To crop or not to crop: Comparing whole‐image and cropped classification on a large dataset of camera trap images DOI Creative Commons

Tomer Gadot,

Ștefan Istrate, HyungWon Kim

et al.

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

Published: Nov. 24, 2024

Abstract Camera traps facilitate non‐invasive wildlife monitoring, but their widespread adoption has created a data processing bottleneck: camera trap survey can create millions of images, and the labour required to review those images strains resources conservation organisations. AI is promising approach for accelerating image review, tools are imperfect; in particular, classifying small animals remains difficult, accuracy falls off outside ecosystems which model was trained. It been proposed that incorporating an object detector into analysis pipeline may help address these challenges, benefit detection not systematically evaluated literature. In this work, authors assess hypothesis cropped from using species‐agnostic yields better than whole images. We find stage classification macro‐average F1 improvement around 25% on large, long‐tailed dataset; reproducible large public dataset smaller benchmark dataset. The describe architecture performs well both detector‐cropped demonstrate state‐of‐the‐art accuracy.

Language: Английский

WildARe-YOLO: A lightweight and efficient wild animal recognition model DOI Creative Commons
Sibusiso Reuben Bakana, Yongfei Zhang, Bhekisipho Twala

et al.

Ecological Informatics, Journal Year: 2024, Volume and Issue: 80, P. 102541 - 102541

Published: Feb. 23, 2024

Language: Английский

Citations

15
HPB3C-3PG algorithm: A new hybrid global optimization algorithm and its application to plant classification DOI Creative Commons
Sukanta Ghosh, Amar Singh,

Shakti Kumar

et al.

Ecological Informatics, Journal Year: 2024, Volume and Issue: 81, P. 102581 - 102581

Published: March 31, 2024

This paper proposes a hybrid bio-inspired search and optimization algorithm that combines the strengths of PB3C (Parallel Big Bang Crunch) 3PGA (3 Parent Genetic Algorithm) algorithms. The employs single population-based evolutionary coupled with multi-population parallel processing techniques to address problems. proposed is implemented in MATLAB software. We evaluate performance on CEC2021 standard test bench suite. approach compared other nine comparative analysis shows algorithms performed better than Furthermore, this chapter an HPB3C-3PGA-based evolve near-optimal architecture CNN. plant image classification Python 12 approaches. achieved accuracy 98.96% Mendeley dataset 98.97% CVIP100 dataset. outperforms all approaches for leaf problem. research significantly contributes overcoming limitations existing approaches, providing robust solution problems tasks.

Language: Английский

Citations

5
An automatic identification method of common species based on ensemble learning DOI Creative Commons
Haoxuan Li, Mei Zhang,

De-Yao Meng

et al.

Ecological Informatics, Journal Year: 2025, Volume and Issue: unknown, P. 103046 - 103046

Published: Jan. 1, 2025

Language: Английский

Citations

0
Hierarchical image classification using transfer learning to improve deep learning model performance for amazon parrots DOI Creative Commons
Jung-Il Kim,

Jong-Won Baek,

Chang-Bae Kim

et al.

Scientific Reports, Journal Year: 2025, Volume and Issue: 15(1)

Published: Jan. 30, 2025

Abstract Numerous studies have proven the potential of deep learning models for classifying wildlife. Such can reduce workload experts by automating species classification to monitor wild populations and global trade. Although typically perform better with more input data, available wildlife data are ordinarily limited, specifically rare or endangered species. Recently, citizen science programs helped accumulate valuable but such is still not enough achieve best performance compared benchmark datasets. Recent applied hierarchical a given dataset improve model accuracy. This study transfer Amazon parrot Specifically, hierarchy was built based on diagnostic morphological features. Upon evaluating performance, outperformed non-hierarchical in detecting parrots. Notably, achieved mean Average Precision (mAP) 0.944, surpassing mAP 0.908 model. Moreover, improved accuracy between morphologically similar The outcomes this may facilitate monitoring trade parrots conservation purposes.

Language: Английский

Citations

0
Metric learning unveiling disparities: A novel approach to recognize false trigger images in wildlife monitoring DOI Creative Commons
Rui Zhu, Enting Zhao, Chunhe Hu

et al.

Ecological Informatics, Journal Year: 2025, Volume and Issue: 87, P. 103091 - 103091

Published: March 5, 2025

Language: Английский

Citations

0
Automated near real-time monitoring in ecology: Status quo and ways forward DOI Creative Commons
Anna M. Davison, Koen de Koning, Franziska Taubert

et al.

