Vincent Miele
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Vincent Miele has a total of 196 co-authors. Here are the top co-authors by number of shared publications:
- Stéphane Dray
Universite Claude Bernard Lyon 1, LBBE, UMR 5558, CNRS, VAS – Villeurbanne, France
10 shared publications - Simon Chamaillé-Jammes
CEFE, CNRS, Univ Montpellier, EPHE, IRD, Montpellier, France
9 shared publications - Gaspard Dussert
Laboratoire de Biométrie et Biologie Evolutive Université Lyon 1, CNRS Villeurbanne France
7 shared publications Franck Picard
UMR CNRS 5558 Université Lyon 1
5 shared publicationsBruno Spataro
Université Lyon 1, CNRS, Villeurbanne, France
4 shared publicationsCatherine Matias
Laboratoire de Probabilités, Statistique et Modélisation, Centre National de la Recherche Scientifique, Sorbonne Université et Université de Paris, Paris, France
4 shared publications
Latest Publications
- 30 October 2026
- Published by Wiley in Mammal Review
- Vol. 56 (4)
More Tools Than Users: A Scoping Review of Image‐Processing Automation in Mammal Camera Trapping
Introduction: Camera trapping has become a cornerstone of mammal monitoring, yet the volume of images frequently exceeds the capacity for manual review. Automated image‐processing tools—like AI‐based species classifiers and on‐device alerting systems—offer scalable solutions, but uptake in applied research (field‐based ecological studies) remains poorly characterised. Aims: We conducted a scoping review of AI in mammal camera‐trapping studies (2014–2026), complemented by a corresponding‐author survey, to assess adoption and synthesise the critical domains advancing the field. Methods: We screened 984 records from a stratified random sample of 8725 records identified through Web of Science and Scopus, retaining 654 publications, and surveyed the corresponding authors of 121 applied studies. Results: AI‐methods publications have outnumbered applied studies each year since 2023, yet only 9.1% (11 of 121) of surveyed researchers reported using AI in their applied work. Adoption emerged only after 2022 and was restricted to blank‐frame filtering alone (9.1%) or paired with species classification (90.9%). Other AI‐enabled tasks such as individual re‐identification, behaviour classification and real‐time alerting were absent from all 288 applied papers reviewed and survey responses. Their prevalence in applied work is unlikely to exceed 1.04% (97.5% upper bound). Synthesis and Outlook: Most image‐processing tools, blank filtering, species classification and real‐time alerting are well developed; adoption in applied research lags, as expected for emerging technology. Implementation also hinges on two still‐underdeveloped capacities: propagating classifier uncertainty into population estimates and automated camera‐to‐animal distance measurement for density estimation of unmarked populations. Bridging these gaps requires standardised reporting, interdisciplinary collaboration and uncertainty‐aware workflows.
- 30 October 2026
- Published by Wiley in Mammal Review
- Vol. 56 (4)
More Tools Than Users: A Scoping Review of Image‐Processing Automation in Mammal Camera Trapping
Introduction: Camera trapping has become a cornerstone of mammal monitoring, yet the volume of images frequently exceeds the capacity for manual review. Automated image‐processing tools—like AI‐based species classifiers and on‐device alerting systems—offer scalable solutions, but uptake in applied research (field‐based ecological studies) remains poorly characterised. Aims: We conducted a scoping review of AI in mammal camera‐trapping studies (2014–2026), complemented by a corresponding‐author survey, to assess adoption and synthesise the critical domains advancing the field. Methods: We screened 984 records from a stratified random sample of 8725 records identified through Web of Science and Scopus, retaining 654 publications, and surveyed the corresponding authors of 121 applied studies. Results: AI‐methods publications have outnumbered applied studies each year since 2023, yet only 9.1% (11 of 121) of surveyed researchers reported using AI in their applied work. Adoption emerged only after 2022 and was restricted to blank‐frame filtering alone (9.1%) or paired with species classification (90.9%). Other AI‐enabled tasks such as individual re‐identification, behaviour classification and real‐time alerting were absent from all 288 applied papers reviewed and survey responses. Their prevalence in applied work is unlikely to exceed 1.04% (97.5% upper bound). Synthesis and Outlook: Most image‐processing tools, blank filtering, species classification and real‐time alerting are well developed; adoption in applied research lags, as expected for emerging technology. Implementation also hinges on two still‐underdeveloped capacities: propagating classifier uncertainty into population estimates and automated camera‐to‐animal distance measurement for density estimation of unmarked populations. Bridging these gaps requires standardised reporting, interdisciplinary collaboration and uncertainty‐aware workflows.
