In the field of ecology, there is a growing interest in machine (including deep) learning for processing and automatizing repetitive analyses on large amounts of images collected from camera traps, drones and smartphones, among others. These analyses include species or individual recognition and classification, counting or tracking individuals, detecting and classifying behavior. By saving countless times of manual work and tapping into massive amounts of data that keep accumulating with technological advances, machine learning is becoming an essential tool for ecologists. We refer to recent papers for more details on machine learning for ecology and evolution (Besson et al. 2022, Borowiec et al. 2022, Christin et al. 2019, Goodwin et al. 2022, Lamba et al. 2019, Nazir & Kaleem 2021, Perry et al. 2022, Picher & Hartig 2023, Tuia et al. 2022, Wäldchen & Mäder 2018).

In their paper, Nakagawa et al. (2023) conducted a systematic review of the literature on machine learning for wildlife imagery. Interestingly, the authors used a method unfamiliar to ecologists but well-established in medicine called rapid review, which has the advantage of being quickly completed compared to a fully comprehensive systematic review while being representative (Lagisz et al., 2022). Through a rigorous examination of more than 200 articles, the authors identified trends and gaps, and provided suggestions for future work. Listing all their findings would be counterproductive (you’d better read the paper), and I will focus on a few results that I have found striking, fully assuming a biased reading of the paper. First, Nakagawa et al. (2023) found that most articles used neural networks to analyze images, in general through collaboration with computer scientists. A challenge here is probably to think of teaching computer vision to the generations of ecologists to come (Cole et al. 2023). Second, the images were dominantly collected from camera traps, with an increase in the use of aerial images from drones/aircrafts that raise specific challenges. Third, the species concerned were mostly mammals and birds, suggesting that future applications should aim to mitigate this taxonomic bias, by including, e.g., invertebrate species. Fourth, most papers were written by authors affiliated with three countries (Australia, China, and the USA) while India and African countries provided lots of images, likely an example of scientific colonialism which should be tackled by e.g., capacity building and the involvement of local collaborators. Last, few studies shared their code and data, which obviously impedes reproducibility. Hopefully, with the journals’ policy of mandatory sharing of codes and data, this trend will be reversed. 

REFERENCES

Besson M, Alison J, Bjerge K, Gorochowski TE, Høye TT, Jucker T, Mann HMR, Clements CF (2022) Towards the fully automated monitoring of ecological communities. Ecology Letters, 25, 2753–2775. https://doi.org/10.1111/ele.14123

Borowiec ML, Dikow RB, Frandsen PB, McKeeken A, Valentini G, White AE (2022) Deep learning as a tool for ecology and evolution. Methods in Ecology and Evolution, 13, 1640–1660. https://doi.org/10.1111/2041-210X.13901

Christin S, Hervet É, Lecomte N (2019) Applications for deep learning in ecology. Methods in Ecology and Evolution, 10, 1632–1644. https://doi.org/10.1111/2041-210X.13256

Cole E, Stathatos S, Lütjens B, Sharma T, Kay J, Parham J, Kellenberger B, Beery S (2023) Teaching Computer Vision for Ecology. https://doi.org/10.48550/arXiv.2301.02211

Goodwin M, Halvorsen KT, Jiao L, Knausgård KM, Martin AH, Moyano M, Oomen RA, Rasmussen JH, Sørdalen TK, Thorbjørnsen SH (2022) Unlocking the potential of deep learning for marine ecology: overview, applications, and outlook†. ICES Journal of Marine Science, 79, 319–336. https://doi.org/10.1093/icesjms/fsab255

Lagisz M, Vasilakopoulou K, Bridge C, Santamouris M, Nakagawa S (2022) Rapid systematic reviews for synthesizing research on built environment. Environmental Development, 43, 100730. https://doi.org/10.1016/j.envdev.2022.100730

Lamba A, Cassey P, Segaran RR, Koh LP (2019) Deep learning for environmental conservation. Current Biology, 29, R977–R982. https://doi.org/10.1016/j.cub.2019.08.016

Nakagawa S, Lagisz M, Francis R, Tam J, Li X, Elphinstone A, Jordan N, O’Brien J, Pitcher B, Sluys MV, Sowmya A, Kingsford R (2023) Rapid literature mapping on the recent use of machine learning for wildlife imagery. EcoEvoRxiv, ver. 4 peer-reviewed and recommended by Peer Community in Ecology.  https://doi.org/10.32942/X2H59D

