Skip to content
Frontiers
Search

Mini Review

Frontiers in Conservation Science ·

Volume 7 - 2026 | https://doi.org/10.3389/fcosc.2026.1837914

Bridging the edge–cloud gap: adaptive AI for robust image and audio wildlife monitoring

  • Delia Velasco-Montero1*corresponding author
  • Jorge Fernández-Berni1
  • Rajesh Sankaran2
  • Javier Bustamante3
  • Peter Beckman4
  1. 1Institute of Microelectronics of Seville (IMSE-CNM), Universidad de Sevilla-CSIC, Seville, Spain
  2. 2Argonne National Laboratory, Lemont, IL, United States
  3. 3Estación Biológica de Doñana (EBD), CSIC, Seville, Spain
  4. 4Northwestern University, Evanston, IL, United States
Article metrics0 viewsView details
PDFCiteMetrics

Abstract

Artificial intelligence (AI) is transforming wildlife monitoring through automated analysis of images and acoustic recordings for tasks such as detection, filtering irrelevant events, and species identification. Many current approaches rely on large models trained on extensive datasets and deployed in the cloud, including systems such as MegaDetector or SpeciesNet for camera-trap imagery and BirdNET for avian acoustics. Although accurate under well-represented conditions, their performance often degrades when applied to new locations, species communities, or recording environments, highlighting persistent challenges in model generalization. Consequently, researchers increasingly rely on species- or site-specific models and adaptive strategies such as calibration, domain adaptation, and continual learning. At the same time, there is growing interest in moving computation closer to the sensor. Edge deployments using lightweight models on embedded platforms such as Raspberry Pi, Nvidia Jetson Nano, or AudioMoth enable real-time inference in remote environments, but introduce constraints related to computation, memory, and energy consumption. These trade-offs motivate hybrid edge-cloud architectures in which edge devices perform local filtering while more complex models and analysis remain in the cloud. This mini-review synthesizes advances in AI-based wildlife monitoring across image and audio modalities, focusing on generalization, data imbalance, and deployment on resource-constrained devices. We review emerging solutions including adaptive calibration, continual learning, and self-supervised representation learning, and discuss how multimodal AI and hybrid edge–cloud systems may enable scalable, robust, and context-aware ecological monitoring.

Keywords

1 Introduction

Artificial intelligence (AI) has transformed the way animal populations are monitored in the field, analyzed, and protected. Over the past decade, advances in both computer vision and ecoacoustics have enabled the automatic detection, classification, and behavioral analysis of wildlife using data collected from sensors deployed in the wild (Stowell, 2022; Kershenbaum et al., 2024; Napier et al., 2024; Xie et al., 2023; Schneider et al., 2018a; Kitzes et al., 2025). However, these sensors – such as camera traps or audiologgers – generate massive volumes of unlabeled data, making AI essential for ecological research and biodiversity assessment. For instance, a typical regional monitoring campaign can accumulate millions of images and tens of thousands of audio hours annually, which translates into terabytes of raw data that congest storage systems and audio loggers human pipelines. In this context, deep learning (DL) models have achieved remarkable success in tasks such as species recognition, filtering irrelevant data (e.g., empty images), detecting animal presence, and identifying vocalizations from audio, thereby providing scalable tools to understand population dynamics (Xie et al., 2023; Lamba et al., 2019; Rajagukguk et al., 2025; Christin et al., 2019; Dufourq et al., 2021; Raza et al., 2025). These methods can reduce field effort by up to 95–99% compared with observer-based monitoring (Hoefer et al., 2025).

State-of-the-art AI models are typically large networks requiring substantial compute to train and deploy. This has led to two diverging paradigms: maximizing accuracy through large cloud-based models versus prioritizing efficiency and autonomy through edge deployment. On the one hand, general-purpose wildlife classifiers are trained on extensive datasets and achieve high accuracy under well-represented conditions (Kahl et al., 2021; Morris et al., 2019; Beery et al., 2019b). These models are typically executed in the cloud or on local high-performance GPU-enabled computing systems. Cloud infrastructures provide access to powerful GPUs and virtually unlimited storage, enabling the deployment of large foundation models but at the expense of communication latency, bandwidth requirements, and energy costs associated with data transmission. On the other hand, there is growing interest in moving computation closer to the sensor, toward the edge (Shi et al., 2016; Liu et al., 2015). Edge AI – running models directly on embedded devices such as Raspberry Pi Upton and Halfacree (2016) or Nvidia Jetson Nano NVIDIA Corporation (2019) – offers critical advantages for field applications, including reduced latency, enabling real-time decisions, independence from network connectivity, and lower power consumption (Jolles, 2021). Despite these benefits, edge-based implementations remain limited due to hardware constraints. Hybrid architectures exploit the complementary strengths of both paradigms, enabling scalable infrastructures for large-scale wildlife monitoring (Belcastro et al., 2025).

However, even generalist models often fail to generalize to new geographic locations, environmental conditions, background variability, or unrepresented species – a limitation that has been repeatedly demonstrated across both image- and audio-based studies (Beery et al., 2018; Lauha et al., 2022; Kitzes et al., 2025; Schneider et al., 2020a; Fairbairn et al., 2025; Lauha et al., 2025). Achieving robust and efficient monitoring thus requires new adaptive and advanced AI strategies capable of updating or recalibrating models in-situ as environmental conditions and data distributions change (Velasco-Montero et al., 2024; Rastikerdar et al., 2025; Jeon et al., 2025). Furthermore, the emergence of large language models (LLMs) and multimodal frameworks opens new opportunities to integrate heterogeneous data sources – ncluding image, audio, and contextual metadata (Höchst et al., 2022; Robinson et al., 2025; Liu et al., 2015).

This mini-review examines recent advances in image- and audio-based AI for wildlife monitoring, emphasizing generalization limitations and deployment constraints. While several surveys review AI for camera traps or ecoacoustics individually, comparatively fewer studies analyze both modalities together or examine how deployment constraints influence model design. In particular, the interaction between model generalization, edge deployment, and adaptive learning strategies remains underexplored. This review addresses this gap by synthesizing advances across image and acoustic monitoring (covered in Sects. 2 and 3, respectively) while focusing on three cross-cutting themes: (i) generalization and domain shift, (ii) edge deployment constraints, and (iii) adaptive and multimodal AI frameworks. By integrating insights from both modalities, Sect. 4 highlights shared challenges, research gaps, and emerging directions toward adaptive, multimodal, and edge-aware wildlife monitoring systems.

2 Image-based AI for animal monitoring

2.1 The era of large-scale cloud classifiers

Camera trapping has emerged as a non-invasive and cost-effective method for long-term wildlife monitoring, enabling the systematic collection of vast amounts of data across diverse ecosystems (Rovero and Zimmermann, 2016). Consequently, the use of DL for wildlife image analysis has become a cornerstone of modern ecological monitoring (Schneider et al., 2018a; Lamba et al., 2019). Convolutional neural networks (CNNs) became the de facto approach, enabling the automatic detection, classification, sex determination, individual identification and counting of animals from large-scale camera trap datasets (Petso et al., 2025), substantially reducing manual annotation effort and accelerating ecological analysis (Norouzzadeh et al., 2018; Tabak et al., 2019). Computer vision tasks such as species classification (Whytock et al., 2021), blank-image filtering (Cunha et al., 2021; Willi et al., 2019; Zhu et al., 2025), object detection (Parham et al., 2018; Singh et al., 2020; Schneider et al., 2018b; Magaldi et al., 2025; Mustafic et al., 2024), and individual re-identification (Schneider et al., 2020b; Brust et al., 2017) have been successfully addressed using supervised DL and weakly-supervised methods (de la Rosa et al., 2023; Xie et al., 2025). More recently, transformer-based architectures and hybrid CNN-transformer models have emerged in ecological vision tasks, motivated by their ability to model long-range dependencies and improve robustness to complex backgrounds and small-object detection (Raza et al., 2025). All of these advances have been supported by the release of large, curated ecological datasets (LILA BC, 2015, 2018, 2019, n.d), enabling the training of general-purpose models with broad taxonomic coverage. As a result, several widely-adopted frameworks – as MegaDetector (Beery et al., 2019b), DeepFaune (Rigoudy et al., 2023), SpeciesNet (Gadot et al., 2024), and a variety of open-source platforms (Vélez et al., 2023; Petroni et al., 2024; Falzon et al., 2020) – have emerged as standard tools for camera-trap data processing. These systems are typically deployed on cloud or high-performance computing infrastructures, and have become essential for large-scale monitoring studies where millions of images must be processed efficiently and consistently.