- Open Access
- 4 March 2026
- Published by Wiley in Methods in Ecology and Evolution
- Vol. 17 (4) , 1248-1258
Paying attention to other animal detections improves camera trap classification models
In ecological studies, automated species classification models are increasingly used to process large volumes of camera trap images. Most current classification models rely on a two‐step pipeline: a detector first locates and crops animals, followed by a classifier that predicts species independently for each crop. While effective, these models still struggle under challenging conditions and ignore temporal context or information from nearby animals available in sequences of camera trap images, whereas human annotators often use it in difficult cases. We propose to leverage self‐attention, a core mechanism of Large Language Models and Vision Transformers, to enable the model to learn relationships between crops occurring in similar contexts. Our self‐attention module operates directly on the set of crop embeddings, producing new representations enriched with information from other crops, improving species classification. The module fits into the two‐step pipeline without requiring structural change and adds only minimal computational overhead. To address the lack of annotated multi‐species sequences, we develop a training strategy that synthetically generates multi‐species sequences from mono‐specific ones. Compared with an independent classifier baseline, our multi‐crop model achieves higher accuracy on mono‐specific sequences, both real and synthetic. In multi‐species settings, evaluated using synthetic test sets, we also observe a substantial improvement in accuracy. Using real but weakly annotated multi‐species sequences, we reformulate the task as multi‐label set classification and conduct a visual analysis to highlight the benefits of our approach. By leveraging the information brought by all detections of animals in the image and others of the same sequence, our approach reduces species misclassifications and enables more accurate estimates for downstream ecological analyses focusing on, for instance, species richness, occupancy and species interaction.
- 4 March 2026
- Published by Wiley in Methods in Ecology and Evolution
- Vol. 17 (4) , 1248-1258
Open AccessPaying attention to other animal detections improves camera trap classification models
In ecological studies, automated species classification models are increasingly used to process large volumes of camera trap images. Most current classification models rely on a two‐step pipeline: a detector first locates and crops animals, followed by a classifier that predicts species independently for each crop. While effective, these models still struggle under challenging conditions and ignore temporal context or information from nearby animals available in sequences of camera trap images, whereas human annotators often use it in difficult cases. We propose to leverage self‐attention, a core mechanism of Large Language Models and Vision Transformers, to enable the model to learn relationships between crops occurring in similar contexts. Our self‐attention module operates directly on the set of crop embeddings, producing new representations enriched with information from other crops, improving species classification. The module fits into the two‐step pipeline without requiring structural change and adds only minimal computational overhead. To address the lack of annotated multi‐species sequences, we develop a training strategy that synthetically generates multi‐species sequences from mono‐specific ones. Compared with an independent classifier baseline, our multi‐crop model achieves higher accuracy on mono‐specific sequences, both real and synthetic. In multi‐species settings, evaluated using synthetic test sets, we also observe a substantial improvement in accuracy. Using real but weakly annotated multi‐species sequences, we reformulate the task as multi‐label set classification and conduct a visual analysis to highlight the benefits of our approach. By leveraging the information brought by all detections of animals in the image and others of the same sequence, our approach reduces species misclassifications and enables more accurate estimates for downstream ecological analyses focusing on, for instance, species richness, occupancy and species interaction.
