Monchy, C. | bioRxiv Science

bioRxiv2026

Monte Carlo simulations to propagate the uncertainty of machine-learning classification into ecological models

Deep learning (DL) is increasingly integrated into quantitative ecology, particularly for automating the classification of sensor data in biodiversity monitoring. In addition to substantially reducing data processing effort, DL models often achieve high classification performance. However, despite ongoing improvements, certain species or classification tasks remain challenging, and predictions are rarely error-free. Manual verification is frequently included into data processing pipelines to mitigate misclassifications, but this approach may mask rather than quantify uncertainty. Here, we propose directly incorporating classification uncertainty into ecological inference, rather than filtering it afterwards. Specifically, we treat model predictions as probabilistic outputs rather than fixed class assignments, by using Monte Carlo simulations to propagate uncertainty from the classification process into downstream ecological models. We illustrate this approach using two case studies. The first estimates stochastic population growth rates for a penguin population using detection time series derived from Radio Frequency IDentification (RFID). The second propagates uncertainty in species identification from camera trap images into occupancy estimates. For both, we compare results obtained propagating classification uncertainty with those from conventional single-class attribution at two confidence score thresholds. Our findings show that propagating uncertainty typically leads to higher, more optimistic ecological estimates compared to the single-class confidence approach. Importantly, this method expands the total uncertainty interval by explicitly introducing the confidence score produced by automatic classification as a representation of uncertainty. Quantifying this uncertainty in parameter estimation allows for more informed and reliable ecological interpretations. Monte Carlo simulations offer a flexible and accessible means to integrate classification uncertainty into diverse ecological modelling workflows.

ecology↗

bioRxiv2024

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

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

ecology↗