2.2 The generalization gap and local adaptation

Despite these advances, image-based models remain highly domain-dependent. When applied to unseen environments, lighting conditions, camera technologies, or ecological contexts, performance often degrades substantially (Beery et al., 2018; Greenberg, 2020). Models trained at one location frequently perform poorly elsewhere – even for the same species – due to the learning of background-specific features rather than intrinsic animal characteristics (Whytock et al., 2021; Shepley et al., 2021). In extreme cases, error rates in classification tasks can increase by more than 90% under cross-site evaluation (Beery et al., 2018). Recent studies show that even state-of-the-art YOLO and transformer-based architectures experience mean average precision drops of 3–10% under cross-ecosystem evaluation (Raza et al., 2025). Model robustness is also affected by real-world ecological factors such as camouflage, occlusion, weather patterns, and small-fauna visibility (Raza et al., 2025; Pestell et al., 2025; Greenberg, 2020).

Several mitigation strategies have been proposed, including synthetic data generation, intensive data augmentation, and habitat-specific calibration (Beery et al., 2018; Gomez Villa et al., 2017; Beery et al., 2019a; Mesnage et al., 2025; Dussert et al., 2025a). However, these methods provide only partial mitigation because performance is influenced by both spatial and temporal changes in the visual input [e.g., new locations or seasonal variation) during long-term deployments – referred to as domain shift (Bayram et al., 2022)]. Over time, most current models are static, meaning they degrade silently in real-world environments. Consequently, models generally require retraining when environmental conditions change, underscoring the need for more dynamic and adaptive AI strategies (Jeon et al., 2025).

2.3 Edge intelligence: efficiency and constraints

There is an increasing shift from cloud-centric processing toward edge computing, where models are executed directly on embedded hardware near the sensor. Edge intelligence reduces latency, enables autonomous operation in remote or connectivity-limited areas, and minimizes the energy costs associated with continuous data transmission (Shi et al., 2016). Recent prototypes demonstrate the feasibility of running lightweight CNNs on embedded platforms (Curtin and Matthews, 2019; Cunha et al., 2021; Zualkernan et al., 2022; Pishdast et al., 2025; Dertien et al., 2023; C.K and E., 2019; Liu et al., 2019; Islam and Valles, 2020; Velasco-Montero et al., 2024; Viswanathan et al., 2026). These systems can perform preliminary blank-image filtering – which in some datasets accounts for more than 75% of images (Zhu et al., 2025) – or event detection before transmitting only relevant data to the cloud, thereby achieving substantial energy savings. In addition, metric learning approaches implemented on Jetson Nano have outperformed YOLO-series models in recognizing false triggers (Zhu et al., 2025), and individual re-identification has been successfully deployed on microcontrollers using CNNs and visual transformers Chen et al., 2025).

However, only a few studies have actually evaluated long-term edge deployments under real-world field conditions (Whytock et al., 2023; Gardiner et al., 2025; Viswanathan et al., 2026). These studies have revealed that edge models are sensitive to local environmental drift, highlighting the need for specific-location model adaptation (Rastikerdar et al., 2025; Velasco-Montero et al., 2024).

In addition, this paradigm introduces new technical challenges. Hardware limitations constrain model size and memory, forcing computational efficiency through techniques such as pruning, quantization and knowledge distillation (Han et al., 2015; Gupta et al., 2015; Hinton et al., 2015). While transformer-based architectures offer improved representation capacity, many variants remain computationally demanding for resource-constrained edge platforms, whereas CNN-based detectors currently dominate real-time edge deployments (Raza et al., 2025).

2.4 Key insights from the image domain and emerging challenges

Across image-based monitoring research, three patterns consistently emerge. First, dataset diversity is critical: performance correlates strongly with the ecological and geographical coverage of training data while the inherently long-tail distribution of species and visual conditions further limits model performance (Schneider et al., 2020a; Mustafic et al., 2024; Cunha et al., 2023). Second, robust monitoring systems must maintain accuracy across both spatial shifts (e.g., a new camera angle, variations in lighting, vegetation) and temporal shifts (seasonal changes, weather patterns). Moving toward model adaptation, domain-specific calibration and continual learning (CL) – where models evolve alongside the ecosystems they monitor (Lesort et al., 2020) – is crucial to achieve field accuracies near to those claimed at training stage (Jeon et al., 2025). This is particularly relevant for long-term edge deployment, for which new in-situ fine-tuning frameworks can adapt to local background drift (e.g., through synthetic image generation) (Velasco-Montero et al., 2024; Rastikerdar et al., 2025). Third, edge intelligence is increasingly feasible – with many efforts put on deploying robust embedded intelligence – and represents a potentially transformative direction (Liu et al., 2015).

3 Audio-based AI for animal monitoring

3.1 Global models for ecoacoustics

The deployment of autonomous recording units (ARUs) has established passive acoustic monitoring (PAM) as a widely-adopted, non-invasive method for biodiversity assessment (Sugai et al., 2018a; Winiarska et al., 2024). These devices enable continuous observation of vocalizing species across large spatial and temporal scales (van Merriënboer et al., 2024; Sugai et al., 2018b). Consequently, recent advances in AI have set the standard for vocalization detection and classification, significantly reducing manual effort, with dominant approaches based on CNN architectures operating on mel-scaled spectrograms corresponding to fixed-length audio segments (Stowell, 2022; Kitzes et al., 2025; Lamba et al., 2019). General-purpose acoustic classifiers, such as BirdNET (Kahl et al., 2021) and Perch (Ghani et al., 2023), have been trained on large, globally distributed datasets (Foundation and Center, n.d; Rauch et al., 2025; Chasmai et al., 2024; Pérez-Granados et al., 2025b) and achieve high accuracy for well-represented species. Model performance has benefited from improved pre-processing, spectrogram parameter optimization, data augmentation, and post-processing strategies (Stowell, 2022; Napier et al., 2024; Dufourq et al., 2021; Thompson et al., 2024; Márquez-Rodríguez et al., 2025; Pérez-Granados et al., 2025a). Similar to image-based approaches, these state-of-the-art acoustic models are computationally intensive and are therefore typically deployed in the cloud or high-performance desktop machines.

3.2 Local models and acoustic embeddings

Despite their success, generalist acoustic models suffer from species coverage gaps, as many regional taxa are often excluded from their training sets. In addition, these models frequently struggle to generalize across heterogeneous soundscapes – mirroring the domain shift identified in the image domain (Sect. 2.2). However, in the audio domain, this issue is further magnified by geographic vocalization variability, particularly in migratory species that exhibit different acoustic repertoires between breeding and wintering areas. To address such generalization gaps, multiple studies have adopted local fine-tuning, which consistently outperforms global models, particularly for rare species or sound types underrepresented in large datasets (Lauha et al., 2025, 2022; Huus et al., 2025). Complementarily, species-specific decision thresholds have also been shown to improve precision for local communities, reflecting the need for adaptive calibration to habitat-specific noise profiles and species assemblages (Ware et al., 2023; Scanferla et al., 2025; Funosas et al., 2024; Thompson et al., 2024; Fairbairn et al., 2025; Iglesias-Merchan et al., 2026).

Transfer learning has become a dominant paradigm in ecoacoustics, with pre-trained models used as feature extractors or initialization points (Allen-Ankins et al., 2025). Rather than training from scratch, researchers leverage acoustic embeddings from large-scale models (e.g., BirdNET) for techniques such as few-shot learning (FSL), clustering, or species-presence modeling – strategies which are particularly effective for rare species with limited labeled data (Williams et al., 2025; Ghani et al., 2023; McGinn et al., 2023; Tolkova et al., 2021; Dufourq et al., 2022; Nolasco et al., 2023; Kath et al., 2024; Dumoulin et al., 2025).

Finally, recent breakthroughs in self-supervised learning (SSL) and unsupervised learning allow models to learn robust representations from unlabeled audio – e.g., animal2vec (Schäfer-Zimmermann et al., 2025), Perch 2.0 (van Merriënboer et al., 2026) –, significantly improving performance under diverse acoustic conditions without requiring extensive manual annotation (Dematties et al., 2024; Denton et al., 2022a, b; Michaud et al., 2023; Kather et al., 2025).

3.3 Edge computing in ecoacoustics

The ecoacoustic domain is uniquely positioned for edge deployment due to the success of low-power ARUs such as AudioMoth Hill et al. (2019). While traditional ARUs generally store raw audio for later offline processing, a new generation of intelligent sensors is emerging. Recent research has explored ondevice inference using lightweight CNNs and neural architectures optimized for embedded hardware – e.g., Bird@Edge (Höchst et al., 2022), ACDNet (Mohaimenuzzaman et al., 2023), ESC-NAS (Ranmal et al., 2024) –, demonstrating real-time inference on microcontrollers and low-power platforms (Vandendriessche et al., 2021; Sethi et al., 2018; da Silva et al., 2019; Vidaña-Vila et al., 2020; Velasco-Montero et al., 2026). These edge-oriented systems empower a potential hybrid pipeline (conceptually detailed in Sect. 4 and Figure 1): lightweight models can filter empty audio or detect candidate events locally, while complex species identification can be offloaded to the cloud or triggered only upon detection (Delgado-Rajó and Travieso-Gonzalez, 2025; Delgado-Rajó et al., 2026). This strategy drastically reduces bandwidth requirements and power consumption, enabling long-term monitoring in remote areas where data transmission is costly or impossible.

