Hierarchical image classification using transfer learning to improve deep learning model performance for amazon...

Area of Science:

• Wildlife biology
• Computer science
• Machine learning

Background:

• Deep learning models show promise for wildlife classification, aiding population monitoring and trade analysis.
• Limited data, especially for rare species, hinders optimal deep learning model performance.
• Hierarchical classification has emerged as a strategy to enhance model performance with limited datasets.

Purpose of the Study:

• To apply hierarchical classification via transfer learning for improved identification of Amazon parrot species.
• To evaluate the effectiveness of a hierarchical model against a non-hierarchical model for Amazon parrot classification.

Main Methods:

• Developed a classification hierarchy based on diagnostic morphological features of Amazon parrots.
• Employed transfer learning techniques within the hierarchical framework.
• Evaluated model performance using metrics such as mean Average Precision (mAP).

Main Results:

• The hierarchical model achieved a higher mean Average Precision (mAP) of 0.944 compared to the non-hierarchical model's mAP of 0.908.
• Hierarchical classification demonstrated improved accuracy in distinguishing between morphologically similar Amazon parrot species.
• The proposed method effectively addressed challenges posed by limited wildlife data.

Conclusions:

• Hierarchical classification by transfer learning is a superior approach for identifying Amazon parrot species.
• This methodology offers a valuable tool for enhancing the monitoring of wild populations and global trade for conservation.
• The study highlights the potential of structured classification approaches in wildlife AI applications.