Collaborative Environmental Modelling

The Challenge

Managing threats to biodiversity in nearshore marine systems is particularly challenging. The majority of the biodiversity in marine systems is in nearshore environments, including seagrass beds and coral reefs. However, these highly productive environments are also within easy reach of recreational, commercial and subsistence fishers worldwide. These fishers are having significant impacts. Marine reserves and fisheries closures are often seen as the most effective management tool in nearshore systems. However, the mix of poaching and use of destructive fishing methods threaten to negate the effect of reserves and other spatial management tools.

Monitoring compliance and implementing effective enforcement is an essential ingredient in making coastal marine reserves more than paper parks. However, one of the central challenges for management agencies is establishing an effective monitoring and surveillance system, particularly in the context of large number of small vessels operating from informal coastal launch sites. This challenge is present from contexts as different as recreational fishing on Australia’s Great Barrier Reef to use of fertilizer-based explosives by coastal fishermen in Southeast Asia.

A key barrier in effective monitoring, control and surveillance in this context is the difficulty in monitoring small scale vessel activities. These vessels are typically too small to be detected using satellites. Tracking systems frequently used on industrial scale vessels are typically not applicable on small scale vessels, as they are often too basic to support the technology, and are exempted from regulations requiring tracking devices. One alternative approach is to shift from monitoring vessels to monitoring contexts, for instance using sensing technologies in and around marine protected areas.

Underwater sound is an ideal tool for place-based monitoring and surveillance in marine systems. The dynamics of underwater sound are relatively well understood, and allow for detection and geolocation of events, such as poaching in reserves or use of explosives for fishing. Sound can be monitored using underwater microphones, known as hydrophones, which are not visible at the surface. This has the benefit of avoiding theft or vandalism, and allowing enforcement authorities to respond to events without revealing the source of their knowledge. Sound carries a large amount of information and is produced continuously by activities. For instance, underwater sound can be used to identify types of engines, movement patterns, equipment such as compressors, and prohibited activities such as fishing with explosives.

While the hardware to record this sound is readily available, two key barriers prevent its widespread use: automated high-quality processing of sound data and reliable inexpensive communication of information. Overcoming these barriers is challenging. Machine learning can improve on classical signal processing to automate analysis, but both the sound environment and the target sounds vary, and thus it is not possible to deploy a static model. Similarly, communications are constrained by power, time available, and bandwidth. Thus, decision-makers need to be able to interact with the sensors in a step-wise conversation, exploring the data and information provided by the sensor, and shifting its behavior as their decision-making evolves. These challenges touch on key issues in the CINTEL FSP, as they require a conversation between the sensors and their users.

The MCS Analytics Team in O&A has developed a hydrophone system, based on a low-cost commercially available hydrophone (Figure 1). The system involves reprogramming the firmware on the hydrophone to do on-board processing, along with connection to a custom external power and communications unit. The team has also developed a prototype buoyancy engine, which can move the entire unit to the surface for communications and recovery, then sink it below the surface for ongoing monitoring.

The system is designed to be modular, and can operate in three modes: archival, real-time, and real-time covert. In archival mode the hydrophone is attached to a subsurface mooring, and deployed for a fixed period with subsequent recovery and processing (Figure 2a).

In real-time mode, the power and communications unit are attached to the hydrophone, and provides onboard processing and alerts on activities via satellite communications (Figure 2b). In real-time covert mode, the real-time unit ascends to the surface to transmit alerts and then descends back to depth for ongoing monitoring. This modular model allows maximum flexibility across different contexts, which facilitates minimum cost applications tailored to particular end users.

Sound data is currently processed in two ways. Archival data is processed using a recurrent neural network which takes in multiple frequency bands at a single time, and is able to identify sound profiles of different activities of interest. The system is embedded in a user environment with a graphical interface which allows a non-expert user to import the data, visualize the sound files, train the machine learning algorithm, and process the data for events of interest (Figure 3).

The real-time units process data onboard using digital signal processing, comparing the intensity of the sound at different frequencies across time to identify events of interest. The team is in the process of moving from more traditional signal processing on the real time units to implementing a simplified version of the DeepSound system on a circuit board designed for machine learning, and integrating that into the real time unit.

CSIRO has currently deployed hydrophones in South Africa, Australia, Malaysia, Indonesia, the UK overseas territories, and Niue. There are planned deployments in the Solomon Islands, Fiji, and Belize. These deployments targeted a range of issues, including fishing with explosives (Figure 4), poaching in reserves, incursions by foreign fishing vessels, and measurement of the intensity of users in marine environments.

exploring ways to enrich information transmission using methods from interpretable machine learning literature and interaction with end users and experts; and
Develop a dynamic decision framework for switching among the multiple deep learning models developed by the team.

A project structured along these lines could be completed on the benchtop, using the existing system and actual sound data from the field. The Postdoctoral Fellow would work directly with current users in Indonesia and Australia to research end user interactions and requirements, and design the structure of the dynamic behaviour and information compression to meet the needs of end users in scenarios drawn from real-world applications. The development of the embedded system would be supported by the existing technical team, allowing the Postdoctoral Fellow to focus on the research aspects of the two topics above.