DEAL - DEcentrAlised Learning for automated image analysis and biodiversity monitoring

As recognised in the UN Decade of Ocean Science for Sustainable Development (2021-2030) programme (https://oceandecade.org/), frequent and high-quality ocean observations are critical for effective marine management and decision-making. However, at the current time, marine biodiversity remains poorly observed, with sampling methods heavily reliant on infrequent and expensive ship-based observations.
Significant advances have been made in developing marine autonomous imaging platforms that collect data for specific organisms, including microscopic plankton that sit at the base of the ocean food web, and sea floor biota. Such platforms can generate millions of images that have the potential to revolutionise our understanding of marine biodiversity, and to facilitate a step change in marine biodiversity monitoring, allowing fine-scale spatial and temporal trends to be resolved. However, for this revolution to be realised, first high-throughput and high-efficacy classification and analysis tools must also be developed.
Fully supervised machine learning models that are given sufficient, high-quality, labelled data have been demonstrated to achieve reliable, high efficacy classification results. This is a pre-requisite for any operational, automated classification system. The main challenge encountered in the field of marine biodiversity monitoring is obtaining sufficient labelled data to employ a supervised learning approach. At the present time, work on automated image classifiers is highly fragmented, and there is an absence of common standards. Many research teams working on plankton act as individuals, building bespoke classifiers which are trained against limited image data, often from a single instrument. The problem with this approach is that individual researchers are ultimately limited by the amount of data they have been able to collate and label, which leads to results with higher biases that inhibit their incorporation into operational biodiversity monitoring platforms. The models are also incapable of detecting new or rare organisms that weren’t present within the original training data, meaning they are unsuitable for studying important processes such as the emergence of invasive species. Meanwhile, centralised services which require users to share their image data with a single custodian by uploading it to a central server before the images are classified, raise concerns over privacy and ownership of data. The process is also highly inefficient, as it relies on the transfer of potentially large volumes of data and its duplication on multiple servers.
In this project, we will develop a web-based application which addresses these problems by using the Swarm Learning Framework which has been developed by our project partner Hewitt Packard Enterprise (HPE). Swarm Learning allows users to participate in a decentralised, collaborative network without a central server. It makes it possible for users to benefit from each other’s data and learnings, and to collaborate in the building of better classification models with lower biases. At the same time, the system preserves data privacy and reduces inefficiencies and carbon costs associated with the transfer and duplication of large volumes of data. By further partnering with world leaders in plankton and sea floor imaging, we will deliver two operational networks for classifying plankton and sea floor image data. We intend the tool and the initial networks we form to act as catalysts, helping to build communities in which data producers coalesce around a set of shared standards, and cooperate in making marine image data suitable for operational biodiversity monitoring.