Identifying Habitat Elements from Bird Images Using Deep Convolutional Neural Networks.

Area of Science:

• Ornithology
• Computer Vision
• Ecological Informatics

Background:

• Digital technology has increased bird image data, posing processing challenges.
• Deep convolutional neural networks (DCNNs) show promise for automated image analysis.
• Habitat element recognition in bird images is crucial for ecological research but underexplored.

Purpose of the Study:

• To demonstrate the effectiveness of DCNNs for recognizing bird habitat elements.
• To introduce the Habitat Elements of Bird Images (HEOBs-10) dataset for benchmarking.
• To evaluate the performance of different DCNN models on this task.

Main Methods:

• Four DCNN models were trained end-to-end on image data.
• A novel dataset, HEOBs-10, comprising 10 categories of habitat elements was curated.
• Model performance was evaluated based on test accuracy.

Main Results:

• All tested DCNN models achieved good results in recognizing habitat elements.
• ResNet-152-based models achieved the highest test accuracy (95.52%).
• AlexNet-based models achieved the lowest test accuracy (89.48%).

Conclusions:

• DCNNs are efficient tools for automatically identifying bird habitat elements from images.
• The HEOBs-10 dataset provides a valuable benchmark for future research.
• Automated habitat element recognition can significantly aid studies on bird-habitat relationships.