Are you an ecologist who uses a computer or know someone that does? Even if your research doesn’t rely heavily on advanced computational techniques, it likely hasn’t escaped your attention that computers are increasingly being used to analyse field data and make predictions about the consequences of environmental change. So before artificial intelligence and robots take over from scientists, now is great time to read about how experts think computers could make your life easier and lead to innovations in ecological research. In “Data-based, synthesis-driven: setting the agenda for computational ecology”, Poisot and colleagues [1] provide a brief history of computational ecology and offer their thoughts on how computational thinking can help to bridge different types of ecological knowledge. In this wide-ranging article, the authors share practical strategies for realising three main goals: (i) tighter integration of data and models to make predictions that motivate action by practitioners and policy-makers; (ii) closer interaction between data-collectors and data-users; and (iii) enthusiasm and aptitude for computational techniques in future generations of ecologists. The key, Poisot and colleagues argue, is for ecologists to “engage in meaningful dialogue across disciplines, and recognize the currencies of their collaborations.” Yes, this is easier said than done. However, the journey is much easier with a guide and when everyone involved serves to benefit not only from the eventual outcome, but also the process.

References

[1] Poisot, T., Labrie, R., Larson, E., & Rahlin, A. (2018). Data-based, synthesis-driven: setting the agenda for computational ecology. BioRxiv, 150128, ver. 4 recommended and peer-reviewed by PCI Ecology. doi: 10.1101/150128

Tick-borne diseases are an important burden on public health all over the globe, making accurate forecasts of tick population a key ingredient in a successful public health strategy. Over long time scales, tick populations can undergo complex dynamics, as they are sensitive to many non-linear effects due to the complex relationships between ticks and the relevant (numerical) features of their environment.

But luckily, capturing complex non-linear responses is a task that machine learning thrives on. In this contribution, Manley et al. (2023) explore the use of Gradient Boosted Trees to predict the distribution (presence/absence) and abundance of ticks across New York state.

This is an interesting modelling challenge in and of itself, as it looks at the same ecological question as an instance of a classification problem (presence/absence) or of a regression problem (abundance). In using the same family of algorithm for both, Manley et al. (2023) provide an interesting showcase of the versatility of these techniques. But their article goes one step further, by setting up a multi-class categorical model that estimates jointly the presence and abundance of a population. I found this part of the article particularly elegant, as it provides an intermediate modelling strategy, in between having two disconnected models for distribution and abundance, and having nested models where abundance is only predicted for the present class (see e.g. Boulangeat et al., 2012, for a great description of the later).

One thing that Manley et al. (2023) should be commended for is their focus on opening up the black box of machine learning techniques. I have never believed that ML models are more inherently opaque than other families of models, but the focus in this article on explainable machine learning shows how these models might, in fact, bring us closer to a phenomenological understanding of the mechanisms underpinning our observations.

There is also an interesting discussion in this article, on the rate of false negatives in the different models that are being benchmarked. Although model selection often comes down to optimizing the overall quality of the confusion matrix (for distribution models, anyway), depending on the type of information we seek to extract from the model, not all types of errors are created equal. If the purpose of the model is to guide actions to control vectors of human pathogens, a false negative (predicting that the vector is absent at a site where it is actually present) is a potentially more damaging outcome, as it can lead to the vector population (and therefore, potentially, transmission) increasing unchecked.

Boulangeat I, Gravel D, Thuiller W. Accounting for dispersal and biotic interactions to disentangle the drivers of species distributions and their abundances: The role of dispersal and biotic interactions in explaining species distributions and abundances. Ecol Lett. 2012;15: 584-593.
https://doi.org/10.1111/j.1461-0248.2012.01772.x

Manley W, Tran T, Prusinski M, Brisson D. (2023) Modeling tick populations: An ecological test case for gradient boosted trees. bioRxiv, 2023.03.13.532443, ver. 3 peer-reviewed and recommended by Peer Community in Ecology. https://doi.org/10.1101/2023.03.13.532443

In the article, "Are the more flexible great-tailed grackles also better at behavioral inhibition?", Logan and colleagues (2021) are setting an excellent standard for cognitive research on wild-caught animals. Using a decent sample (N=18) of wild-caught birds, they set out to test the ambiguous link between behavioral flexibility and behavioral inhibition, which is supported by some studies but rejected by others. Where this study is more thorough and therefore also more revealing than most extant research, the authors ran a battery of tests, examining both flexibility (reversal learning and solution switching) and inhibition (go/no go task; detour task; delay of gratification) through multiple different test series. They also -- somewhat accidentally -- performed their experiments and analyses with and without different criteria for correctness (85%, 100%). Their mistakes, assumptions and amendments of plans made during preregistration are clearly stated and this demonstrates the thought-process of the researchers very clearly.

Logan et al. (2021) show that inhibition in great-tailed grackles is a multi-faceted construct, and demonstrate that the traditional go/no go task likely tests a very different aspect of inhibition than the detour task, which was never linked to any of their flexibility measures. Their comprehensive Bayesian analyses held up the results of some of the frequentist statistics, indicating a consistent relationship between flexibility and inhibition, with more flexible individuals also showing better inhibition (in the go/no go task). This same model, combined with inconsistencies in the GLM analyses (depending on the inclusion or exclusion of an outlier), led them to recommend caution in the creation of arbitrary thresholds for "success" in any cognitive tasks. Their accidental longer-term data collection also hinted at patterns of behaviour that shorter-term data collection did not. Of course, researchers have to decide on success criteria in order to conduct experiments, but in the same way that frequentist statistics are acknowledged to have flaws, the setting of success criteria must be acknowledged as inherently arbitrary. Where possible, researchers could reveal novel, biologically salient patterns by continuing beyond the point where a convenient success criterion has been reached. This research also underscores that tests may not be examining the features we expected them to measure, and are highly sensitive to biological and ecological variation between species as well as individual variation within populations.

