Spoilt for choice: An evaluation of machine learning for wildlife classification in tropical forest settings

Spoilt for choice: An evaluation of machine learning for wildlife classification in tropical forest settings | Authorea try { document.documentElement.classList.add('js'); } catch (e) { } var _gaq = _gaq || []; _gaq.push(['_setAccount', 'G-8VDV14Y67G']); _gaq.push(['_trackPageview']); (function() { var ga = document.createElement('script'); ga.type = 'text/javascript'; ga.async = true; ga.src = ('https:' == document.location.protocol ? 'https://ssl' : 'http://www') + '.google-analytics.com/ga.js'; var s = document.getElementsByTagName('script')[0]; s.parentNode.insertBefore(ga, s); })(); Skip to main content Preprints Collections Wiley Open Research IET Open Research Ecological Society of Japan All Collections About About Authorea FAQs Contact Us Quick Search anywhere Search for preprint articles, keywords, etc. Search Search ADVANCED SEARCH SCROLL This is a preprint and has not been peer reviewed. Data may be preliminary. 25 March 2026 V1 Latest version Share on Spoilt for choice: An evaluation of machine learning for wildlife classification in tropical forest settings Authors : Paulina Kukofka 0009-0000-6552-4591 [email protected] , Barbara Fruth , Anna-Marie Broska , Roland Cleva , Alain Mussa , Nadia Balduccio , and Mattia Bessone 0000-0002-8066-6413 Authors Info & Affiliations https://doi.org/10.22541/au.177443678.86881395/v1 250 views 96 downloads Contents Abstract Information & Authors Metrics & Citations View Options References Figures Tables Media Share Abstract In times of species decline and biodiversity loss, large-scale, multi-species surveys are essential for effective conservation planning. Camera traps have become indispensable for wildlife monitoring, providing detailed insights into species abundance and distribution. However, manual processing of large camera trap datasets remains time-consuming and resource-intensive. Although machine learning algorithms offer a scalable solution, their performance on data collected in densely vegetated tropical forests remains to be validated. Here, we evaluated machine learning algorithms for species classification, using 5,886 videos and 1,369 images from 55 camera traps deployed in the rainforest near Salonga National Park, Democratic Republic of the Congo, between September 2023 and March 2024. We compared performance metrics across species and algorithms and assessed how factors such as model confidence, distance, and body size influenced accuracies. Overall classification accuracies were lower than officially reported but comparable across region-specific algorithms. Specialist algorithms outperformed a generalist algorithm for classifications on the species level. Performance varied strongly between species, particularly for rare and small-bodied animals. Labelling confidence-based thresholding improved accuracies but resulted in substantial loss of data. Algorithmic confidence score was the strongest predictor of classification accuracy in region-specific algorithms, while distance negatively affected detection and correct classifications. Our findings underscore the importance of critical validation when applying machine learning algorithms, especially in challenging environments such as tropical forests. Machine learning performance is highly location- and species-dependent, with errors potentially propagating to derived ecological metrics. We outline practical steps to integrate manual validation into machine learning workflows to help practitioners avoid pitfalls and ensure robust and rigorous wildlife monitoring that strengthens conservation efforts. Spoilt for choice: An evaluation of machine learning for wildlife classification in tropical forest settings Authors Paulina Kukofka 1,2 *, Barbara Fruth 2,3,4 *, Anna-Marie Broska 5 , Roland Cleva 3 , Alain Omari Mussa 3 , Nadia Balduccio 2,3,6 , Mattia Bessone 2,3 Affiliations 1. Department of Biology, University of Konstanz, Universitätsstraße 10, 78464 Konstanz, Germany. 2. Department for the Ecology of Animal Societies, Max Planck Institute of Animal Behavior, Bücklestraße 5, 78467 Konstanz, Germany. 3. Centre for the Advanced Study of Collective Behaviour, Department of Biology, University of Konstanz, Universitätsstraße 10, 78464 Konstanz, Germany. 4. Centre for Research and Conservation, Royal Zoological Society of Antwerp, Antwerp, Belgium. 5. Department of Biology, University of Bielefeld, Universitätsstraße 25, 33615 Bielefeld, Germany. 6. International Max Planck Research School for Quantitative Behavior, Ecology and Evolution, Am Obstberg 1, 78315, Radolfzell, Germany. *Corresponding author: [email protected] ; [email protected] Running headline: Validating machine learning in tropical forests This research was funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy (EXC 2117 – 422037984, grant L22-01). It was endorsed by the Max Planck-Project for Bonobo Behavioural Ecology through the long-term support of the Max Planck Society, Germany, the Centre for Research and Conservation of the Royal Zoological Society of Antwerp, Belgium, the Ouwehands Zoo Foundation (OZF), and Bonobo Alive e.V. We thank our partners and collaborators in the DRC: the Institut Congolais pour la Conservation de la Nature (ICCN), the Ministère de la Recherche Scientifique, the Institut National pour la Recherche Biomedicale (INRB), the Unité de Gestion du Parc National de la Salonga (UGPNS), the Faculty of Agricultural and Environmental Sciences at the University of Kinshasa (UNIKIN), the Ambassade d’Allemagne and the Fondation Hanns-Seidel in Kinshasa. We are grateful to M. Crofoot, G. Hohmann, P. Pereboom, S. Strauff, K. Wendt, M. Wikelski, and I. Razik for institutional and scientific support. We thank the communities of the groupement Bolongo and all individuals involved in data collection. A special thank you goes to Justin Bkata’nkoy, Ramane Boke Lokomo, Bwela Ezelio, Mapiano Buadi, Kamuley Isapfu, Eto’o Jodel, Pascal Naky and Tartis Longwango. Author contributions Conceptualization: MB, BF, PK Data collection: RC, AOM Data curation & analysis: PK, MB, AMB, NB Funding acquisition: BF Methodology: MB, PK Project administration: BF, MB Supervision: MB, BF Visualization: PK Writing – original draft: PK Writing – review & editing: all authors Data availability statement The data and code used in this study is publicly available on GitHub: https://github.com/livingingroups/species-classification-ml-evaluation. Conflict of interest statement The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. Abstract In times of species decline and biodiversity loss, large-scale, multi-species surveys are essential for effective conservation planning. Camera traps have become indispensable for wildlife monitoring, providing detailed insights into species abundance and distribution. However, manual processing of large camera trap datasets remains time-consuming and resource-intensive. Although machine learning algorithms offer a scalable solution, their performance on data collected in densely vegetated tropical forests remains to be validated. Here, we evaluated machine learning algorithms for species classification, using 5,886 videos and 1,369 images from 55 camera traps deployed in the rainforest near Salonga National Park, Democratic Republic of the Congo, between September 2023 and March 2024. We compared performance metrics across species and algorithms and assessed how factors such as model confidence, distance, and body size influenced accuracies. Overall classification accuracies were lower than officially reported but comparable across region-specific algorithms. Specialist algorithms outperformed a generalist algorithm for classifications on the species level. Performance varied strongly between species, particularly for rare and small-bodied animals. Labelling confidence-based thresholding improved accuracies but resulted in substantial loss of data. Algorithmic confidence score was the strongest predictor of classification accuracy in region-specific algorithms, while distance negatively affected detection and correct classifications. Our findings underscore the importance of critical