Perspectives in machine learning for wildlife conservation

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

An interdisciplinary community of computer scientists and ecologists is emerging.

“However, an interdisciplinary community of computer scientists and ecologists is emerging, which we hope will tackle this technological and societal challenge together.”

May combine two claims.

Central challenges to efficient data analysis are the sheer volume of data generated by modern collection methods and the heterogeneous nature of many ecological datasets, which preclude the use of simple automated analysis techniques.

“Central challenges to efficient data analysis are the sheer volume of data generated by modern collection methods and the heterogeneous nature of many ecological datasets, which preclude the use of simple automated analysis techniques26.”

• cites [26] Farley, S. S., Dawson, A., Goring, S. J. & Williams, J. W. Situating ecology as a big-data science: current advances, challenges, and solutions. BioScience 68, 563–576 (2018). sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Current processing approaches inefficiently distill sensor data into relevant information.

“but are limited by current processing approaches which inefficiently distill data into relevant information”

Probably not a substantial claim: may not pass S4 (based).

Inexpensive and accessible sensors are accelerating data acquisition in animal ecology.

“Inexpensive and accessible sensors are accelerating data acquisition in animal ecology.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Deep learning models have shown superior performance across the majority of machine learning use cases.

“a family of highly versatile ML models based on artificial neural networks that have shown superior performance across the majority of ML use cases”

• cites [32] LeCun, Y., Bengio, Y. & Hinton, G. Deep learning. Nature 521, 436–444 (2015). sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Hybrid models leveraging both deep learning and bio-physical models tend to be less data-intensive, avoid incoherent predictions, and are generally more interpretable than purely data-driven models.

“such hybrid models tend to be less data-intensive, avoid incoherent predictions, and are generally more interpretable than purely data-driven models.”

• cites [48] Reichstein, M. et al. Deep learning and process understanding for data-driven Earth system science. Nature 566, 195–204 (2019). sourceNot checked yet

  Not checked in the free run.

• cites [49] Camps-Valls, G., Tuia, D., Zhu, X. X. & Reichstein, M. Deep Learning for the Earth Sciences: A Comprehensive Approach to Remote Sensing, Climate Science and Geosciences (Wiley & Sons, 2021). sourceNot checked yet

  Not checked in the free run.

• cites [50] Karpatne, A. et al. Theory-guided data science: A new paradigm for scientific discovery from data. IEEE Trans. Knowl. Data Eng. 29, 2318–2331 (2017). sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Compared to camera traps, passive acoustic monitoring is mostly unaffected by light and weather conditions, senses the environment omnidirectionally, and tends to be cost-effective when data needs to be collected at large spatiotemporal scales with high resolution.

“Compared to camera traps, PAM is mostly unaffected by light and weather conditions (some factors like wind still play a role), senses the environment omnidirectionally, and tends to be cost-effective when data needs to be collected at large spatiotemporal scales with high resolution60.”

• cites [60] Wood, C. M. et al. Detecting small changes in populations at landscape scales: a bioacoustic site-occupancy framework. Ecol. Indic. 98, 492–507 (2019). sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The viability of community science services may be limited depending on the task and oftentimes substantial efforts are required to verify volunteer-annotated data.

“However, it is prudent to note that the viability of community science services may be limited depending on the task and that oftentimes substantial efforts are required to verify volunteer-annotated data.”

May combine two claims.

Various supervised machine learning methods have shown potential in automating behavior analysis from accelerometer data, identifying behavioral state from trajectories, and predicting animal movement.

“Various supervised ML methods have shown their potential in automating behavior analysis from accelerometer data68,69, identifying behavioral state from trajectories70, and predicting animal movement71.”

• cites [68] Browning, E. et al. Predicting animal behaviour using deep learning: GPS data alone accurately predict diving in seabirds. Methods Ecol. Evolution 9, 681–692 (2018). sourceNot checked yet

  Not checked in the free run.

• cites [69] Liu, Z. Y.-C. et al. Deep learning accurately predicts white shark locomotor activity from depth data. Anim. Biotelemetry 7, 1–13 (2019). sourceNot checked yet

  Not checked in the free run.

• cites [70] Wang, G. Machine learning for inferring animal behavior from location and movement data. Ecol. Inform. 49, 69–76 (2019). sourceNot checked yet

  Not checked in the free run.

May combine two claims.

While machine learning has been extensively applied to camera trap images, its application to long-term passive acoustic monitoring datasets is still in its infancy and the first deep-learning-based studies are only starting to appear.

“While ML has been extensively applied to camera trap images, its application to long-term PAM datasets is still in its infancy and the first DL-based studies are only starting to appear”

• cites [61] Kahl, S., Wood, C. M., Eibl, M. & Klinck, H. Birdnet: a deep learning solution for avian diversity monitoring. Ecol. Inform. 61, 101236 (2021). sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Flying at lower altitudes, for example lower than 150 m, can have a significant impact on group and individual behavior of mammals, although the severity of wildlife disturbance from drone deployments depends heavily on the focal species, the equipment used, and characteristics of the drone flight.

“A recent study80 showed that flying at lower altitudes (e.g., lower than 150 m) can have a significant impact on group and individual behavior of mammals, although the severity of wildlife disturbance from drone deployments will depend heavily on the focal species, the equipment used, and characteristics of the drone flight (such as approach speed and altitude)81”

• cites [80] Schroeder, N. M., Panebianco, A., Gonzalez Musso, R. & Carmanchahi, P. An experimental approach to evaluate the potential of drones in terrestrial mammal research: a gregarious ungulate as a study model. R. Soc. open Sci. 7, 191482 (2020). sourceNot checked yet

  Not checked in the free run.

