This study supports the conclusion that the chromatophore patterns on the head of delta smelt are more stable from winter to spring than fall to winter. Therefore, the use of chromatophores as natural marks in delta smelt is more feasible within the adult stage than between subadult and adult stages. Unlilke Castillo et al. (2018), the present study considered a larger interval between photo sessions and two greatly different outdoor ambient light scenarios across all photo sessions, hence the present study is more likely to include the range of ambient light experienced by wild delta smelt. Given the limited individual marking and tagging methods available for this species, the stability of natural marks throughout the adult stage of delta smelt under a wide range of ambient light levels is a potentially useful finding. Moreover, over the winter-spring period, automated recognition of natural marks was 100% accurate in all fish for which matching was possible, with better matching success for the low light treatment (92%) than the high light treatment (62%).
After considering the time interval between photo sessions, light treatments, gender, and the life stages of fish throughout the study period, our analyses support the conclusion that ontogenetic changes between the subadult and adult stages is the likely reason for the higher matching success of images between winter and spring (adult stage only) compared to fall and winter or fall and spring (subadult to adult stages). Such conclusion is consistent with the counterintuitive finding that the three month interval between photo sessions 1–2 (fall–winter) did not result in greatly higher matching success compared to the six month interval between sessions 1–3 (fall–spring; Fig. 3a-b; Table 2).
Despite the reported higher frequency of breeding tubercles on the head of male delta smelt during the spawning season (Wang 2007), fish gender did not influence the visual match quality, which is consistent with the findings of Castillo et al. (2018), who also indicated that changes in spawning condition (spawning vs. non-spawning) between photo sessions did not influence the visual match quality. Hence, our results support the conclusion that chromatophores become stable once adult reach pre-spawning condition around January. Such conclusion is consistent with the 100% visual matching success reported for adult delta smelt between January and May 2013 (Castillo et al. 2018), which was achieved despite a 100-fold increase in ambient light occurring between the last two photo sessions. Surprisingly, no differences in the general appearance of chromatophores were apparent to testers between the fish held at low and high ambient light scenarios. Besides delta smelt, to our knowledge, only Atlantic salmon Salmo salar has been tested using visual recognition of natural marks in the eye and jaw areas and such study achieved a 100% matching success over a period of 2 months (Garcia de Leaniz et al. 1994).
The fact that the TinEye API only mistakenly recognized one fish throughout the three photo sessions (i.e., 0.16% of the fish identified) indicates its algorithm only reports matches when the probability of image recognition is extremely high. Yet, the selectivity of such algorithm resulted in the majority of the individuals being unmatched in session pairs 1–2 and 1–3. The output of the TinEye API only included match quality for matched images. This explains why the reported automated match quality between photo sessions (Fig. 4a) did not differ as much as it would be expected based on the corresponding differences in match success (Fig. 3a). Although we did not compare our TinEye results with other automated image recognition methods, and found no such comparative studies in other species of fish, individual matching success varied up to 27% among three automated image recognition methods tested in the freshwater turtle Pseudemys gorzugi (Suriyamongkol and Mali 2018). Given the large differences in the duration of stable natural marks suggested by automated matching in our study, and other species (e.g., Van Tienhoven et al. 2007; Merz et al. 2012), comparisons across different automated methods for a given species should be ideally performed using the same set of images, or at least the same size ranges and life stages.
Based on the percent of match success, our results and those of Castillo et al. (2018) show the stability of natural marks was consistently higher for visual recognition compared to automated recognition. The 100% visual matching success for the adult stage in both light treatments considered in the present study suggest the potential for practical use of natural marks in delta smelt is highest during winter and spring. Although visual recognition further showed natural marks are stable enough to recognize over 70% of the fish between the subadult and adult stages, a higher matching success is required to consistently apply natural marks between the subadult and adult life stages. The differences in matching success between our visual and automated results, and the advantages of automated methods using large sets of images suggest that additional progress in automated image recognition may prove beneficial. For example, through the use of biometrics (Kühl and Burghardt 2013), in combination with several image identification techniques to improve matching success (e.g., Das et al. 2015). Although preliminary recognition for uncropped head images of delta smelt using TinEye showed the matching success did not significantly increase compared to images of cropped head areas B − C (Tien-Chieh Hung, UC Davis FCCL, unpubl. data), further assessments of the combined potential of natural marks and features such as the contour of the head, could prove valuable, particularly if the depth of field is adjusted to improve image resolution over larger body areas.
