Creating synthetic and site specific ML training datasets for conservation

Background / Projects / Team / Contact

Background

While many land agricultural producers wish to act as responsible stewards for biodiversity, there is often a paucity of data available on wildlife populations on their land. The United States Department of Agriculture (USDA), through the National Resource Conservation Service (NRCS), is seeking tractable techological and methodological solutions to this challenge. NRCS has provided support for this project in order to incentivize wildlife conservation by producers and stewards on working lands. Hiring specialist biologists to survey these lands is often prohibitively costly and can complicate activities on working lands. Our chosen solution is to provide site-specific year-long monitoring by pairing machine learning tools to sort images from an automated camera that can be deployed for up to 1 year.

Description

Project GitHub site contains all relevant information//methdology for the creation of synthetic training data using preserved herpteofauna specimen.

Project Github site includes detailed standard operating procedure (SOP) for creating a synthetic training data using Adobe Photoshop and Adobe After Effects. Also incuded in the SOP is a procedure for training a machine learning model on the Google Vertex API, and exporting all results to tabular format. Project Zenodo site includes example image datsets created for the project.

Citation

Fraser, S., C. Woodman, and C. J. Evelyn. 2022. Animating the dead. Github. 25 August 2022. https://github.com/Automated-Wildlife-CIG/animating-the-dead

Fraser, S., C. Woodman, and C. J. Evelyn. 2022. Creating Site Specific Synthetic ML Training Datasets for Conservation: Sample Synthetic Training Dataset [Resized]. Zenodo. 1 June 2022. DOI:10.5281/zenodo.6603940