• cites [81] Bennitt, E., Bartlam-Brooks, H. L. A., Hubel, T. Y. & Wilson, A. M. Terrestrial mammalian wildlife responses to Unmanned Aerial Systems approaches. Sci. Rep. 9, 1–10 (2019). sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

On-animal sensors are the most common remote sensing devices deployed in animal ecology.

“On-animal sensors are the most common remote sensing devices deployed in animal ecology10.”

• cites [10] Kays, R., Crofoot, M. C., Jetz, W. & Wikelski, M. Terrestrial animal tracking as an eye on life and planet. Science 348, aaa2478 (2015). sourceNot checked yet

  Not checked in the free run.

May combine two claims.

The inherent complexity of soundscapes requires noise-robust algorithms that generalize well and can separate and identify many animal sounds of interest from confounding natural and anthropogenic signals in a wide variety of acoustic environments.

“Second, the inherent complexity of soundscapes requires noise-robust algorithms that generalize well and can separate and identify many animal sounds of interest from confounding natural and anthropogenic signals in a wide variety of acoustic environments62.”

• cites [62] Stowell, D., Wood, M. D., Pamuła, H., Stylianou, Y. & Glotin, H. Automatic acoustic detection of birds through deep learning: the first bird audio detection challenge. Methods Ecol. Evolution 10, 368–380 (2019). sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable).

Integrating ecological domain knowledge into machine learning methods is essential to designing models that are accurate in the way they describe animal life.

“integrating ecological domain knowledge into ML methods is essential to designing models that are accurate in the way they describe animal life.”

Probably not a substantial claim: may not pass S2 (checkable).

The challenging nature of ecological data, compounded by the size of datasets generated by novel sensors and the ever-increasing complexity of state-of-the-art machine learning methods, favors a collaborative approach that harnesses the expertise of both the machine learning and animal ecology communities, rather than an application of off-the-shelf machine learning methodologies to ecological挑战.

“We believe that the challenging nature of ecological data, compounded by the size of the datasets generated by novel sensors and the ever-increasing complexity of state-of-the-art ML methods, favor a collaborative approach that harnesses the expertise of both the ML and animal ecology communities, rather than an application of off-the-shelf ML methodologies to ecological challenges.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Increased scale of available data due to de-siloing efforts from organizations like Wildlife Insights and LILA.science will help increase machine learning accuracy and robustness across regions and taxa.

“The increased scale of available data due to de-siloing efforts from organizations like Wildlife Insights (www.wildlifeinsights.org) and LILA.science (www.lila.science) will help increase ML accuracy and robustness across regions and taxa.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The Great Grevy's Rally was hosted for the first time in January 2016 and has become a regular event held every other year.

“Hosted for the first time in January 2016, it has become a regular event, held every other year.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

Sensor data has the potential to unlock ecological understanding on a scale difficult to imagine in the recent past.

“The sensor data described in the previous section has the potential to unlock ecological understanding on a scale difficult to imagine in the recent past.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Hundreds of people capture approximately 50,000 images over the 2-day Great Grevy's Rally event.

“Hundreds of people, from school children and park rangers, to Nairobi families and international tourists, embark on a mission to photograph Grevy’s zebras across its range in Kenya, capturing ~50,000 images over the 2-day event.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Wildbook data has become the basis for IUCN Red List global population numbers for several species.

“Wildbook data has become the basis for the IUCN Red List global population numbers for several species”

May combine two claims.

Using volunteer-acquired images provides a substantial increase in the number of images and the chances of photographing species or single individuals in different regions, poses, and viewing angles.

“it provides a substantial increase in the number of images and the chances of photographing species or single individuals in different regions, poses, and viewing angles.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The Great Grevy's Rally is a successful demonstration of the power of community science-based wildlife monitoring.

“The Great Grevy’s Rally, a community science-based wildlife census effort occurring every 2 years in Kenya92, is a successful demonstration of the power of community science-based wildlife monitoring.”

• cites [92] Parham, J., Crall, J., Stewart, C., Berger-Wolf, T. & Rubenstein, D. I. Animal population censusing at scale with citizen science and photographic identification. in AAAI Spring Symposium-Technical Report (2017).Not checked yet

  Not checked in the free run.

May combine two claims.

Wildbook supports collaborative mark-recapture, molecular ecology, and social ecology studies, especially where community science and artificial intelligence can help scale-up projects.

“Wildbook supports collaborative mark-recapture, molecular ecology, and social ecology studies, especially where community science and artificial intelligence can help scale-up projects.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

The Great Grevy's Rally has become the foundation of the Kenya Wildlife Service's Grevy's zebra endangered species management policy and generates the official IUCN Red List population numbers for the species.

“The Great Grevy’s Rally has become the foundation of the Kenya Wildlife Service’s Grevy’s zebra endangered species management policy and generates the official IUCN Red List population numbers for the species.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Annotator errors include misdetected or mislabeled animals due to annotator fatigue or insufficient knowledge, and systematic errors from adversarial annotators.

“This is due to annotator errors, including misdetected or mislabeled animals due to annotator fatigue or insufficient knowledge about the annotation task, as well as systematic errors from adversarial annotators.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Wildbook is an open-source software platform that blends structured wildlife research with artificial intelligence, community science, and computer vision to speed population analysis and develop new insights to help conservation.

“Wildbook, a project of the non-profit Wild Me, is an open-source software platform that blends structured wildlife research with artificial intelligence, community science, and computer vision to speed population analysis and develop new insights to help conservation”

May combine two claims.

For simple tasks like animal detection, community science effort can be open to the public, while for more complex tasks such as identifying bird species with subtle appearance differences, communities of experts are needed to provide accurate labels.

