Counting using deep learning regression gives value to ecological surveys.

Area of Science:

• Ecological image analysis
• Machine learning applications in ecology
• Automated object counting

Background:

• Ecological studies often require manual counting of objects, which is time-consuming and limits field or lab work.
• Digital imagery offers potential for automating counting tasks.
• Existing image-level annotations can be leveraged for automated counting without labeling individual objects.

Purpose of the Study:

• To develop and demonstrate a machine learning approach for automated object counting using image-level annotations.
• To apply deep learning regression to diverse ecological counting tasks: fish otolith daily growth rings and aerial seal counts.
• To assess the method's accuracy and potential for utilizing historical ecological data.

Main Methods:

• Implemented a deep learning regression model for automated counting.
• Applied the model to microscopic images of fish otoliths for growth ring counts.
• Applied the model to aerial imagery of seals for population counts, including targeted relabeling for performance improvement.

Main Results:

• Achieved high accuracy for fish otolith daily growth ring counts (RMSE of 3.40, R² of 0.92).
• Initial seal count performance showed lower accuracy (RMSE of 23.46, R² of 0.72) due to training data limitations.
• Performance significantly improved for seal counts (RMSE of 19.03, R² of 0.77) after relabeling 100 images based on model predictions.

Conclusions:

• The deep learning regression approach provides accurate and direct counts for ecological research.
• This method effectively automates counting tasks, reducing manual effort and time constraints.
• The approach adds value to historical ecological datasets by enabling automated analysis without individual object labeling.