Fish exposed to either high or low ambient light had a corresponding small to moderate increase in chromatophores size between sessions 1–2 and 1–3. In contrast, a significantly decreased expression of chromatophores was observed between sessions 2–3 in both treatments. Such constriction of chromatophores is likely due to the seasonal increase in sunlight hours from mid-winter to mid-spring, and this could also account for the observed interaction between photo sessions and light treatments. Similarly, Castillo et al. (2018) reported reduced chromatophore expression between early-spring to mid-spring, but unlike the present study, the increased sunlight in that study entailed the transfer of fish from indoors to outdoors tanks in early spring. Importantly, in neither of these studies the smaller size of chromatophores attributed to increased sunlight influenced the visual matching success despite a 100-fold difference in ambient light scenarios. However, such large differences in illuminance could explain the lower success of automated recognition for the high light treatment in sessions 2–3 (Fig. 3a). Despite the limitations of the two image recognition methods evaluated in our study, these results suggest natural marks could also be feasible in wild delta smelt at least during the adult stage. Nevertheless, the skin coloration of wild delta smelt can become lighter following field capture and handling (V. Afentoulis, California Department of Water Resources; pers. comm.). Compared to the chromatophores expression seen in the present study in both light treatments, the skin coloration of wild adult delta smelt was significantly lighter for all fish photographed following capture by surface trawl nets in the upper San Francisco Estuary (G. Castillo, U.S. Fish and Wildlife Service, Lodi, California; unpubl. data). Moreover, wild delta smelt experience higher handling stress compared to cultured delta smelt (Afentoulis et al. 2013). Thus, it is conceivable that stress-induced chromatophore constriction could be more evident in wild delta smelt (e.g., physiological color change, after Sugimoto 2002). Additional studies are needed to evaluate the potential effect of stress on chromatophore expression and on the post-handling survival of wild delta smelt.
Only one person was required to acquire all images used to evaluate natural marks using visual and automated methods. Assuming 500 fish are used per photo session, it would take about 21 h to acquire all images (i.e., 2.5 min per fish). Unlike automated recognition, visual recognition did not require editing images but the time required for visually matching one fish in a pair of 30 images ranged between 16 and 28 min per tester. In contrast, the time required for automated matching was only about 2 min per fish for a pair of 200 images (including 300 min editing images for two photo sessions, 30 min of upload time and 10 min to run 200 automated comparisons). However, when considering the lowest automated matching success of TinEye (sessions 1–3), the required time per reported fish match would be about 8 min. Thus, the extra time required to edit images prior to obtaining automated matching results may outweigh the time requirements of visual matching per fish even when matching success is low. If the goal is to maximize matching success for small sample sizes, our results suggest that visual recognition still may have advantages over automated recognition.
Although VIA tagging could be the most practical method to identify cultured adult delta smelt (Sandford et al. 2019), smaller fish experience increased tag shedding (Lindberg et al. 2013). In addition to tag loss, VIA tagging may not ensure individual identification due to difficulty in reading tag codes in some species (e.g., Barriga et al. 2015). As in the case of natural marks, the feasibility of VIA tags on wild delta smelt has not been evaluated. Although the range of VIA tag shedding in the present study (0.0 to 1.2%) was similar to that reported for tagging conducted in 2013 (0.3 to 1.7%; Castillo et al. 2018), cultured subadult delta smelt in the present study were effectively tagged at a smaller average size (about 8 mm smaller) than the adult fish in the referred 2013 study. Moreover, post-tagging survival during session 3 was higher in the present study (84–87%; Table 1) than in the 2013 study (78%). These results suggest VIA tags could also be considered to evaluate the effectiveness of natural marks in subadult and adult wild delta smelt.