“For simple tasks like animal detection, community science effort can be open to the public, while for more complex ones such as identifying bird species with subtle appearance differences (“fine-grained classification”, also see the glossary), communities of experts are needed to provide accurate labels.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Out of more than 120,000 species monitored by the IUCN Red List of Threatened Species, up to 17,000 have a Data deficient status.

“out of more than 120,000 species monitored by the IUCN Red List of Threatened Species, up to 17,000 have a ‘Data deficient’ status2.”

• cites [2] Committee, T. I. R. L. The IUCN Red List of Threatened Species - Strategic Plan 2017-2020. Tech. Rep., IUCN (2017).Not checked yet

  Not checked in the free run.

May combine two claims.

Volunteer-annotated data services may be limited depending on the task and often require substantial verification effort.

“services may be limited depending on the task and that oftentimes substantial efforts are required to verify volunteer-annotated data.”

May combine two claims.

With the ability to identify individual animals in images, Wildbook can enable an accurate population census and track population trends over time.

“With the ability to identify individual animals in those images, Wildbook can enable an accurate population census and track population trends over time.”

May combine two claims.

Using volunteer-acquired images sacrifices control over image acquisitions and is prone to significant noise, for example through low-quality imagery.

“Although this approach sacrifices control over image acquisitions and is likewise prone to inducing significant noise to datasets, for example through low-quality imagery”

Probably not a substantial claim: may not pass S3 (load-bearing).

There are currently Wildbooks for over 50 species, from sea dragons to zebras, spanning the entire planet.

“There are currently Wildbooks for over 50 species, from sea dragons to zebras, spanning the entire planet.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

More than 80 scientific publications have been enabled by Wildbook.

“More than 80 scientific publications have been enabled by Wildbook.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Wildbook supported the change in conservation status for whale sharks from vulnerable to endangered.

“supported the change in conservation status for whale sharks from “vulnerable” to “endangered”.”

May combine two claims.

Wildbook image analysis can start with images from any source and use ML and computer vision to detect multiple animals, classify their species, and identify individual animals.

“The image analysis of Wildbook can start with images from any source—scientists, camera traps, drones, community scientists, or social media—and use ML and computer vision to detect multiple animals in the images100 to not only classify their species, but identify individual animals applying a suite of different algorithms101,147.”

• cites [100] Parham, J. et al. An animal detection pipeline for identification. in IEEE Winter Conference on Applications of Computer Vision (WACV), 1075–1083 (IEEE, 2018). sourceNot checked yet

  Not checked in the free run.

• cites [101] Weideman, H. et al. Extracting identifying contours for African elephants and humpback whales using a learned appearance model. in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2020). sourceNot checked yet

  Not checked in the free run.

• cites [147] Arzoumanian, Z., Holmberg, J. & Norman, B. An astronomical pattern-matching algorithm for computer-aided identification of whale sharks Rhincodon typus. J. Appl. Ecol. 42, 999–1011 (2005). sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable).

To unlock that potential, sensor data must be interpreted and converted to usable information for ecological research.

“But to do so, it must be interpreted and converted to usable information for ecological research.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The measures yielded by this conversion are sometimes referred to as animal biometrics.

“The measures yielded by this conversion, reviewed in this section, are also sometimes referred to as animal biometrics93.”

• cites [93] Kühl, H. S. & Burghardt, T. Animal biometrics: quantifying and detecting phenotypic appearance. Trends Ecol. Evolution 28, 432–441 (2013). sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Computational restrictions in these models limit the number of training examples, likely reducing generalization to new sensor deployments or regions.

“Second, computational restrictions in these models limit the number of training examples, which is likely to have detrimental effects on variations in data (temporal, seasonal, etc.), thus reducing the generalization capabilities to new sensor deployments or regions.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

WPS used the MegaDetector to analyze over 900K images last year, which comes out to 2.5K images per day.

“WPS used the MegaDetector to analyze over 900K images last year alone, which comes out to 2.5K images per day.”

Probably not a substantial claim: may not pass S4 (based).

Conservation efforts of endangered species require knowledge on how many individuals of the species in question are present in a study area.

“Conservation efforts of endangered species require knowledge on how many individuals of the species in question are present in a study area.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

WPS have over 400 connected cameras deployed in 18 different countries, and that number is growing rapidly.

“They have over 400 connected cameras deployed in 18 different countries, and that number is growing rapidly.”

Probably not a substantial claim: may not pass S4 (based).

Imaging sensors paired with ML models that detect and count individuals in imagery can provide the most significant input to these observations.

“It is these observations where imaging sensors (camera traps, UAVs, etc.), paired with ML models that detect and count individuals in the imagery, can provide the most significant input.”

May combine two claims.

MegaDetector is a generic, global-scale human, animal, and vehicle detection model that works off-the-shelf for most camera trap data.

“This generic, global-scale human, animal, and vehicle detection model works off-the-shelf for most camera trap data”

• cites [36] Beery, S., Morris, D. & Yang, S. Efficient pipeline for camera trap image review. in Proceedings of the Workshop Data Mining and AI for Conservation, Conference for Knowledge Discovery and Data Mining (2019).Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Tasks involved in different approaches show similarities with traditional tasks in ML and computer vision, such as detection, localization, identification, and pose estimation.

“Interestingly, the tasks involved in the different approaches show similarities with traditional tasks in ML and computer vision (e.g., detection, localization, identification, pose estimation), for which we provide a matching example in animal ecology in Fig. 5.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Such estimations are conventionally realized with statistical occurrence models informed by sample-based species observations.

“Such estimations are conventionally realized with statistical occurrence models that are informed by sample-based species observations.”

May combine two claims.

