Método para la detección de aves en espacios naturales y no naturales basado en técnicas de machine learning
Cargando...
Autores
Director
Tipo de contenido
Editor
Document language:
Español
Fecha
Título de la revista
ISSN de la revista
Título del volumen
Documentos PDF
Resumen
Abstract
Bird observation and monitoring play a fundamental role in the study and conservation of biodiversity, as these species act as sensitive indicators of environmental change. However, reliably detecting birds from images remains a challenging task. Birds may appear in highly diverse scenarios, ranging from dense forests and open natural landscapes to urban environments dominated by human-made structures. In addition to this contextual diversity, factors such as variations in lighting, the small size of many species, partial occlusions, and visual similarity to the background further complicate automatic detection using conventional methods. In recent years, machine learning techniques—and particularly Deep Learning models—have demonstrated significant potential for addressing complex computer vision problems. Nevertheless, in real-world environmental monitoring applications, important challenges remain related to model robustness, generalization across different environments, and the balance between accuracy and computational cost. In this context, there is a growing need for approaches that do not rely on a single model, but instead combine different perspectives of visual analysis to achieve more reliable decisions. This work is framed within this need and presents a Machine Learning–based method for bird detection in images from both natural and non-natural environments. The proposed approach is based on a two-stage complementary strategy. In the first stage, the entire image is broadly analyzed to locate regions that may contain birds, using an object detector of the YOLO (You Only Look Once) family, specifically the YOLOv8n version. In the second stage, these candidate regions are examined in greater detail using a binary convolutional neural network, whose purpose is to confirm or reject the presence of a bird. This division of the process makes it possible to take advantage of the speed and coverage of the initial detector while reinforcing the reliability of the final decision. For the development and evaluation of the method, images from the bird subset of the Open Images V7 dataset were used, balanced between natural and non-natural contexts. In order to coherently integrate the information provided by both models, their probabilistic outputs are calibrated and combined into a single prediction per image using an ensemble scheme. This integration is designed to promote system stability in visually complex scenarios, reduce errors in challenging cases, and maintain consistent performance across different types of environments. The evaluation of the method follows a reproducible protocol and relies on metrics widely accepted in the literature, enabling a systematic analysis of system behavior and comparison with relevant approaches reported in previous studies. The results show that combining a broad localization stage with a subsequent verification step improves the robustness of the detection process and offers an appropriate balance between result quality and computational efficiency. Quantitatively, in the main validation stage, the Bidirectional Gating ensemble achieved an Accuracy of 0.9783 (97.83%), Precision of 0.9714, Recall of 0.9855, and an F1-score of 0.9784, with only 3 errors over 138 evaluated images (2 false positives and 1 false negative), while maintaining an AUC close to 0.93. In addition, in an independent external validation (200 images from CUB-200-2011 and MS-COCO), the system obtained an Accuracy of 0.945, Precision of 0.901, Recall of 1.000, an F1-score of 0.948, and an AUC of 0.997, supporting the method’s generalization capability. Overall, this work aims to contribute a practical and well-founded approach for automatic bird detection in images, with potential applications in real-world environmental monitoring, biodiversity conservation, and computer-assisted ecological analysis.
Palabras clave propuestas
Descripción
ilustraciones, gráficas, tablas