What is a BRUVS?
BRUVS stands for, Baited Remote Underwater Video Station
Why are BRUVS important?
Monitoring fish biodiversity and biomass are the best indicator of ecological health
Fisheries and aquaculture* $2.5b in 2013-14
Economy* $100 bn
Biodiversity* 13.86 mil km2
Why does BRUVS need to be automated?
BRUVS have become widely adopted as the standard tool for non-invasive fish sampling in both Australia (GlobalArchive) and across the world (Globalfinprint). An estimated 4000 hours of BRUVS imagery are collected annually in Australia, taking approximately 12000 hours to manually extract usable data, at a cost of approximately $1.4 million. Inefficiency in the current approach creates an analysis time-lag - and significant bottlenecks - between recording and the delivery of numerical information to researchers, managers and policy makers. Limited volumes of collected and analysed data remains a major disincentive in the adoption of video based techniques around the world, where major ecological events may be missed.
What is AFID?
A new Automated Fish Identification (AFID) platform is being developed for active machine learning and integrated into EventMeasure: AFID will be an active learning feedback system by taking expertly curated data from EventMeasure to train the latest machine learning algorithms. When a fish ecologist annotates BRUVS, EventMeasure will automatically query AFID for identified species, potential Max-N frames as well as a running MaxN count. The user will then have the ability to accept AFID classification and length measurements or correct any errors. When the user has completed their analysis, EventMeasure will report the results to AFID which will automatically re-train and relearn from the corrected annotations. This will result in AFID becoming more accurate over time as more analysis is completed and more data generated for machine learning algorithms.
Why do we need AFID?
The key challenges of AFID will be addressed and solved through the combination of machine learning algorithms with novel sensor technology in order to harness the expertise of trained ecologists. Our solution will automate fish length measurements, pre-screen videos, find potential MaxN frames, total fish count estimates and group similarly looking fish for fast bulk labelling. The outputs from machine learning and sensor technology will be incorporated into EventMeasure, ensuring continuity with current BRUVS workflows and future proofing through adaptation to emerging machine learning technologies with minimal disruption to current BRUVS workflows
How is AFID different?
AFID is an opensource Django based, Autonomous Fish Identification AFID platform. AFID can be run as a stand alone desktop application or hosted on a cloud platform such as Amazon Web Services (AWS). AFID will handle all of the autonomous machine learning analysis. AFID will include a plug-in system that will allow different machine learning algorithms to be used and customised as required by the customer, who may wish to design their own. AFID will be designed to work closely with EventMeasure through an Application Programming Interface (API). EventMeasure will continue to be adapted to incorporate AFID generated classifications and length measurements as well as export human based classifications back to AFID in order for AFID to retrain, relearn and improve - achieving active machine learning.
Why is opensource important?
As the members of our consortium have a background in research and see the value of making BRUVS research accessible to all scientific researchers. AFID and machine learning code will remain open source for research and non-profit use. This will prevent ‘vendor lock in’ and allow both ecologists and data scientists to use, contribute and improve the ML algorithms. Our open source approach is the major reason we have gained the support and commitment of time from the Australian BRUVS working group.
How does it work?
EventMeasure has been adapted to work with AFID and machine learning models. This raises new opportunities with regards to how ecologists (users) label fish in an active learning environment. The new proposed workflow is summarised:
User uploads BRUVS video to AFID
As a background task, AFID runs machine learning methods over video and autonomously:
Choses the most appropriate ML model based on metadata
Classifies all fish in every frame
Calculates MaxN, running MaxN & Average MaxN
Calculates Fish Lengths
Finds relative abundance
AFID Generates a summary report from step 2.
User opens EventMeasure and uses the summary report to skip empty frames and to target frames reported MaxN frames and species of interest.
EventMeasure queries AFID usingthe AFID API to download ML labelled frame data.
User accepts and corrects AFID results as required and adds any missing classifications or length measurements
User saves results which AFID consumes and triggers a new training event and the ML model is re-trained on newly updated data
Users have the ability to export data from AFID - results and video to Global Archive ( or similar ) adding to new community data sets.
How can I contribute?
AFID is currently in the feasibility stage of the BRII - AIMS challenge. If successfully funded, and you have data to contribute to public datasets, time to contribute to machine learning algorithms, or would like to be part of our stakeholder group, please email Dan Marrable