An evaluation of platforms for processing camera‐trap data using artificial intelligence

Methods in Ecology and Evolution, Journal Year: 2022, Volume and Issue: 14(2), P. 459 - 477

Published: Dec. 28, 2022

Abstract Camera traps have quickly transformed the way in which many ecologists study distribution of wildlife species, their activity patterns and interactions among members same ecological community. Although they provide a cost‐effective method for monitoring multiple species over large spatial temporal scales, time required to process data can limit efficiency camera‐trap surveys. Thus, there has been considerable attention given use artificial intelligence (AI), specifically deep learning, help data. Using learning these applications involves training algorithms, such as convolutional neural networks (CNNs), particular features images automatically detect objects (e.g. animals, humans, vehicles) classify species. To overcome technical challenges associated with CNNs, several research communities recently developed platforms that incorporate easy‐to‐use interfaces. We review key characteristics four AI platforms—Conservation AI, MegaDetector, MLWIC2: Machine Learning Wildlife Image Classification Insights—and two auxiliary platforms—Camelot Timelapse—that output processing compare software programming requirements, features, management tools format. also R code from our own work demonstrate how users evaluate model performance. found classifications Conservation MLWIC2 Insights generally had low moderate recall. Yet, precision some higher taxonomic groups was high, MegaDetector high recall when classifying either ‘blank’ or ‘animal’. These results suggest most will need predictions, but improve camera‐trap‐data by allowing filter dataset into subsets certain blanks) be verified using bulk actions. By reviewing popular AI‐powered sharing an open‐source GitBook illustrates manage performance, we hope facilitate ecologists'

Language: Английский

---

Animal Detection and Classification from Camera Trap Images Using Different Mainstream Object Detection Architectures

Animals, Journal Year: 2022, Volume and Issue: 12(15), P. 1976 - 1976

Published: Aug. 4, 2022

Camera traps are widely used in wildlife surveys and biodiversity monitoring. Depending on its triggering mechanism, a large number of images or videos sometimes accumulated. Some literature has proposed the application deep learning techniques to automatically identify camera trap imagery, which can significantly reduce manual work speed up analysis processes. However, there few studies validating comparing applicability different models for object detection real field monitoring scenarios. In this study, we firstly constructed image dataset Northeast Tiger Leopard National Park (NTLNP dataset). Furthermore, evaluated recognition performance three currently mainstream architectures compared training day night data separately versus together. experiment, selected YOLOv5 series (anchor-based one-stage), Cascade R-CNN under feature extractor HRNet32 two-stage), FCOS extractors ResNet50 ResNet101 (anchor-free one-stage). The experimental results showed that day-night joint is satisfying. Specifically, average result our was 0.98 mAP (mean precision) animal 88% accuracy video classification. One-stage YOLOv5m achieved best accuracy. With help AI technology, ecologists extract information from masses imagery potentially quickly efficiently, saving much time.

Language: Английский

---

A gentle introduction to computer vision‐based specimen classification in ecological datasets

Journal of Animal Ecology, Journal Year: 2024, Volume and Issue: 93(2), P. 147 - 158

Published: Jan. 17, 2024

Abstract Classifying specimens is a critical component of ecological research, biodiversity monitoring and conservation. However, manual classification can be prohibitively time‐consuming expensive, limiting how much data project afford to process. Computer vision, form machine learning, help overcome these problems by rapidly, automatically accurately classifying images specimens. Given the diversity animal species contexts in which are captured, there no universal classifier for all use cases. As such, ecologists often need train their own models. While numerous software programs exist support this process, fundamental understanding computer vision works select appropriate model workflows based on specific case, types, computing resources desired performance capabilities. Ecologists may also face characteristic quirks datasets, such as long‐tail distributions, ‘unknown’ species, similarity between polymorphism within impact efficacy vision. Despite growing interest ecology, few available challenges they likely encounter. Here, we present gentle introduction using In manuscript associated GitHub repository, demonstrate prepare training data, basic procedures, methods evaluation selection. Throughout, explore considerations should make when models, domains, feature extractors class imbalances. With basics, adjust achieve research goals and/or account uncertainty downstream analysis. Our goal provide guidance getting started or improving learning visual tasks.

Language: Английский

---

Human impacts on mammals in and around a protected area before, during, and after COVID‐19 lockdowns

Conservation Science and Practice, Journal Year: 2022, Volume and Issue: 4(7)

Published: June 7, 2022

The dual mandate for many protected areas (PAs) to simultaneously promote recreation and conserve biodiversity may be hampered by negative effects of on wildlife. However, reports these are not consistent, presenting a knowledge gap that hinders evidence-based decision-making. We used camera traps monitor human activity terrestrial mammals in Golden Ears Provincial Park the adjacent University British Columbia Malcolm Knapp Research Forest near Vancouver, Canada, with objective discerning relative various forms cougars (

Language: Английский

---

Automated visitor and wildlife monitoring with camera traps and machine learning

Remote Sensing in Ecology and Conservation, Journal Year: 2023, Volume and Issue: 10(2), P. 236 - 247

Published: Aug. 30, 2023

Abstract As human activities in natural areas increase, understanding human–wildlife interactions is crucial. Big data approaches, like large‐scale camera trap studies, are becoming more relevant for studying these interactions. In addition, open‐source object detection models rapidly improving and have great potential to enhance the image processing of from wildlife activities. this study, we evaluate performance model MegaDetector cross‐regional monitoring using traps. The at detecting counting humans, animals vehicles evaluated by comparing results with manual classifications than 300 000 images three study regions. Moreover, investigate structural patterns misclassification typical temporal analyses conducted ecological research. Overall, accuracy was very high 96.0% animals, 93.8% persons 99.3% vehicles. Results reveal systematic misclassifications that can be automatically identified removed. show readily used count people on underestimating −0.05, −0.01 counts per image. Most importantly, pattern a long‐term time series manually classified highly correlated classification (Pearson's r = 0.996, p < 0.001) diurnal kernel densities were almost equivalent automated classification. thus prove overall applicability process studies without further intervention. Besides acceleration speed, also suitable allows reproducibility scientific while complying privacy regulations.

