Performance Comparison of Bird Classification Models in Dongtan, China Based on YOLOv7, SVC and XGBoost

Abstract

This paper mainly explores the performance differences of different models in the classification of birds in Chongming Dongtan, Shanghai. The research object of this paper is a data set composed of 2,900 images from three categories (Anas formosa, Anas platyrhynchos and swan), which are among the most abundant and representative bird species in Dongtan. I constructed VGG model, VGG-SVC fusion model and VGG-XGBoost fusion model where I input the results of the last hidden layer of VGG into the SVC and XGBoos. The experimental results show that the bird classification model constructed has achieved high accuracy in the test set. Among them, VGG model has the best performance, with its accuracy of 79.92%, followed by VGG-SVC model with the accuracy of 74.62% and VGG- XGBoost fusion model with the accuracy of 70.45%. Then, I integrated these three models into a ensemble classifier by soft voting method, and its accuracy rate was 81.82%. Finally, I compared the ensemble classifier with YOLOv7, and found that the accuracy of YOLOv7 was 91.26% which was quite good. The results show that both the deep learning model and the traditional machine learning model can be used for bird classification, and the combination of feature extractor and classifier can further improve the accuracy of the model. The research results of this paper provide useful reference for the practical application of bird classification in Dongtan and even around the world.