Article: McClintock, B. T., Moreland, E. E., London, J. M., Dahle, S. P., Brady, G. M., Richmond, E. L., Yano, K. M. and Boveng, P. L. (2015), Quantitative assessment of species identification in aerial transect surveys for ice-associated seals. Marine Mammal Science. doi: 10.1111/mms.12206
There are several species of seals that utilize ice platforms in order to mate, rear their young and in the spring, molt. These ice seals have been a focus of conservation due to the effects of global climate change on shifting ice patterns. Distribution and abundance data for ice seals is time consuming and very difficult, therefore until recently little information about ice seals has been known. In order to streamline conservation measures, a lot more information is needed about the different seal species and while aerial and satellite images have been used to detect their distribution, misidentification is a huge problem. This results in unreliable information and thus conservation management is difficult to implement.
Seal species are difficult to distinguish from a distance because they have similar habitats, distribution patterns, body shapes, colors, and movements (Fig. 1, Fig. 2). For example, the characteristic torpedo shape of a spotted seal isn’t easily distinguished by satellite or plane because ice location and lighting can cause the torpedo shape to look more rounded. This can lead to the seal being misidentified as a ringed seal.
McClintock et al. assessed species and age class identification of four different ice seal species using aerial photographs. The common errors aerial surveys detect were identified by comparing a software analysis program to observational identification of ice seals.
Aerial transect surveys were performed on ice seals in the eastern Bearing Sea. Using systematic random sampling, 716 images were selected to represent the entire study. Four seal biologists were then assigned to 600 photographs each, and only two biologists were assigned to the remaining 159 photographs. The observer was told to identify the seal as either a bearded seal, a ribbon seal, a ringed seal, a spotted seal, or unknown. For each identification their confidence level was then recorded (guess; <50%, likely; 51-99%, or positive; 100%). Age class (pup, non-pup, unknown) and their confidence levels on this was also recorded. For each photograph the observer was also told to record any specific characteristic found. A list of these traits are seen in Table 1. A model was then set up to compare all of this data into a statistically representative sample. This model was used to compare observer analysis to software analysis on species and age identifications.
What this showed and why we care
Estimates on true species identification and age class were identical between software analysis and observational identification (Table 2). Both the observer and non-observer (software) analysis showed that species and age class misidentification occurred on all species and age and was not targeted. Within the observer analysis, the spotted seal was the seal most misidentified (11% mistaken for ribbon seals), while the other three seal species misidentification remained relatively low. When spotted seals’ distinguished pelage pattern was obscured, their body position and size is easily mistaken for a ribbon seal, but misidentification could have been caused by the high abundance of observer unease. Only 39% of correct spotted seal observations were in the positive confidence level. Certain traits were proven to be strong predictors for species identification, showing observer analysis to be a stronger species identification indicator than the software analysis program. Between the observer and non-observer bias, pups were slightly more likely to be mistaken for non-pups, but overall age class misidentification remained the same throughout any age bracket.
Strict sampling methods and observer training have been put in place to eliminate species misidentification, but results from this study prove that even with strict observer protocols misidentification does occur. Although the misidentification rates were low, the fact that they still exist demonstrates how species distribution and abundances can be miscalculated and therefore conservation strategies can be misguided. In order to better understand the ice seal population and distribution patterns it is best to have repeat sampling and include observer analysis along with a framework model to evaluate aerial images. Misidentification of a specific species is just as likely to occur with a trained observer as it is with a computer model, but knowledge of seal behavior can reduce the overall misidentification. For large databases, an automated recognition software program will reduce time and energy, but the findings in this study show how complicated a process image analysis really is. An automated classification system using both visual clues (color and body shape) and behavioral context will provide the best chance at correct identification from aerial photographs.