Conceptual framework of the adaptive edge-cloud continuum for wildlife monitoring. It comprises three layers: (1) the input layer, where multimodal sensors capture raw environmental data at deployments; (2) the edge layer, composed of low-power platforms for light-weight real-time inference and data filtering; and (3) the cloud layer, managing large-scale storage, human-in-the-loop labeling, and large model inference/training. The feedback loop enables periodic model weight updates for *in-situ* adaptation to environmental drift.
Figure 1. Conceptual framework of the adaptive edge-cloud continuum for wildlife monitoring. It comprises three layers: (1) the input layer, where multimodal sensors capture raw environmental data at deployments; (2) the edge layer, composed of low-power platforms for light-weight real-time inference and data filtering; and (3) the cloud layer, managing large-scale storage, human-in-the-loop labeling, and large model inference/training. The feedback loop enables periodic model weight updates for in-situ adaptation to environmental drift.

Despite this progress, robust long-term edge deployments remain rare, and relatively few studies have successfully integrated algorithms into field-ready ARUs instead of general-purpose embedded platforms (Prince et al., 2019; Ciapponi et al., 2026; Velasco-Montero et al., 2026). As in the image domain, practical edge deployments underscore the importance of adaptive mechanisms, such as domain-specific training, or threshold adjustment and post-processing strategies (Singer et al., 2024; Scanferla et al., 2025).

3.4 Key insights from the audio domain and emerging trends

Research in ecoacoustic AI highlights modality-specific challenges as well as emerging methodological trends. First, unlike images, acoustic signals originate from variable distances and frequently overlap, resulting in complex soundscapes where multiple species and noise sources coexist. This acoustic complexity amplifies previously discussed issues of performance degradation across geographic and environmental heterogeneity (van Merriënboer et al., 2024; Napier et al., 2024; Kitzes et al., 2025). Second, the acoustic domain is undergoing a shift toward SSL and embedding-based approaches, which often outperform traditional supervised training in data-scarce scenarios (Robinson et al., 2025; Michaud et al., 2023; Dematties et al., 2024; Schäfer-Zimmermann et al., 2025). For instance, emerging models like BioLingual (Robinson et al., 2024) and NatureLM-audio Robinson et al., 2025) leverage audio-language alignment, and support zero-shot learning (ZSL) for improved generalization. Third, edge intelligence represents a promising direction, as lightweight architectures are approaching practical performance levels for real-time inference (Sethi et al., 2018).

4 Discussion

4.1 Cross-domain challenges in image- and audio-based monitoring

Despite modality-specific differences, image- and audio-based monitoring systems share common structural challenges. Table 1 summarizes these challenges and highlights representative mitigation strategies proposed in the literature.

Table 1. Cross-domain challenges in AI-based wildlife monitoring and representative strategies.

ChallengeImage-based monitoringAudio-based monitoringStrategies
Domain shiftVariations in lighting, background, habitat, weather, and vegetation (Raza et al., 2025; Beery et al., 2018; Greenberg, 2020; Norman et al., 2023).Habitat-dependent soundscapes, geographic variation in vocalizations, sensor sensitivity, and weather conditions (Kahl et al., 2021; Wood and Kahl, 2024; Stowell, 2022).In-situ calibration, threshold adaptation, continual learning (CL), and zero-shot learning (ZSL) (Dussert et al., 2025a; Velasco-Montero et al., 2024; Vyskočil and Picek, 2025; Santamaria et al., 2025; Miao et al., 2025; Jeon et al., 2025; Rastikerdar et al., 2025; Dussert et al., 2025b).
AmbiguityOcclusion, camouflage, and detection of small fauna (Raza et al., 2025; Pestell et al., 2025; Greenberg (2020).Overlapping calls, variable source distance, and low signal-to-noise ratio (Stowell, 2022; Napier et al., 2024).Self-supervised learning (SSL), multimodal fusion (image, audio, context), and ZSL (Buxton et al., 2018; Miao et al., 2025).
Data imbalanceRare species in long-tail datasets and unseen species (Norouzzadeh et al., 2018; Cunha et al., 2023; Schneider et al., 2020a).Infrequent calls masked by overlapping sources (Stowell, 2022).Data augmentation, re-sampling and re-weighting, SSL, few-shot learning (FSL), and foundation models (Cunha et al., 2023; Mesnage et al., 2025; Beery et al., 2019a; Robinson et al., 2025; Schäfer-Zimmermann et al., 2025; McGinn et al., 2023; Cultrera et al., 2023).
Hardware limitationsHigh computational demand, especially for vision transformers (ViTs) on edge devices (Raza et al., 2025; Chen et al., 2025).Computational and memory constraints on microcontrollers for real-time passive acoustic monitoring (PAM) (Vandendriessche et al., 2021; Höchst et al., 2022).Model pruning, quantization, knowledge distillation, and efficient model or pipeline design (Viswanathan et al., 2026; Gardiner et al., 2025; Vandendriessche et al., 2021; Ranmal et al., 2024; Rastikerdar et al., 2025; Velasco-Montero et al., 2026).

Generalization loss, data ambiguity, and domain shift. A fundamental issue across both modalities is performance degradation under varying environmental conditions and sensor conditions. In image-based systems, domain shift manifests as sensitivity to lighting, background, vegetation, and camera placement, and fine-grain intra-species variation (age, sex, individual; while in the acoustic domain, it arises from habitat-dependent soundscapes, varying microphone sensitivity, and variation in vocalizations. Beyond environmental noise, ambiguity caused by occlusions in images or overlapping vocalizations in audio further complicates inference. To achieve model generalization, current approaches increasingly rely on habitat-specific calibration, species-specific decision thresholds, and SSL and ZSL techniques.

Long-tail distributions and data imbalance. Both domains struggle with the long-tailed distribution of species, with rare species or infrequent behaviors underrepresented in the datasets. This bias favors common species, leading to high accuracy on benchmarks but poor performance in ecological validity (Cunha et al., 2023; Willi et al., 2019; Beery et al., 2019a). While data augmentation and synthetic data generation can mitigate this, the scarcity of high-quality data labeled by experts for rare events remains a significant barrier for training generalist models.

The labeling bottleneck and citizen science. The scarcity of labeled data is a universal constraint. Manual annotation is time-consuming, costly, and prone to human error. While citizen science platforms like xeno-canto have revolutionized data availability (Lehikoinen et al., 2023; Gutiérrez-Zapata et al., 2024; Sharpe et al., 2025), the lack of standardization across diverse data streams remains a major governance challenge for the field (Napier et al., 2024).

Hardware and operational constraints. Moving AI to the edge introduces a significant trade-off between accuracy and energy usage. Edge computing presents constraints regarding computational capacity, memory and energy budget that can limit the execution of large models, such as ViTs.

Explainability and failure detection. The ‘black-box’ nature of DL remains a concern for ecological research. Models often fail silently, providing confident but incorrect predictions under noise. Incorporating Explainable AI (XAI) techniques (Kamath and Liu (2021), such as saliency maps, Grad-CAM visualizations, or attention-based explanations, is a requirement for transparent, interpretable, and responsible use of AI in ecology (Miao et al., 2019).

4.2 Future directions and research opportunities

Current strategies to address the challenges outlined above are summarized in the Strategies column of Table 1. To synthesize these trends, we propose a conceptual framework for adaptive wildlife monitoring based on an edge-cloud continuum, illustrated in Figure 1. This framework integrates the following emerging pillars:

Adaptive calibration and continual learning. Moving from static models to systems that update incrementally is crucial to overcome the “generalization wall” across spatial and temporal scales. As introduced in Sect. 2.4, online CL (where models evolve alongside the ecosystems they monitor) has been proposed as a potential solution (Mulimani and Mesaros, 2024; Velasco-Montero et al., 2024; Rastikerdar et al., 2025). Specifically, CL techniques include experience replay, regularization-based and parameter isolation methods (De Lange et al., 2022). Techniques such as in-situ fine-tuning and online threshold adjustment allow models to recalibrate to site-specific noise and seasonal shifts. Recent benchmarks in CL (Jeon et al., 2025) demonstrate how models accumulate knowledge over time without catastrophic forgetting on previously learned taxa. Furthermore, the integration of robust out-of-distribution (OOD) detection methods (Cultrera et al., 2023; Muthivhi et al., 2025) can potentially ensure that systems can correctly manage novel or uncertain events, rather than producing silent failures.