validation when applying machine learning algorithms, especially in challenging environments such as tropical forests. Machine learning performance is highly location- and species-dependent, with errors potentially propagating to derived ecological metrics. We outline practical steps to integrate manual validation into machine learning workflows to help practitioners avoid pitfalls and ensure robust and rigorous wildlife monitoring that strengthens conservation efforts. Keywords Machine learning, camera traps, wildlife monitoring, computer vision, species classification, tropical forest Introduction Throughout the world, human activities are driving unprecedented species decline and biodiversity loss (Ceballos et al., 2015; Ripple et al., 2019). Over the past 50 years, wildlife populations have declined by 73% ( WWF Living Planet Report , 2024) and more than 47,000 species are threatened with extinction, 27% of them mammals (IUCN, 2025). Informed conservation action requires identifying key areas of concern through extensive and regular surveys capable of capturing population dynamics across broad spatial and temporal scales (Nguimdo et al., 2025; Stephenson et al., 2022). This makes robust long-term wildlife monitoring more critical than ever (Langhammer et al., 2024; Nichols & Williams, 2006). This need is particularly acute in tropical forests, global biodiversity hotspots that support more than 60% of terrestrial vertebrate species (Myers et al., 2000; Pillay et al., 2022). Despite their ecological importance, these ecosystems face intense anthropogenic pressures, including deforestation, habitat fragmentation, and overhunting (Bourgoin et al., 2024; Brodie et al., 2015; Fischer et al., 2021). At the same time, tropical forests present unique challenges for wildlife monitoring (Zwerts et al., 2021). Dense vegetation and the cryptic behaviour of many species complicate their detection (Bessone et al., 2020). In these environments, traditional field methods relying on human observers, such as sampling of direct sightings or indirect animal signs, are slow, resource-intensive, and impractical for large-scale, multi-species censuses (Devos et al., 2008; Manley et al., 2004; Michel et al., 2026; Sunde & Jessen, 2013). Camera traps have emerged as a popular alternative to direct human observation. They are economical, minimally invasive tools that can collect data simultaneously over extended periods and across large areas (Delisle et al., 2021; McCarthy et al., 2019; Rovero & Zimmermann, 2016). The recorded observations enable the estimation of important ecological parameters, including animal density, distribution, and activity patterns (Howe et al., 2017; Rich et al., 2016; Rowcliffe et al., 2014). However, the huge volume of image and video data generated by camera traps often creates a processing bottleneck (Haucke et al., 2022; Magaldi et al., 2025). Manual extraction of information, such as species labels, can require weeks to months depending on dataset size and available human resources, considerably slowing down the information chain from data collection to interpretation, and consequently delaying timely conservation interventions (Ahumada et al., 2020; Haucke et al., 2022). The rise of machine learning algorithms for image processing promises to alleviate these constraints by automating specific tasks and thus substantially reducing processing times for camera trap datasets (Beery et al., 2019; Fergus et al., 2024; Vélez et al., 2023). Advances in deep learning, ranging from convolutional neural networks that extract layered features from raw images (Girshick et al., 2014; Redmon et al., 2016) to transformer‑based architectures that model long‑range spatial relationships (Dosovitskiy et al., 2021), have improved algorithm performance and applicability to real-world data. Current algorithms can filter images without detections (‘blanks’; Beery et al., 2019), count the number of individuals in an image (Norouzzadeh et al., 2018), identify species (e.g., Gadot et al., 2024; Norouzzadeh et al., 2018; Whytock et al., 2021; Yu et al., 2013), individuals (Bogucki et al., 2019; Hou et al., 2020), or behaviors (Chan et al., 2025), and estimate distances between camera and animal (Aamir et al., 2025; Haucke et al., 2022; Johanns et al., 2022). Machine learning pipelines for species classification usually employ a two-stage design that separates object detection from species identification (Dorne et al., 2025; Gadot et al., 2024; Magaldi et al., 2025). An object detector model localizes animals within images and generates bounding boxes around them. The most widely used detector is MegaDetector (Beery et al., 2019). Images are typically cropped to a bounding box region to minimize background interference before being processed by a classifier model trained to identify specific species (Gadot et al., 2024; Wäldchen & Mäder, 2018). Most classifiers operate on individual frames and temporal relationships between successive images are ignored (Dorne et al., 2025; Magaldi et al., 2025). Video-based platforms handle frame extraction internally and often pass only a subset of frames with high detection probabilities to the classifier model to reduce the risk of false positive detections and computational load (Dorne et al., 2025; Magaldi et al., 2025). The predictions are then aggregated into a single video-level classification (Dorne et al., 2025; Magaldi et al., 2025). Few models are truly video native and capable of capturing motion and interframe relationships (e.g., SlowFast, Feichtenhofer et al., 2019). Species classifiers vary in scope from broad taxonomic and geographic coverage (e.g., SpeciesNet; Gadot et al., 2024) to region-specific models trained on local species assemblages (e.g., European species (DeepFaune), Rigoudy et al., 2023; Central African forest species (Gabon), Whytock et al., 2021). Machine learning algorithms show great potential to support conservation efforts, with a growing number now accessible through user-friendly interfaces available to field practitioners and people without programming experience (e.g., wildlifeinsights.org, okala.io, agouti.io). However, while automated workflows can substantially reduce camera trap processing time, uncritical adoption of their outputs without any form of validation may lead to biased inference (Bevan et al., 2026; Huebner et al., 2024). This may be particularly the case in challenging habitats such as tropical forest (Fergus et al., 2024; Zwerts et al., 2021). To date, no study has comprehensively evaluated algorithm application in these environments or assessed whether conditions that challenge human reviewers, such as detecting small or occluded species and distinguishing morphologically similar taxa (Zett et al., 2022), similarly compromise machine learning performance (Haucke, Kühl, Hoyer, et al., 2022; Westworth et al., 2022). Object detection algorithms may fail to identify animals that are partially obscured by vegetation or that are small and distant, while vegetation itself may be misclassified as fauna (Leorna & Brinkman, 2022; Westworth et al., 2022). Additional external factors, including image coloration, lighting conditions, and animal orientation, can further affect performance (Westworth et al., 2022). Since reliable population estimates are critical for wildlife conservation and depend on precise species classification, it is essential to 1) identify the factors that influence detection and classification accuracy, 2) evaluate whether machine learning algorithms can overcome location-specific challenges, and 3) use this information to guide decisions about algorithm selection and integration into processing workflows. In this study, we evaluated the performance of four machine learning algorithms for the classification of camera trap species in an African tropical forest setting. We evaluated region-specific models from Zamba ( African species and African forest ; Dorne et al., 2025) and MbazaAI (Central African forests; Whytock et al., 2021) , as well as the global Google SpeciesNet model via the Wildlife Insights platform (Ahumada et al., 2020; Gadot et al., 2024) . Although the underlying code for all algorithms is open source, we deploy them through their respective web-based platforms or standalone software to reflect typical field practitioner workflows. We compare manually annotated and expert-validated data with model predictions to quantify algorithms’ performance and assess how metrics vary between species and algorithms. We then explore how ecological, image level and algorithmic factors influence detection and classification accuracy. We predict that 1. Region-specific models will outperform the globally trained model, particularly for morphologically similar species (Magaldi et al., 2025; Okuley et al., 2025) . 