About iBenthos

Why we need iBenthos ?

The extent and condition of seagrass meadows around Australia and across nations of the Indo-Pacific is poorly quantified but increasingly recognised as a key asset providing key ecosystem service, including carbon storage and sequestration. Many national and regional organisations and local communities including indigenous ranger groups now routinely undertake seagrass monitoring.

However, lack of appropriate and scalable technologies is a key constraint for comprehensive mapping, monitoring and assessment of seagrass meadows. Remote sensing approaches – including the use of satellites and aerial and submersible drones – show promise but are often limited in optically deep or complex (turbid) waters and require on-site validation to ensure accuracy of mapping products.

IBenthos is a publicly accessible cloud-based software platform that has been developed for users across Australia and Indo-Pacific nations to assist in the mapping and monitoring of the extent, condition and ecosystem services provided by seagrass meadows.

How iBenthos Works

Underwater optical imaging has become an essential tool for studying and managing many aspects of coastal and marine ecosystems, such as enabling the rapid and comprehensive collection of information for mapping benthic habitats like seagrass. Recent advances in the automated image analysis techniques to classify objects of interest (e.g., seagrass) now allow rapid processing of images enabling comprehensive repeatable and timely actionable information that can be scaled to geographic areas of interest.

iBenthos harnesses machine learning (ML) models trained with regionally relevant image datasets to automatically detect seagrass, identify morphotypes and measure percent cover in underwater imagery.

iBenthos supports the analysis of data collected from subtidal diver and drop camera surveys, intertidal surveys, and towed and autonomous vehicles.

A user-friendly interface enables users to upload images and other data from field trips and analyse them to produce key metrics such as seagrass percent cover, composition, biomass, and carbon content. iBenthos also provides tools including map-based exploration, summary reporting tools and is developing AI based agents to automate reporting.

IBenthos enables and encourages the sharing, collaboration and release of datasets. Permissions (with optional user agreements), and mechanisms for the public release of data are also provided. Public datasets are licensed under a Creative Commons “CC BY 4.0” license.

What iBenthos does and doesn’t do

While the name implies a broader application, at present iBenthos is focused on providing an effective workflow for the analysis and mapping of seagrass across Indo-Pacific. With time we envisage adding other key benthic habitats, including corals.

With ML models constantly evolving – as better methods and more comprehensive datasets become available – our aim is to make a number of these models available to users through a consistent workflow and user interface.

In iBenthos we aim to provide some level of analytical capability at both the image and aggregate (for example site) levels. Exports functions and links to API’s are available such as through Google Earth Engine for producing and validating maps.

As the models available in iBenthos can be improved significantly through access to more images to train the models, we encourage users to make their datasets available for this purpose. Making the datasets available for training purposes is not the same as making the dataset available for public access which also encourages improvements in the model. If you are interested in optimising performance for your specific project/location, we encourage you to reach out and discuss your requirements with us.

Note that the models are publicly accessible, and a condition of use is that images uploaded by users may be utilised for model refinement. This ensures continual improvement of model performance.

There are some things iBenthos doesn’t including at present:

• Classification/detection of marine fauna
• Regionalised models by fine-tuning of seagrass models

A simple manual

iBenthos provides a straight forward workflow for scientists to organise survey data and process it with machine learning models to obtain regional insights about seagrass stattistics. A minimalist workflow would contain: Add a project -> upload images -> view and validate the uploaded images -> run Machine Learning jobs on the uploads -> view Machine Learning prediction results -> add a collection for reporting -> create Google Earth Engine mapping. The following sections provide more details on how each of the components involved works on iBenthos.

Project management

A Project is meant to capture multiple Uploads within a certain proximity. In practice, a project can be used to host a survey dataset captured in a study site, within which a user can create multiple uploads that correspond to transects. That said, one can use the Project/Upload structure to capture whatever spatial hierarchy as they see fit.

Survey image data upload and preliminary validation

Once on the Porject Summary page, the user can click the Upload button to access the web-based drag-n-drop upload tool to upload a new batch of image samples. Note that it’s a requirement for the images to be geotagged and timestamped in the exif as the minimum requirement for metadata parsing. The Author field in the exif is also recommended as it is used to record the creator of the image sample. with the Uploads tab on the left pane, one can access the list of historical uploads. Clicking on one of the uploads will bring the user to the data preview page where individual images and their metadata can be viewed and validated.

Machine learning model prediction and results viewing

On the Uploads tab, one can choose a Machine Learning model and start a prediction job for an upload by clicking the Classify button. Different models provide different analytical modalities, which include binary Presence/Absence, multi-class percentage cover estimates, or percentage cover level (low/medium/high) depending on the choice of model. The system will provide a rough time estimate for the job. Once started, the job will be running in the background and its status will be trackable on the Jobs page.

Once completed, clicking on a specific job on the job list will bring up the job details including job creator, time, sample size, model, and upload-level status aggregated from individual image-level results. Clicking View Results on the top-right will show all individual images, and clicking on an image will lead to the individual image-level results page, where results from multiple models for the same image can be viewed and compared if multiple models have been used to analyse the image.

Collection-based anaysis

Collection is another feature that iBenthos provides for flexible analysis and reporting for (regional) insights aggreagated from multiple projects. By clicking on the Add Collection button, one can create a Collection of images-level predictoin results from chosen projects, into an aggregated report. Projects chosen for the collection are required to have the same analytical modality/modalities to generate a meaningful aggregated report.

Model Garden and Playground

A Model Garden is provided to host all available Machine Learning models available on the platform for the user to run against their datasets, which includes two main features:

• Model Card: a model card include comprehensive information about a model’s lineage in terms of training dataset, annotation method, model architecture, performance metrics, input/output format, results interpretation, etc. This is meant to provide transparency about the a model’s strengths and weaknesses and promote informed/responsible usage of Machine Learning.
• Playground: from a model’s model card, one can click the Open in Playground to start an ad-hoc test on the model’s suitability for their usecase/data. It allows the user to direct upload images to be processed by the chosen model and see the results without going through the whole process of project or data management workflow. This is meant to be used a an initial assessment because one commits into uploading a large volume of images and using the full functionality of the iBenthos platform.

Large-scale Mapping with Google Earth Engine (GEE)

GEE provides the key ingredient for bridging the gap between field-collected sparse samples and a dense prediction at the regional scale. GEE mapping follows the following workflow: by taking in a group of training labels in the format of <lat,lon,label>, which in this case are the predicted presence/absence or coverage labels by iBenthos model from field image sample, GEE trains a model that associates these labels with corresponding satellite images pixels observed at a range of bands and time period. GEE then applies this model on satellite image pixels across the entire chosen region, hence generalising the predictions to pixels that don’t have field samples associated with and scaling up the analysis. Note that the quality of such generalisation process is highly dependent on sample size and heterogenuity. GEE results for regions closer/more similar to the orginal region covered by field samples will be more reliable than those for farther/more dissimilar regions.

Acknowledgment and acknowledging project

iBenthos was developed by CSIRO with the support of Google LLC and the Australian Government and following extensive surveys of seagrasses with many collaborators both around Australia, including JCU and UQ, and in- country partners across the Indo-Pacific; Indonesia, Thailand Fiji, and Timor Leste.

To cite the project please use the following paper:
Publication to be added