Using MegaDetector allowed IDFG to label all data before the start of the next monitoring season, whereas manual labeling previously resulted in a lag of approximately 5 years from image collection to completion of labeling.

“allowed IDFG to label all data before the start of the next monitoring season, whereas manual labeling previously resulted in a lag of ~5 years from image collection to completion of labeling.”

May combine two claims.

IDFG relies heavily on camera traps to estimate and monitor the wolf population, and needs to process data each year before the next season to make informed policy changes or conservation interventions.

“IDFG relies heavily on camera traps to estimate and monitor this wolf population, and needs to process the data collected each year before the start of the next season in order to make informed policy changes or conservation interventions.”

May combine two claims.

The standard for identification of animal species and identity is DNA profiling, which can be difficult to scale to large, distributed populations.

“The standard for identification of animal species and identity is DNA profiling97, which can be difficult to scale to large, distributed populations54,93.”

• cites [97] Avise, J. C. Molecular Markers, Natural History and Evolution (Springer Science & Business Media, 2012).Not checked yet

  Not checked in the free run.

• cites [54] Schneider, S., Taylor, G. W., Linquist, S. & Kremer, S. C. Past, present and future approaches using computer vision for animal re-identification from camera trap data. Methods Ecol. Evolution 10, 461–470 (2019). sourceNot checked yet

  Not checked in the free run.

• cites [93] Kühl, H. S. & Burghardt, T. Animal biometrics: quantifying and detecting phenotypic appearance. Trends Ecol. Evolution 28, 432–441 (2013). sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Conversion of sensor data can take the form of abundance mapping, individual animal re-identification, herd tracking, or digital reconstruction of the environment.

“For example, such conversion can take the form of abundance mapping, individual animal re-identification, herd tracking, or digital reconstruction (three-dimensional, phenotypical) of the environment the animals live in.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The publicly-hosted MegaDetector API has been integrated into wildlife monitoring workflows of over 30 organizations worldwide.

“the publicly-hosted MegaDetector API has been integrated into the wildlife monitoring workflows of over 30 organizations worldwide, including the Wildlife Conservation Society, San Diego Zoo Global, and Island Conservation.”

May combine two claims.

In that study, humans achieved identification accuracy between 20% (novices) and 42% (experts) with annotation time between 1 and 2 hours, while the DL model achieved 84% in seconds.

“where humans achieved identification accuracy between 20% (novices) and 42% (experts) with an annotation time between 1 and 2 h, the DL model achieved an identification accuracy of 84% in a matter of seconds.”

• cites [38] Schofield, D. et al. Chimpanzee face recognition from videos in the wild using deep learning. Sci. Adv. 5, eaaw0736 (2019). sourceNot checked yet

  Not checked in the free run.

May combine two claims.

A study of 23 chimpanzees in Guinea applied facial recognition to a multi-year dataset of 1 million images and achieved greater than 90% accuracy.

“a study of a small group of 23 chimpanzees in Guinea applied facial recognition techniques to a multi-year dataset comprising 1 million images and achieved >90% accuracy38.”

• cites [38] Schofield, D. et al. Chimpanzee face recognition from videos in the wild using deep learning. Sci. Adv. 5, eaaw0736 (2019). sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S4 (based).

The 3D shape of an individual can be related to its health, age, or reproductive status.

“The 3D shape of an individual can be related to its health, age, or reproductive status”

Probably not a substantial claim: may not pass S3 (load-bearing).

As an alternative to gene-based identification, manual tagging can be used to keep track of individual animals.

“As an alternative to gene-based identification, manual tagging can be used to keep track of individual animals10,93.”

• cites [10] Kays, R., Crofoot, M. C., Jetz, W. & Wikelski, M. Terrestrial animal tracking as an eye on life and planet. Science 348, aaa2478 (2015). sourceNot checked yet

  Not checked in the free run.

• cites [93] Kühl, H. S. & Burghardt, T. Animal biometrics: quantifying and detecting phenotypic appearance. Trends Ecol. Evolution 28, 432–441 (2013). sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Particular challenges for animal re-identification in wild populations include difficulty of manual curation, larger populations, changes in appearance, few sightings per individual, and frequent addition of new individuals, creating an open-set problem.

“The particular challenges for animal (re-)identification in wild populations are related to the difficulty of manual curation, larger populations, changes in appearance (e.g., due to scars, growth), few sightings per individual, and the frequent addition of new individuals that may enter the system due to birth or immigration, therefore creating an “open-set” problem104 wherein the model must deal”

• cites [104] Bendale, A. & Boult, T. E. Towards open set deep networks. in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 1563–1572 (2016). sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Recent work defines the SMAL model, a 3D articulated shape model for a set of quadruped families.

“Recent work defines the SMAL (Skinned Multi-Animal Linear) model, a 3D articulated shape model for a set of quadruped families107.”

• cites [107] Zuffi, S., Kanazawa, A., Jacobs, D. W. & Black, M. J. 3D menagerie: modeling the 3D shape and pose of animals. in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 6365–6373 (2017). sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

3D shape models have recently been defined also for birds.

“3D shape models have been recently defined also for birds111.”

• cites [111] Wang, Y., Kolotouros, N., Daniilidis, K. & Badger, M. Birds of a feather: capturing avian shape models from images. in 2021 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 14739–14749 (2021). sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

3D reconstructions of kill sites for lions in South Africa revealed novel insights into predator-prey relationships and their connection to ecosystem stability and functioning.

“3D reconstructions of kill sites for lions in South Africa revealed novel insights into the predator-prey relationships and their connection to ecosystem stability and functioning114”

• cites [114] Davies, A. B., Tambling, C. J., Kerley, G. I. H. & Asner, G. P. Effects of vegetation structure on the location of lion kill sites in African thicket. PLoS ONE 11, e0149098 (2016). sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

In ref. 110, the SMAL model was used in a DL approach to predict 3D shape and pose of the Grevy's zebra from images.