2. Performance metrics will be lower for rare species due to an imbalanced representation in the models’ training data (Schneider et al., 2020) . 3. factors influencing labelling accuracy will be similar across algorithms (Westworth et al., 2022) . While our analysis focuses on a specific set of existing algorithms, and new models continue to emerge rapidly, our aim is to provide practical validation steps applicable to any machine learning algorithm for species classification in wildlife monitoring. Materials and Methods Study Area The study area lies within the Congo basin, in the western periphery of Salonga National Park (block South), Democratic Republic of the Congo, the largest expanse of protected lowland rainforest in Africa. The site covers 273 km 2 of community forest of three villages situated north of the Lokoro I river (Fig. 1). The landscape is characterized by dense tropical lowland forests, interspersed with small grassy savannahs and shrublands (Fruth, 2011). The area is home to a diverse range of vertebrates, including emblematic and rare species such as the bonobo ( Pan paniscus ), pangolin ( Smutsia spp. ), African golden cat ( Caracal aurata) and leopard ( Panthera pardus pardus ) (Bessone et al., 2020; Hohmann & Fruth, 2003). Ethical Statement The protocols used for the collection of data were approved by the Comité National d’Ethique de la Santé (No 442/CNES/BN/PMMF/2023) and by the University of Kinshasa, Faculty of Agricultural Sciences (UNIKIN/FACAGRO/VDR/MMM/011/RKF/2023). The study was carried out in accordance with the requirements and guidelines of the Congolese Wildlife Authority (Institut National pour la Conservation de la Nature; ICCN) and adheres to the legal requirements of the host country, DRC. Figure 1. Study site and survey design. a) Democratic Republic of the Congo (DRC) with Salonga National Park (green) and b) the community forests of three villages (red) west of Salonga National Park, block south. Camera traps are included (orange) and excluded (green) due to failure or theft. Data collection We placed 55 infrared camera traps (27 Bushnell Core DS-4K / S-4K units and 29 Browning Aggressor HD No Glow units) across the study area on a systematic grid generated from a random starting point (Fig. 1b). The cameras were spaced two kilometres apart, mounted at a height of 50 cm, aligned parallel to the ground, and oriented at 0 ° or 180 ° to minimize false triggers due to sunset or sunrise. They were deployed between September 2023 and March 2024 in continuous video mode, 24 hours per day, with an inter-trigger lag time of one second and high sensitivity. Each camera was set to record 15 to 20 second videos, except for two cameras that recorded images. Due to theft (n = 3) and malfunctions, the station deployment durations ranged from 1.7 to 91.7 days (mean = 40.2 days, SD = 21.8 days). Species classification Two expert observers manually annotated a dataset of 5,886 videos and 1,369 images at the species level, using Kingdon (2018) for taxonomic reference. We used these manually assigned labels as a benchmark to evaluate the performance of machine learning algorithms. We selected four algorithms for automated species classification (Table 1): the African species (time_distributed) and African forest (slowfast) models from Zamba (v2.2.1; Dorne et al., 2025), accessed via the Zamba Cloud platform (zambacloud.com); the Central African forests (Gabon) model from MbazaAI (version 2.1.1; Whytock et al., 2021) through its offline software (appsilon.com/data-for-good/mbaza-ai); and the SpeciesNet model (Gadot et al., 2024) via the Wildlife Insights platform (wildlifeinsights.org; accessed in 2025). As our dataset consisted primarily of videos, but only Zamba Cloud supported direct video uploads, we extracted one frame per second from each video to maintain a manageable data volume. Each algorithm outputted a species label along with a confidence score ranging from 0 (low confidence) to 1 (high confidence). To assign a species label at the video level, we selected the non- ‘blank’ (that is, false trigger) label with the highest confidence score, unless all frames were classified as ‘blank’. MbazaAI and Zamba Cloud also provide top-2 and top-3 predictions, which can be useful for flagging rare species. However, since our aim was to assess model performance without manual post-hoc validation, we disregarded alternative predictions from our analysis. Table 1. Overview of platforms and models for automated species classification. Platform type Web-based platform Offline software Web-based platform Accessibility Zambacloud.com appsilon.com/data-for-good/mbaza-ai wildlifeinsights.org Source github.com/drivendataorg/zamba github.com/Appsilon/mbaza github.com/google/cameratrapai Offline capability No Yes No Upload support Images & Videos Images only Images only Strengths of platform/software Option to train custom models and to submit corrections Optimized for low-bandwidth environments Facilitates manual review and correction of automated labels Model tested on the platform African species (time_distributed) African forest (slowfast) Central African forests (Gabon) SpeciesNet ensemble Native input type Image Video Image Video Model Architecture 1. Object detection 2. Species classification CNN-based ensemble 1. MegaDetectorLite 1 2. TimeDistributedEfficientNetV2 2,3 CNN-based ensemble 1. MegaDetectorLite 1 2. SlowFast backbone 4 Deep CNN 1.+2. ResNet50 5 CNN-based ensemble 1. MegaDetector 6 2. EfficientNetV2 2 Geography Central and West Africa Central and West Africa Central Africa Global Number of species/groups 32 32 26 1,295 Training data ~ 250,000 videos ~ 15,000 videos ~ 350,000 images > 35,000,000 images Geographical coverage of training data Cameroon, Central African Republic, Democratic Republic of the Congo, Gabon, Guinea, Liberia, Mozambique, Nigeria, Republic of the Congo, Senegal, Tanzania, Uganda Cameroon, Central African Republic, Democratic Republic of the Congo, Gabon, Guinea, Liberia, Mozambique, Nigeria, Republic of the Congo, Senegal, Tanzania, Uganda Cameroon, Gabon, Republic of the Congo Global Output Top-3 predicted species labels, model confidence scores for all possible labels Top-3 predicted species labels, model confidence scores for all possible labels Top-3 predicted species labels, model confidence scores for all possible labels, number of individuals One species label, model confidence score, reviewer info, blank/non-blank, number of individuals Strength of model Well suited for tropical forests Potentially better at detecting small species Well suited for tropical forests Worldwide species; trained on largest dataset to date 1 Dorne et al., 2025; 2 Tan & Le, 2021; 3 docs.fast.ai/layers.html#timedistributed; 4 Feichtenhofer et al., 2019; 5 He et al., 2016; 6 Beery et al., 2019. Performance metrics of classification algorithms We evaluated machine learning performance by comparing model outputs to the manually annotated dataset. Initial exploration revealed frequent misclassification by Wildlife Insights at the species level, probably due to the platform’s inability to geofence predictions to the study region. To improve predictive capacity, we aggregated its classifications at higher taxonomic levels or by shared visual features (for example, grouping all duiker and antelope-like species into a single ’antelope / duiker’ label). To ensure comparability across algorithms and account for differences in taxonomic predictions, we further grouped species into 10 categories: antelope/duiker, bird, bonobo (the only non-human great ape