- Open Access
Paying Attention to Other Animal Detections Improves Camera Trap Classification Models
In ecological studies, automated species classification models are increasingly used to process large volumes of camera trap images. Most current classification models rely on a two-step pipeline: a detector first locates and crops animals, followed by a classifier that predicts species independently for each crop. While effective, these models still struggle under challenging conditions and ignore temporal context or information from nearby animals available in sequences of camera trap images, whereas human annotators often use it in difficult cases. We propose to leverage self-attention, a core mechanism of Large Language Models and Vision Transformers, to enable the model to learn relationships between crops occurring in similar contexts. Our self-attention module operates directly on the set of crop embeddings, producing new representations enriched with information from other crops, improving species classification. The module fits into the two-step pipeline without requiring structural change and adds only minimal computational overhead. To address the lack of annotated multi-species sequences, we develop a training strategy that synthetically generates multi-species sequences from monospecific ones. Compared to an independent classifier baseline, our multi-crop model achieves higher accuracy on mono-specific sequences, both real and synthetic. In multispecies settings, evaluated using synthetic test sets, we also observe a substantial improvement in accuracy. Using real but weakly annotated multi-species sequences, we reformulate the task as multi-label set classification, and conduct a visual analysis, to highlight the benefits of our approach. By leveraging the information brought by all detections of animals in the image and others of the same sequence, our approach reduces species misclassifications and enables more accurate estimates for downstream ecological analyses focusing on, for instance, species richness, occupancy, and species interaction.
Open AccessPaying Attention to Other Animal Detections Improves Camera Trap Classification Models
In ecological studies, automated species classification models are increasingly used to process large volumes of camera trap images. Most current classification models rely on a two-step pipeline: a detector first locates and crops animals, followed by a classifier that predicts species independently for each crop. While effective, these models still struggle under challenging conditions and ignore temporal context or information from nearby animals available in sequences of camera trap images, whereas human annotators often use it in difficult cases. We propose to leverage self-attention, a core mechanism of Large Language Models and Vision Transformers, to enable the model to learn relationships between crops occurring in similar contexts. Our self-attention module operates directly on the set of crop embeddings, producing new representations enriched with information from other crops, improving species classification. The module fits into the two-step pipeline without requiring structural change and adds only minimal computational overhead. To address the lack of annotated multi-species sequences, we develop a training strategy that synthetically generates multi-species sequences from monospecific ones. Compared to an independent classifier baseline, our multi-crop model achieves higher accuracy on mono-specific sequences, both real and synthetic. In multispecies settings, evaluated using synthetic test sets, we also observe a substantial improvement in accuracy. Using real but weakly annotated multi-species sequences, we reformulate the task as multi-label set classification, and conduct a visual analysis, to highlight the benefits of our approach. By leveraging the information brought by all detections of animals in the image and others of the same sequence, our approach reduces species misclassifications and enables more accurate estimates for downstream ecological analyses focusing on, for instance, species richness, occupancy, and species interaction.
- Open Access
Limits and promises of Earth Observation Foundation Models in Predicting Multi-Trophic Soil Biodiversity
Soil biodiversity is essential for terrestrial ecosystems, influencing nutrient cycling, carbon sequestration, agricultural productivity, and resilience to environmental changes. Yet, it faces significant threats from land-use changes, pollution, agricultural intensification, and climate change. Effective predictive modeling tools are urgently needed to inform conservation and management strategies. Although species distribution models (SDMs) have been successful for aboveground biodiversity, their application to soil biodiversity is limited by scarce large-scale datasets with spatial mismatches between environmental data and soil habitats. Recent advances, including environmental DNA (eDNA) metabarcoding, now allow extensive multi-taxa assessments of soil biodiversity. Simultaneously, remote sensing technologies provide high-resolution spatial data, potentially overcoming traditional coarse-gridded environmental limitations. This study evaluates Earth Observation Foundation (EOF) models, deep learning models pretrained on massive remote sensing datasets to summarize earth observation images into embeddings, to predict multi-trophic soil biodiversity in the French Alps. We compare models using EOF-derived embeddings from orthophotos with coarse-gridded and high-quality in-situ variables. We modeled relative abundance for 51 trophic groups across seven taxa using Random Forest, Light Gradient Boosting Machine, and Artificial Neural Networks, evaluating four data configurations: coarse-gridded environmental data, high-quality in-situ data, EOF embeddings, and a hybrid embedding-tabular approach. High-quality in-situ climate and soil data consistently delivered the highest predictive accuracy, especially for microbial and fungal groups. EOF embeddings provided valuable spatial context but did not surpass in-situ data performance, showing partial redundancy. Integrating remote sensing data can enhance biodiversity modeling in areas lacking detailed in-situ measurements, underscoring their complementary role in ecological assessments.
