Zhongqi Miao

Passive acoustic monitoring enables continuous, non-invasive biodiversity assessment across diverse ecosystems. The scale of these datasets has driven the adoption of machine learning, with supervised approaches showing strong performance. However, supervised methods require time-resolved annotated datasets, which remain scarce, especially in complex tropical soundscapes. We present PteroSet, a curated dataset of strongly annotated Neotropical bird vocalizations recorded in Puerto Asís (Putumayo) and Pivijay (Magdalena), Colombia, between 2023 and 2025. The dataset comprises 563 recordings (73.62 h) and 15,372 time–frequency annotations, including 6,702 events identified to the species level across 168 species. We release the annotations in a COCO-inspired JSON schema that unifies audio files, taxonomic categories, and labels for machine learning workflows. Beyond providing annotated data, PteroSet serves as a realistic benchmark that highlights key characteristics of tropical soundscapes, including acoustic co-occurrence and domain shift across recording sites. We provide a deep learning baseline for binary bird detection, demonstrating PteroSet’s usability and the challenges it presents.

Effective conservation of the endangered Cook Inlet beluga whale ( Delphinapterus leucas ) requires comprehensive spatiotemporal data, yet monitoring efforts remain spatially biased, underrepresenting important southern habitats. Passive acoustic monitoring (PAM) provides the necessary broad‐scale coverage, but its expansion introduces substantial computational challenges, including high soundscape variability, rarity of target species' signals relative to background noise, and multi‐species signal interference that can compromise classifier performance. We present an open‐source deep learning framework designed to improve robustness, adaptability, and domain generalization of PAM analyses for this population. Building on a beluga‐focused binary classifier, we implemented a dual‐stage model that separates signal detection from species classification and expanded the framework to a multi‐species context that includes killer whales ( Orcinus orca ) and humpback whales ( Megaptera novaeangliae ). Contrastive audio‐language models were used to efficiently increase annotation coverage for previously underrepresented species, while active learning enabled iterative refinement of model performance on new data. The framework was applied to PAM datasets from management and ecologically significant regions of lower Cook Inlet. Results demonstrated improved detection of rare species occurrences and increased confidence in daily beluga presence estimates, strengthening the role of PAM in informing recovery efforts and management decisions. This transferable workflow supports continued advancement of large‐scale, long‐term marine mammal monitoring programs.

The DeepFaune New England model classifies wildlife species in trail camera images, identifying 24 taxa from northeastern North America with high (97%) accuracy. The model was adapted from the DeepFaune model for identifying European wildlife, demonstrating the practicality of transfer learning across continents. The majority of training data is openly licensed, and the model itself is open source, enabling easy integration into camera trapping workflows. The open source software is available at ( https://code.usgs.gov/vtcfwru/deepfaune‐new‐england ), and has been further integrated into the PyTorch‐Wildlife framework.

Abandoned, lost, or otherwise discarded fishing gears (ALDFG), commonly referred to as ghost nets, pose a persistent global threat to marine biodiversity. Constructed from durable synthetic polymers, ghost nets remain intact for decades, continuing to entangle and kill marine organisms while damaging habitats and imposing economic burdens on fisheries and coastal communities. Despite their ecological significance, ghost nets are notoriously difficult to detect due to oceanic dispersal, submersion, and burial in sediment. Side-scan sonar has emerged as a powerful detection tool, but its high cost and limited spatial coverage constrain its large-scale application. In this study, we evaluate the feasibility of applying modern computer vision and AI techniques to sonar-derived imagery for automated ghost net detection. In our experiments, we achieved an approximately 90% ghost net detection rate in data collected from the Baltic Sea and the Puget Sound regions. To operationalize this approach, we developed GhostNetZero, a human-in-the-loop web platform that integrates AI predictions with expert review, streamlining validation workflows and enabling iterative model refinement. Our results highlight the promise of AI-assisted sonar analysis in scaling ghost net detection, complementing costly manual surveys and supporting targeted removal operations. By advancing automated detection methods, this study contributes to global efforts to mitigate the impacts of ghost gear and safeguard marine biodiversity.

The integration of artificial intelligence (AI) into biodiversity research and conservation is growing rapidly, demonstrating great potential in reducing the intensive human labour required for data preprocessing, thereby, facilitating larger data collections that offer ecological insights at unprecedented scales. However, most of these AI applications for biodiversity are still in the early stages of development, hindered by challenges inherent in real‐world datasets and the limited accessibility of these technologies to practitioners without extensive programming knowledge. The recent advent of multimodal language models, which can process and generate multiple data modalities, has significantly expanded the realm of possible AI applications in biodiversity research. These models have demonstrated the ability to classify species and recognize more complex concepts, such as animal postures and orientations, without prior exposure during training. Multimodal language models can also provide explanations for their predictions and interact with humans in natural language, thereby making them more transparent, intuitive and accessible to non‐specialists. Despite these advancements, the use of multimodal language models for biodiversity still needs to overcome unique barriers to application, including high computational and financial demands, reliance on prompt engineering for consistent model performance on large datasets and insufficient open‐source sharing of state‐of‐the‐art methods. This paper explores the transformative potential of multimodal language models for biodiversity research and discusses several possible applications in biodiversity research. We also discuss challenges to implement these models in real‐world conservation scenarios and propose directions for future research to overcome these hurdles. Our goal is to encourage robust discussions and research into the integration of multimodal language models to advance AI for biodiversity research and conservation.

