Background

Feed makes up a large fraction of the total costs in Atlantic salmon (Salmo salar L.) cage culture. With feed loss from commercial farming sites at 5-7%, minimizing the amount of wasted feed is important both for economic reasons and to reduce the environmental impact of salmon farming. To achieve this, a better understanding of the dynamic processes involved in the feeding process is needed. For this purpose, the dynamics of the feeding process are studied through further development of a model published by Alver et al. (2004), where the feed concentration is calculated in a 3D grid. The model takes into account the physical properties of the feed, the geometry of the cage, the properties of the feed spreader, environmental factors such as wind and current, as well as the feeding behavior of the fish.

Methods

Feed spreader patterns have been parameterized based on data from Oehme et al. (2012). The data was fitted to skewed normal distributions in the forwards and backwards directions, and a 360 degrees distribution was computed based on interpolation between the two measured directions.

The original pellet distribution model has been further developed into a 3D model, by using a 3D version of the transport equation. Some further modifications were needed to make the model work in three dimensions. Handling of 3D cage geometry was added, such as calculation of feed wastage when feed moves outside of the designated cage volume.

Validation of the model will be done in several steps: 1) Validation of physical pellet spreading properties with no fish present. 2) Validation with fish present in small sea cages. 3) Validation in full scale cages. The first of these steps is ongoing, while step 2 and 3 are planned autumn 2013/winter 2014. In all validation steps instrumentation is required to provide actual measurements of the rate of pellets passing through a given cage volume as a function of time. A pellet sensor based on underwater machine vision technology has been developed to serve this requirement.

Machine vision pellet detection systems have been developed and investigated previously, but for the purpose of high-resolution model validation current solutions seem inadequate. A high accuracy detection system with respect to spatial and temporal distribution of pellets in the water column during a feeding bout is required. As opposed to many commercial systems, which measure pellet loss beneath the feeding depth, this sensor must be able to measure the pellet flux at any location within the cage. 

Results

A prototype machine vision pellet detection system has been designed and implemented. The pellet sensor itself consists of a high definition digital camera, positioned horizontally, and pointed against a uniformly backlit opaque surface. Tests have shown that the backlighting approach yields a high-contrast image of passing pellets while avoiding the adverse effects of ambient light variations and an unpredictable image background. A funnel allows for defining the extent of the horizontal area covered by the pellet sensor, and guides the pellets into the machine vision detection system. All objects in the camera’s field of view are detected using blob-analysis and their motion is individually tracked using a Kalman filter. The image analysis algorithm enables filtering on a range of variables, such as pellet sinking speed, projected area and convexity. This makes it possible for the detection system to separate pellets from foreign objects, as well as detecting overlapping pellets.