Open AccessLimits and promises of Earth Observation Foundation Models in Predicting Multi-Trophic Soil Biodiversity
Soil biodiversity is essential for terrestrial ecosystems, influencing nutrient cycling, carbon sequestration, agricultural productivity, and resilience to environmental changes. Yet, it faces significant threats from land-use changes, pollution, agricultural intensification, and climate change. Effective predictive modeling tools are urgently needed to inform conservation and management strategies. Although species distribution models (SDMs) have been successful for aboveground biodiversity, their application to soil biodiversity is limited by scarce large-scale datasets with spatial mismatches between environmental data and soil habitats. Recent advances, including environmental DNA (eDNA) metabarcoding, now allow extensive multi-taxa assessments of soil biodiversity. Simultaneously, remote sensing technologies provide high-resolution spatial data, potentially overcoming traditional coarse-gridded environmental limitations. This study evaluates Earth Observation Foundation (EOF) models, deep learning models pretrained on massive remote sensing datasets to summarize earth observation images into embeddings, to predict multi-trophic soil biodiversity in the French Alps. We compare models using EOF-derived embeddings from orthophotos with coarse-gridded and high-quality in-situ variables. We modeled relative abundance for 51 trophic groups across seven taxa using Random Forest, Light Gradient Boosting Machine, and Artificial Neural Networks, evaluating four data configurations: coarse-gridded environmental data, high-quality in-situ data, EOF embeddings, and a hybrid embedding-tabular approach. High-quality in-situ climate and soil data consistently delivered the highest predictive accuracy, especially for microbial and fungal groups. EOF embeddings provided valuable spatial context but did not surpass in-situ data performance, showing partial redundancy. Integrating remote sensing data can enhance biodiversity modeling in areas lacking detailed in-situ measurements, underscoring their complementary role in ecological assessments.
- Open Access
- 20 May 2025
- Published by Wiley in Methods in Ecology and Evolution
- Vol. 16 (7) , 1460-1472
Zero‐shot animal behaviour classification with vision‐language foundation models
Understanding the behaviour of animals in their natural habitats is critical to ecology and conservation. Camera traps are a powerful tool to collect such data with minimal disturbance. They however produce a very large quantity of images, which can make human‐based annotation cumbersome or even impossible. While automated species identification with artificial intelligence has made impressive progress, the automatic classification of animal behaviours in camera trap images remains a developing field. Here, we explore the potential of foundation models, specifically vision‐language models (VLMs), to perform this task without the need to first train a model, which would require some level of human‐based annotation. Using two datasets, of alpine and African fauna, we investigate the zero‐shot capabilities of different kinds of recent VLMs to predict behaviours and estimate behaviour‐specific diel activity patterns in three ungulate species. By comparing our predictions to behaviours annotated by participatory science, our results show that using these automatic methods, it is possible to achieve F1‐score as high as 86.39% in behaviour classification and produce activity patterns that closely align with those derived from participatory science data (overlap indexes between 84% and 90%). These findings demonstrate the potential of foundation models and vision‐language models in ecological research. Ecologists are encouraged to adopt these new methods and leverage their full capabilities to facilitate ecological studies.
