Glossary

Definitions of common terms used in ReefCloud and associated documents

ReefCloud Data Portal

The ReefCloud Data Portal is the free-to-use website which supports coral reef monitoring teams globally to store, manage and analyse coral reef benthic imagery using machine learning. Once a user account is set-up, the Data Portal allows teams to build an AI model with an informative label set, import and annotate benthic photo data, and visualise and export results within projects. Project permissions are managed within the Data Portal, where varied levels of access to users within a team can be applied.

ReefCloud Public Dashboard

The ReefCloud Public Dashboard is a publicly viewable website where synthesised data on hard coral, macroalgae and benthic composition from select ReefCloud Projects is displayed. Given appropriate permissions, ReefCloud aims to share global trend data on coral reefs derived from ReefCloud Projects broadly using the Public Dashboard.

Project

A Project is a private space inside the ReefCloud Data Portal created for you and your collaborators to store and manage benthic imagery and metadata, annotate photo data, and visualise and export results related to your specific coral reef monitoring needs.

Site

A Site is a fixed geographic location (defined by a latitude and longitude) selected by users to make repeated measures over time as part of a coral reef monitoring sampling design. Within ReefCloud, a site is the minimum georeferenced spatial unit for a collection of photos. Sites remain fixed inside a Project, even if just visited once. Sites may contain multiple transects, at multiple depths, and may be sampled multiple times (see Survey below).

Survey

A Survey for the purposes of ReefCloud is a collection (such as a transect) of benthic facing photos nested within a Site, occurring at a specific defined point in time. Often, coral reef monitoring involves multiple surveys at a single site, and these may be repeated over time.

Annotation

Annotation means marking or labelling parts of an image to show what's there; in the case of ReefCloud, it refers to identifying what is under a sampling point (e.g., "coral", "algae", "sand"). Each of the labels is an annotation: by collecting lots of these labelled points, the AI learns how to recognise organisms automatically. "Annotating your photos" means working through your images to label human annotatable points.

Label Sets

A collection of labels (or codes) used to classify benthic taxa from images used by the human and machine. Each label represents a benthic category, group or substrate type. Label Sets can be customised per project and are essential for ensuring consistency in annotations.

Human Annotatable Points

The total number of Human Annotatable Points represents the number of point coordinates overlaid per image multiplied by the total number of images in each Project dataset. These Human Annotatable Points are annotated by the user to train the Model within a Project given a defined Label Set. Tip: users should aim to annotate between 10-15% of the available Human Annotatable Points (across surveys) to adequately train the Model.

The number of Human Annotatable Points and their arrangement across each photo are defined during the initial Project set-up and remain fixed for your Project.

To view and annotate these points within a Project, click the "Classify Images" tab. A point (or multiple points) can be selected and annotated with the appropriate label, enabling machine learning. How points are viewed (including size and opacity, as well as annotation status as Unclassified, Human Classified, Machine Classified) can be adjusted by the user using the 'paint bucket' button on the top right.

Transfer Learning Points

The total number of points annotated in any given image in ReefCloud is always 50. This number includes the Human Annotatable Points defined by the user when creating a Project. For example, if a Project is set-up to include 5 Human Annotatable Points per image, the machine will annotate an additional 45 randomly dispersed points per image, totalling 50 Transfer Learning Points per image (and so on). This allows for data standardisation across ReefCloud analysis and reporting.

Therefore, the total number of Transfer Learning Points represents the total number of points annotated by the machine multiplied by the total number of images in a given Project dataset.

The additional machine annotatable points cannot be seen by the user on a given image, but the information will be available in the classification data exported.

Model

Each ReefCloud AI Model is Project specific, building on prior learning from large benthic imagery datasets to become 'smarter' through Machine Learning given knowledge gained from Human Annotatable Points labelled by the user.

Tags

A tag can be used to give context to an data point within the Classifier without training the model. Tags can be applied to human annotated points, and can be filtered for during classification and export of data. Tags do not contribute to machine learning but can be used to group and manage annotation data based on user-defined criteria.

Quality Control

A tool used to review, confirm or improve annotations within a project for accuracy, completeness, and consistency.

Model Validation

The process of assessing the performance of the project's AI model by comparing its predictions against human annotations.

Machine is another term sometimes used to describe the AI model.

Each project has a label set which defines the labels used to classify benthic taxa within a project. Label set templates are provided when selecting your label set when creating a project, which might be appropriate for different questions, taxonomy experience or other considerations such as how much detail can be ascertained from a given image due to how it was collected (i.e., depth, image quality, field of view).

Each label within a project label set must be mapped to a Global Benthic Label Set. This allows for data to be visualised within the data portal, and data to be aggregated within the public dashboard.