“In ref. 110, the SMAL model has been used in a DL approach to predict 3D shape and pose of the Grevy’s zebra from images.”

• cites [110] Zuffi, S., Kanazawa, A., Berger-Wolf, T. & Black, M. J. Three-D safari: learning to estimate zebra pose, shape, and texture from images" in the wild". in 2019 IEEE International Conference on Computer Vision (ICCV), 5359–5368 (2019). sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Spatial reconstructions of the environment can be extracted using dedicated sensors such as LiDAR or reconstructed from multiple images via mosaicking or structure from motion/SLAM.

“Such spatial reconstructions of the environment can either be extracted by using dedicated sensors such as LiDAR116 or can be reconstructed from multiple images, either by stitching the images into a unified two-dimensional panorama (e.g., mosaicking117) or by computing the three-dimensional environment from partially overlapping images (e.g., structure from motion118 or simultaneous localization”

• cites [116] Risse, B., Mangan, M., Stürzl, W. & Webb, B. Software to convert terrestrial LiDAR scans of natural environments into photorealistic meshes. Environ. Model. Softw. 99, 88–100 (2018). sourceNot checked yet

  Not checked in the free run.

• cites [117] Haalck, L. & Risse, B. Embedded dense camera trajectories in multi-video image mosaics by geodesic interpolation-based reintegration. in 2021 IEEE Winter Conference on Applications of Computer Vision (WACV), 1849–1858 (2021). sourceNot checked yet

  Not checked in the free run.

• cites [118] Schonberger, J. L. & Frahm, J.-M. Structure-from-motion revisited. in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 4104–4113 (2016). sourceNot checked yet

  Not checked in the free run.

May combine two claims.

The 3D pose of the body can provide finer information on posture attributes and allows kinematic and behavioral analyses.

“the 3D pose of the body can provide finer information with respect to posture attributes and allows, for instance, kinematic as well as behavioral analyses.”

May combine two claims.

Satellite remote sensing has become an integral part to study animal habitats, biological diversity, and spatiotemporal changes of abiotic conditions, since it allows mapping land cover, soil moisture, or temperature at scale.

“Satellite remote sensing has become an integral part to study animal habitats, biological diversity, and spatiotemporal changes of abiotic conditions113, since it allows to map quantities like land cover, soil moisture, or temperature at scale.”

• cites [113] Pettorelli, N. et al. Satellite remote sensing for applied ecologists: opportunities and challenges. J. Appl. Ecol. 51, 839–848 (2014). sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Image-based 3D pose and shape estimation methods provide rich information about individuals but require accurate shape models and prior knowledge about the animal's 3D motion.

“Image-based 3D pose and shape estimation methods provide rich information about individuals but require, in addition to accurate shape models, prior knowledge about the animal’s 3D motion.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Biggs et al. built on SMAL for 3D shape and motion of dogs from video and for recovery of dog shape and pose across many breeds.

“Biggs et al. built on this work for 3D shape and motion of dogs from video108 and for recovery of dog shape and pose across many different breeds109.”

• cites [108] Biggs, B., Roddick, T., Fitzgibbon, A. & Cipolla, R. Creatures great and smal: recovering the shape and motion of animals from video. in 2018 Asian Conference on Computer Vision (ACCV), 3–19 (Springer, 2018). sourceNot checked yet

  Not checked in the free run.

• cites [109] Biggs, B., Boyne, O., Charles, J., Fitzgibbon, A. & Cipolla, R. Who left the dogs out? 3D animal reconstruction with expectation maximization in the loop. in 2020 European Conference on Computer Vision (ECCV), 195–211 (Springer, 2020). sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Wildlife behavior and conservation cannot be dissociated from the environment animals evolve and live in.

“Wildlife behavior and conservation cannot be dissociated from the environment animals evolve and live in.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Various user-friendly toolboxes are available to extract the 2D posture of animals from videos, while the user can define which body parts should be estimated.

“Various user-friendly toolboxes are available to extract the 2D posture of animals from videos (Fig. 1d, e), while the user can define which body parts should be estimated (reviewed in ref. 76).”

• cites [76] Mathis, A., Schneider, S., Lauer, J. & Mathis, M. W. A primer on motion capture with deep learning: principles, pitfalls, and perspectives. Neuron 108, 44–65 (2020). sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Marker-less DL methods for pose estimation have tremendously improved over the last years and already impacted biology.

“For pose estimation, marker-less methods based on DL have tremendously improved over the last years and already impacted biology105.”

• cites [105] Mathis, M. W. & Mathis, A. Deep learning tools for the measurement of animal behavior in neuroscience. Curr. Opin. Neurobiol. 60, 1–11 (2020). sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Graph regression techniques were deployed to reconstruct species interaction networks in a community of European birds with promising results, including better causality estimates of the relations in the graph.

“More recently, graph regression techniques were deployed to reconstruct species interaction networks in a community of European birds with promising results, including better causality estimates of the relations in the graph131.”

• cites [131] Faisal, A., Dondelinger, F., Husmeier, D. & Beale, C. M. Inferring species interaction networks from species abundance data: a comparative evaluation of various statistical and machine learning methods. Ecol. Inform. 5, 451–464 (2010). sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Non-parametric ML techniques like Random Forest have been successfully used to predict species richness and have shown significant error reduction with respect to traditional counterparts used in ecology, for example in fishes, spiders, and small mammals.

“Non-parametric ML techniques like Random Forest126 have been successfully used to predict species richness and have shown significant error reduction with respect to the traditional counterparts used in ecology, for example in the estimation of richness distributions of fishes127,128, spiders 129, and small mammals130.”