- 20 May 2025
- Published by Wiley in Methods in Ecology and Evolution
- Vol. 16 (7) , 1460-1472
Open AccessZero‐shot animal behaviour classification with vision‐language foundation models
Understanding the behaviour of animals in their natural habitats is critical to ecology and conservation. Camera traps are a powerful tool to collect such data with minimal disturbance. They however produce a very large quantity of images, which can make human‐based annotation cumbersome or even impossible. While automated species identification with artificial intelligence has made impressive progress, the automatic classification of animal behaviours in camera trap images remains a developing field. Here, we explore the potential of foundation models, specifically vision‐language models (VLMs), to perform this task without the need to first train a model, which would require some level of human‐based annotation. Using two datasets, of alpine and African fauna, we investigate the zero‐shot capabilities of different kinds of recent VLMs to predict behaviours and estimate behaviour‐specific diel activity patterns in three ungulate species. By comparing our predictions to behaviours annotated by participatory science, our results show that using these automatic methods, it is possible to achieve F1‐score as high as 86.39% in behaviour classification and produce activity patterns that closely align with those derived from participatory science data (overlap indexes between 84% and 90%). These findings demonstrate the potential of foundation models and vision‐language models in ecological research. Ecologists are encouraged to adopt these new methods and leverage their full capabilities to facilitate ecological studies.
- Open Access
Changes in Muscle Activation and Joint Motion During Walking After Transtibial Amputation with Sensory Feedback from Spinal Cord Stimulation: A Case Study
Background: The goal of this study was to examine the effects of spinal cord stimulation (SCS) on muscle activity during walking after lower-limb amputation. Amputation results in a loss of sensory feedback and alterations in gait biomechanics, including co-contractions of antagonist muscles about the knee and ankle, and reduced pelvic obliquity range-of-motion and pelvic drop. SCS can restore sensation in the missing limb, but its effects on muscle activation and gait biomechanics have not been studied in people with lower-limb amputation. Methods: This case study included a participant with transtibial amputation who was implanted percutaneously with SCS electrodes over the lumbosacral enlargement for 84 days. SCS was used during in-lab experiments to provide somatosensory feedback from the missing limb, relaying a sense of plantar pressure when the prosthesis was in the stance phase of the gait cycle. We used electromyography (EMG) to record muscle activity from the residual and intact limbs, and 3D motion capture to measure pelvic obliquity. EMG signals were recorded during walking with and without SCS at early and late time points across the implant duration. Results: During walking, co-contraction of knee antagonist muscles was reduced following multiple sessions of SCS-mediated sensory restoration. Additionally, the activation of the hip abductor (tensor fasciae latae) muscle increased activity during gait with SCS-mediated sensory restoration, which corresponded to an increase in pelvic obliquity range-of-motion and pelvic drop, towards normal. Conclusions: Restoring sensation in the missing limb using SCS modulated muscle activity during walking leading to improved coordination and pelvic motion in an individual with lower-limb amputation.
Open AccessChanges in Muscle Activation and Joint Motion During Walking After Transtibial Amputation with Sensory Feedback from Spinal Cord Stimulation: A Case Study
Background: The goal of this study was to examine the effects of spinal cord stimulation (SCS) on muscle activity during walking after lower-limb amputation. Amputation results in a loss of sensory feedback and alterations in gait biomechanics, including co-contractions of antagonist muscles about the knee and ankle, and reduced pelvic obliquity range-of-motion and pelvic drop. SCS can restore sensation in the missing limb, but its effects on muscle activation and gait biomechanics have not been studied in people with lower-limb amputation. Methods: This case study included a participant with transtibial amputation who was implanted percutaneously with SCS electrodes over the lumbosacral enlargement for 84 days. SCS was used during in-lab experiments to provide somatosensory feedback from the missing limb, relaying a sense of plantar pressure when the prosthesis was in the stance phase of the gait cycle. We used electromyography (EMG) to record muscle activity from the residual and intact limbs, and 3D motion capture to measure pelvic obliquity. EMG signals were recorded during walking with and without SCS at early and late time points across the implant duration. Results: During walking, co-contraction of knee antagonist muscles was reduced following multiple sessions of SCS-mediated sensory restoration. Additionally, the activation of the hip abductor (tensor fasciae latae) muscle increased activity during gait with SCS-mediated sensory restoration, which corresponded to an increase in pelvic obliquity range-of-motion and pelvic drop, towards normal. Conclusions: Restoring sensation in the missing limb using SCS modulated muscle activity during walking leading to improved coordination and pelvic motion in an individual with lower-limb amputation.
