2021-07-29 –, Green
Aquaculture, or the farmed production of fish and shellfish, has grown rapidly, from supplying just 7% of fish for human consumption in 1974 to more than half in 2016. Sustaining this rapid expansion requires data-driven management of the production process and environmental impacts. This talk presents a machine-learning-based exploration of environmental and fish behaviour datasets collected at three salmon farms in Norway, Scotland, and Canada using AutoML tools in Julia.
Data generated on modern aquaculture farms extend across a wide variety of forms. In situ sensors sample large numbers of environmental variables such as temperature, current velocity, dissolved oxygen (DO), chlorophyll and salinity. Remotely-sensed environmental data can sample much larger spatial domains and can be at the bay-scale – from land-based sensors such as CODAR-type HF radar – or at the global scale from satellite-based monitoring system. Informing on farm operations also requires sampling of animal variables such as size, clustering behaviour, and movement, and this is typically done using underwater technologies such as hydroacoustic technology, video monitoring, and aerial drone imagery. Further, there are large datasets of pertinent variables that are generated by numerical models such as weather or ocean circulation products. These datasets constitute huge data volumes with distinct characteristics. Integrating and extracting information from these disparate data sources (in scalable manner) are key to encapsulating the full dynamics of the farm environment and enabling effective management.
This paper presents an analysis of environmental and fish behaviour datasets collected at three salmon farms in Norway, Scotland, and Canada. Information on fish behaviour were collected using hydroacoustic sensors that sampled the vertical distribution of fish in a cage at high spatial and temporal resolution, while a network of environmental sensors characterised local site conditions. We present an analysis of the environmental and hydroacoustic datasets using the Julia open-source packages we developed: data were preprocessed and curated into time-aligned matrix form using TSML (https://github.com/IBM/TSML.jl), and machine learning pipelines were identified and implemented using Lale (https://github.com/IBM/Lale.jl).
Analysis enabled a quantitative investigation of the effects of environmental conditions on fish response together with information on drivers of anomalous fish response. Results demonstrated pronounced temporal variations in fish distribution as dictated by factors such as diurnal patterns, dynamics (currents and winds), and oxygen and temperature variations. Diurnal patterns driven by natural changes in light intensity were broadly similar across sites although this trend was ameliorated at the Norwegian site which was located inside the Arctic circle and experience 24 hours of daylight during summer months. Generally, fish occupied a deeper position in the cage during the day and were more tightly clustered; while at night, fish utilised more of the cage volume and were at a higher average position.
Analysis indicated that temperature was the primary environmental driver at two of the three sites. Temperature in the warmer summer months exhibited pronounced stratification before returning to a well-mixed temperature profile in September and October. During these stratified periods there was a tendency for fish to cluster to the warmer, upper portion of the cage and avoid colder temperatures. On the other hand, in reasonably homogeneous environments where temperature varies little with depth (such as at the Canada site during autumn), temperature did not influence the vertical distribution of salmon.
Variation in oxygen levels were most pronounced at the Canada site which showed consistently lower values than at other sites. Feature importance analysis indicated that dissolved oxygen values were the most important contributor to fish behaviour and in particular during periods of lower oxygen levels, a pronounced response was noted. Analysis indicated that fish moved towards the surfaces when values drop below 7mgL-1 which is in line with literature which reports reduced appetites and feeding in Atlantic salmon when values drop below this threshold.
Results presented in this paper indicate pronounced differences between sites and the need to consider these variations for farm management. One could readily use this analysis to quantify the difference between sites, and further to identify the fundamental drivers to these variations. This could be particularly valuable when comparing different farm systems such as inshore and offshore and the associated operational implications.
I am a research scientist at the IBM Research working on the following areas: AutoML, AutoAI, RL/ML Optimization, and Decision Optimization.
Fearghal O’Donncha is a research scientist at IBM Research — Ireland. His work focuses on applying simulation models, analytics, and machine learning techniques to assist industry operations. This encompasses developing and deploying simulation-based models, integrating sensor data from a variety of IoT platforms, developing AI-based models that extract value from sensor or expert data, and optimizing these tools to a variety of HPC and cloud-based platforms. He is an adjunct faculty member at the National University of Ireland, Galway.