Language: Английский

---

Mammalian predator and prey responses to recreation and land use across multiple scales provide limited support for the human shield hypothesis

Ecology and Evolution, Journal Year: 2023, Volume and Issue: 13(9)

Published: Sept. 1, 2023

Outdoor recreation is widespread, with uncertain effects on wildlife. The human shield hypothesis (HSH) suggests that could have differential predators and prey, predator avoidance of humans creating a spatial refuge 'shielding' prey from people. generality the HSH remains to be tested across larger scales, wherein shielding may prove generalizable, or diminish variability in ecological contexts. We combined data 446 camera traps 79,279 sampling days 10 landscapes spanning 15,840 km

Language: Английский

---

UAV equipped with infrared imaging for Cervidae monitoring: Improving detection accuracy by eliminating background information interference

Ecological Informatics, Journal Year: 2024, Volume and Issue: 81, P. 102651 - 102651

Published: May 24, 2024

Wild Cervidae(deer and their relatives) play a crucial role in maintaining ecological balance are integral components of ecosystems. However, factors such as environmental changes poaching behaviors have resulted habitat degradation for Cervidae. The protection wild Cervidae has become urgent, monitoring is one the key means to ensure effectiveness protection. Object detection algorithms based on deep learning offer promising potential automatically detecting identifying animals. when those used inference unseen background environments, there will be significant decrease accuracy, especially situation that certain type images collected from single scene algorithm training. In this paper, two-stage localization classification pipeline proposed. effectively reduces interference enhances accuracy. first stage, YOLOv7 network designed locate UAV infrared images, while implementing improved bounding box regression through α-IoU loss function enables more accurately. Then, Cevdidae objects extracted eliminate information. second named CA-Hybrid, Convolutional Neural Networks(CNN) Vision Transformer(ViT), well Channel Attention Mechanism(CAM) expression features, constructed accurately identify categories. Experimental results indicate method achieves an Average Precision (AP) 95.9% location top-1 accuracy 77.73% identification. This research contributes comprehensive accurate Cervidae, provides valuable references subsequent UAV-based wildlife monitoring.

Language: Английский

---

Beyond observation: Deep learning for animal behavior and ecological conservation

Ecological Informatics, Journal Year: 2024, Volume and Issue: unknown, P. 102893 - 102893

Published: Nov. 1, 2024

Language: Английский

---

IoT-Based Object-Detection System to Safeguard Endangered Animals and Bolster Agricultural Farm Security

Future Internet, Journal Year: 2023, Volume and Issue: 15(12), P. 372 - 372

Published: Nov. 21, 2023

Significant threats to ecological equilibrium and sustainable agriculture are posed by the extinction of animal species subsequent effects on farms. Farmers face difficult decisions, such as installing electric fences protect their farms, although these measures can harm animals essential for maintaining equilibrium. To tackle issues, our research introduces an innovative solution in form object-detection system. In this research, we designed implemented a system that leverages ESP32-CAM platform conjunction with YOLOv8 model. Our proposed aims identify endangered harmful within farming environments, providing real-time alerts farmers wildlife integrating cloud-based alert train model effectively, meticulously compiled diverse image datasets featuring agricultural settings, subsequently annotating them. After that, tuned hyperparameter enhance performance The results from optimized auspicious. It achieves remarkable mean average precision (mAP) 92.44% impressive sensitivity rate 96.65% unseen test dataset, firmly establishing its efficacy. achieving optimal result, employed IoT when detects presence animals, it immediately activates audible buzzer. Additionally, was utilized notify neighboring effectively potential danger. This research’s significance lies drive conservation while simultaneously mitigating damage inflicted animals.

Language: Английский

---

Removing Human Bottlenecks in Bird Classification Using Camera Trap Images and Deep Learning

Remote Sensing, Journal Year: 2023, Volume and Issue: 15(10), P. 2638 - 2638

Published: May 18, 2023

Birds are important indicators for monitoring both biodiversity and habitat health; they also play a crucial role in ecosystem management. Declines bird populations can result reduced services, including seed dispersal, pollination pest control. Accurate long-term of birds to identify species concern while measuring the success conservation interventions is essential ecologists. However, time-consuming, costly often difficult manage over long durations at meaningfully large spatial scales. Technology such as camera traps, acoustic monitors drones provide methods non-invasive monitoring. There two main problems with using traps monitoring: (a) cameras generate many images, making it process analyse data timely manner; (b) high proportion false positives hinders processing analysis reporting. In this paper, we outline an approach overcoming these issues by utilising deep learning real-time classification automated removal trap data. Images classified Faster-RCNN architecture. transmitted 3/4G processed Graphical Processing Units (GPUs) conservationists key detection metrics, thereby removing requirement manual observations. Our models achieved average sensitivity 88.79%, specificity 98.16% accuracy 96.71%. This demonstrates effectiveness automatic

Language: Английский

---