- Open Access
- 16 June 2024
- Published by Wiley in Remote Sensing in Ecology and Conservation
- Vol. 11 (1) , 88-99
Being confident in confidence scores: calibration in deep learning models for camera trap image sequences
In ecological studies, machine learning models are increasingly being used for the automatic processing of camera trap images. Although this automation facilitates and accelerates the identification step, the results of these models may lack interpretability and their immediate applicability to ecological downstream tasks (e.g. occupancy estimation) remains questionable. In particular, little is known about their calibration, a property that allows confidence scores to be interpreted as probabilities that model's predictions are true. Using a large and diverse European camera trap dataset, we investigate whether deep learning models for species classification in camera trap images are well calibrated. Additionally, as camera traps are often configured to take multiple photos of the same event, we also explore the calibration of predictions aggregated across sequences of images. Finally, we study the effect and the practicality of a post‐hoc calibration method, i.e. temperature scaling, for predictions made at image and sequence levels. Based on five established models and three independent test sets, we show that averaging the logits over the sequence, selecting an appropriate architecture, and optionally using temperature scaling can produce well‐calibrated models. Our findings have clear implication for, for instance, the calculation of error rates or the selection of confidence score thresholds in ecological studies making use of artificial intelligence models.
- 16 June 2024
- Published by Wiley in Remote Sensing in Ecology and Conservation
- Vol. 11 (1) , 88-99
Open AccessBeing confident in confidence scores: calibration in deep learning models for camera trap image sequences
In ecological studies, machine learning models are increasingly being used for the automatic processing of camera trap images. Although this automation facilitates and accelerates the identification step, the results of these models may lack interpretability and their immediate applicability to ecological downstream tasks (e.g. occupancy estimation) remains questionable. In particular, little is known about their calibration, a property that allows confidence scores to be interpreted as probabilities that model's predictions are true. Using a large and diverse European camera trap dataset, we investigate whether deep learning models for species classification in camera trap images are well calibrated. Additionally, as camera traps are often configured to take multiple photos of the same event, we also explore the calibration of predictions aggregated across sequences of images. Finally, we study the effect and the practicality of a post‐hoc calibration method, i.e. temperature scaling, for predictions made at image and sequence levels. Based on five established models and three independent test sets, we show that averaging the logits over the sequence, selecting an appropriate architecture, and optionally using temperature scaling can produce well‐calibrated models. Our findings have clear implication for, for instance, the calculation of error rates or the selection of confidence score thresholds in ecological studies making use of artificial intelligence models.
- Open Access
Zero-shot animal behavior classification with vision-language foundation models
1. Understanding the behavior of animals in their natural habitats is critical to ecology and conservation. Camera traps are a powerful tool to collect such data with minimal disturbance. They however produce very a large quantity of images, which can make human-based annotation cumbersome or even impossible. While automated species identification with artificial intelligence has made impressive progress, automatic classification of animal behaviors in camera trap images remains a developing field. 2. Here, we explore the potential of foundation models, specifically Vision Language Models (VLMs), to perform this task without the need to first train a model, which would require some level of human-based annotation. Using an original dataset of alpine fauna with behaviors annotated by participatory science, we investigate the zero-shot capabilities of different kind of recent VLMs to predict behaviors and estimate behavior-specific diel activity patterns in three ungulate species. 3. Our results show that using these methods, it is possible to achieve accuracies over 91% in behavior classification and produce activity patterns that closely align with those derived from participatory science data (overlap indexes between 84% and 90%). 4. These findings demonstrate the potential of foundation models and vision-language models in ecological research. Ecologists are encouraged to adopt these new methods and leverage their full capabilities to facilitate ecological studies.
