Benchmarking wild bird detection in complex forest scenes

Ecological Informatics, Journal Year: 2024, Volume and Issue: 80, P. 102466 - 102466

Published: Jan. 9, 2024

Camera traps are widely used for wildlife monitoring and making informed conservation land-management decisions, but the resulting 'big data' laborious to process. Deep learning-based methods have been adopted detection in camera traps. However, these detect large mammals uncomplicated scenes, where powerful deep-learning models work effectively. Few studies conducted develop artificial intelligence recognizing wild birds that live complicated field scenes with protective colors small sizes. Here we a dataset of 9717 images from 15 bird species based on test 8 object algorithms (Faster RCNN, Cascade RetinaNet, FCOS, RepPoints, ATSS, Deformable-DETR, Sparse RCNN) assess their performance. We also explored effect different backbones model accuracy. Among them, RCNN performs best, mAP 0.693 capabilities. Models perform differently certain species, significantly affect accuracy model. utilizing Swin-T backbone is best-performing combination, 0.704. This study could help researchers identify efficiently inspires research recognition complex ecological settings.

Language: Английский

---

Object classification and visualization with edge artificial intelligence for a customized camera trap platform

Ecological Informatics, Journal Year: 2024, Volume and Issue: 79, P. 102453 - 102453

Published: Jan. 2, 2024

The camera traps have revolutionized the image and video capture in ecology are often used to monitor record animal presence. With miniaturization of low power electronic devices, better battery technologies, software advancements, it has become possible use edge such as Raspberry Pi that can not only images videos, but also enable sophisticated processing, off-site communications. These developments help provide near real-time insights reduce manual processing images. on-board classification visualization is facilitated by advancements Deep Neural Networks (DNN), transfer learning approaches, libraries. This paper provides an investigation with approaches using pre-trained DNN models, visualizations Explainable Artificial Intelligence (XAI) techniques on Zero (RPi-Z) device. MobileNetV2 model was for Florida-Part1 dataset obtaining results precision, recall, F1-score 0.95, 0.96, 0.95 respectively. We compared performance MobileNetV2, EfficientNetV2B0, MobileViT models Extinction best 0.97, 0.96 respectively, obtained EfficientNetV2B0 model. Two XAI techniques, Gradient-weighted-Class Activation Mapping (Grad-CAM) Occlusion Sensitivity were through heatmaps, highlight relative importance areas contributing model's prediction, understand bias. practical case scenarios utilizing optimization deployment ecological research.

Language: Английский

---

Development of a cost-efficient automated wildlife camera network in a European Natura 2000 site

Basic and Applied Ecology, Journal Year: 2024, Volume and Issue: 79, P. 141 - 152

Published: June 27, 2024

Modern approaches with advanced technology can automate and expand the extent resolution of biodiversity monitoring. We present development an innovative system for automated wildlife monitoring in a coastal Natura 2000 nature reserve Netherlands 65 wireless 4G cameras which are deployed autonomously field 12 V/2A solar panels, i.e. without need to replace batteries or manually retrieve SD cards. The transmit images automatically (through mobile network) sensor portal, contains PostgreSQL database functionalities task scheduling data management, allowing scientists site managers via web interface view remotely monitor performance (e.g. number uploaded files, battery status card storage cameras). camera trap sampling design combines grid-based stratified by major habitats placement along traditional route, experimental set-up inside outside large herbivore exclosures. This provides opportunities studying distribution, habitat use, activity, phenology, population structure community composition species allows comparison novel approaches. Images transferred application programming interfaces external services identification long-term storage. A deep learning model was tested showed promising results identifying focal species. Furthermore, detailed cost analysis revealed that establishment costs higher but annual operating much lower than those trapping, resulting being >40 % more cost-efficient. developed end-to-end pipeline demonstrates continuous networks is feasible cost-efficient, multiple benefits extending current methods. be applied open other reserves network coverage.

Language: Английский

---

Harnessing Artificial Intelligence for Wildlife Conservation

Conservation, Journal Year: 2024, Volume and Issue: 4(4), P. 685 - 702

Published: Nov. 11, 2024

The rapid decline in global biodiversity demands innovative conservation strategies. This paper examines the use of artificial intelligence (AI) wildlife conservation, focusing on Conservation AI platform. Leveraging machine learning and computer vision, detects classifies animals, humans, poaching-related objects using visual spectrum thermal infrared cameras. platform processes these data with convolutional neural networks (CNNs) transformer architectures to monitor species, including those that are critically endangered. Real-time detection provides immediate responses required for time-critical situations (e.g., poaching), while non-real-time analysis supports long-term monitoring habitat health assessment. Case studies from Europe, North America, Africa, Southeast Asia highlight platform’s success species identification, monitoring, poaching prevention. also discusses challenges related quality, model accuracy, logistical constraints outlining future directions involving technological advancements, expansion into new geographical regions, deeper collaboration local communities policymakers. represents a significant step forward addressing urgent offering scalable adaptable solution can be implemented globally.