Automatically detecting sound events with Artificial Intelligence (AI) has become increas- ingly popular in the field of bioacoustics, ecoacoustics, and soundscape ecology, particularly for wildlife monitoring and conservation. Conventional methods predominantly employ supervised learning techniques that depend on substantial amounts of manually annotated bioacoustic data. However, manual annotation in bioacoustics is tremendously resource- intensive in terms of both human labor and financial resources, and it requires considerable domain expertise. Moreover, the supervised learning framework limits the application scope to predefined categories within a closed setting. The recent advent of Multi-Modal Language Models has markedly enhanced the versatility and possibilities within the realm of AI appli- cations, as this technique addresses many of the challenges that inhibit the deployment of AI in real-world applications. In this paper, we explore the potential of Multi-Modal Language Models in the context of bioacoustics through a case study. We aim to showcase the potential and limitations of Multi-Modal Language Models in bioacoustic applications. In our case study, we applied an Audio-Language Model–—a type of Multi-Modal Language Model that aligns language with audio / sound recording data—–named CLAP (Contrastive Language–Audio Pretraining) to eight bioacoustic benchmarks covering a wide variety of sounds previously unfamiliar to the model. We demonstrate that CLAP, after simple prompt engineering, can effectively recognize group-level categories such as birds, frogs, and whales across the benchmarks without the need for specific model fine-tuning or additional training, achieving a zero-shot transfer recognition performance comparable to supervised learning baselines. Moreover, we show that CLAP has the potential to perform tasks previously unattainable with supervised bioacoustic approaches, such as estimating relative distances and discovering unknown animal species. On the other hand, we also identify limitations of CLAP, such as the model’s inability to recognize fine-grained species-level categories and the reliance on manually engineered text prompts in real-world applications.

The integration of Artificial Intelligence (AI) into biodiversity research and conservation is growing rapidly, demonstrating great potential in reducing the intensive human labor required for data preprocessing, thereby, facilitating larger data collections that offer ecological insights at unprecedented scales. However, most of these AI applications for biodiversity are still in the early stages of development, hindered by challenges inherent in real-world datasets and the limited accessibility of these technologies to practitioners without extensive programming knowledge. The recent advent of multimodal language models, which can process and generate multiple data modalities, has significantly expanded the realm of possible AI applications in biodiversity research. These models have demonstrated the ability to classify species and recognize more complex concepts, such as animal postures and orientations, without prior exposure during training. Multimodal language models can also provide explanations for their predictions and interact with humans in natural language, thereby making them more transparent, intuitive, and accessible to non-specialists. Despite these advancements, the use of multimodal language models for biodiversity still needs to overcome unique barriers to application, including high computational and financial demands, reliance on prompt engineering for consistent model performance on large datasets, and insufficient open-source sharing of state-of-the-art methods. This paper explores the transformative potential of multimodal language models for biodiversity research, compared with traditional machine learning methods, and discusses several potential applications in biodiversity research. We also discuss challenges to implementing these models in real-world conservation scenarios and propose directions for future research to overcome these hurdles. Our goal is to encourage robust discussions and research into the integration of multimodal language models to advance AI for biodiversity research and conservation.

Due to deteriorating environmental conditions and increasing human activity, conservation efforts directed towards wildlife is crucial. Motion-activated camera traps constitute an efficient tool for tracking and monitoring wildlife populations across the globe. Supervised learning techniques have been successfully deployed to analyze such imagery, however training such techniques requires annotations from experts. Reducing the reliance on costly labelled data therefore has immense potential in developing large-scale wildlife tracking solutions with markedly less human labor. In this work we propose WildMatch, a novel zero-shot species classification framework that leverages multimodal foundation models. In particular, we instruction tune vision-language models to generate detailed visual descriptions of camera trap images using similar terminology to experts. Then, we match the generated caption to an external knowledge base of descriptions in order to determine the species in a zero-shot manner. We investigate techniques to build instruction tuning datasets for detailed animal description generation and propose a novel knowledge augmentation technique to enhance caption quality. We demonstrate the performance of WildMatch on a new camera trap dataset collected in the Magdalena Medio region of Colombia.