• cites [126] Breiman, L. Random forests. Mach. Learn. 45, 5–32 (2001). sourceNot checked yet

  Not checked in the free run.

• cites [127] Parravicini, V. et al. Global patterns and predictors of tropical reef fish species richness. Ecography 36, 1254–1262 (2013). sourceNot checked yet

  Not checked in the free run.

• cites [128] Smoliński, S. & Radtke, K. Spatial prediction of demersal fish diversity in the baltic sea: comparison of machine learning and regression-based techniques. ICES J. Mar. Sci. 74, 102–111 (2017). sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Spatially explicit linear regression models have been conventionally used to predict species and community distribution based on explanatory variables such as climate and topography.

“Spatially explicit linear regression models have been conventionally used to predict species and community distribution based on explanatory variables such as climate and topography124,125.”

• cites [124] Guisan, A. & Zimmermann, N. E. Predictive habitat distribution models in ecology. Ecol. Model. 135, 147–186 (2000). sourceNot checked yet

  Not checked in the free run.

• cites [125] Lehmann, A., Overton, J. M. & Austin, M. P. Regression models for spatial prediction: their role for biodiversity and conservation. Biodivers. Conserv. 11, 2085–2092 (2002). sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Analyses of biodiversity, represented by species abundance and richness, are foundational to much ecological research and many conservation initiatives.

“Analyses of biodiversity, represented by such measures as species abundance and richness, are foundational to much ecological research and many conservation initiatives.”

May combine two claims.

Tree-based methods are well-suited to these tasks because they perform explicit feature ranking and are able to model nonlinear relationships between covariates and species distribution.

“Tree-based methods are well-suited to these tasks because they perform explicit feature ranking (and thus feature selection) and are able to model nonlinear relationships between covariates and species distribution.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Security of rangers themselves is at stake; for example, the flight path of drones might be backtracked to reveal their location.

“Security of rangers themselves is also at stake; for example, the flight path of drones might be backtracked to reveal their location.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Population counts without uncertainty estimation can lead to erroneous and potentially devastating conclusions.

“population counts without uncertainty estimation can lead to erroneous and potentially devastating conclusions.”

May combine two claims.

Biases are not limited to geographical provenance: the type of sensors used, the species they depict, and the imbalance in the number of individuals observed per species must also be considered when training or using models to avoid potentially catastrophic drop-offs in accuracy.

“Biases are not limited to the geographical provenance of images: the type of sensors used (RGB vs. infrared or thermal), the species they depict, and the imbalance in the number of individuals observed per species55,132 must also be considered when training or using models to avoid potentially catastrophic drop-offs in accuracy”

• cites [55] Beery, S., Van Horn, G. & Perona, P. Recognition in terra incognita. in 2018 European Conference on Computer Vision (ECCV), 456–473 (2018). sourceNot checked yet

  Not checked in the free run.

• cites [132] Van Horn, G. et al. The inaturalist species classification and detection dataset. in 2018 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 8769–8778 (2018). sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

The universality of ML is now strongly questioned and the inductive bias of traditional DL models is challenged by new approaches that bridge domain knowledge, fundamental laws, and data science.

“Such universality of ML is now strongly questioned and the inductive bias of traditional DL models is challenged by new approaches that bridge domain knowledge, fundamental laws, and data science.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Machine and deep learning are becoming necessary accelerators for wildlife research and conservation actions in natural reserves.

“Machine and deep learning are becoming necessary accelerators for wildlife research and conservation actions in natural reserves.”

Probably not a substantial claim: may not pass S2 (checkable).

Without quality control measures, relying on model-based results is risky and could have difficult-to-evaluate impacts on research in animal ecology, as incorrect results hidden in a suboptimally trained model will become more difficult to detect.

“Without such quality control measures, relying on model-based results is risky and could have difficult-to-evaluate impacts on research in animal ecology, as incorrect results hidden in a suboptimally trained model will become more and more difficult to detect.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Errors by models can lead to erroneous decisions on site that can be catastrophic for the population they are trying to preserve or for populations at the border of human/wildlife conflicts.

“Computer scientists must be aware that errors by their models can lead to erroneous decisions on site that can be catastrophic for the population they are trying to preserve or for the populations that live at the border of human/wildlife conflicts.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Cloud computing services exist that forgo the need of hardware management, but they pose per-resource costs that quickly scale to several thousands of dollars per month for a single virtual machine.

“Although cloud computing services exist that forgo the need of hardware management, they likewise pose per-resource costs that quickly scale to several thousands of dollars per month for a single virtual machine.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Models currently used in animal ecology are far from such a carbon footprint.

“Even though the models currently used in animal ecology are far from such a carbon footprint”

Probably not a substantial claim: may not pass S3 (load-bearing).

Recently, it has been estimated that large, state-of-the-art models for understanding natural language emit as much carbon as several cars in their entire lifetime.

“recently, it has been estimated that large, state-of-the-art models for understanding natural language emit as much carbon as several cars in their entire lifetime135.”

• cites [135] Strubell, E., Ganesh, A. & McCallum, A. Energy and policy considerations for deep learning in NLP. in Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 3645–3650 (2019). sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

Opening up existing datasets, especially private-owned images acquired by non-professionals, is both a necessary and difficult challenge for the near future.

“However, opening up existing datasets (and especially when using private-owned images acquired by non-professionals as in ref. 92) is both a necessary and difficult challenge for the near future.”

