migrations (Figures 6, 7). Around 75% of the fish swam step-wise around the start of the dusk descent. Although descending step-wise tracks dominated throughout this period of population descent, some ascending individuals were also recorded. During the latter part of the night, the proportion of step-wise swimming behavior increased (Figure 7), reaching a maximum of 87% at the end of the dawn ascent, but then also with a high proportion (>25%) of step-wise descending individuals. The net vertical relocation speed was ∼0.9 cm s^–1, while the actual swimming speed during the vertical steps was ∼2–3 cm s^–1 (Table 2).

Vertically stationary tracks dominated (>50%) the nocturnal records throughout most of the night. These tracks usually lasted for > 1 min, with the track duration being restricted by the number of co-occurring tracks (Figure 6C). In most of the cases where M. muelleri relocated vertically, they moved step-wise (Figures 6, 7). Additional vertical displacement was caused by apparent internal waves. In a subsample of randomly selected days, internal wave amplitudes were around 0.7 m (range 0–1.8 m) and periods about 10 min (range 0–30 min). Internal waves led to vertical transitions of up to 0.6 cm s^–1, but usually around 0.2 cm s^–1, in otherwise stationary animals.

Even though there were main migration periods at dusk and dawn and dominance of stationary tracks at night, there was a clear pattern of the majority of step-wise relocating individuals descending before midnight and ascending after midnight (Figure 8). This pattern was accentuated toward the end of the study period (Figure 8A). The proportion of descending and ascending tracks depended on depth (Figure 8B). During the first half of the night, descending tracks dominated at all depths, nevertheless we observed the highest proportion of descending tracks in shallow layers of 20–40 m depth. In layers between 70 and 80 m, there always was a strong majority of ascending tracks subsequent to midnight. Correspondingly, the average displacement was downward before midnight and upward after midnight. Individual velocities reached > 1 cm s^–1 during the dusk and dawn migrations, while velocities in the middle of the night were mostly lower. Maximum velocities increased over winter.

The individual behavior during the overall population’s upward migration in the afternoon and descent in the morning, differed from that observed at night. During these main migration periods, most pearlsides schooled and individuals could not be resolved. However, individuals could occasionally be detected close to these schools. The pause duration in these individuals was short and some of these individuals refrained from pausing altogether and either ascended or descended directly (Figure 8). The highest vertical velocities were achieved in such direct tracks, with means of 1.0 and 1.8 cm s^–1 in ascending and descending tracks, respectively (Table 2).