Ecological Informatics, Journal Year: 2025, Volume and Issue: 89, P. 103157 - 103157

Published: April 18, 2025

Language: Английский

Citations

0
A systematic study on transfer learning: Automatically identifying empty camera trap images using deep convolutional neural networks DOI Creative Commons

Dengqi Yang,

De-Yao Meng,

Haoxuan Li

et al.

Ecological Informatics, Journal Year: 2024, Volume and Issue: 80, P. 102527 - 102527

Published: Feb. 17, 2024

Transfer learning is extensively utilized for automatically recognizing and filtering out empty camera trap images that lack animal presence. Current research uses transfer identifying typically solely updates the fully connected layer of models, they usually select a pre-trained source model only based on its relevance to target task. However, do not consider optimization update selection, nor investigate effect sample size class number domain data set used construct performance model. Both these are issues worth exploring. We answered two using three different datasets ResNext-101 Our experimental results showed when 20,000 training samples from ImageNet dataset Snapshot Serengeti dataset, our proposed optimal layers improved accuracy 92.9% 95.5% (z = −7.087, p < 0.001, N 8118) compared existing method updating layer. A similar improvement was observed transferring Lasha Mountain dataset. Additionally, indicated increasing binary-class build 100,000 1 million, 90.4% 93.5% −3.869, 8948). Similar were obtained constructing ten classifications. Based results, we drew following conclusions: (1) instead commonly can significantly improve model's performance. (2) The varied transferred same (3) classes in did impact positively correlated with performance, there might be threshold effect.

Language: Английский

Citations

3
Enabling people-centric climate action using human-in-the-loop artificial intelligence: a review DOI Creative Commons
Ramit Debnath, Nataliya Tkachenko, Malay Bhattacharyya

et al.

Current Opinion in Behavioral Sciences, Journal Year: 2025, Volume and Issue: 61, P. 101482 - 101482

Published: Jan. 24, 2025

Language: Английский

Citations

0
Being confident in confidence scores: calibration in deep learning models for camera trap image sequences DOI Creative Commons
Gaspard Dussert, Simon Chamaillé‐Jammes, Stéphane Dray

et al.

Remote Sensing in Ecology and Conservation, Journal Year: 2024, Volume and Issue: unknown

Published: June 16, 2024

Abstract In ecological studies, machine learning models are increasingly being used for the automatic processing of camera trap images. Although this automation facilitates and accelerates identification step, results these may lack interpretability their immediate applicability to downstream tasks (e.g. occupancy estimation) remains questionable. particular, little is known about calibration, a property that allows confidence scores be interpreted as probabilities model's predictions true. Using large diverse European dataset, we investigate whether deep species classification in images well calibrated. Additionally, traps often configured take multiple photos same event, also explore calibration aggregated across sequences Finally, study effect practicality post‐hoc method, i.e. temperature scaling, made at image sequence levels. Based on five established three independent test sets, show averaging logits over sequence, selecting an appropriate architecture, optionally using scaling can produce well‐calibrated models. Our findings have clear implication for, instance, calculation error rates or selection score thresholds studies making use artificial intelligence

Language: Английский

Citations

3
A versatile, semi-automated image analysis workflow for time-lapse camera trap image classification DOI Creative Commons
Gerardo Celis, Peter S. Ungar, Aleksandr Sokolov

et al.

Ecological Informatics, Journal Year: 2024, Volume and Issue: 81, P. 102578 - 102578

Published: March 26, 2024

Camera traps are a powerful, practical, and non-invasive method used widely to monitor animal communities evaluate management actions. However, camera trap arrays can generate thousands millions of images that require significant time effort review. Computer vision has emerged as tool accelerate this image review process. We propose multi-step, semi-automated workflow which takes advantage site-specific generalizable models improve detections consists (1) automatically identifying removing low-quality in parallel with classification into animals, humans, vehicles, empty, (2) cropping objects from classifying them (rock, bait, species), (3) manually inspecting subset images. trained evaluated approach using 548,627 46 cameras two regions the Arctic: "Finnmark" (Finnmark County, Norway) "Yamal" (Yamalo-Nenets Autonomous District, Russia). The automated steps yield accuracies 92% 90% for Finnmark Yamal sets, respectively, reducing number required manual inspection 9.2% set 3.9% set. amount invested developing would be offset by saved automation after 960 thousand have been processed. Researchers modify multi-step process develop their own meet other needs monitoring surveying wildlife, balancing acceptable levels false negatives positives.

Language: Английский

Citations

2