• cites [92] Parham, J., Crall, J., Stewart, C., Berger-Wolf, T. & Rubenstein, D. I. Animal population censusing at scale with citizen science and photographic identification. in AAAI Spring Symposium-Technical Report (2017).Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

In some cases, it is imperative that the privacy of the data be preserved, for instance to avoid giving poachers access to locations of animals in near-real-time.

“In some cases, it is imperative that the privacy of the data be preserved, for instance to avoid giving poachers access to locations of animals in near-real-time134.”

• cites [134] Lennox, R. J. et al. A novel framework to protect animal data in a world of ecosurveillance. BioScience 70, 468–476 (2020). sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Purchasing prices of such machines alone are often prohibitively high especially for budget-constrained conservation organizations; programming, running, and maintenance costs further add to the bill.

“Purchasing prices of such machines alone are often prohibitively high especially for budget-constrained conservation organizations; programming, running, and maintenance costs further add to the bill.”

Classified as probably not a substantial claim. May combine two claims.

Smaller models are not only less expensive to train and use, their lighter computational costs allow them to be run on smaller devices, opening opportunities for real-time ML on the edge.

“Smaller models are not only less expensive to train and use, their lighter computational costs allow them to be run on smaller devices, opening opportunities for real-time ML “on the edge”—i.e., within the sensors themselves.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Fostering a culture of individual and cross-institutional data sharing in ecology will allow ML approaches to improve in robustness and accuracy.

“Fostering a culture of individual and cross-institutional data sharing in ecology will allow ML approaches to improve in robustness and accuracy.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based). May combine two claims.

Estimations of models remain approximations and need to be treated as such.

“The estimations of models remain approximations and need to be treated as such”

Probably not a substantial claim: may not pass S2 (checkable).

Transparency around the training data and the intended model usage is a necessity.

“transparency around the training data and the intended model usage is a necessity133.”

• cites [133] Copas, K. et al. Training machines to improve species identification using GBIF-mediated datasets. in AGU Fall Meeting Abstracts, Vol. 2019, IN53C–0758 (2019).Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

State-of-the-art ML models are now being applied to many tasks in animal ecology and wildlife conservation.

“State-of-the-art ML models are now being applied to many tasks in animal ecology and wildlife conservation.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Environmental costs of AI are often disregarded, as energy consumption of large calculations is still considered an endless resource, assuming that the money to pay for it is available.

“environmental costs of AI are often disregarded, as energy consumption of large calculations is still considered an endless resource (assuming that the money to pay for it is available).”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Training and running models with millions of parameters on large volumes of data requires powerful, somewhat specialized hardware.

“Training and running models with millions of parameters on large volumes of data requires powerful, somewhat specialized hardware.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

The ML and DL fields have focused mainly on black box models that learn correlations from data directly, and domain knowledge has been repeatedly ignored in favor of generic approaches.

“The ML and DL fields have focused mainly on black box models that learn correlations from data directly, and domain knowledge has been repeatedly ignored in favor of generic approaches that could fit to any kind of dataset.”

May combine two claims.

Many open imagery repositories such as Artportalen.se, Naturgucker.de, and Waarneming.nl collect images from specific regions, and most contributions on iNaturalist come from the Northern hemisphere.

“For example, many open imagery repositories such as Artportalen.se, Naturgucker.de, and Waarneming.nl collect images from specific regions, and most contributions on iNaturalist132 (see Table 2) come from the Northern hemisphere.”

• cites [132] Van Horn, G. et al. The inaturalist species classification and detection dataset. in 2018 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 8769–8778 (2018). sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Laboratories rightly serve as the ultimate development space for such in-the-wild applications.

“Thus, laboratories rightly serve as the ultimate development space for such in-the-wild applications.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

This line of interdisciplinary research is small but growing, with several studies published in recent years.

“This line of interdisciplinary research is small but growing, with several studies published in recent years.”

May combine two claims.

Ref. 139 incorporates knowledge about the distribution of species and photographer biases into a DL model for species classification in images and reports accuracy improvements of up to 12% in iNaturalist over a baseline without such priors.

“Similarly ref. 139, incorporate knowledge about the distribution of species as well as photographer biases into a DL model for species classification in images and report accuracy improvements of up to 12% in iNaturalist over a baseline without such priors.”

• cites [139] Mac Aodha, O., Cole, E. & Perona, P. Presence-only geographical priors for fine-grained image classification. in Proceedings of the IEEE/CVF International Conference on Computer Vision, 9596–9606 (2019). sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

The hybrid models paradigm is one of the most exciting avenues in modern ML and promises real collaboration between domains of application and ML, especially when coupled with algorithmic designs that allow interpretation and understanding of visual cues.

“This “hybrid models” paradigm48,50 is one of the most exciting avenues in modern ML and promises real collaboration between domains of application and ML, especially when coupled with algorithmic designs that allow interpretation and understanding of the visual cues that are being used136.”

• cites [48] Reichstein, M. et al. Deep learning and process understanding for data-driven Earth system science. Nature 566, 195–204 (2019). sourceNot checked yet

  Not checked in the free run.

• cites [50] Karpatne, A. et al. Theory-guided data science: A new paradigm for scientific discovery from data. IEEE Trans. Knowl. Data Eng. 29, 2318–2331 (2017). sourceNot checked yet

  Not checked in the free run.

• cites [136] Samek, W., Montavon, G., Vedaldi, A., Hansen, L. K. & Müller, K.-R. Explainable AI: Interpreting, Explaining and Visualizing Deep Learning, Vol. 11700 (Springer Nature, 2019). sourceNot checked yet

  Not checked in the free run.

May combine two claims.

The high level of control afforded by laboratory settings facilitated the rapid development of such tools, many of which are now being adopted for use in field studies of free-moving animals in complex natural environments.