Self-supervised and foundation models. By exploiting vast amounts of unlabeled ecological data, SSL reduces dependence on expensive manual annotations by training models on pretext tasks derived directly from unlabeled data, which enables the extraction of robust feature representations from massive, uncurated datasets ) (Zhang et al., 2025). At the same time, the emergence of multimodal language models (MLMs) represents a major methodological development (Miao et al., 2025). The next frontier lies in vision-language (VLMs) and audio-language models (ALMs), which have already achieved over 86% accuracy in zero-shot animal behavior classification (Dussert et al., 2025b).

By aligning spectrogram or image embeddings with textual descriptions in a shared latent space,

foundation models support ZSL to extract fine-grain ecological data including intra-species patterns (e.g., age, sex, feather coloration, behavior), environmental context characterization, and hierarchical clustering at taxonomic level (family, genus, species) (Markoff et al., 2026; Gu et al., 2025). These foundation models may increase robustness to domain shift and data scarcity. Although still in early stages, these approaches suggest a potential shift toward foundation models capable of integrating heterogeneous ecological data streams.

Multimodal fusion. The integration of image and audio signals within monitoring systems provides complementary information, increasing detection reliability. For instance, audio sensors may detect species hidden by dense vegetation where cameras fail (Buxton et al., 2018). This synergy highlights the potential of multimodal monitoring systems – including images, audio, and contextual variables (spatiotemporal and environmental metadata) – to resolve ambiguities that remain challenging for unimodal approaches and to enable more robust ecological inference (Miao et al., 2025).

Edge-cloud continuum. The balance between cloud-based accuracy and edge-based efficiency represents a central design trade-off for scalable monitoring. Cloud deployments support large models and facilitate centralized updates, whereas edge deployments enable real-time, autonomous operation. Hybrid edgecloud architectures provide a compromise allowing scalable, accurate, and efficient operation (Belcastro et al., 2025). This paradigm is particularly important for long-term deployments, where environmental conditions, species composition, and sensor configurations evolve over time.

In summary, as depicted in Figure 1, this adaptive edge-cloud continuum integrating emerging approaches represents a promising pathway for sustainable, large-scale monitoring. In this three-level architecture, (1) sensors collect data at deployments; (2) edge devices perform real-time inference and local filtering; and low-confidence or OOD instances are transmitted to the cloud (3), where multimodal fusion, human-in-theloop approaches, or inference based on foundation models can be performed. The resulting knowledge is then pushed back to the edge as model updates. In this sense, generalization should no longer be viewed solely as a model property, but as a system-level capability emerging from the interaction between adaptive learning, multimodal sensing, and edge-cloud collaboration.

The convergence of edge computing, adaptive learning, and multimodal AI represents a significant opportunity for next-generation ecological monitoring systems. Moving beyond static cloud-based models toward adaptive edge-cloud architectures will be essential for enabling scalable, robust, and autonomous biodiversity monitoring in complex and dynamic natural environments.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Rosa M. Badia, Barcelona Supercomputing Center, Spain

Reviewed by: Shilpa Sonawani, Dr. Vishwanath Karad MIT World Peace University, India

Statements

Conflict of interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Funding

The author(s) declared that financial support was received for this work and/or its publication. This publication has been funded by grant USECHIP (TSI-069100-2023-001), project funded by the Secretary of State for Telecommunications and Digital Infrastructure, Ministry for the Digital Transformation and Civil Service and by the European Union Next GenerationEU/PRTR. The work was also supported by the DISCOVER-US project under the European Union’s Horizon Europe programme (Grant Agreement No. 101135064); by the U.S. National Science Foundation under Award No. 2436842 (“SAGE GRANDE: An Open Artificial Intelligence Testbed for Edge Computing and Intelligent Sensing”); and by the project “Plan Complementario de I+D+i en el área de Biodiversidad (PCBIO)”, funded by the European Union (NextGenerationEU, Recovery, Transformation and Resilience Plan) and the Regional Government of Andalucia (Ref: BIOD22_00033_3_PPCB). Any opinions, findings, and conclusions or recommendations expressed are those of the authors and do not necessarily reflect the views of the funding agencies.

Author contributions

DV-M: Conceptualization, Investigation, Writing – original draft. JF-B: Writing – review & editing. RS: Writing – review & editing. JB: Writing – review & editing. PB: Writing – review & editing.