in the study area), carnivore, human, monkey, pangolin, red river hog, rodent/shrew, and blank (false-trigger) (Table S1). We calculated the overall accuracy as the proportion of correct predictions relative to the total of predictions. For species-specific performance metrics, we coded each classification for each species as: ’True Positive’ (TP) when the manual and predicted presence of the species matched; ’True Negative’ (TN) when the manual and predicted absence of the species matched; ’False Positive’ (FP) when the algorithm incorrectly predicted the species as present; and ’False Negative’ (FN) when the algorithm incorrectly predicted the species as absent. We then calculated (1) precision as TP / (TP + FP), reflecting the proportion of correct predictions among all predictions for a given species; and (2) recall as TP / (TP+FN), measuring the proportion of actual occurrences correctly identified. Finally, we investigated how these metrics varied across different thresholds of model confidence scores (i.e., retaining only predictions above specified confidence levels). Factors influencing species detection and classification To investigate which factors influence object detection and species classification, we fitted Bayesian multinomial logistic regression models with three outcome categories: (1) correct classification, (2) false-species (animal detected but misclassified) and (3) false-blank predictions (animal not detected). We implemented separate models for each algorithm in Stan (version 2.32.2) through the RStan package (version 2.32.7, Stan Development Team, 2025) in R (version 4.5.1, R Core Team, 2025). We evaluated five factors that we expected to influence algorithmic accuracy based on the available literature (Table 2). Additionally, we also tested whether distance effects vary by body size category (small, medium, large). We accounted for nonindependence by including random intercepts for camera stations (n = 30) and species (n = 30), which allowed us to estimate within-camera and within-species variation while pooling information across groups. Table 2. Variables included in the models and corresponding predictions. Confidence score Continuous Standardized algorithm confidence score (0-1) higher confidence correlates with increased accuracy 1,2 Distance to camera Continuous Standardized distance in meters; mean across 2-second intervals for videos Animals are more likely to be missed at larger distances 3,4 Animals are harder to identify at larger distances 3,4 Body size Continuous Standardized species’ average body mass (kg) Small animals are missed more frequently 4,5 Small animals are harder to identify 3 Animal orientation Categorical Frontal, lateral, dorsal, multiple (videos only) - Lateral passes (the most common angle) yield higher accuracy 6 Image coloration Categorical Grayscale, colour Stronger contrasts positively influence accuracy 4 1 Whytock et al., 2021; 2 Willi et al., 2019; 3 Leorna & Brinkman, 2022; 4 Westworth et al., 2022; 5 Dorne et al., 2025; 6 Gomez Villa et al., 2017. We specified weakly informative priors for all parameters (Lemoine, 2019) and fitted four Markov Chain Monte Carlo (MCMC) chains of 2,000 iterations each. We evaluated convergence using R-hat statistics (all < 1.01) and evaluated the fit of the model using posterior predictive checks, comparing the observed outcome frequencies to the distributions of the simulated frequencies drawn from the posterior predictive distribution. We modelled outcome probabilities using the softmax function and report posterior means and 95% credible intervals (CIs). We evaluated posterior uncertainty and effect direction using the “p_direction” function of the bayestestR package (Makowski et al., 2019). Results Between September 2023 and March 2024, 55 camera stations recorded 46 species across 4,405 camera days, yielding 24,633 videos and 1,951 images. We evaluated machine learning algorithms using a random subset of 5,886 videos and 1,369 images. Inter-observer reliability for manual species classification was 93%. Performance metrics of classification algorithms We processed 110,276 video frames in MbazaAI and Wildlife Insights and 5,886 videos on Zamba Cloud. After aggregating predictions to one label per video, Wildlife Insights managed to assign labels to 99 % of the data, but frequently returned generic classifications (19 % ’Animal’, 13 % ’Mammal’, Fig. 2). Overall classification accuracy was 66 % for Zamba African species , 60 % for Zamba African forest, 59 % for MbazaAI and 38 % for Wildlife Insights. After aggregating species into broader groups, Zamba African species reached an accuracy of 69 %, Zamba African forest 64 %, MbazaAI Central African forests 66 %, and Wildlife Insights 58 %. When looking at algorithm accuracy for individual species, Zamba African species achieved the highest precision for humans (97 %) and red river hogs (94 %), with the best recall for humans (79 %), but performed poorly for monkeys (precision: 19 %, recall: 36 %) (Table S2). Zamba African forest showed high precision for antelopes/duikers (84 %) but low precision for monkeys (4 %) and carnivores (5 %), with recall ranging from 43 % (birds) to 73 % (antelopes/duikers) (Table S3). MbazaAI achieved high precision for antelope/duikers (83 %) but poorly classified bonobos (14 %) and monkeys (3 %). Recall was highest for red river hogs and antelope/duikers (both 79%), but only 4 % for carnivores and bonobos (Table S4). Wildlife Insights achieved perfect precision for bonobos (100 %) but only 10 % recall, with high precision for humans (99 %) and lowest for monkeys (70 %). Recall was highest in humans (83 %) (Table S5). All algorithms failed to classify pangolins (0 % recall and precision); however, only five pangolin instances were present in the data set (Fig. 2). Applying minimum confidence thresholds improved precision and recall for some species, and particularly in the region-specific models, while Wildlife Insights performance remained relatively stable (Fig. 3, S2). However, higher thresholds consistently reduced the amount of usable data (Fig. 3, S2). Detection of small animals (that is, rodents and shrews; regardless of correct identification) was highest for MbazaAI (91%) and Zamba African forest (89 %), followed by Wildlife Insights (75 %) and Zamba African species (53 %). On the contrary, blank image detection was lowest for MbazaAI (50 %) and highest for Wildlife Insights (93 %) and Zamba African species (91 %), with Zamba African forest at 70 % (Fig. 2, 3). Figure 2. Recall confusion matrix [%] of manual vs. predicted species labels for a) Zamba - African species , b) Zamba - African forest , c) MbazaAI – Central African forests , and d) Wildlife Insights - SpeciesNet . Each matrix compares manually annotated species labels (rows) with algorithm predictions (columns), with cell values representing percentages. Rows sum to 100 %, indicating the distribution of predictions for each species. Figure 3. Species-level performance metrics for Zamba - African species , b) Zamba - African forest , c) MbazaAI – Central African forests , and d) Wildlife Insights - SpeciesNet across model confidence thresholds. Blue = recall, orange = precision, and gray = proportion of usable data retained from the original dataset after thresholding. For the visualization of all species groups, see Fig. S2. Factors influencing species detection and classification We tested how model confidence score, distance from the camera, body size, orientation to the camera, and image coloration affect model performance in the four algorithms (Table S6-9). When we investigated the interaction between distance and body size, we found weak support for an effect and therefore excluded it from the final models (Fig. S1). Certainty showed the strongest and most consistent effects across all algorithms (Fig. 4). Higher certainty strongly increased the probability of correct classification (Zamba African species : β = 2.88, 95% CI [1.64, 4.10], p _direction = 1.00; Zamba African forest : β = 2.86 [1.59, 4.16], p _direction = 1.00; MbazaAI: β = 2.67 [1.36, 4.03], p _direction = 1.00), except for Wildlife Insights, which showed a weak negative trend (β = -0.55 [-2.08, 0.88], p _direction = 0.76). In contrast, higher certainty decreased misclassifications (’false species’) in all algorithms (Zamba African species : β = -0.90, [-2.16, 0.33], p _direction = 0.92; Zamba African forest : β = -2.03 [-3.33, -0.73], p _direction = 1.00; MbazaAI: β = -2.60 [-4.01, -1.24], p _direction = 1.00; Wildlife Insights: β = -2.98 [-4.44, -1.49], p _direction = 1.00). Higher certainty strongly decreased the probability of missed detections (’false blank’) in Zamba African species (β = -1.99 [-3.27, -0.75], p _direction = 1.00), but this pattern was reversed for Wildlife Insights (β = 3.52 [1.71, 5.30], p _direction = 1.00). Distance showed credible negative effects on correct classification of Zamba African species (β = -1.02 [-2.19, 0.13], p _direction = 0.96), with similar but weaker trends in other algorithms (Zamba African forest : β = -0.54 [-1.69, 0.61], p _direction = 0.82; MbazaAI: β = -0.76 [-1.92, 0.37], p _direction = 0.91; Wildlife Insights: β = -0.81 [-1.95, 0.32], p _direction = 0.92) (Fig. 4). Similarly, missed detections increased at larger distances across all algorithms, with the strongest effect in Wildlife Insights (β = 1.47 [0.26, 2.68], p _direction = 0.99), followed by MbazaAI (β = 0.90 [-0.31, 2.06], p _direction = 0.93), Zamba African species (β = 0.80 [-0.38, 1.94], p _direction = 0.91), and Zamba African forest (β = 0.55 [-0.66, 1.77], p _direction = 0.82). We found a weak negative trend for distance in misclassifications in Wildlife Insights (β = -0.67 [-1.80, 0.45], p _direction = 0.88), but no credible effects for the other algorithms. No algorithm showed credible effects of body size (Fig. 4), animal orientation (Fig. S3), or image coloration (Fig. S4) on any outcome category, with all 95% credible intervals overlapping zero. Figure 4. Posterior distributions of effects of confidence score, distance, and body size on classification outcomes for different algorithms. Points show posterior means, thick lines show 50% credible intervals, thin lines show 95% credible intervals. Discussion Our study evaluated machine learning algorithm performance for species classification in tropical forests and investigated factors that influence detection and classification accuracy. Our findings demonstrate that current algorithms vary in their performance for different species, but that they are well-developed to handle a range of potentially impeding external factors, with comparable effects across algorithms. However, manual validation remains essential to ensure scientific integrity and generate reliable metrics of wildlife population to inform conservation efforts (Bevan et al., 2026). Performance metrics of classification algorithms Overall classification accuracy was comparable among region-specific algorithms but lower for the generalist Wildlife Insights, which frequently misclassified native species as visually similar nonnative species. Although grouping species by shared visual features or higher taxonomic levels improved Wildlife Insights performance to levels comparable with region-specific models, this approach requires extensive post hoc species-level validation, reducing its suitability for multispecies surveys and large datasets (Magaldi et al., 2025; Vélez et al., 2023). However, generalist algorithms may detect invasive or rare species absent from region-specific species lists (Gadot et al., 2024). Overall accuracy was notably lower than officially reported figures for Zamba African species (-16%; zambacloud.com/#how-accurate-is-zamba-cloud) and MbazaAI Central African forests (-19%; Whytock et al., 2021). Both Zamba and MbazaAI models use whole-image classification rather than cropping to animal regions, and can thus be affected by background interference, particularly in areas characterized by dense vegetation, such as our study site (Gadot et al., 2024; Schneider et al., 2020). Species-level recall and precision varied considerably between both species groups and algorithms. Zamba African species and Wildlife Insights performed comparably for blank image and human detection, enabling efficient filtering of nontarget images in early stages of analysis. Although Wildlife Insights employs the highly accurate but computationally intensive MegaDetector (Beery et al., 2019), Zamba African species appears to compensate for the potential limitations of MegaDetectorLite (Dorne et al., 2025) with effective recognition through its classifier. On the contrary, the Zamba African forest performed slightly worse at blank detection, possibly because its motion-based pipeline generates false positives from non-animal movement. However, it outperformed Zamba African species and Wildlife Insights in detecting small-bodied animals, albeit with more frequent misclassifications. MbazaAI, which integrates object detection and classification within one model, showed low blank recall, suggesting conservative prediction thresholds. However, this approach resulted in high detection of small-bodied animals. As expected in imbalanced datasets (Schneider et al., 2020), recall and precision were low across all algorithms for rare and emblematic species, such as the African golden cat, bonobo, and pangolin, with the exception of Wildlife Insights, which performed well for carnivores. Confidence thresholds improved region-specific model performance but substantially reduced usable data, which is particularly problematic for rare or elusive species (Whytock et al., 2021; Willi et al., 2019). Nevertheless, thresholds may offer viable alternatives when manual validation is infeasible. Our findings confirm that machine learning performance is highly location and species dependent, even when models were trained on datasets closely resembling the study site (Beery et al., 2019; Schneider et al., 2020; Vélez et al., 2023). Factors influencing species detection and classification Model confidence score emerged as the strongest predictor of classification performance for region-specific models, confirming that confidence thresholds reliably improve predictive accuracy (Whytock et al., 2021; Willi et al., 2019) . This effect corresponds to the observed increase in accuracies at higher confidence thresholds (see above). In particular, Wildlife Insights did not show an effect of the confidence score on the correct classification and, conversely, higher scores increased missed detections. This counterintuitive pattern may reflect Wildlife Insights’ prefiltering approach, which only outputs predictions above 0.5 confidence, leaving portions of data unlabelled. Practitioners should therefore exercise caution when applying additional thresholds to Wildlife Insights output and rather opt for alternative validation methods. Distance to camera negatively impacted correct classification across all algorithms, with missed detections increasing at greater distances, particularly for Wildlife Insights. Previous studies confirm that distance reduces detection probability and classification accuracy, both in manual and automated approaches (Leorna & Brinkman, 2022; Westworth et al., 2022) . Although MegaDetector can detect objects up to 222m in open habitat (Leorna & Brinkman, 2022) , in forests with dense understory, animals are likely to be occluded at much shorter distances, hampering detection of diagnostic features facilitating species identification (Gomez Villa et al., 2017; Westworth et al., 2022) . We predicted that this effect would correlate with body size (Bukombe et al., 2016) but found no evidence for body size effects, even when in interaction with distance from the camera. Small animals may simply be invisible beyond certain distances and fully obscured by vegetation, likely not even triggering the camera trap (Hofmeester et al., 2017). Although some algorithms showed reduced recall for small animals, this may also reflect differences in detection pipelines and prediction thresholds rather than inherent size-based limitations. Similarly, animal orientation and image coloration did not show consistent effects, suggesting that these variables are either less influential than expected or adequately represented in model training. Application and outlook Machine learning