Language: Английский

---

An Overview of AI Applications in Wildlife Conservation

Advances in environmental engineering and green technologies book series, Journal Year: 2025, Volume and Issue: unknown, P. 19 - 48

Published: Jan. 10, 2025

The integration of artificial intelligence (AI) into wildlife conservation has revolutionized methodologies for monitoring species, enhancing habitat management, and combating poaching. This chapter examines various AI applications that contribute to the protection preservation biodiversity. Remote sensing technologies, powered by machine learning algorithms, assist in assessing health tracking changes over time. AI-driven image recognition tools enable identification individual animals from camera trap photos, facilitating more accurate population estimates behavioral studies. Moreover, predictive analytics play a crucial role forecasting human-wildlife conflicts informing proactive management strategies. synthesis technologies demonstrates their potential enhance efforts, optimize resource allocation, ultimately foster effective initiatives. ongoing advancement this field promises create innovative solutions some most pressing challenges.

Language: Английский

---

Recognition of European mammals and birds in camera trap images using deep neural networks

IET Computer Vision, Journal Year: 2024, Volume and Issue: unknown

Published: July 3, 2024

Abstract Most machine learning methods for animal recognition in camera trap images are limited to mammal identification and group birds into a single class. Machine visually discriminating birds, turn, cannot discriminate between mammals not designed images. The authors present deep neural network models recognise both bird species They train classification as well predicting the taxonomy, that is, genus, family, order, group, class names. Different architectures, including ResNet, EfficientNetV2, Vision Transformer, Swin ConvNeXt, compared these tasks. Furthermore, investigate approaches overcome various challenges associated with image analysis. authors’ best achieve mean average precision (mAP) of 97.91% on validation data set mAPs 90.39% 82.77% test sets recorded forests Germany Poland, respectively. Their taxonomic reach mAP 97.18% 94.23% 79.92% two sets,

Language: Английский

---

AI-Driven Real-Time Monitoring of Ground-Nesting Birds: A Case Study on Curlew Detection Using YOLOv10

Remote Sensing, Journal Year: 2025, Volume and Issue: 17(5), P. 769 - 769

Published: Feb. 23, 2025

Effective monitoring of wildlife is critical for assessing biodiversity and ecosystem health as declines in key species often signal significant environmental changes. Birds, particularly ground-nesting species, serve important ecological indicators due to their sensitivity pressures. Camera traps have become indispensable tools nesting bird populations, enabling data collection across diverse habitats. However, the manual processing analysis such are resource-intensive, delaying delivery actionable conservation insights. This study presents an AI-driven approach real-time detection, focusing on curlew (Numenius arquata), a experiencing population declines. A custom-trained YOLOv10 model was developed detect classify curlews chicks using 3/4G-enabled cameras linked Conservation AI platform. The system processes camera trap real time, significantly enhancing efficiency. Across 11 sites Wales, achieved high performance, with 90.56%, specificity 100%, F1-score 95.05% detections 92.35%, 96.03% chick detections. These results demonstrate capability systems deliver accurate, timely assessments, facilitating early interventions advancing use technology research.

Language: Английский

---

Smart camera traps and computer vision improve detections of small fauna

Ecosphere, Journal Year: 2025, Volume and Issue: 16(3)

Published: March 1, 2025

Abstract Limited data on species' distributions are common for small animals, impeding conservation and management. Small especially ectothermic taxa, often difficult to detect, therefore require increased time resources survey effectively. The rise of technology has enabled researchers monitor animals in a range ecosystems longer periods than traditional methods (e.g., live trapping), increasing the quality cost‐effectiveness wildlife monitoring practices. We used DeakinCams, custom‐built smart camera traps, address three aims: (1) To including ectotherms, evaluate performance customized computer vision object detector trained SAWIT dataset automating classification; (2) At same field sites using commercially available we evaluated how well MegaDetector—a freely detection model—detected images containing animals; (3) complementarity these two different approaches monitoring. collected 85,870 videos from DeakinCams 50,888 commercial cameras. For with data, yielded 98% Precision but 47% recall, species classification, varied by 0% Recall birds 26% 14% spiders. detections trap images, MegaDetector returned 99% Recall. found that only detected nocturnal ectotherms invertebrates. Making use more diverse datasets training models as advances machine learning will likely improve like YOLO novel environments. Our results support need continued cross‐disciplinary collaboration ensure large environmental train test existing emerging algorithms.

Language: Английский

---

Biodiversity monitoring for biocredits: a case study comparing acoustic, eDNA, and traditional methods

Biodiversity and Conservation, Journal Year: 2025, Volume and Issue: unknown

Published: May 9, 2025

Language: Английский

---

An Improved Bird Detection Method Using Surveillance Videos from Poyang Lake Based on YOLOv8

Animals, Journal Year: 2024, Volume and Issue: 14(23), P. 3353 - 3353

Published: Nov. 21, 2024

Poyang Lake is the largest freshwater lake in China and plays a significant ecological role. Deep-learning-based video surveillance can effectively monitor bird species on lake, contributing to local biodiversity preservation. To address challenges of multi-scale object detection against complex backgrounds, such as high density severe occlusion, we propose new model known YOLOv8-bird model. First, use Receptive-Field Attention convolution, which improves model's ability capture utilize image information. Second, redesign feature fusion network, termed DyASF-P2, enhances network's small features reduces target information loss. Third, lightweight head designed reduce size without sacrificing precision. Last, Inner-ShapeIoU loss function proposed localization challenge. Experimental results PYL-5-2023 dataset demonstrate that achieves precision, recall, [email protected], [email protected]:0.95 scores 94.6%, 89.4%, 94.8%, 70.4%, respectively. Additionally, outperforms other mainstream models terms accuracy. These indicate well-suited for counting tasks, enable it support monitoring environment Lake.

Language: Английский

---