“The high level of control (e.g., of lighting conditions, sensor calibration, and environment) afforded by laboratory settings facilitated the rapid development of such tools, many of which are now being adopted for use in field studies of free-moving animals in complex natural environments75,142.”

• cites [75] Graving, J. M. et al. DeepPoseKit, a software toolkit for fast and robust animal pose estimation using deep learning. Elife 8, e47994 (2019). sourceNot checked yet

  Not checked in the free run.

• cites [142] Joska, D. et al. AcinoSet: a 3D pose estimation dataset and baseline models for Cheetahs in the wild. 2021 IEEE International Conference on Robotics and Automation (ICRA) Preprint at https://arxiv.org/abs/2103.13282 (IEEE, Xi’an, China, 2021). sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Ref. 140 used expert knowledge of park rangers to augment sparse and noisy records of poaching activity, improving predictions of poaching occurrence and enabling more efficient use of limited patrol resources in a Chinese nature reserve.

“Finally ref. 140, used expert knowledge of park rangers to augment sparse and noisy records of poaching activity, thereby improving predictions of poaching occurrence and enabling more efficient use of limited patrol resources in a Chinese nature reserve.”

• cites [140] Gurumurthy, S. et al. Exploiting Data and Human Knowledge for Predicting Wildlife Poaching. in Proceedings of the 1st ACM SIGCAS Conference on Computing and Sustainable Societies, 1–8, https://doi.org/10.1145/3209811.3209879 (ACM, 2018). sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Statistical models for species richness and diversity are routinely used to estimate abundances and study species co-occurrence and interactions.

“Statistical models for species richness and diversity are routinely used to estimate abundances and study species co-occurrence and interactions.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Algorithmic insights gained in the lab can be transferred back into the wild-studies on short videos or camera traps can leverage lab-generated data that is arguably less diverse, but easier to control.

“In addition, algorithmic insights gained in the lab can be transferred back into the wild—studies on short videos or camera traps can leverage lab-generated data that is arguably less diverse, but easier to control.”

Probably not a substantial claim: may not pass S3 (load-bearing).

This opens interesting research opportunities for the adaptation of lab-generated simulation to real-world conditions, similar to what has been observed in image synthesis for self driving and robotics.

“This opens interesting research opportunities for the adaptation of lab-generated simulation to real-world conditions, similar to what has been observed in the field of image synthesis for self driving143 and robotics144 in the last decade.”

• cites [143] Chen, Q. & Koltun, V. Photographic image synthesis with cascaded refinement networks. in 2019 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 1511–1520 (2017). sourceNot checked yet

  Not checked in the free run.

• cites [144] Lee, J., Hwangbo, J., Wellhausen, L., Koltun, V. & Hutter, M. Learning quadrupedal locomotion over challenging terrain. Sci. Robot. 5, eabc5986 (2020). sourceNot checked yet

  Not checked in the free run.

May combine two claims.

These approaches challenge the dogma of ML models learning exclusively from data and achieve more efficient model learning and enhanced plausibility of solutions.

“These approaches challenge the dogma of ML models learning exclusively from data and achieve more efficient model learning (since base knowledge is available from the start and does not have to be re-learnt) and enhanced plausibility of the solutions (because the solution space can be constrained to a range of ecologically plausible outcomes).”

Probably not a substantial claim: may not pass S4 (based).

Computer scientists have yet to integrate ecological knowledge such as underlying biological processes into machine learning models.

“Computer scientists have yet to integrate ecological knowledge such as underlying biological processes into ML models, and the lack of transparency of current DL models has so far been a major obstacle to incorporating ML into ecological research.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Open challenges in species distribution modeling include increasing the scale of joint models geospatially, temporally, and taxonomically; building methods that can leverage multiple data types despite bias from non-uniform sampling strategies; incorporating ecological knowledge such as species dispersal and community composition; and expanding methods for evaluation.

“Open challenges include increasing the scale of joint models geospatially, temporally, and taxonomically, building methods that can leverage multiple data types despite bias from non-uniform sampling strategies, incorporating ecological knowledge such as species dispersal and community composition, and expanding methods for the evaluation of these models.”

• cites [146] Beery, S., Cole, E., Parker, J., Perona, P. & Winner, K. Species distribution modeling for machine learning practitioners: a review. in Proceedings of the 4th ACM SIGCAS Conference on Computing and Sustainable Societies (2021). sourceNot checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

The vision described in the paper cannot be achieved without interdisciplinary thinking.

“Finally, we wish to re-emphasize that the vision described here cannot be achieved without interdisciplinary thinking: for all these exciting opportunities, processing big ecological data is necessitating analytical techniques of such complexity that no single ecologist can be expected to have all the technical expertise (plus domain knowledge) required to carry out groundbreaking studies65.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The GeoLifeCLEF task is complex because only positive counts are provided, no absence data are available, and predictions are counted as correct if the ground truth species is among the 30 predicted with highest confidence.

“The task is complex: only positive counts are provided, no absence data are available, and predictions are counted as correct if the ground truth species is among the 30 predicted with highest confidence.”

Probably not a substantial claim: may not pass S2 (checkable).

Improvements from machine learning and new generation sensors require ever-closer cooperation between ecologists and machine learning specialists, since recent approaches are more complex than ever and require strict quality control and detailed design knowledge.

“Although often spectacular, such improvements require ever-closer cooperation between ecologists and ML specialists, since recent approaches are more complex than ever and require strict quality control and detailed design knowledge.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

The lack of transparency of current deep learning models has so far been a major obstacle to incorporating machine learning into ecological research.

“Computer scientists have yet to integrate ecological knowledge such as underlying biological processes into ML models, and the lack of transparency of current DL models has so far been a major obstacle to incorporating ML into ecological research.”