has the potential to enhance biodiversity monitoring and contribute to more efficient and data-driven conservation strategies. Automated workflows can substantially accelerate camera trap data processing up to 500 % using MegaDetector (Fennell et al., 2022), and species classification algorithms can save 55-79 hours per 10,000 images or videos (Bak et al., 2025; Norouzzadeh et al., 2018). Hybrid approaches that combine machine learning models with citizen science classifications have reduced human effort by 43 % (Willi et al., 2019), while fully automated workflows could achieve time savings of up to 690 hours per 100,000 videos (Bak et al., 2025). Upload and processing times vary by platform and algorithm, but remain substantially shorter than manual annotations and can often run in the background. The field of machine learning for image classification continues to evolve rapidly. Today, many platforms offer data management support beyond species classification, such as metadata extraction, efficient manual review and annotations, and data export, further streamlining workflows (Vélez et al., 2023). For African forests, the recently published DeepForestVision algorithm has reported higher accuracies than the MbazaAI and Zamba African species models (Magaldi et al., 2025). In particular, users can select which species from the model list to include during labelling, thus restricting classifications to regionally relevant species. By providing accessible advanced algorithms even for practitioners with limited programming experience, these tools reduce barriers in wildlife monitoring and conservation (Bak et al., 2025). However, we also emphasize the need for cautious implementation: automated tools used without critical validation may introduce biases or errors with knock-on consequences on the resulting ecological parameters (Bevan et al., 2026). The integration of machine learning into conservation practices should therefore be guided by transparency, validation, and close collaboration between ecologists, data scientists, and local stakeholders. When choosing an appropriate algorithm, practitioners should consider the type of their data (images or videos), the species of interest, and site-specific challenges, rather than relying solely on reported performance metrics. Regardless of the machine learning algorithm used, manual validation remains critical for maintaining scientific rigor. This may involve testing a subset of algorithm outputs against human annotations or conducting visual inspections, particularly for classifications that appear incorrect or counterintuitive (e.g., species unlikely to be present in the study region, frequent occurrences of rare species labels). An example of a semiautomated workflow is outlined in Fig. 5. By following the steps highlighted therein, practitioners can avoid the pitfalls inherent to machine learning algorithms, ensuring robust and rigorous wildlife monitoring, with the potential to inform conservation actions. Figure 5. Semiautomated workflow for species classification. A combination of automated species classification with validation steps to enhance data accuracy. Graphic by I. Razik. References Aamir, M., Wijers, M., Loveridge, A., & Markham, A. (2025). A robust metric distance and height estimation pipeline for wildlife camera trap imagery. Ecological Informatics , 92 , 103520. https://doi.org/10.1016/j.ecoinf.2025.103520Ahumada, J. A., Fegraus, E., Birch, T., Flores, N., Kays, R., O’Brien, T. G., Palmer, J., Schuttler, S., Zhao, J. Y., Jetz, W., Kinnaird, M., Kulkarni, S., Lyet, A., Thau, D., Duong, M., Oliver, R., & Dancer, A. (2020). Wildlife Insights: A Platform to Maximize the Potential of Camera Trap and Other Passive Sensor Wildlife Data for the Planet. Environmental Conservation , 47 (1), 1–6. https://doi.org/10.1017/S0376892919000298Bak, T., Camp, R. J., Burt, M. D., & Vogt, S. (2025). Automated methods for processing camera trap video data for distance sampling. Pacific Conservation Biology , 31 (4). https://doi.org/10.1071/pc25008Beery, S., Morris, D., & Yang, S. (2019). Efficient Pipeline for Camera Trap Image Review (Version 1). arXiv. https://doi.org/10.48550/ARXIV.1907.06772Bessone, M., Kühl, H. S., Hohmann, G., Herbinger, I., N’Goran, K. P., Asanzi, P., Da Costa, P. B., Dérozier, V., Fotsing, E. D. B., Beka, B. I., Iyomi, M. D., Iyatshi, I. B., Kafando, P., Kambere, M. A., Moundzoho, D. B., Wanzalire, M. L. K., & Fruth, B. (2020). Drawn out of the shadows: Surveying secretive forest species with camera trap distance sampling. Journal of Applied Ecology , 57 (5), 963–974. https://doi.org/10.1111/1365-2664.13602Bevan, P. A., Pantazis, O., Pringle, H. A. I., Ferreira, G. B., Ingram, D. J., Madsen, E. K., Thomas, L., Thanet, D. R., Silwal, T., Rayamajhi, S., Brostow, G. J., Mac Aodha, O., & Jones, K. E. (2026). Deep learning‐based ecological analysis of camera trap images is impacted by training data quality and quantity. Remote Sensing in Ecology and Conservation , rse2.70052. https://doi.org/10.1002/rse2.70052Bogucki, R., Cygan, M., Khan, C. B., Klimek, M., Milczek, J. K., & Mucha, M. (2019). Applying deep learning to right whale photo identification. Conservation Biology , 33 (3).Bourgoin, C., Ceccherini, G., Girardello, M., Vancutsem, C., Avitabile, V., Beck, P. S. A., Beuchle, R., Blanc, L., Duveiller, G., Migliavacca, M., Vieilledent, G., Cescatti, A., & Achard, F. (2024). Human degradation of tropical moist forests is greater than previously estimated. Nature , 631 (8021), 570–576. https://doi.org/10.1038/s41586-024-07629-0Brodie, J. F., Giordano, A. J., Zipkin, E. F., Bernard, H., Mohd‐Azlan, J., & Ambu, L. (2015). Correlation and persistence of hunting and logging impacts on tropical rainforest mammals. Conservation Biology , 29 (1), 110–121. https://doi.org/10.1111/cobi.12389Bukombe, J., Senzota, R. B., Fryxell, J. M., Kittle, A., Kija, H., Hopcraft, J. G. C., Mduma, S., & Sinclair, A. R. E. (2016). Do animal size, seasons and vegetation type influence detection probability and density estimates of Serengeti ungulates? African Journal of Ecology , 54 (1), 29–38. https://doi.org/10.1111/aje.12255Ceballos, G., Ehrlich, P. R., Barnosky, A. D., García, A., Pringle, R. M., & Palmer, T. M. (2015). Accelerated modern human–induced species losses: Entering the sixth mass extinction. Science Advances , 1 (5), e1400253. https://doi.org/10.1126/sciadv.1400253Chan, A. H. H., Putra, P., Schupp, H., Köchling, J., Straßheim, J., Renner, B., Schroeder, J., Pearse, W. D., Nakagawa, S., Burke, T., Griesser, M., Meltzer, A., Lubrano, S., & Kano, F. (2025). YOLO-Behaviour: A simple, flexible framework to automatically quantify animal behaviours from videos. Methods in Ecology and Evolution , 16 (4), 760–774. https://doi.org/10.1111/2041-210x.14502Delisle, Z. J., Flaherty, E. A., Nobbe, M. R., Wzientek, C. M., & Swihart, R. K. (2021). Next-Generation Camera Trapping: Systematic Review of Historic Trends Suggests Keys to Expanded Research Applications in Ecology and Conservation. Frontiers in Ecology and Evolution , 9 . https://doi.org/10.3389/fevo.2021.617996Devos, C., Sanz, C., Morgan, D., Onononga, J., Laporte, N., & Huynen, M. (2008). Comparing ape densities and habitats in northern Congo: Surveys of sympatric gorillas and chimpanzees in the Odzala and Ndoki regions. American Journal of Primatology , 70 (5), 439–451. https://doi.org/10.1002/ajp.20514Dorne, E., Qi, J., Bull, P., Stephens, C., Bessone, M., Debetencourt, B., Fruth, B., Morgan, D., Palmer, M. S., Sanz, C., Wendefeuer, J., Crockford, C., Deschner, T., Langergraber, K. E., Piel, A. K., Robbins, M., Sommer, V., Stewart, F. A., Wittig, R. M., … Arandjelovic, M. (2025). Zamba: Computer vision for wildlife conservation . 