Avaliação de redes neurais convolucionais e monitoramento de fauna urbana na Universidade Federal do Amapá campus Marco Zero: desafios da automação em ambiente de transição amazônico

Abstract

The monitoring of stray animals in urban areas is a strategic component of One Health, aiming at zoonosis control and the mitigation of impacts on native biodiversity. However, manual analysis of large volumes of data from camera traps constitutes a methodological bottleneck. This study aimed to develop and evaluate the performance of a Deep Learning prototype, based on the ResNet-50 architecture, for the automated classification of fauna at the Marco Zero campus of the Federal University of Amapá (UNIFAP), comparing it to the established DeepFaune and AddaxAI models. A total of 19,190 images were collected in an Amazonian urban environment. The prototype was implemented via the Torch for R ecosystem using fine-tuning techniques. Results revealed a global accuracy of 29.99% and a Kappa index of 0.0981 for the prototype, while the DeepFaune and AddaxAI models reached accuracy levels of ~65%. The analysis evidenced that the low performance stems from domain shift and the Terra Incognita effect, exacerbated by vegetation density and the loss of chromatic descriptors in nocturnal (infrared) records. In addition to the technological bias, the study identified significant anthropogenic pressure, with 16.68% of records consisting of domestic dogs (Canis lupus familiaris) and cats (Felis catus). It is concluded that, in the current scenario, the full automation of environmental surveillance at UNIFAP is unfeasible, and manual classification by specialists remains a mandatory and indispensable step to ensure data reliability, positioning artificial intelligence as a preliminary screening tool.