Publicação
Automatic animal recognition in wildlife conservation programmes using deep convolutional neural networks
| dc.contributor.author | Sousa, Susana Teixeira de | |
| dc.contributor.institution | Faculty of Sciences | |
| dc.contributor.supervisor | Pesquita, Cátia Luísa Santana Calisto | |
| dc.contributor.supervisor | Fonseca, Manuel João Caneira Monteiro da | |
| dc.date.accessioned | 2026-02-11T16:15:01Z | |
| dc.date.available | 2026-02-11T16:15:01Z | |
| dc.date.issued | 2025 | |
| dc.description | Tese de Mestrado, Bioinformática e Biologia Computacional, 2025, Universidade de Lisboa, Faculdade de Ciências | |
| dc.description.abstract | Biodiversity conservation increasingly depends on innovative approaches to monitor wildlife communities in the face of environmental change and anthropogenic pressures. This dissertation evaluates the use of Convolutional Neural Networks (CNNs) and transfer learning to automate the classification of camera-trap images, with a specific focus on carnivorous species in Mediterranean ecosystems. The study draws upon three datasets comprising more than 650,000 images collected between 2013 and 2022 in two protected areas: Companhia das Lezírias (Santarém) and Grândola (Setúbal). A combination of experimental approaches was employed, including models trained from scratch and transfer learning strategies leveraging pre-trained architectures. Cross-validation was implemented to ensure robustness, and ensemble methods were tested to further stabilize performance. Results revealed that VGG16 consistently achieved a good level of performance, with precision and recall surpassing 97% in binary classification tasks (animal vs. empty; carnivore vs. other). However, a significant limitation emerged in the form of poor cross-site generalization, highlighting the effects of environmental variability and domain shift. To mitigate this, incremental retraining experiments using only 10–20% of annotated images from new environments demonstrated substantial performance improvements, confirming the viability of adaptation under resource-constrained conditions. This dissertation concludes that CNN-based models with transfer learning constitute a powerful, adaptable tool for wildlife monitoring. Rather than proposing a fixed model, it advances a replicable workflow that can be tailored to local datasets. The findings reinforce the potential of artificial intelligence to complement ecological expertise, reducing manual workload and strengthening the capacity for long-term biodiversity monitoring. | en |
| dc.format | application/pdf | |
| dc.identifier.uri | http://hdl.handle.net/10400.5/117013 | |
| dc.language.iso | eng | |
| dc.subject | Ecological informatics | |
| dc.subject | Wildlife monitoring | |
| dc.subject | Camera-trap images | |
| dc.subject | Convolutional Neural Networks (CNNs) | |
| dc.subject | Computer vision | |
| dc.title | Automatic animal recognition in wildlife conservation programmes using deep convolutional neural networks | en |
| dc.type | master thesis | |
| dspace.entity.type | Publication | |
| rcaap.rights | openAccess |
Ficheiros
Principais
1 - 1 de 1
A carregar...
- Nome:
- TM_Susana_Sousa.pdf
- Tamanho:
- 3.03 MB
- Formato:
- Adobe Portable Document Format