Nazir S, Kaleem M (2021) Advances in image acquisition and processing technologies transforming animal ecological studies. Ecological Informatics, 61, 101212. https://doi.org/10.1016/j.ecoinf.2021.101212

Perry GLW, Seidl R, Bellvé AM, Rammer W (2022) An Outlook for Deep Learning in Ecosystem Science. Ecosystems, 25, 1700–1718. https://doi.org/10.1007/s10021-022-00789-y

Pichler M, Hartig F Machine learning and deep learning—A review for ecologists. Methods in Ecology and Evolution, n/a. https://doi.org/10.1111/2041-210X.14061

Tuia D, Kellenberger B, Beery S, Costelloe BR, Zuffi S, Risse B, Mathis A, Mathis MW, van Langevelde F, Burghardt T, Kays R, Klinck H, Wikelski M, Couzin ID, van Horn G, Crofoot MC, Stewart CV, Berger-Wolf T (2022) Perspectives in machine learning for wildlife conservation. Nature Communications, 13, 792. https://doi.org/10.1038/s41467-022-27980-y

Wäldchen J, Mäder P (2018) Machine learning for image-based species identification. Methods in Ecology and Evolution, 9, 2216–2225. https://doi.org/10.1111/2041-210X.13075

Capture-Mark-Recapture (CMR) data are commonly used to estimate ecological variables such as abundance, survival probability, or transition rates from one state to another (e.g. from juvenile to adult, or migration from one site to another). Many studies have shown how estimations can be affected by neglecting one aspect of the population under study (e.g. the heterogeneity in survival between individuals) or one limit of the methodology itself (e.g. the fact that observers might not detect an individual although it is still alive). Strikingly, very few studies have yet assessed the robustness of one fundamental assumption of all CMR-based inferences: marks are supposed definitive and immutable. If they are not, how are estimations affected? Addressing this issue is the main goal of the paper by Touzalin et al. (2023), and they did a very nice work. But, because the answer is not that simple, it also calls for further investigations.

When and why would mark loss bias estimation? In at least two situations. First, when estimating survival rates: if an individual loses its mark, it will be considered as dead, hence death rates will be overestimated. Second, more subtly, when estimating transition rates: if one individual loses its mark at the specific moment where its state changes, then a transition will be missed in data. The history of the marked individual would then be split into two independent CMR sequences as if there were two different individuals, including one which died.

Touzalin et al. (2023) thoroughly studied these two situations by estimating ecological parameters on 1) well-thought simulated datasets, that cover a large range of possible situations inspired from a nice compilation of hundreds of estimations from fish and bats studies, and 2) on their own bats dataset, for which they had various sources of information about mark losses, i.e. different mark types on the same individuals, including mark based on genotypes, and marks found on the soil in the place where bats lived. Their main findings from the simulated datasets are that there is a general trend for underestimation of survival and transition rates if mark loss is not accounting for in the model, as it would be intuitively expected. However, they also showed from the bats dataset that biases do not show any obvious general trend, suggesting complex interactions between different ecological processes and/or with the estimation procedure itself.

The results by Touzalin et al. (2023) strongly suggest that mark loss should systematically be included in models estimating parameters from CMR data. In addition to adapt the inferential models, the authors also recommend considering either a double marking, or even a single but ‘permanent’ mark such as one based on the genotypes. However, the potential gain of a double marking or of the use of genotypes is still to be evaluated both in theory and practice, and it seems to be not that obvious at first sight. First because double marking can be costly for experimenters but also for the marked animals, especially as several studies showed that marks can significantly affect survival or recapture rates. Second because multiple sources of errors can affect genotyping, which would result in wrong individual assignations especially in populations with low genetic diversity or high inbreeding, or no individual assignation at all, which would increase the occurrence of missing data in CMR datasets. Touzalin et al. (2023) supposed in their paper that there were no genotyping errors, but one can doubt it to be true in most situations. They have now important and interesting other issues to address.

References

Frédéric Touzalin, Eric J. Petit, Emmanuelle Cam, Claire Stagier, Emma C. Teeling, Sébastien J. Puechmaille (2023) Mark loss can strongly bias demographic rates in multi-state models: a case study with simulated and empirical datasets. BioRxiv, ver. 3 peer-reviewed and recommended by Peer Community in Ecology. https://doi.org/10.1101/2022.03.25.485763