Charley Gros

The Tasmanian masked owl ( Tyto novaehollandiae subsp. castanops ) and swift parrot ( Lathamus discolor ) rely on old forest features that are declining across their ranges in Tasmania, Australia. Under the Australian Environment Protection and Biodiversity Conservation Act 1999, the swift parrot is listed as Critically Endangered, and the Tasmanian masked owl as Vulnerable. Their elusive behaviour and high mobility make monitoring difficult, hindering conservation actions. Passive acoustic monitoring can greatly increase spatial and temporal survey coverage, though the identification of the species' vocalisations within large audio datasets remains challenging. We deployed automated recording units at 101 sites in Tasmania's native forests to collect a large and representative acoustic dataset. We trained a ResNet convolutional neural network model to automate call detection for both the Tasmanian masked owl and the swift parrot. Our model demonstrated high performance, with a recall of 97.7% and precision of 97.8% for the Tasmanian masked owl, and a recall of 87.5% and precision of 88.4% for the swift parrot, outperforming both BirdNET and Kaleidoscope Lite. Through two real‐world applications, we illustrated how our method provides detailed quantitative insights into habitat use patterns over extended spatial and temporal scales. Our deep learning model for automated call detection provides a framework to reliably and efficiently analyse acoustic datasets across large spatiotemporal scales, enhancing PAM's application to conservation planning.

Aim: To test whether vocal characteristics of little‐known and ‘lost’ owl species can be inferred from environmentally driven morphological and vocal correlations, using the Australian Masked Owl ( Tyto novaehollandiae ) as a model system. Location: Australasia. Taxon: Australian Masked Owl ( T. novaehollandiae superspecies, Tytonidae). Methods: We fit linear models to examine the relationship between environmental variables (latitude, temperature, elevation) and wing length for 196 Australian Masked Owl specimens. We also examined the relationship between these environmental variables and acoustic characteristics (call duration and mean dominant frequency) from 700 calls recorded via passive acoustic monitoring. These datasets were then integrated to predict vocalisations for the Papuo‐Moluccan taxa Tyto aurantia , Tyto manusi , Tyto sororcula and Tyto novaehollandiae calabyi . Results: Wing length of Australian Masked Owls was greater in females than in males ( R ² = 0.480, p < 0.001) and increased with cooler temperatures ( R ² = 0.240, p < 0.001) and higher latitudes ( R ² = 0.211, p < 0.001). Call duration decreased with decreasing latitude ( R ² = 0.347, p = 0.013) and temperature ( R ² = 0.347, p = 0.036). Mean dominant frequency increased at lower latitudes ( R ² = 0.402, p < 0.001) and warmer temperatures ( R ² = 0.391, p < 0.001) and decreased with elevation ( R ² = 0.429, p = 0.032). Main Conclusions: Environmental factors influencing morphology correlate with vocal traits, enabling predictions for species lacking reference calls. By correlating small datasets of data‐deficient species with larger datasets from better‐known relatives, our approach provides insights into the expected vocalisations of ‘lost’ Papuo‐Moluccan masked owls, challenges taxonomic boundaries, and offers a broadly applicable framework for biogeographic and taxonomic inference.

Antarctic seafloor biodiversity is unique, more than half of its species are endemic. However, the distribution of this biodiversity is poorly understood, and missing baseline data is hindering informed management and conservation. Here we combine information extracted from seafloor images with environmental variables and use statistical models to, for the first time, predictively map the distribution of seafloor biodiversity on the entire Antarctic continental shelf. We identify hotspots of morphospecies richness and epifaunal cover, and show biodiversity hotspots are most prevalent in shallow, cold, food-rich environments with strong seafloor currents and little sedimentation. The colder half of the Antarctic continental shelf hosts 84% of seafloor biodiversity hotspots, highlighting its potential vulnerability to ocean warming. We assess spatial management and show only 14% of hotspots occur within existing marine protected areas. Identifying areas considered important for multiple components of the marine ecosystem, a key consideration in spatial management, reveals predicted seafloor biodiversity hotspots overlap 21-times more than expected at random with pelagic conservation priority areas identified from seabird and mammal tracks. Our analysis reveals new seafloor biodiversity hotspots and provides managers and stakeholders with new information critical to conserve one of the most unique marine ecosystems on Earth.

