Using informative priors to account for identifiability issues in occupancy models with identification errors

Peer Community Journal, Journal Year: 2025, Volume and Issue: 5

Published: Jan. 20, 2025

Non-invasive monitoring techniques like camera traps, autonomous recording units and environmental DNA are increasingly used to collect data for understanding species distribution. These methods have prompted the development of statistical models suit specific sampling designs get reliable ecological inferences. Site occupancy estimate occurrence patterns, accounting possibility that target may be present but unobserved. Here, two key processes crucial: detection, when a leaves signs its presence, identification where these accurately recognized. While both prone error in general, wrong identifications often considered as negligible with situ observations. When applied passive bio-monitoring data, characterized by datasets requiring automated processing, this second source can no longer ignored misclassifications at steps lead significant biases estimates. Several model extensions been proposed address potential errors. We propose an extended accounts process addition detection. Similar other recent attempts account false positives, our suffer from identifiability issues, which usually require another perfect resolve them. As alternative such unavailable, we leveraging existing knowledge within Bayesian framework incorporating through informative prior. Through simulations, compare different prior choices encode varying levels information, ranging cases is available, instances accurate metrics on performance identification, scenarios based generally accepted assumptions. demonstrate that, compared using default prior, integrating information about reduces bias parameter Overall, approach mitigates estimation bias, minimizes requirements. In conclusion, provide method applicable various designs, trap, bioacoustics, or eDNA surveys, alongside non-invasive technologies, produce outcomes inform conservation decisions.

Language: Английский

---

An automatic identification method of common species based on ensemble learning

Ecological Informatics, Journal Year: 2025, Volume and Issue: unknown, P. 103046 - 103046

Published: Jan. 1, 2025

Language: Английский

---

Habitat and Predator Influences on the Spatial Ecology of Nine-Banded Armadillos

Diversity, Journal Year: 2025, Volume and Issue: 17(4), P. 290 - 290

Published: April 19, 2025

Mesopredator suppression has implications for community structure, biodiversity, and ecosystem function, but mesopredators with physical defenses may not avoid apex predators. We investigated nine-banded armadillos (Dasypus novemcinctus) in southwestern Oklahoma (USA) to evaluate if a species was influenced by dominant predator, the coyote (Canis latrans). sampled coyotes motion-activated cameras. used single-species conditional two-species occupancy models assess influences of environmental factors on armadillo occurrence site-use intensity (i.e., detection). camera-based detections characterize diel activity each their overlap. Nine-banded greater at sites closer cover, lower slopes, further from water, whereas space use higher elevations; both were positively associated recent burns. coyotes, suppressed presence coyotes. (strictly nocturnal) (predominantly had high overlap summer activity. are engineers often considered threat concern and/or nuisance. Thus, understanding role interspecific interactions important conservation management.

Language: Английский

---

Improving AI performance in wildlife monitoring through species and environment-specific training: A case study on desert Bighorn sheep

Ecological Informatics, Journal Year: 2025, Volume and Issue: unknown, P. 103179 - 103179

Published: May 1, 2025

Language: Английский

---

A classification‐occupancy model based on automatically identified species data

Ecology, Journal Year: 2025, Volume and Issue: 106(5)

Published: May 1, 2025

Occupancy models estimate a species' occupancy probability while accounting for imperfect detection, but often overlook the issue of false-positive detections. This problem false positives has gained attention recently with rapid advancement automated species detection tools using artificial intelligence (AI), which generate continuous confidence scores each detection. Novel have been introduced that integrate these to identify positives, require thorough assessments diagnosis and validation. Here, we propose new model based solely on AI-detected data. We conducted simulations examine inferential predictive accuracies known true parameters analyzed data test practical usefulness through goodness-of-fit tests evaluation external Our proposed mostly outperformed alternative ignore or error probabilities in terms accuracy simulation analyses case study, not discrimination metrics The aids understanding species-habitat relationships developing biodiversity monitoring workflows by both false-negative errors.

Language: Английский

---

Being confident in confidence scores: calibration in deep learning models for camera trap image sequences

Remote Sensing in Ecology and Conservation, Journal Year: 2024, Volume and Issue: unknown

Published: June 16, 2024

Abstract In ecological studies, machine learning models are increasingly being used for the automatic processing of camera trap images. Although this automation facilitates and accelerates identification step, results these may lack interpretability their immediate applicability to downstream tasks (e.g. occupancy estimation) remains questionable. particular, little is known about calibration, a property that allows confidence scores be interpreted as probabilities model's predictions true. Using large diverse European dataset, we investigate whether deep species classification in images well calibrated. Additionally, traps often configured take multiple photos same event, also explore calibration aggregated across sequences Finally, study effect practicality post‐hoc method, i.e. temperature scaling, made at image sequence levels. Based on five established three independent test sets, show averaging logits over sequence, selecting an appropriate architecture, optionally using scaling can produce well‐calibrated models. Our findings have clear implication for, instance, calculation error rates or selection score thresholds studies making use artificial intelligence

Language: Английский

---

Using informative priors to account for identifiability issues in occupancy models with identification errors

bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2024, Volume and Issue: unknown

