ZenRiver game concept: accelerating creation of machine learning imagery training datasets using citizen science Active
We aim to develop a web-based game where players use human-assisted image segmentation to produce annotated “meditation drawing” images of surface water sites to accelerate the creation of machine learning imagery training datasets. The game will also public education and outreach opportunities.
The USGS Observing Systems Division is exploring the use of image-based machine learning (ML) models for environmental monitoring, using imagery collected from networks of cameras. Image segmentation is a ML technique for labeling parts of images into groups of related pixels (e.g., all pixels representing water in an image of a river) to make further processing possible. To train image-based ML models, collections of annotated images, called training datasets are needed. The process of annotating images is currently a laborious task, resulting in a dearth of labeled images available to train USGS ML models. This project will prototype a web-based crowdsourcing tool using a game concept that will allow “players” to produce labeled images that can be used in the ML segmentation process. This could greatly accelerate the creation of these datasets. The tool will be advertised to internal colleagues and stakeholders and will present an opportunity to engage the public on this exciting task.
We aim to develop a web-based game where players use human-assisted image segmentation to produce annotated “meditation drawing” images of surface water sites to accelerate the creation of machine learning imagery training datasets. The game will also public education and outreach opportunities.
The USGS Observing Systems Division is exploring the use of image-based machine learning (ML) models for environmental monitoring, using imagery collected from networks of cameras. Image segmentation is a ML technique for labeling parts of images into groups of related pixels (e.g., all pixels representing water in an image of a river) to make further processing possible. To train image-based ML models, collections of annotated images, called training datasets are needed. The process of annotating images is currently a laborious task, resulting in a dearth of labeled images available to train USGS ML models. This project will prototype a web-based crowdsourcing tool using a game concept that will allow “players” to produce labeled images that can be used in the ML segmentation process. This could greatly accelerate the creation of these datasets. The tool will be advertised to internal colleagues and stakeholders and will present an opportunity to engage the public on this exciting task.