The Tasmanian masked owl ( Tyto novaehollandiae subsp. castanops ) and swift parrot ( Lathamus discolor ), both endangered species, rely on old forest features that are declining across their ranges in Tasmania, Australia. Their elusive behavior and high mobility make monitoring difficult, hindering conservation actions. Passive acoustic monitoring can greatly increase spatial and temporal survey coverage, though the identification of the species’ vocalisations within large audio datasets remains challenging. We deployed automated recording units at 108 sites in Tasmania’s native forests to collect a large and representative acoustic dataset. We trained a neural network model to automate call detection for both the Tasmanian masked owl and the swift parrot. Our model demonstrated high performance, with a sensitivity of 97.7% and specificity of 96.5% for the Tasmanian masked owl, and a sensitivity of 87.5% and specificity of 83.3% for the swift parrot. Through two real-world applications, we illustrated how our method provides detailed quantitative insights into habitat use patterns over extended spatial and temporal scales. This innovative approach enhances both site-specific and population monitoring, enabling more effective and targeted conservation actions for these endangered species.

Marine imagery is a comparatively cost-effective way to collect data on seafloor organisms, biodiversity and habitat morphology. However, annotating these images to extract detailed biological information is time-consuming and expensive, and reference libraries of consistently annotated seafloor images are rarely publicly available. Here, we present the Antarctic Seafloor Annotated Imagery Database (AS-AID), a result of a multinational collaboration to collate and annotate regional seafloor imagery datasets from 19 Antarctic research cruises between 1985 and 2019. AS-AID comprises of 3,599 georeferenced downward facing seafloor images that have been labelled with a total of 615,051 expert annotations. Annotations are based on the CATAMI (Collaborative and Automated Tools for Analysis of Marine Imagery) classification scheme and have been reviewed by experts. In addition, because the pixel location of each annotation within each image is available, annotations can be viewed easily and customised to suit individual research priorities. This dataset can be used to investigate species distributions, community patterns, it provides a reference to assess change through time, and can be used to train algorithms to automatically detect and annotate marine fauna.

Medical tasks are prone to inter-rater variability due to multiple factors such as image quality, professional experience and training, or guideline clarity. Training deep learning networks with annotations from multiple raters is a common practice that mitigates the model's bias towards a single expert. Reliable models generating calibrated outputs and reflecting the inter-rater disagreement are key to the integration of artificial intelligence in clinical practice. Various methods exist to take into account different expert labels. We focus on comparing three label fusion methods: STAPLE, average of the rater's segmentation, and random sampling of each rater's segmentation during training. Each label fusion method is studied using both the conventional training framework and the recently published SoftSeg framework that limits information loss by treating the segmentation task as a regression. Our results, across 10 data splittings on two public datasets, indicate that SoftSeg models, regardless of the ground truth fusion method, had better calibration and preservation of the inter-rater rater variability compared with their conventional counterparts without impacting the segmentation performance. Conventional models, i.e., trained with a Dice loss, with binary inputs, and sigmoid/softmax final activate, were overconfident and underestimated the uncertainty associated with inter-rater variability. Conversely, fusing labels by averaging with the SoftSeg framework led to underconfident outputs and overestimation of the rater disagreement. In terms of segmentation performance, the best label fusion method was different for the two datasets studied, indicating this parameter might be task-dependent. However, SoftSeg had segmentation performance systematically superior or equal to the conventionally trained models and had the best calibration and preservation of the inter-rater variability.

Medical tasks are prone to inter-rater variability due to multiple factors such as image quality, professional experience and training, or guideline clarity. Training deep learning networks with annotations from multiple raters is a common practice that mitigates the model’s bias towards a single expert. Reliable models generating calibrated outputs and reflecting the inter-rater disagreement are key to the integration of artificial intelligence in clinical practice. Various methods exist to take into account different expert labels. We focus on comparing three label fusion methods: STAPLE, average of the rater’s segmentation, and random sampling of each rater’s segmentation during training. Each label fusion method is studied using both the conventional training framework and the recently published SoftSeg framework that limits information loss by treating the segmentation task as a regression. Our results, across 10 data splittings on two public datasets (spinal cord gray matter challenge, and multiple sclerosis brain lesion segmentation), indicate that SoftSeg models, regardless of the ground truth fusion method, had better calibration and preservation of the inter-rater rater variability compared with their conventional counterparts without impacting the segmentation performance. Conventional models, i.e., trained with a Dice loss, with binary inputs, and sigmoid/softmax final activate, were overconfident and underestimated the uncertainty associated with inter-rater variability. Conversely, fusing labels by averaging with the SoftSeg framework led to underconfident outputs and overestimation of the rater disagreement. In terms of segmentation performance, the best label fusion method was different for the two datasets studied, indicating this parameter might be task-dependent. However, SoftSeg had segmentation performance systematically superior or equal to the conventionally trained models and had the best calibration and preservation of the inter-rater variability. SoftSeg has a low computational cost and performed similarly in terms of uncertainty to ensembles which require multiple models and forward passes. Our code is available at https://ivadomed.org.