References

  1. Allen-Ankins S., Hoefer S., Bartholomew J., Brodie S., Schwarzkopf L. (2025). The use of BirdNET embeddings as a fast solution to find novel sound classes in audio recordings. Front. Ecol. Evol. 12. doi: doi:10.3389/fevo.2024.1409407
  2. Bayram F., Ahmed B. S., Kassler A. (2022). From concept drift to model degradation: An overview on performance-aware drift detectors. Knowledge-Based Syst. 245, 108632. doi: doi:10.1016/j.knosys.2022.108632
  3. Beery S., Liu Y., Morris D., Piavis J., Kapoor A., Meister M.. (2019 a). “ Synthetic examples improve generalization for rare classes,” in 2020 IEEE Winter Conference on Applications of Computer Vision (WACV) (USA: IEEE), 852 – 862.
  4. Beery S., Morris D., Yang S. (2019 b). Efficient pipeline for camera trap image review. arXiv. doi: doi:10.48550/arXiv.1907.06772
  5. Beery S., Van Horn G., Perona P. (2018). “ Recognition in terra incognita ”, in Computer Vision – ECCV 2018 (Switzerland: Springer).
  6. Belcastro L., Marozzo F., Orsino A., Talia D., Trunfio P. (2025). Navigating the edge-cloud continuum: A state-of-practice survey. 14, 40622 – 40647. doi: doi:10.1109/access.2026.3673012
  7. Brust C.-A., Burghardt T., Groenenberg M., Kading C., Kühl H. S., Manguette M. L.. (2017). “ Towards automated visual monitoring of individual gorillas in the wild ”, in Proceedings of the IEEE International Conference on Computer Vision Workshops (ICCVW) (Piscataway, NJ: IEEE), 2820 – 2830. doi: doi:10.1109/ICCVW.2017.333
  8. Buxton R. T., Lendrum P. E., Crooks K. R., Wittemyer G. (2018). Pairing camera traps and acoustic recorders to monitor the ecological impact of human disturbance. Global Ecol. Conserv. 16, e00493. doi: doi:10.1016/j.gecco.2018.e00493
  9. Chasmai M., Shepard A., Maji S., Van Horn G. (2024). The iNaturalist sounds dataset. Adv. Neural Inf. Process. Syst. 37, 132524 – 132544. doi: doi:10.52202/079017-4213
  10. Chen Y., Zhao D., Koh Y. S., Xu T. (2025). Animal re-identification on microcontrollers.
  11. Christin S., Hervet E., Lecomte N. (2019). Applications for deep learning in ecology. Methods Ecol. Evol. 10, 1632 – 1644. doi: doi:10.1111/2041-210X.13256
  12. Ciapponi S., Mannini L., Scanferla J., Anderle M., Farella E. (2026). " Enabling multi-species bird classification on low-power bioacoustic loggers." in ICASSP 2026 - 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 15212 – 15216. doi: 10.1109/ICASSP55912.2026.35111463679
  13. C.K S., E. S. (2019). “ Automated wildlife monitoring using deep learning ”, in Proceedings of the International Conference on Systems, Energy & Environment (ICSEE) 2019.
  14. Cultrera L., Seidenari L., Del Bimbo A. (2023). “ Leveraging visual attention for out-of-distribution detection ”, in 2023 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW) (Piscataway, NJ: IEEE), 4449 – 4458. doi: doi:10.1109/ICCVW60793.2023.00479
  15. Cunha F., dos Santos E. M., Barreto R. S., Colonna J. G. (2021). " Filtering Empty Camera Trap Images in Embedded Systems." in 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), (Nashville, TN), 2438 – 2446. doi: doi:10.1109/CVPRW53098.2021.00276
  16. Cunha F., dos Santos E. M., Colonna J. G. (2023). Bag of tricks for long-tail visual recognition of animal species in camera-trap images. Ecol. Inf. 76, 102060. doi: doi:10.1016/j.ecoinf.2023.102060
  17. Curtin B. H., Matthews S. J. (2019). “ Deep learning for inexpensive image classification of wildlife on the Raspberry Pi ”, in: 2019 IEEE 10th Annual Ubiquitous Computing, Electronics & Mobile Communication Conference (UEMCON) (Piscataway, NJ: IEEE), 0082 – 0087.
  18. da Silva B., W. Happi A., Braeken A., Touhafi A. (2019). Evaluation of classical machine learning techniques towards urban sound recognition on embedded systems. Appl. Sci. 9, 3885. doi: doi:10.3390/app9183885
  19. De Lange M., Aljundi R., Masana M., Parisot S., Jia X., Leonardis A.. (2022). A continual learning survey: Defying forgetting in classification tasks. IEEE Trans. Pattern Anal. Mach. Intell. 44, 3366 – 3385. doi: doi:10.1109/TPAMI.2021.3057446
  20. de la Rosa D., Rivera A. J., del Jesus M. J., Charte F. (2023). PARDINUS: Weakly supervised discarding of photo-trapping empty images based on autoencoders.
  21. Delgado-Rajó F. A., Travieso-Gonzalez C. M. (2025). Flexible hybrid edge computing IoT architecture for low-cost bird songs detection system. Ecol. Inf. 90, 103231. doi: doi:10.1016/j.ecoinf.2025.103231
  22. Delgado-Rajó F. A., Travieso-González C. M., Hernández-López R. (2026). Scalable and lowpower edge architecture with Wi-Fi HaLow and on-device spectrograms generation for flexible urban bioacoustics monitoring. Internet Things 36, 101864. doi: doi:10.1016/j.iot.2025.101864
  23. Dematties D., Rajani S., Sankaran R., Shahkarami S., Raut B., Collis S.. (2024). Acoustic fingerprints in nature: A self-supervised learning approach for ecosystem activity monitoring. Ecol. Inf. 83, 102823. doi: doi:10.1016/j.ecoinf.2024.102823
  24. Denton T., Wisdom S., Hershey J. R. (2022 a). “ Improving bird classification with unsupervised sound separation ”, in ICASSP 2022–2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (Piscataway, NJ: IEEE), 636 – 640. doi: doi:10.1109/ICASSP43922.2022.9747202
  25. Denton T., Wisdom S., Hershey J. R. (2022 b). “ Improving bird classification with unsupervised sound separation ”, in ICASSP 2022–2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (Piscataway, NJ: IEEE), 636 – 640. doi: doi:10.1109/ICASSP43922.2022.9747202
  26. Dertien J. S., Negi H., Dinerstein E., Krishnamurthy R., Negi H. S., Gopal R.. (2023). Mitigating human–wildlife conflict and monitoring endangered tigers using a real-time camera-based alert system. BioScience 73, 748 – 757. doi: doi:10.1093/biosci/biad076
  27. Dufourq E., Batist C., Foquet R., Durbach I. (2022). Passive acoustic monitoring of animal populations with transfer learning. Ecol. Inf. 70, 101688. doi: doi:10.1016/j.ecoinf.2022.101688
  28. Dufourq E., Durbach I., Hansford J. P., Hoepfner A., Ma H., Bryant J. V.. (2021). Automated detection of Hainan gibbon calls for passive acoustic monitoring. Remote Sens. Ecol. Conserv. 7, 475 – 487. doi: doi:10.1002/rse2.201
  29. Dumoulin V., Stretcu O., Hamer J., Harrell L., Laber R., Larochelle H.. (2025). The search for Squawk: Agile modeling in bioacoustics. arXiv. doi: doi:10.48550/arXiv.2505.03071
  30. Dussert G., Chamaillé-Jammes S., Dray S., Miele V. (2025 a). Being confident in confidence scores: Calibration in deep learning models for camera trap image sequences. Remote Sens. Ecol. Conserv. 11, 88 – 99. doi: doi:10.1002/rse2.412
  31. Dussert G., Miele V., Van Reeth C., Delestrade A., Dray S., Chamaillé-Jammes S. (2025 b). Zero-shot animal behaviour classification with vision-language foundation models. Methods Ecol. Evol. 16, 1460 – 1472. doi: doi:10.1111/2041-210X.70059
  32. Fairbairn A. J., Burmeister J.-S., Weisser W. W., Meyer S. T. (2025). BirdNET can be as good as experts for acoustic bird monitoring in a European city. PloS One 20, 1 – 12. doi: doi:10.1371/journal.pone.0330836
  33. Falzon G., Lawson C., Cheung K.-W., Vernes K., Ballard G. A., Fleming P. J. S.. (2020). ClassifyMe: A field-scouting software for the identification of wildlife in camera trap images. Animals 10, 58. doi: doi:10.3390/ani10010058
  34. Foundation and Center, N. B (n. d). Xeno-Canto. Sharing Wildlife Sounds From Around the World. Available online at: https://xeno-canto.org/ (Accessed August 16, 2026).
  35. Funosas D., Barbaro L., Schillé L., Elger A., Castagneyrol B., Cauchoix M. (2024). Assessing the potential of BirdNET to infer European bird communities from large-scale ecoacoustic data. Ecol. Indic. 164, 112146. doi: doi:10.1016/j.ecolind.2024.112146
  36. Gadot T., Istrate S., Kim H., Morris D., Beery S., Birch T.. (2024). To crop or not to crop: Comparing whole-image and cropped classification on a large dataset of camera trap images. IET Comput. Vision 18, 1193 – 1208. doi: doi:10.1049/cvi2.12318
  37. Gardiner R. J., Rowlands S., Simmons B. I. (2025). Towards scalable insect monitoring: Ultralightweight CNNs as on-device triggers for insect camera traps. Methods Ecol. Evol 17, 357 – 370. doi: doi:10.1111/2041-210X.70098
  38. Ghani B., Denton T., Kahl S., Klinck H. (2023). Global birdsong embeddings enable superior transfer learning for bioacoustic classification. Sci. Rep. 13, 22876. doi: doi:10.1038/s41598-023-49989-z
  39. Gomez Villa A., Salazar A., Vargas F. (2017). Towards automatic wild animal monitoring: Identification of animal species in camera-trap images using very deep convolutional neural networks. Ecol. Inf. 41, 24 – 32. doi: doi:10.1016/j.ecoinf.2017.07.004
