Scholarlink

<p>Accurate categorization of bird calls is essential for effective biodiversity surveillance and for assessing ecosystem conditions. It also enables monitoring of avian species distributions and detecting population changes. This is particularly important for species that are difficult to observe visually, but whose presence can still be identified through their distinctive calls. Moreover, in the long term, collecting audio samples can help capture changes in populations that may indicate shifts in ecosystem health. This research investigates the use of a two-level framework composed of a convolutional neural network (CNN), Categorical Boosting (CatBoost), and the Light Gradient Boosting Machine (LightGBM) for accurate sound classification of six South American bird species. Additionally, this study proposes a modified version of the variable neighborhood search (VNS) algorithm, employed to tune both layers of the framework for this specific task. A thorough comparative analysis was conducted against other powerful optimizers, accompanied by rigorous statistical evaluation of the results and SHapley Additive exPlanations (SHAP) analysis of the best models. The proposed approach achieved a superior classification accuracy of nearly 76.5% with a Matthews correlation coefficient of 0.722354, significantly outperforming the other evaluated methods. These results demonstrate both strong predictive performance and the suitability of the framework for classifying calls from different bird species.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Classifying birds of South America via audio analysis using convolutional networks and boosting models optimized by metaheuristics

  • 期刊论文

  • 发表日期：

    2026年1月3日

查看全文
  • Tea Dogandzic,
  • Miodrag Zivkovic,
  • Luka Jovanovic,
  • Jasmina Perisic,
  • Marina Milovanovic,
  • Milos Antonijevic,
  • Mohamed Hammad,
  • Nebojsa Bacanin

摘要

Accurate categorization of bird calls is essential for effective biodiversity surveillance and for assessing ecosystem conditions. It also enables monitoring of avian species distributions and detecting population changes. This is particularly important for species that are difficult to observe visually, but whose presence can still be identified through their distinctive calls. Moreover, in the long term, collecting audio samples can help capture changes in populations that may indicate shifts in ecosystem health. This research investigates the use of a two-level framework composed of a convolutional neural network (CNN), Categorical Boosting (CatBoost), and the Light Gradient Boosting Machine (LightGBM) for accurate sound classification of six South American bird species. Additionally, this study proposes a modified version of the variable neighborhood search (VNS) algorithm, employed to tune both layers of the framework for this specific task. A thorough comparative analysis was conducted against other powerful optimizers, accompanied by rigorous statistical evaluation of the results and SHapley Additive exPlanations (SHAP) analysis of the best models. The proposed approach achieved a superior classification accuracy of nearly 76.5% with a Matthews correlation coefficient of 0.722354, significantly outperforming the other evaluated methods. These results demonstrate both strong predictive performance and the suitability of the framework for classifying calls from different bird species.