ReefNetAI 0.3.0 Beta: A stronger workflow from source data to insight

ReefNetAI 0.3.0 Beta is now available. This release strengthens the path from bringing reef survey data into the platform to reviewing annotations, generating masks, and training source-specific models. It also makes important product concepts clearer so teams can understand what data they are working with and what each action will do.

CoralNet data is now reusable

CoralNet import is now a dedicated source-data workflow. An import creates a reusable dataset with its own identity, status, image and annotation counts, label inventory, available scopes, artifacts, and import history. Teams can inspect what landed before choosing how to use it.

• Attach an imported dataset to an existing SiteVisit or create a new SiteVisit from it.
• Choose confirmed annotations when available, with an honest fallback to all available annotations.
• Share a dataset with viewer or editor access without sharing private CoralNet credentials.
• Check CoralNet for updates while preserving the last successful snapshot and attachment history.

Guided classification and segmentation

Once a dataset is attached to a SiteVisit, 0.3.0 Beta offers clearer starting points for downstream work. Teams can review imported point annotations in classification, generate segmentation masks from imported points, or start a clean workflow when imported annotations are not the right input. Coverage checks explain when selected images have partial or missing imported points before a job begins.

More dependable labelsets and assistive labeling

Imported label provenance is preserved in an immutable snapshot, while each SiteVisit receives an editable working labelset. Label grouping and mapping versions are carried through downstream workflows so training and reporting use the vocabulary the team expects.

Assistive mode has also been hardened across point classification and mask-cutout labeling. The interface now distinguishes live assistive suggestions from custom source-model predictions, respects active labelsets, and applies stronger read-only protections for viewers. Confirmed work can improve later suggestions within the correct context without changing unrelated projects.

Custom source models with two training profiles

SiteVisit-centered workflows can now train source-specific models using two product profiles: Standard for a lighter-weight path and High quality for a more intensive model. Teams can follow training progress, review evaluation results, and explicitly approve a trained version before using it. Approved models can support both classification predictions and segmentation mask-cutout suggestions.

The goal of 0.3.0 Beta is not simply to add more tools. It is to make the full workflow more reusable, understandable, and trustworthy.

Try 0.3.0 Beta

ReefNetAI remains in Beta, and feedback from researchers, monitoring teams, and conservation practitioners is essential. Explore the updated documentation, try the new workflows, and join us on Discord to share what works and where the platform can improve.