  40. Greenberg S. (2020). Automated Image Recognition for Wildlife Camera Traps: Making it Work for You (Calgary, Alberta: University of Calgary). Technical report.
  41. Gu J., Stevens S., Campolongo E. G., Thompson M. J., Zhang N., Wu J.. (2025). Bioclip 2: emergent properties from scaling hierarchical contrastive learning. Adv. Neural. Inf. Process. Syst. 38, 102778 – 102811. doi: doi:10.52202/085713-3436
  42. Gupta S., Agrawal A., Gopalakrishnan K., Narayanan P. (2015). “ Deep learning with limited numerical precision ”, in Proceedings of the 32nd International Conference on Machine Learning (PMLR, Vol. 37). (Lille: PMLR), 1737 – 1746.
  43. Gutiérrez-Zapata S., Santoro S., Gegundez-Arias M. E., Selva N., Calzada J. (2024). Dog invasions in protected areas: A case study using camera trapping, citizen science and artificial intelligence. Global Ecol. Conserv. 54, e03109. doi: doi:10.1016/j.gecco.2024.e03109
  44. Han S., Pool J., Tran J., Dally W. J. (2015). “ Learning both weights and connections for efficient neural networks ”, in: Proceedings of the 29th International Conference on Neural Information Processing Systems (NIPS 2015) (Cambridge, MA, USA: MIT Press), 1135 – 1143.
  45. Hill A. P., Prince P., Snaddon J. L., Doncaster C. P., Rogers A. (2019). AudioMoth: A low-cost acoustic device for monitoring biodiversity and the environment. HardwareX 6, e00073. doi: doi:10.1016/j.ohx.2019.e00073
  46. Hinton G., Vinyals O., Dean J. (2015). Distilling the knowledge in a neural network. arXiv. doi: doi:10.48550/arXiv.1503.02531
  47. Höchst J., Bellafkir H., Lampe P., Vogelbacher M., Mühling M., Schneider D.. (2022). “ Bird@Edge: Bird species recognition at the edge ”, in Proceedings of 10th International Conference on Networked Systems (NETYS) (Berlin, Heidelberg: Springer-Verlag), 69 – 86. doi: doi:10.1007/978-3-031-17436-0_6
  48. Hoefer S., Allen-Ankins S., McKnight D. T., Nordberg E. J., Schwarzkopf L. (2025). Sensors versus surveyors: Comparing passive acoustic monitoring, camera trapping and observer-based monitoring for terrestrial mammals. Methods Ecol. Evol. 16, 2603 – 2621. doi: doi:10.1111/2041-210X.70169
  49. Huus J., Kelly K. G., Bayne E. M., Knight E. C. (2025). HawkEars: A regional, high-performance avian acoustic classifier. Ecol. Inf. 87, 103122. doi: doi:10.1016/j.ecoinf.2025.103122
  50. Iglesias-Merchan C., Sanchez-Torres R., Alonso R. (2026). Analysis of birdnet configuration and performance applied to the acoustic monitoring of a restored quarry. Environments 13, 31. doi: doi:10.3390/environments13010031
  51. Islam S. B., Valles D. (2020). “ Identification of wild species in Texas from camera-trap images using deep neural network for conservation monitoring ”, in 2020 10th Annual Computing and Communication Workshop and Conference (CCWC) (Piscataway, NJ: IEEE), 0537 – 0542. doi: doi:10.1109/CCWC47524.2020.9031190
  52. Jeon S., Mai Z., Tian H., Bakshi V., Wang L., Hou J.. (2025). “ Continually adapt or not (CAN)? A continual learning benchmark of camera trap species classification over time ”, in: NeurIPS 2025 Workshop on Imageomics: Discovering Biological Knowledge from Images Using AI. (San Diego, CA: Neural Information Processing Systems).
  53. Jolles J. W. (2021). Broad-scale applications of the Raspberry Pi: A review and guide for biologists. Methods Ecol. Evol. 12, 1562 – 1579. doi: doi:10.1111/2041-210X.13652
  54. Kahl S., Wood C. M., Eibl M., Klinck H. (2021). BirdNET: A deep learning solution for avian diversity monitoring. Ecol. Inf. 61, 101236. doi: doi:10.1016/j.ecoinf.2021.101236
  55. Kamath U., Liu J. (2021). Explainable Artificial Intelligence: An Introduction to Interpretable Machine Learning (Switzerland: Springer). doi: doi:10.1007/978-3-030-83356-5
  56. Kath H., Serafini P. P., Campos I. B., Gouveaˆ T. S., Sonntag D. (2024). Leveraging transfer learning and active learning for data annotation in passive acoustic monitoring of wildlife. Ecol. Inf. 82, 102710. doi: doi:10.1016/j.ecoinf.2024.102710
  57. Kather VS, Ghani B, Stowell D (2025). Clustering and novel class recognition: evaluating bioacoustic deep learning feature extractors. in Proceedings of the 11th Convention of the European Acoustics Association, 2153 – 2160. doi: doi:10.61782/fa.2025.0231
  58. Kershenbaum A., C¸aglar˘ Akc¸ay, Babu-Saheer L., Barnhill A., Best P., Cauzinille J.. (2024). Automatic detection for bioacoustic research: a practical guide from and for biologists and computer scientists. Biol. Rev 100, 620 – 646. doi: doi:10.1111/brv.13155
  59. Kitzes J., Chronister L., Czarnecki C., Fiss C., Freeland-Haynes L., Goodman B. D.. (2025). Integrating AI models into ecological research workflows: The case of terrestrial bioacoustics. Methods Ecol. Evol 17, 238 – 256. doi: doi:10.1111/2041-210X.70133
  60. Lamba A., Cassey P., Segaran R. R., Koh L. P. (2019). Deep learning for environmental conservation. Curr. Biol. 29, R977 – R982. doi: doi:10.1016/j.cub.2019.08.016
  61. Lauha P.. (2025). “ Bird Sounds Global – model builder: An end-to-end workflow for building locally fine-tuned bird classifiers,” in Authorea Preprints. doi: doi:10.22541/au.176599468.88578746/v1
  62. Lauha P., Somervuo P., Lehikoinen P., Geres L., Richter T., Seibold S.. (2022). Domain-specific neural networks improve automated bird sound recognition already with small amount of local data. Methods Ecol. Evol. 13, 2799 – 2810. doi: doi:10.1111/2041-210X.14003
  63. Lehikoinen P., Rannisto M., Camargo U., Aintila A., Lauha P., Piirainen E.. (2023). A successful crowdsourcing approach for bird sound classification. Citizen Science: Theory Pract. 8, 16. doi: doi:10.5334/cstp.556
  64. Lesort T., LoMonaco V., Stoian A., Maltoni D., Filliat D., Díaz-Rodríguez N. (2020). Continual learning for robotics: Definition, framework, learning strategies, opportunities and challenges. Inf. Fusion 58, 52 – 68. doi: doi:10.1016/j.inffus.2019.12.004
  65. LILA BC (2015). Snapshot Serengeti. Available online at: http://lila.science/datasets/snapshot-serengeti (Accessed August 16, 2026).
  66. LILA BC (2018). Caltech Camera Traps. Available online at: http://lila.science/datasets/caltech-camera-traps (Accessed August 16, 2026).
  67. LILA BC (2019). North American Camera Trap Images. Available online at: http://lila.science/datasets/nacti (Accessed August 16, 2026).
  68. LILA BC (n. d). Camera Traps Datasets. Available online at: http://lila.science/category/camera-traps/ (Accessed August 16, 2026).
  69. Liu X., Jia Z., Hou X., Fu M., Ma L., Sun Q. (2019). “ Real-time marine animal images classification by embedded system based on Mobilenet and transfer learning ”, in 2019 IEEE OCEANS – Marseille (Piscataway, NJ: IEEE), 1 – 5. doi: doi:10.1109/OCEANSE.2019.8867190
  70. Liu X., Yang T., Yan B. (2015). “ Internet of Things for wildlife monitoring ”, in 2015 IEEE/CIC International Conference on Communications in China - Workshops (CIC/ICCC) (Piscataway, NJ: IEEE), 62 – 66. doi: doi:10.1109/ICCChinaW.2015.7961581
  71. Magaldi H., Cornette R., Tibesigwa J. J., Katumba R., Rugonge H., Amarasekaran B.. (2025). DeepForestVision: Automated wildlife identification for camera traps of African tropical forests. Ecol. Solutions Evidence 6, e70167. doi: doi:10.1002/2688-8319.70167
  72. Markoff H., Bengtson S. H., Ørsted M. (2026). Vision transformers for zero-shot clustering of animal images: A comparative benchmarking study. arXiv. doi: doi:10.48550/arXiv.2602.03894
  73. Márquez-Rodríguez A., Ángel Mohedano-Munoz M., Marín-Jiménez M. J., Santamaría-García E., Bastianelli G., Jordano P.. (2025). A bird song detector for improving bird identification through deep learning: A case study from Doñana. Ecol. Inf. 90, 103254. doi: doi:10.1016/j.ecoinf.2025.103254
  74. McGinn K., Kahl S., Peery M. Z., Klinck H., Wood C. M. (2023). Feature embeddings from the BirdNET algorithm provide insights into avian ecology. Ecol. Inf. 74, 101995. doi: doi:10.1016/j.ecoinf.2023.101995
  75. Mesnage C. S., Corbett A., Curry J., Rowlands S., Dutta A., Everson R.. (2025). Copy-paste augmentation improves automatic species identification in camera trap images. Ecol. Evol. 15, e72357. doi: doi:10.1002/ece3.72357
  76. Miao Z., Gaynor K. M., Wang J., Liu Z., Muellerklein O., Norouzzadeh M. S.. (2019). Insights and approaches using deep learning to classify wildlife. Sci. Rep. 17, 238 – 256. doi: doi:10.1038/s41598-019-44565-w
  77. Miao Z., Zhang Y., Fabian Z., Hernandez Celis A., Beery S., Li C.. (2025). New frontiers in artificial intelligence for biodiversity research and conservation with multimodal language models. Methods Ecol. Evol. doi: doi:10.1111/2041-210X.70120