Published: May 10, 2024

Abstract Non-invasive monitoring techniques like camera traps, autonomous recording units and environmental DNA are increasingly used to collect data for understanding species distribution. These methods have prompted the development of statistical models suit with specific sampling designs get reliable ecological inferences. Site occupancy estimate occurrence patterns, accounting possibility that target may be present but unobserved. Here, two key processes crucial: detection, when a leaves signs its presence, identification where these accurately recognized. While both prone error in general, wrong identifications often considered as negligible situ observations. When applied passive bio-monitoring data, characterized by datasets requiring automated processing, this second source can no longer ignored misclassifications at steps lead significant biases estimates. Several model extensions aim address potential errors. We propose an extended accounts process addition detection. Similar other recent attempts account false positives, our suffer from identifiability issues, which usually require another perfect resolve them. As alternative such unavailable, we leveraging existing knowledge within Bayesian framework incorporating through informative prior. Through simulations, compare different prior choices encode varying levels information, ranging cases is available, instances accurate metrics on performance identification, scenarios based generally accepted assumptions. demonstrate that, compared using default prior, integrating information about reduces bias parameter Overall, approach mitigates estimation bias, minimizes requirements. In conclusion, provide method applicable various designs, trap, bioacoustics, or eDNA surveys, alongside non-invasive technologies, produce outcomes inform conservation decisions.

Language: Английский

---

Being confident in confidence scores: calibration in deep learning models for camera trap image sequences

bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2023, Volume and Issue: unknown

Published: Nov. 13, 2023

Abstract In ecological studies, machine learning models are increasingly being used for the automatic processing of camera trap images. Although this automation facilitates and accelerates identification step, results these may lack interpretability their immediate applicability to downstream tasks (e.g occupancy estimation) remain questionable. particular, little is known about calibration, a property that guarantees confidence scores can be reliably interpreted as probabilities model’s predictions true. Using large diverse European dataset, we investigate whether deep species classification in images well calibrated, or contrast over/under-confident. Additionally, traps often configured take multiple photos same event, also explore calibration at sequence level. Finally, study effect practicality post-hoc method, i.e. temperature scaling, made image levels. Based on five established three independent test sets, our findings show that, using right methodology, it possible enhance scores, with clear implication for, instance, calculation error rates selection score thresholds studies making use artificial intelligence models.

Language: Английский

---

An Ecologist‐Friendly R Workflow for Expediting Species‐Level Classification of Camera Trap Images

Ecology and Evolution, Journal Year: 2024, Volume and Issue: 14(12)

Published: Dec. 1, 2024

ABSTRACT Camera trapping has become increasingly common in ecological studies, but is hindered by analyzing large datasets. Recently, artificial intelligence (deep learning models particular) emerged as a promising solution. However, applying deep for images processing complex and often requires programming skills Python, reducing its accessibility. Some authors addressed this issue with user‐friendly software, further progress was the transposition of to R, statistical language frequently used ecologists, enhancing flexibility customization without advanced computer expertise. We aimed develop workflow based on R scripts streamline entire process, from selecting classifying camera trap images. Our integrates MegaDetector object detector labelling custom training state‐of‐the‐art YOLOv8 model, together potential offline image augmentation manage imbalanced Inference results are stored database compatible Timelapse quality checking model predictions. tested our collected within project targeting medium mammals Central Italy, obtained an overall precision 0.962, recall 0.945, mean average 0.913 set only 1000 pictures per species. Furthermore, achieved 91.8% correct species‐level classifications unclassified images, reaching 97.1% those classified > 90% confidence. YOLO, fast light architecture, enables application even resource‐limited machines, integration makes it useful during early stages data collection. All pretrained available enable adaptation other contexts, plus development.

Language: Английский

---

Improving the integration of artificial intelligence into existing ecological inference workflows

Methods in Ecology and Evolution, Journal Year: 2024, Volume and Issue: unknown

Published: Dec. 26, 2024

Abstract Artificial intelligence (AI) has revolutionised the process of identifying species and individuals in audio recordings camera trap images. However, despite developments sensor technology, machine learning statistical methods, a general AI‐assisted data‐to‐inference pipeline yet to emerge. We argue that this is, part, due lack clarity around several decisions existing workflows, including: choice classifier used (e.g. semi‐ vs. fully automated); how confidence scores are interpreted; availability selection appropriate methods for drawing ecological inferences. Here, we attempt conceptualise workflow associated with automated tools ecology. motivate perspective using our experiences occupancy modelling monitoring data collected through passive acoustic trapping, priority areas future developments. offer an accessible guide support community navigating capitalising on rapid technological methodological advances. describe different error types arise from both sensor‐based classifiers themselves; handled at each stage workflow; finally, implications opportunities deciding step pipeline. recommend ‘black box’ like neural network classification algorithms should be embraced ecology, but widespread uptake requires more formal integration AI into inference workflows. Like broadly, however, successful development new pipelines is multidisciplinary endeavour input everyone invested collecting, processing, analysing data.

Language: Английский

---