To me, this study is an excellent argument for pre-registration of research (registered as Logan et al. 2019 and accepted by Vogel 2019), as the authors did not end up cherry-picking only those results or methods that worked. The fact that some of the tests did not "work", but was still examined, added much value to the study. The current paper is a bit densely written because of the comprehensiveness of the research. Some editorial polishing would likely make for more elegant writing. However, the arguments are clear, the results novel, and the questions thoroughly examined. The results are important not only for cognitive research on birds, but are potentially valuable to any cognitive scientist. I recommend this article as excellent food for thought.

References

Logan CJ, McCune K, Johnson-Ulrich Z, Bergeron L, Seitz B, Blaisdell AP, Wascher CAF. (2019) Are the more flexible individuals also better at inhibition? http://corinalogan.com/Preregistrations/g_inhibition.html  In principle acceptance by PCI Ecology of the version on 6 Mar 2019

Logan CJ, McCune KB, MacPherson M, Johnson-Ulrich Z, Rowney C, Seitz B, Blaisdell AP, Deffner D, Wascher CAF (2021) Are the more flexible great-tailed grackles also better at behavioral inhibition? PsyArXiv, ver. 7 peer-reviewed and recommended by Peer community in Ecology. https://doi.org/10.31234/osf.io/vpc39

Vogel E (2019) Adapting to a changing environment: advancing our understanding of the mechanisms that lead to behavioral flexibility. Peer Community in Ecology, 100016. https://doi.org/10.24072/pci.ecology.100016 

Biodiversity monitoring increasingly relies on modern technologies such as sensor networks and environmental DNA. These high-throughput methods allow biodiversity assessments with unprecedented detail and are especially useful to detect rare and secretive species that are otherwise difficult to observe with traditional survey-based methods. False negatives through imperfect detection are a typical problem in survey data and depend on intrinsic characteristics of the species, site characteristics of the survey site as well as survey characteristics (Guillera 2017). While imperfect detection might be reduced in modern sensor data and eDNA data, also these types of data are by no means error-free and may bare other challenges. In particular, the bioinformatics and image classification approaches used for species identification from these data can induce a higher rate of false positives than would be expected in expert-based survey data (Hartig et al. 2024).

Occupancy models (or occupancy-detection models) have been widely used to map species distributions by fitting a hierarchical model that estimates the paramaters of both the species-environment relationship and an observation submodel. They account for false negatives by inferring detectability from the detection history of a survey location, for example from replicate visits or multiple observers (Guillera 2017). These basic occupancy-detection models assume no false positive errors in the data. Other authors have proposed extensions for false positives that typically rely on unambiguous (known truth) information for some sites or observations (Chambert et al. 2015).

In their preprint, Monchy et al. (2024) propose an extension of classic occupancy models that considers a two-step observation process modelling the detection probability at occupied sites and the associated identification probability, separated into the true positive identification rate and the true negative identification rate. Using a simulation approach, the authors compare the effectiveness of a frequentist (maximum likelihood-based) and Bayesian approach for parameter estimation and identifiability, and additionally test the effectiveness of different priors (from non-informative to highly informative). Results of the maximum-likelihood approach indicated biased parameter estimates and identifiability problems. In the Bayesian approach, inclusion of prior information greatly reduces biases in parameter estimates, especially in detection and positive identification rate.

Importantly, informative priors for the identification process are a by-product of the classifiers that are developed for processing the eDNA data or sensor data. For example, species identification from acoustic sensors is based on image classifiers trained on labelled bird song spectrograms (Kahl et al. 2021) and as part of the evaluation of the classifier, the true positive rate (sensitivity) is routinely being estimated and could thus be readily used in occupancy models accounting for false positives. Thus, the approach proposed by Monchy et al. (2024) is not only highly relevant for biodiversity assessments based on novel sensor and eDNA data but also provides very practical solutions that do not require additional unambiguous data but recycle data that are already available in the processing pipeline. Applying their framework to real-world data will help reducing biases in biodiversity assessments and through improved understanding of the detection process it could also help optimising the design of sensor networks.

References

Thierry Chambert,  David A. W. Miller,  James D. Nichols (2015), Modeling false positive detections in species occurrence data under different study designs. Ecology, 96: 332-339. https://doi.org/10.1890/14-1507.1

Gurutzeta Guillera-Arroita (2017) Modelling of species distributions, range dynamics and communities under imperfect detection: advances, challenges and opportunities. Ecography, 40: 281-295. https://doi.org/10.1111/ecog.02445

Florian Hartig, Nerea Abrego, Alex Bush, Jonathan M. Chase, Gurutzeta Guillera-Arroita, Mathew A. Leibold,  Otso Ovaskainen, Loïc Pellissier, Maximilian Pichler, Giovanni Poggiato, Laura Pollock, Sara Si-Moussi, Wilfried Thuiller, Duarte S. Viana, David I. Warton, Damaris Zurell D, Douglas W. Yu (2024) Novel community data in ecology - properties and prospects. Trends in Ecology & Evolution, 39: 280-293. https://doi.org/10.1016/j.tree.2023.09.017

Stefan Kahl, Connor M. Wood, Maximilian Eibl, Holger Klinck (2021) BirdNET: A deep learning solution for avian diversity monitoring. Ecological Informatics, 61: 101236. https://doi.org/10.1016/j.ecoinf.2021.101236

Célian Monchy, Marie-Pierre Etienne, Olivier Gimenez (2024) Using informative priors to account for identifiability issues in occupancy models with identification errors. bioRxiv, ver.3 peer-reviewed and recommended by PCI Ecology https://doi.org/10.1101/2024.05.07.592917