  78. Michaud F., Sueur J., Le Cesne M., Haupert S. (2023). Unsupervised classification to improve the quality of a bird song recording dataset. Ecol. Inf. 74, 101952. doi: doi:10.1016/j.ecoinf.2022.101952
  79. Mohaimenuzzaman M., Bergmeir C., West I., Meyer B. (2023). Environmental sound classification on the edge: A pipeline for deep acoustic networks on extremely resource-constrained devices. Pattern Recognit. 133, 109025. doi: doi:10.1016/j.patcog.2022.109025
  80. Morris D.. (2019). Megadetector. Available online at: https://github.com/agentmorris/MegaDetector (Accessed August 16, 2026).
  81. Mulimani M., Mesaros A. (2024). “ Online domain-incremental learning approach to classify acoustic scenes in all locations ”, in 2024 32nd European Signal Processing Conference (EUSIPCO) (Piscataway, NJ: IEEE), 96 – 100. doi: doi:10.23919/EUSIPCO63174.2024.10715156
  82. Mustafic S., Dachs D., Prüller R., Schöggl F., Perko R. (2024). “ Improving wildlife management with AI: Species detection and classification from camera trap data ”, in 4th International Workshop on Camera Traps, AI, and Ecology (CamTraps).
  83. Muthivhi M., Huo J., Gustafsson F., van Zyl T. L. (2025). Improving wildlife out-of-distribution detection: Africa’s big five. arXiv. doi: doi:10.48550/arXiv.2506.06719
  84. Napier T., Ahn E., Allen-Ankins S., Schwarzkopf L., Lee I. (2024). Advancements in preprocessing, detection and classification techniques for ecoacoustic data: A comprehensive review for large-scale passive acoustic monitoring. Expert Syst. Appl. 252, 124220. doi: doi:10.1016/j.eswa.2024.124220
  85. Nolasco I., Singh S., Morfi V., Lostanlen V., Strandburg-Peshkin A., Vidaña-Vila E.. (2023). Learning to detect an animal sound from five examples. Ecol. Inf. 77, 102258. doi: doi:10.1016/j.ecoinf.2023.102258
  86. Norman D. L., Bischoff P. H., Wearn O. R., Ewers R. M., Rowcliffe J. M., Evans B.. (2023). Can CNN-based species classification generalise across variation in habitat within a camera trap survey ? Methods Ecol. Evol. 14, 242 – 251. doi: doi:10.1111/2041-210X.14031
  87. Norouzzadeh M. S., Nguyen A., Kosmala M., Swanson A., Palmer M. S., Packer C.. (2018). Automatically identifying, counting, and describing wild animals in camera-trap images with deep learning. Proc. Natl. Acad. Sci. 115, E5716 – E5725. doi: doi:10.1073/pnas.1719367115
  88. NVIDIA Corporation (2019). Jetson Nano Technical Reference Manual. Available online at: https://developer.nvidia.com/embedded/jetson-nano
  89. Parham J., Stewart C., Crall J., Rubenstein D., Holmberg J., Berger-Wolf T. (2018). “ An animal detection pipeline for identification ”, in Proceedings of the 2018 IEEE Winter Conference on Applications of Computer Vision (WACV) (Piscataway, NJ: IEEE), 1075 – 1083. doi: doi:10.1109/WACV.2018.00123
  90. Pérez-Granados C.. (2025 b). WABAD: A world annotated bird acoustic dataset for passive acoustic monitoring. Res. square. doi: doi:10.21203/rs.3.rs-5729784/v1
  91. Pérez-Granados C., Funosas D., Morant J., Marín Gómez O. H., Mendoza I., Mohedano-Munoz M. A.. (2025 a). Optimization of passive acoustic bird surveys: a global assessment of BirdNET settings. Ibis 168, 785 – 798. doi: doi:10.1111/ibi.70013
  92. Pestell A. J. L., Rendall A. R., Sinclair R. D., Ritchie E. G., Nguyen D. T., Corva D. M.. (2025). Smart camera traps and computer vision improve detections of small fauna. Ecosphere 16, e70220. doi: doi:10.1002/ecs2.70220
  93. Petroni L., Natucci L., Massolo A. (2024). An ecologist-friendly R workflow for expediting species-level classification of camera trap images. Ecol. Evol. 14, e70544. doi: doi:10.1002/ece3.70544
  94. Petso T., Jamisola R. S., Alibhai S., Namoshe M., Mmereki W., Jewell Z. (2025). An investigation of lightweight AI models to classify African ungulate species from tracks. Ecol. Inf. 92, 103393. doi: doi:10.1016/j.ecoinf.2025.103393
  95. Pishdast H., Kalva H., Tye D. (2025). “ AI-enabled smart camera traps for wildlife monitoring in African ecosystems ”, in: Proceedings of the 2025 International Conference on Information Technology for Social Good (New York, NY: Association for Computing Machinery), 107 – 112. doi: doi:10.1145/3748699.3749780
  96. Prince P., Hill A., Piña Covarrubias E., Doncaster P., Snaddon J. L., Rogers A. (2019). Deploying acoustic detection algorithms on low-cost, open-source acoustic sensors for environmental monitoring. Sensors MDPI 19, 553. doi: doi:10.3390/s19030553
  97. Rajagukguk R. A., yeon Lee S., yeon Park J., Daniel K. F., rin Lee C., Chen Z.. (2025). Deep learning for visual animal monitoring (detection, tracking, pose estimation, and behavior classification): A comprehensive review. Smart Agric. Technol. 12, 101539. doi: doi:10.1016/j.atech.2025.101539
  98. Ranmal D., Ranasinghe P., Paranayapa T., Meedeniya D., Perera C. (2024). ESC-NAS: Environment sound classification using hardware-aware neural architecture search for the edge. Sensors 24, 3749. doi: doi:10.3390/s24123749
  99. Rastikerdar MM, Huang J, Guan H, Ganesan D (2025). WildFit: autonomous in-situ model adaptation for resource-constrained IoT systems. in Proceedings of the 2026 ACM/IEEE International Conference on Embedded Artificial Intelligence and Sensing Systems (New York, NY: Association for Computing Machinery), SenSys '26, 517 – 530. doi: doi:10.1145/3774906.3800481
  100. Rauch L., Schwinger R., Wirth M., Heinrich R., Huseljic D., Herde M.. (2025). BirdSet: A large-scale dataset for audio classification in avian bioacoustics. arXiv. doi: doi:10.48550/arXiv.2403.10380
  101. Raza A., Hanif F., Mohammed H. A. (2025). Analyzing the enhancement of CNN-YOLO and transformer based architectures for real-time animal detection in complex ecological environments. Sci. Rep. 15, 39142. doi: doi:10.1038/s41598-025-26645-2
  102. Rigoudy N., Dussert G., Benyoub A., Besnard A., Birck C., Boyer J.. (2023). The DeepFaune initiative: a collaborative effort towards the automatic identification of French fauna in camera-trap images. bioRxiv. doi: doi:10.1101/2022.03.15.484324
  103. Robinson D., Miron M., Hagiwara M., Weck B., Keen S., Alizadeh M.. (2025). NatureLM-audio: an audio-language foundation model for bioacoustics.
  104. Robinson D, Robinson A, Akrapongpisak L (2024). Transferable models for bioacoustics with human language supervision. in ICASSP 2024 – 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (Piscataway, NJ: IEEE), 1316 – 1320. doi: doi:10.1109/ICASSP48485.2024.10447250
  105. Rovero F., Zimmermann F. (2016). Camera Trapping for Wildlife Research (UK: Pelagic Publishing).
  106. Santamaria J., Isaza C., Giraldo J. (2025). “ CATALOG: A camera trap language-guided contrastive learning model Proceedings - 2025 IEEE Winter Conference on Applications of Computer Vision, WACV 2025 ”, in Proceedings - 2025 IEEE Winter Conference on Applications of Computer Vision, WACV 2025 (Piscataway, NJ: IEEE), 1197 – 1206. doi: doi:10.1109/WACV61041.2025.00124
  107. Scanferla J., Brambilla M., Brambilla G., Hilpold A., Marchetti A. E., Puff F.. (2025). Determining species-specific thresholds to improve precision in passive acoustic monitoring. Ecol. Inf. 91, 103423. doi: doi:10.1016/j.ecoinf.2025.103423
  108. Schäfer-Zimmermann J. C., Demartsev V., Averly B., Dhanjal-Adams K. L., Duteil M., Gall G.. (2025). animal2vec and MeerKAT: A self-supervised transformer for rare-event raw audio input and a large-scale reference dataset for bioacoustics. Methods Ecol. Evol 17, 875 – 888. doi: doi:10.1111/2041-210x.70218
  109. Schneider S., Greenberg S., Taylor G. W., Kremer S. C. (2020 a). Three critical factors affecting automated image species recognition performance for camera traps. Ecol. Evol. 10, 3503 – 3517. doi: doi:10.1002/ece3.6147
  110. Schneider S., Taylor G. W., Kremer S. C. (2018 b). “ Deep learning object detection methods for ecological camera trap data ”, in 2018 15th Conference on Computer and Robot Vision (CRV) (Piscataway, NJ: IEEE), 321 – 328.
  111. Schneider S, Taylor GW, Kremer SC (2020 b). Similarity learning networks for animal individual re-identification – beyond the capabilities of a human observer. in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW) (Piscataway, NJ: IEEE), 44 – 52. doi: doi:10.1109/WACVW50321.2020.9096925
  112. Schneider S., Taylor G., Linquist S., Kremer S. (2018 a). Past, present, and future approaches using computer vision for animal re-identification from camera trap data. Methods Ecol. Evol. 10. doi: doi:10.1111/2041-210X.13133