85–111. https://doi.org/10.25080/crcw9835Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., & Houlsby, N. (2021). An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (arXiv:2010.11929). arXiv. https://doi.org/10.48550/arXiv.2010.11929Feichtenhofer, C., Fan, H., Malik, J., & He, K. (2019). SlowFast Networks for Video Recognition (arXiv:1812.03982). arXiv. https://doi.org/10.48550/arXiv.1812.03982Fennell, M., Beirne, C., & Burton, A. C. (2022). Use of object detection in camera trap image identification: Assessing a method to rapidly and accurately classify human and animal detections for research and application in recreation ecology. Global Ecology and Conservation , 35 , e02104. https://doi.org/10.1016/j.gecco.2022.e02104Fergus, P., Chalmers, C., Longmore, S., & Wich, S. (2024). Harnessing Artificial Intelligence for Wildlife Conservation. Conservation , 4 (4), 685–702. https://doi.org/10.3390/conservation4040041Fischer, R., Taubert, F., Müller, M. S., Groeneveld, J., Lehmann, S., Wiegand, T., & Huth, A. (2021). Accelerated forest fragmentation leads to critical increase in tropical forest edge area. Science Advances , 7 (37). https://doi.org/10.1126/sciadv.abg7012Fruth, B. (2011). The CBD in the Democratic Republic of Congo (DRC): The project: “the Cuvette Centrale as reservoir of medicinal plants” in the process of implementation. Curare , 34 , 1–2.Gadot, T., Istrate, Ștefan, Kim, H., Morris, D., Beery, S., Birch, T., & Ahumada, J. (2024). To crop or not to crop: Comparing whole‐image and cropped classification on a large dataset of camera trap images. IET Computer Vision , 18 (8), 1193–1208. https://doi.org/10.1049/cvi2.12318Girshick, R., Donahue, J., Darrell, T., & Malik, J. (2014). Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation . 580–587.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. Ecological Informatics , 41 , 24–32. https://doi.org/10.1016/j.ecoinf.2017.07.004Haucke, T., Kühl, H. S., Hoyer, J., & Steinhage, V. (2022). Overcoming the Distance Estimation Bottleneck in Estimating Animal Abundance with Camera Traps. Ecological Informatics , 68 , 101536. https://doi.org/10.1016/j.ecoinf.2021.101536He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep Residual Learning for Image Recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 770–778. https://doi.org/10.1109/CVPR.2016.90Hofmeester, T. R., Rowcliffe, J. M., & Jansen, P. A. (2017). A simple method for estimating the effective detection distance of camera traps. Remote Sensing in Ecology and Conservation , 3 (2), 81–89. https://doi.org/10.1002/rse2.25Hohmann, G., & Fruth, B. (2003). Lui Kotal-A new site for field research on bonobos in the Salonga National Park. Pan Africa News , 10 , 25–27. https://doi.org/10.5134/143430Hou, J., He, Y., Yang, H., Connor, T., Gao, J., Wang, Y., Zeng, Y., Zhang, J., Huang, J., Zheng, B., & Zhou, S. (2020). Identification of animal individuals using deep learning: A case study of giant panda. Biological Conservation , 242 , 108414. https://doi.org/10.1016/j.biocon.2020.108414Howe, E. J., Buckland, S. T., Després‐Einspenner, M., & Kühl, H. S. (2017). Distance sampling with camera traps. Methods in Ecology and Evolution , 8 (11), 1558–1565. https://doi.org/10.1111/2041-210X.12790Huebner, S. E., Palmer, M. S., & Packer, C. (2024). Human Supervision is Key to Achieving Accurate AI-assisted Wildlife Identifications in Camera Trap Images. Citizen Science: Theory and Practice , 9 (1), 38. https://doi.org/10.5334/cstp.752IUCN. (2025). The IUCN Red List of Threatened Species. Version 2025-1 .Johanns, P., Haucke, T., & Steinhage, V. (2022). Automated Distance Estimation for Wildlife Camera Trapping. Ecological Informatics , 70 , 101734. https://doi.org/10.1016/j.ecoinf.2022.101734Kingdon, J. (2018). The Kingdon Field Guide to African Mammals: Second Edition . Bloomsbury Wildlife.Langhammer, P. F., Bull, J. W., Bicknell, J. E., Oakley, J. L., Brown, M. H., Bruford, M. W., Butchart, S. H. M., Carr, J. A., Church, D., Cooney, R., Cutajar, S., Foden, W., Foster, M. N., Gascon, C., Geldmann, J., Genovesi, P., Hoffmann, M., Howard-McCombe, J., Lewis, T., … Brooks, T. M. (2024). The positive impact of conservation action. Science , 384 (6694), 453–458. https://doi.org/10.1126/science.adj6598Lemoine, N. P. (2019). Moving beyond noninformative priors: Why and how to choose weakly informative priors in Bayesian analyses. Oikos , 128 (7), 912–928. https://doi.org/10.1111/oik.05985Leorna, S., & Brinkman, T. (2022). Human vs. machine: Detecting wildlife in camera trap images. Ecological Informatics , 72 , 101876. https://doi.org/10.1016/j.ecoinf.2022.101876Magaldi, H., Cornette, R., Tibesigwa, J. J., Katumba, R., Rugonge, H., Amarasekaran, B., Anderson, N., Cappelle, N., Cardoso, A. W., Cornélis, D., Deschner, T., Fonteyn, D., Garriga, R. M., Van Lunteren, P., Rufray, X., Vanthomme, H., Zwerts, J. A., & Krief, S. (2025). DeepForestVision: Automated wildlife identification for camera traps of African tropical forests. Ecological Solutions and Evidence , 6 (4), e70167. https://doi.org/10.1002/2688-8319.70167Makowski, D., Ben-Shachar, M., & Lüdecke, D. (2019). bayestestR: Describing Effects and their Uncertainty, Existence and Significance within the Bayesian Framework. Journal of Open Source Software , 4 (40), 1541. https://doi.org/10.21105/joss.01541Manley, P. N., Zielinski, W. J., Schlesinger, M. D., & Mori, S. R. (2004). Evaluation of a multiple-species approach to monitoring species at the ecoregional scale. Ecological Applications , 14 (1), 296–310. https://doi.org/10.1890/02-5249McCarthy, M. S., Després-Einspenner, M.-L., Farine, D. R., Samuni, L., Angedakin, S., Arandjelovic, M., Boesch, C., Dieguez, P., Havercamp, K., Knight, A., Langergraber, K. E., Wittig, R. M., & Kühl, H. S. (2019). Camera traps provide a robust alternative to direct observations for constructing social networks of wild chimpanzees. Animal Behaviour , 157 , 227–238. https://doi.org/10.1016/j.anbehav.2019.08.008Michel, A., Mbeke, J. K., Visando, B. K., Kahlenberg, S., Fawcett, K., & Caillaud, D. (2026). Terrain ruggedness and human activities influence the distribution of Grauer’s gorillas (Gorilla beringei graueri) and eastern chimpanzees (Pan troglodytes schweinfurthii) in the Tayna Nature Reserve, Democratic Republic of Congo. Biological Conservation , 313 , 111641. https://doi.org/10.1016/j.biocon.2025.111641Myers, N., Mittermeier, R. A., Mittermeier, C. G., Da Fonseca, G. A. B., & Kent, J. (2000). Biodiversity hotspots for conservation priorities. Nature , 403 (6772), 853–858. https://doi.org/10.1038/35002501Nguimdo, V. R. V., Abwe, E. E., Morgan, B. J., Ketchen, M. E., Mfossa, D. M., Abwe, A. E., Betobe, N. E., Whytock, R. C., Tobler, M. W., Kamp, J., Balkenhol, N., Waltert, M., & Soofi, M. (2025). Long‐Term Monitoring of Hunting Signs Reveals Complex Spatiotemporal Patterns of Hunting Activities in an Unprotected African Rainforest. Diversity and Distributions , 31 (1). https://doi.org/10.1111/ddi.13951Nichols, J., & Williams, B. (2006). Monitoring for conservation. Trends in Ecology & Evolution , 21 (12), 668–673. https://doi.org/10.1016/j.tree.2006.08.007Norouzzadeh, M. S., Nguyen, A., Kosmala, M., Swanson, A., Palmer, M. S., Packer, C., & Clune, J. (2018). Automatically identifying, counting, and describing wild animals in camera-trap images with deep learning. Proceedings of the National Academy of Sciences , 115 (25). https://doi.org/10.1073/pnas.1719367115Okuley, O. S., Aiello, C. M., Glad, W., Perkins, K., Ianniello, R., Darby, N., & Epps, C. W. (2025). Improving AI performance in wildlife monitoring through species and environment-specific training: A case study on desert Bighorn sheep. Ecological Informatics , 89 , 103179. https://doi.org/10.1016/j.ecoinf.2025.103179Pillay, R., Venter, M., Aragon-Osejo, J., González-del-Pliego, P., Hansen, A. J., Watson, J. E., & Venter, O. (2022). Tropical forests are home to over half of the world’s vertebrate species. Frontiers in Ecology and the Environment , 20 (1), 10–15. https://doi.org/10.1002/fee.2420R Core Team. (2025). R: A Language and Environment for Statistical Computing [Computer software]. R Foundation for Statistical Computing. https://www.R-project.org/Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. (2016). You Only Look Once: Unified, Real-Time Object Detection. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 779–788. https://doi.org/10.1109/CVPR.2016.91Rich, L. N., Miller, D. A. W., Robinson, H. S., McNutt, J. W., & Kelly, M. J. (2016). Using camera trapping and hierarchical occupancy modelling to evaluate the spatial ecology of an African mammal community. Journal of Applied Ecology , 53 (4), 1225–1235. https://doi.org/10.1111/1365-2664.12650Rigoudy, N., Dussert, G., Benyoub, A., Besnard, A., Birck, C., Boyer, J., Bollet, Y., Bunz, Y., Caussimont, G., Chetouane, E., Carriburu, J. C., Cornette, P., Delestrade, A., De Backer, N., Dispan, L., Le Barh, M., Duhayer, J., Elder, J.