Application of Deep Learning techniques in the Classification of Bird Audio for Environmental Monitoring in Doñana

Abstract

Passive acoustic monitoring (PAM) that uses devices like automatic audio recorders has become a fundamental tool in conserving and managing natural ecosystems. However, this practice generates a large volume of unsupervised audio data, and extracting valid information for environmental monitoring is a significant challenge. It is then critically necessary to use methods that leverage Deep Learning techniques for automating species detection. BirdNET is a model trained for bird identification that has succeeded in many study systems, especially in North America or Europe, but it results inadequate for other regions due to insufficient training and its bias on focal sounds rather than entire soundscapes. Another added problem for species detection is that many audios recorded in PAM programs are empty of sounds of species of interest or these sounds overlap. This study presents a multi-stage process for automatically identifying bird vocalizations that includes first a YOLOv8-based Bird Song Detector, and second, a fine-tuned BirdNET for species classification at a local scale with enhanced detection accuracy. As a study case, we applied this Bird Song Detector to audio recordings collected in Do˜nana National Park (SW Spain) as a part of the BIRDeep project. We annotated 461 minutes of audio data from three main habitats across nine different locations within Doñana, resulting in 3749 annotations representing 38 different classes. Mel spectrograms were employed as graphical representations of bird audio data, facilitating the application of image processing methods. Several detectors were trained in different experiments, which included data augmentation and hyperparameter exploration to improve the model’s robustness. The model giving the best results included the creation of synthetic background audios with data augmentation and the use of an environmental sound library. This proposed pipeline using the Bird Song Detector as a preliminary step, significantly improves BirdNET detections by increasing True Positives by approximately 281.97%, and reducing False Negatives by about 62.03%, thus demonstrating a novel and effective approach for bird species identification. Our findings underscore the importance of adapting general-purpose tools to address specific challenges in biodiversity monitoring. The experimental results show that fine-tuning Deep Learning models that account for the unique characteristics of specific ecological contexts can substantially enhance the accuracy and efficiency of PAM’s efforts.