Open AccessZero-shot animal behavior classification with vision-language foundation models
1. Understanding the behavior of animals in their natural habitats is critical to ecology and conservation. Camera traps are a powerful tool to collect such data with minimal disturbance. They however produce very a large quantity of images, which can make human-based annotation cumbersome or even impossible. While automated species identification with artificial intelligence has made impressive progress, automatic classification of animal behaviors in camera trap images remains a developing field. 2. Here, we explore the potential of foundation models, specifically Vision Language Models (VLMs), to perform this task without the need to first train a model, which would require some level of human-based annotation. Using an original dataset of alpine fauna with behaviors annotated by participatory science, we investigate the zero-shot capabilities of different kind of recent VLMs to predict behaviors and estimate behavior-specific diel activity patterns in three ungulate species. 3. Our results show that using these methods, it is possible to achieve accuracies over 91% in behavior classification and produce activity patterns that closely align with those derived from participatory science data (overlap indexes between 84% and 90%). 4. These findings demonstrate the potential of foundation models and vision-language models in ecological research. Ecologists are encouraged to adopt these new methods and leverage their full capabilities to facilitate ecological studies.
- Open Access
Being confident in confidence scores: calibration in deep learning models for camera trap image sequences
In ecological studies, machine learning models are increasingly being used for the automatic processing of camera trap images. Although this automation facilitates and accelerates the identification step, the results of these models may lack interpretability and their immediate applicability to ecological downstream tasks (e.g occupancy estimation) remain questionable. In particular, little is known about their calibration, a property that guarantees that confidence scores can be reliably interpreted as probabilities that a model’s predictions are true. Using a large and diverse European camera trap dataset, we investigate whether deep learning models for species classification in camera trap images are well calibrated, or in contrast over/under-confident. Additionally, as camera traps are often configured to take multiple photos of the same event, we also explore the calibration of predictions at the sequence level. Finally, we study the effect and the practicality of a post-hoc calibration method, i.e. temperature scaling, for predictions made at image and sequence levels. Based on five established models and three independent test sets, our findings show that, using the right methodology, it is possible to enhance the interpretability of the confidence scores, with clear implication for, for instance, the calculation of error rates or the selection of confidence score thresholds in ecological studies making use of artificial intelligence models.
Open AccessBeing confident in confidence scores: calibration in deep learning models for camera trap image sequences
In ecological studies, machine learning models are increasingly being used for the automatic processing of camera trap images. Although this automation facilitates and accelerates the identification step, the results of these models may lack interpretability and their immediate applicability to ecological downstream tasks (e.g occupancy estimation) remain questionable. In particular, little is known about their calibration, a property that guarantees that confidence scores can be reliably interpreted as probabilities that a model’s predictions are true. Using a large and diverse European camera trap dataset, we investigate whether deep learning models for species classification in camera trap images are well calibrated, or in contrast over/under-confident. Additionally, as camera traps are often configured to take multiple photos of the same event, we also explore the calibration of predictions at the sequence level. Finally, we study the effect and the practicality of a post-hoc calibration method, i.e. temperature scaling, for predictions made at image and sequence levels. Based on five established models and three independent test sets, our findings show that, using the right methodology, it is possible to enhance the interpretability of the confidence scores, with clear implication for, for instance, the calculation of error rates or the selection of confidence score thresholds in ecological studies making use of artificial intelligence models.
- Top Cited Papers
- 20 October 2023
- Published by Springer Nature in European Journal of Wildlife Research
- Vol. 69 (6) , 1-12
The DeepFaune initiative: a collaborative effort towards the automatic identification of European fauna in camera trap images
- 20 October 2023
- Published by Springer Nature in European Journal of Wildlife Research
- Vol. 69 (6) , 1-12
Top Cited PapersThe DeepFaune initiative: a collaborative effort towards the automatic identification of European fauna in camera trap images