-F., Fanjul, J.-B., … Chamaillé-Jammes, S. (2023). The DeepFaune initiative: A collaborative effort towards the automatic identification of European fauna in camera trap images. European Journal of Wildlife Research , 69 (6). https://doi.org/10.1007/s10344-023-01742-7Ripple, W. J., Wolf, C., Newsome, T. M., Barnard, P., & Moomaw, W. R. (2019). World Scientists’ Warning of a Climate Emergency .Rovero, F., & Zimmermann, F. (Eds.). (2016). Camera Trapping for Wildlife Research . Pelagic Publishing. JSTOR. https://doi.org/10.2307/jj.28833751Rowcliffe, J. M., Kays, R., Kranstauber, B., Carbone, C., & Jansen, P. A. (2014). Quantifying levels of animal activity using camera trap data. Methods in Ecology and Evolution , 5 (11), 1170–1179. https://doi.org/10.1111/2041-210X.12278Schneider, S., Greenberg, S., Taylor, G. W., & Kremer, S. C. (2020). Three critical factors affecting automated image species recognition performance for camera traps. Ecology and Evolution , 10 (7), 3503–3517. https://doi.org/10.1002/ece3.6147Stan Development Team. (2025). RStan: The R interface to Stan (Version 2.32.7) [Computer software]. https://mc-stan.org/Stephenson, P. J., Londoño-Murcia, M. C., Borges, P. A. V., Claassens, L., Frisch-Nwakanma, H., Ling, N., McMullan-Fisher, S., Meeuwig, J. J., Unter, K. M. M., Walls, J. L., Burfield, I. J., Do Carmo Vieira Correa, D., Geller, G. N., Montenegro Paredes, I., Mubalama, L. K., Ntiamoa-Baidu, Y., Roesler, I., Rovero, F., Sharma, Y. P., … Fumagalli, L. (2022). Measuring the Impact of Conservation: The Growing Importance of Monitoring Fauna, Flora and Funga. Diversity , 14 (10), 824. https://doi.org/10.3390/d14100824Sunde, P., & Jessen, L. (2013). It counts who counts: An experimental evaluation of the importance of observer effects on spotlight count estimates. European Journal of Wildlife Research , 59 (5), 645–653. https://doi.org/10.1007/s10344-013-0717-8Tan, M., & Le, Q. V. (2021). EfficientNetV2: Smaller Models and Faster Training (arXiv:2104.00298). arXiv. https://doi.org/10.48550/arXiv.2104.00298Vélez, J., McShea, W., Shamon, H., Castiblanco‐Camacho, P. J., Tabak, M. A., Chalmers, C., Fergus, P., & Fieberg, J. (2023). An evaluation of platforms for processing camera‐trap data using artificial intelligence. Methods in Ecology and Evolution , 14 (2), 459–477. https://doi.org/10.1111/2041-210X.14044Wäldchen, J., & Mäder, P. (2018). Machine learning for image based species identification. Methods in Ecology and Evolution , 9 (11), 2216–2225. https://doi.org/10.1111/2041-210x.13075Westworth, S. O. A., Chalmers, C., Fergus, P., Longmore, S. N., Piel, A. K., & Wich, S. A. (2022). Understanding External Influences on Target Detection and Classification Using Camera Trap Images and Machine Learning. Sensors , 22 (14), 5386. https://doi.org/10.3390/s22145386Whytock, R. C., Świeżewski, J., Zwerts, J. A., Bara‐Słupski, T., Koumba Pambo, A. F., Rogala, M., Bahaa‐el‐din, L., Boekee, K., Brittain, S., Cardoso, A. W., Henschel, P., Lehmann, D., Momboua, B., Kiebou Opepa, C., Orbell, C., Pitman, R. T., Robinson, H. S., & Abernethy, K. A. (2021). Robust ecological analysis of camera trap data labelled by a machine learning model. Methods in Ecology and Evolution , 12 (6), 1080–1092. https://doi.org/10.1111/2041-210X.13576Willi, M., Pitman, R. T., Cardoso, A. W., Locke, C., Swanson, A., Boyer, A., Veldthuis, M., & Fortson, L. (2019). Identifying animal species in camera trap images using deep learning and citizen science. Methods in Ecology and Evolution , 10 (1), 80–91. https://doi.org/10.1111/2041-210X.13099 WWF Living Planet Report 2024 A system in peril (1. unveränderte engl.-sprachige Gesamtausgabe 2024). (2024). WWF Deutschland.Yu, X., Wang, J., Kays, R., Jansen, P. A., Wang, T., & Huang, T. (2013). Automated identification of animal species in camera trap images. EURASIP Journal on Image and Video Processing , 2013 (1). https://doi.org/10.1186/1687-5281-2013-52Zett, T., Stratford, K. J., & Weise, F. J. (2022). Inter-observer variance and agreement of wildlife information extracted from camera trap images. Biodiversity and Conservation , 31 (12), 3019–3037. https://doi.org/10.1007/s10531-022-02472-zZwerts, J. A., Stephenson, P. J., Maisels, F., Rowcliffe, M., Astaras, C., Jansen, P. A., Van Der Waarde, J., Sterck, L. E. H. M., Verweij, P. A., Bruce, T., Brittain, S., & Van Kuijk, M. (2021). Methods for wildlife monitoring in tropical forests: Comparing human observations, camera traps, and passive acoustic sensors. Conservation Science and Practice , 3 (12), e568. https://doi.org/10.1111/csp2.568 Information & Authors Information Version history V1 Version 1 25 March 2026 Copyright This work is licensed under a Non Exclusive No Reuse License. Keywords comparative ecological experiment population ecology statistical terrestrial vertebrate Authors Affiliations Paulina Kukofka 0009-0000-6552-4591 [email protected] Max Planck Institute of Animal Behavior View all articles by this author Barbara Fruth Max Planck Institute of Animal Behavior View all articles by this author Anna-Marie Broska Universität Bielefeld Fakultät für Biologie View all articles by this author Roland Cleva University of Konstanz View all articles by this author Alain Mussa University of Konstanz View all articles by this author Nadia Balduccio Max Planck Institute of Animal Behavior View all articles by this author Mattia Bessone 0000-0002-8066-6413 Max Planck Institute of Animal Behavior View all articles by this author Metrics & Citations Metrics Article Usage 250 views 96 downloads .FvxKWukQNSOunydq8rnd { width: 100px; } Citations Download citation Paulina Kukofka, Barbara Fruth, Anna-Marie Broska, et al. Spoilt for choice: An evaluation of machine learning for wildlife classification in tropical forest settings. Authorea . 25 March 2026. DOI: https://doi.org/10.22541/au.177443678.86881395/v1 If you have the appropriate software installed, you can download article citation data to the citation manager of your choice. Simply select your manager software from the list below and click Download. For more information or tips please see 'Downloading to a citation manager' in the Help menu . Format Please select one from the list RIS (ProCite, Reference Manager) EndNote BibTex Medlars RefWorks Direct import Tips for downloading citations document.getElementById('citMgrHelpLink').addEventListener('click', function() { popupHelp(this.href); return false; }); $(".js__slcInclude").on("change", function(e){ if ($(this).val() == 'refworks') $('#direct').prop("checked", false); $('#direct').prop("disabled", ($(this).val() == 'refworks')); }); View Options View options PDF View PDF Figures Tables Media Share Share Share article link Copy Link Copied! Copying failed. Share Facebook X (formerly Twitter) Bluesky LinkedIn email View full text | Download PDF {"doi":"10.22541/au.177443678.86881395/v1","type":"Article"} Now Reading: Share Figures Tables Close figure viewer Back to article Figure title goes here Change zoom level Go to figure location within the article Download figure Toggle share panel Toggle share panel Share Toggle information panel Toggle information panel Go to previous graphic Go to next graphic Go to previous table Go to next table All figures All tables View all material View all material xrefBack.goTo xrefBack.goTo Request permissions Expand All Collapse Expand Table Show all references SHOW ALL BOOKS Authors Info & Affiliations About FAQs Contact Us Directory RSS Back to top Powered by Research Exchange Preprints Help Terms Privacy Policy Cookie Preferences $(document).ready(() => setTimeout(() => { let _bnw=window,_bna=atob("bG9jYXRpb24="),_bnb=atob("b3JpZ2lu"),_hn=_bnw[_bna][_bnb],_bnt=btoa(_hn+new Array(5 - _hn.length % 4).join(" ")); $.get("/resource/lodash?t="+_bnt); },4000)); (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'9fe1e1ec88dc4193',t:'MTc3OTE4MDkyNA=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();