The objective of precision aquaculture is to manage the observed status of the farms relative to a defined benchmark (e.g. projected biomass, fish welfare status, ambient dissolved oxygen levels.). Hence a key functionality of the IoT platform is the capability to manage various machine-learning models and integrate with the different data streams coming from sensors, weather data, and other open sources. A range of machine learning and mechanistic models relate to managing aquaculture operations. In particular, we focus on:

·       Mechanistic and data-driven models to predict fish health, biomass, and mortality based on information on feed and environmental stressors

·       Predictive models to inform on outbreaks of parasitic infections

·       Deep learning models to forecast oceanographic condition multiple days/weeks in advance

Fish feed is the most expensive part of aquaculture and causes environmental problems when excess product sinks to the bottom. Optimal supply of feed is a complex selection that includes feed composition, growth stage, biomarkers and environmental conditions. While mechanistic models have been developed that simulate growth rates based on feeding regime and environmental conditions, the nonlinear relationships and sensitivity to external events such as diseases or parasites, have made prediction difficult (Føre et. al. 2016).

A practical solution requires that prediction be based on observed status to maintain accuracy. One approach combines mechanistic models with observations using data assimilation – a mathematical technique that incorporates process knowledge encapsulated in a physics-based model with information from observations describing the current state of the system (described schematically below). As part of the GAIN project, we explore how a sophisticated data assimilation algorithm -- the Kalman Filter -- could help manage the cultivation of trouts in raceway systems

More details on the work are provided in our recent paper and it elucidates the complex marriage of mathematics and data to inform management (Royer et. al.).

A precision aquaculture implementation can be summarised as:

·       dynamic process models for individual fish growth based on feeding regime and environmental conditions are implemented

·       continuous update of model state based on actual fish position and/or biomass as measured by hydroacoustic sensors such as CageEye system described in section A or the Biosonics aquaculture biomass monitor.

Figure 1: Mechanistic models contain errors that increase with time due to model imperfections and deviations of forcing conditions from reality. Data assimilation minimizes these errors by correcting the model stats using new observations (from: Houser (2013))

Data assimilation concepts have seen enormous application since the 1960s as scientists aimed to update models using sparse sensor observations (Martin et. al. 2015). As sensors become more prevalent, data-intensive computing is continuing to transform industries and decisions (Hey et. al. 2009). Leveraging the large datasets being generated on aquaculture farms has multiple advantages, particularly related to extracting insight from highly complex nonlinear processes not amenable to encoding within a set of explanatory equations. An obvious case in aquaculture is fish health and in particular parasitic outbreaks.

Sea lice presence in salmon farms is a complex interplay of hydrodynamics, lice load, temperature, and position of the fish in the water column. Nonlinear, opaque relationships have traditionally made mechanistic modelling impractical. More recently, IBM, in collaboration with industry stakeholders, has implemented a deep learning model that collates data from multiple sources and predicts sea-lice outbreaks, termed “AquaCloud”. The model was fed with data on environmental conditions and lice counts from over 2,000 salmon cages along the Norwegian coast. Combining a dense network of environmental sensors and manual sampling (of lice count), the deep learning model provides a two-week-ahead prediction of lice count with 70% accuracy.

Machine learning-based models for geophysical processes is an active area of research. We recently developed and demonstrated a machine-learning surrogate model for a physics-based ocean-wave model (James et. al. 2018) The machine-learning model yielded enormous speedup (> five-thousand-fold) in computational time while maintaining accuracy that was well within the confidence bounds of the physics-based model. In effect, deep-learning-based approaches enable the transition of complex modelling systems from HPC to edge devices (naturally the training of the models is expensive, but once trained, deployment is cheap).

Much research is taking place at the interface of mechanistic and machine learning. The objective here is to incorporate the robustness and stability of physics-based models with the adaptability and accuracy of data-driven methods. Early implementations since the 1960s considered model and data blending using assimilation methods but today more sophisticated methods are based on flexible modelling platforms and advanced machine learning and data assimilation techniques.