  113. Sethi S. S., Ewers R. M., Jones N. S., Orme C. D. L., Picinali L. (2018). Robust, real-time and autonomous monitoring of ecosystems with an open, low-cost, networked device. Methods Ecol. Evol. 9, 2383 – 2387. doi: doi:10.1111/2041-210X.13089
  114. Sharpe C. R., Hill R. A., Chappell H. M., Green S. E., Holden K., Fergus P.. (2025). Increasing citizen scientist accuracy with artificial intelligence on UK camera-trap data. Remote Sens. Ecol. Conserv. 11, 641 – 655. doi: doi:10.1002/rse2.70012
  115. Shepley A., Falzon G., Meek P., Kwan P. (2021). Automated location invariant animal detection in camera trap images using publicly available data sources. Ecol. Evol. 11, 4494 – 4506. doi: doi:10.1002/ece3.7344
  116. Shi W., Cao J. Q., Li Y., Xu L. (2016). Edge computing: Vision and challenges. IEEE Internet Things J. 3, 637 – 646. doi: doi:10.1109/JIOT.2016.2579198
  117. Singer D., Hagge J., Kamp J., Hondong H., Schuldt A. (2024). Aggregated time-series features boost species-specific differentiation of true and false positives in passive acoustic monitoring of bird assemblages. Remote Sens. Ecol. Conserv. 10, 517 – 530. doi: doi:10.1002/rse2.385
  118. Singh A, Pietrasik M, Natha G, Ghouaiel N, Brizel K, Ray N. (2020). Animal detection in man-made environments. in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (Piscataway, NJ: IEEE), 1427 – 1438. doi: doi:10.1109/WACV45572.2020.9093504
  119. Stowell D. (2022). Computational bioacoustics with deep learning: a review and roadmap. PeerJ 10, e13152. doi: doi:10.7717/peerj.13152
  120. Sugai L. S. M., Silva T. S. F., Ribeiro J., José Wagner, Llusia D. (2018 a). Terrestrial passive acoustic monitoring: review and perspectives. BioScience 69, 15 – 25. doi: doi:10.1093/biosci/biy147
  121. Sugai L. S. M., Silva T. S. F., Ribeiro J., José Wagner, Llusia D. (2018 b). Terrestrial passive acoustic monitoring: review and perspectives. BioScience 69, 15 – 25. doi: doi:10.1093/biosci/biy147
  122. Tabak M. A., Norouzzadeh M. S., Wolfson D. W., Sweeney S. J., Vercauteren K. C., Snow N. P.. (2019). Machine learning to classify animal species in camera trap images: applications in ecology. Methods Ecol. Evol. 10, 585 – 590. doi: doi:10.1111/2041-210X.13120
  123. Thompson M. C., Ducey M. J., Gunn J. S., Rowe R. J. (2024). A post-processing framework for assessing BirdNET identification accuracy and community composition. 167, 530 – 542. doi: doi:10.1111/ibi.13357
  124. Tolkova I., Chu B., Hedman M., Kahl S., Klinck H. (2021). Parsing birdsong with deep audio embeddings. arXiv. doi: doi:10.48550/arXiv.2108.09203
  125. Upton E., Halfacree G. (2016). Raspberry Pi User Guide (England: Wiley).
  126. Vandendriessche J., Wouters N., da Silva B., Lamrini M., Chkouri M. Y., Touhafi A. (2021). Environmental sound recognition on embedded systems: from FPGAs to TPUs. Electronics 10, 2622. doi: doi:10.3390/electronics10212622
  127. van Merriënboer B., Dumoulin V., Hamer J., Harrell L., Burns A., Denton T. (2026). Perch 2.0: the bittern lesson for bioacoustics. arXiv. doi: doi:10.48550/arXiv.2508.04665
  128. van Merriënboer B., Hamer J., Dumoulin V., Triantafillou E., Denton T. (2024). Birds, bats and beyond: evaluating generalization in bioacoustics models. Front. Bird Sci. 3. doi: doi:10.3389/fbirs.2024.1369756
  129. Velasco-Montero D., Fernández-Berni J., Carmona-Galán R., Sanglas A., Palomares F. (2024). Reliable and efficient integration of AI into camera traps for smart wildlife monitoring based on continual learning. Ecol. Inf. 83, 102815. doi: doi:10.1016/j.ecoinf.2024.102815
  130. Velasco-Montero D., Lozano-Pons C., Fernández-Berni J., Bastianelli G. (2026). On-site acoustic identification of bird species based on a shallow neural network. Ecol. Inf. 94, 103687. doi: doi:10.1016/j.ecoinf.2026.103687
  131. Vélez J., McShea W., Shamon H., Castiblanco-Camacho P. J., Tabak M. A., Chalmers C.. (2023). An evaluation of platforms for processing camera-trap data using artificial intelligence. Methods Ecol. Evol. 14, 459 – 477. doi: doi:10.1111/2041-210X.14044
  132. Vidaña-Vila E., Navarro J., Borda-Fortuny C., Stowell D., Alsina-Pages R. M. (2020). Low-cost distributed acoustic sensor network for real-time urban sound monitoring. Electronics 9, 2119. doi: doi:10.3390/electronics9122119
  133. Viswanathan A. S., Bock A., Bent Z., Peyton M. A., Tartakovsky D. M., Santos J. E. (2026). Two-stage wildlife event classification for edge deployment. Sensors 26, 1366. doi: doi:10.3390/s26041366
  134. Vyskočil J., Picek L. (2025). “ Towards zero-shot camera trap image categorization ”, in Computer Vision – ECCV 2024 Workshops (Cham: Springer Nature Switzerland), 37 – 53.
  135. Ware L., Mahon C. L., McLeod L., Jetté J.-F. (2023). Artificial intelligence (BirdNET) supplements manual methods to maximize bird species richness from acoustic data sets generated from regional monitoring. Can. J. Zool. 101, 1031 – 1051. doi: doi:10.1139/cjz-2023-0044
  136. Whytock R. C., Świezewski J., Zwerts J. A., Bara-Slupski T., Koumba Pambo A. F., Rogala M.. (2021). Robust ecological analysis of camera trap data labelled by a machine learning model. Methods Ecol. Evol. 12, 1080 – 1092. doi: doi:10.1111/2041-210X.13576
  137. Whytock R. C., Suijten T., van Deursen T., Świezewski J., Mermiaghe H., Madamba N.. (2023). Real-time alerts from AI-enabled camera traps using the Iridium satellite network: a case-study in Gabon, Central Africa. Methods Ecol. Evol. 14, 867 – 874. doi: doi:10.1111/2041-210X.14036
  138. Willi M., Pitman R. T., Cardoso A. W., Locke C., Swanson A., Boyer A.. (2019). Identifying animal species in camera trap images using deep learning and citizen science. Methods Ecol. Evol. 10, 80 – 91. doi: doi:10.1111/2041-210X.13099
  139. Williams B, van Merrienboer B, Dumoulin V, Hamer J, Triantafillou E, Fleishman AB. (2025). Using tropical reef, bird and unrelated sounds for superior transfer learning in marine bioacoustics. Philos Trans R Soc Lond B Biol Sci 380, 20240280. doi: doi:10.1098/rstb.2024.0280
  140. Winiarska D., Szymański P., Osiejuk T. S. (2024). Detection ranges of forest bird vocalisations: guidelines for passive acoustic monitoring. Sci. Rep. 14, 894. doi: doi:10.1038/s41598-024-51297-z
  141. Wood C. M., Kahl S. (2024). Guidelines for appropriate use of BirdNET scores and other detector outputs. J. Ornithology 165, 777 – 782. doi: doi:10.1007/s10336-024-02144-5
  142. Xie P., Gao R., Lu W., Shen L. (2025). Weakly supervised bounding-box generation for camera-trap image based animal detection. IET Comput. Vision 19, e12332. doi: doi:10.1049/cvi2.12332
  143. Xie J., Zhong Y., Zhang J., Liu S., Ding C., Triantafyllopoulos A. (2023). A review of automatic recognition technology for bird vocalizations in the deep learning era. Ecol. Inf. 73, 101927. doi: doi:10.1016/j.ecoinf.2022.101927
  144. Zhang J., Yang L., Sajjadi Mohammadabadi S. M., Yan F. (2025). A survey on self-supervised learning: recent advances and open problems. Neurocomputing 655, 131409. doi: doi:10.1016/j.neucom.2025.131409
  145. Zhu R., Zhao E., Hu C., Xie J., Zhang J., Hu H. (2025). Metric learning unveiling disparities: a novel approach to recognize false trigger images in wildlife monitoring. Ecol. Inf. 87, 103091. doi: doi:10.1016/j.ecoinf.2025.103091
  146. Zualkernan I. A., Dhou S., Judas J., Sajun A., Gomez B. R., Hussain L. A. (2022). An IoT system using deep learning to classify camera trap images on the edge. Computers 11, 13 – 13. doi: doi:10.3390/computers11010013

Citation

Velasco-Montero, D., et al. (2026). Bridging the edge–cloud gap: adaptive AI for robust image and audio wildlife monitoring. Frontiers in Conservation Science, Volume 7. https://doi.org/
Received
Accepted
Published
Volume
7 - 2026

Copyright

Copyright © 2026 Velasco-Montero, Fernández-Berni, Sankaran, Bustamante and Beckman.

License: https://creativecommons.org/licenses/by/4.0/

*Correspondence: Delia Velasco-Montero

Article metrics

Views
0
Downloads
0
Citations
0
Shares
0

Related articles

Original ResearchHuman-Hippopotamus conflicts and socio-ecological tensions in the Ruzizi plain: a one health and political ecology perspectiveEsther Mufungizi Ahana, Marie Shangalume Fadhili, François Mmunga Assumani, Christian Zagabe Muderhwa, Gustave Nachigera MushagalusaPerspectiveCyclic droughts, flooding and collective trauma: an environmental feedback loop to inform governance and socio-ecological resilience in Kenya’s rift valleyGilbert O. Obwoyere, Scott JonesHypothesis and TheoryThe Arid Zone Pollinator Paradox: hidden resilience erosion and structural-functional decoupling in dryland pollination systemsRachid Sabbahi