Automatic detection of seaweed and mussels in the water column using Python scripts

Aquaculture installations become more abundant in large parts of the oceans, with the tendency to promote dual usage of marine space, combining windmill farms and aquaculture installations. In the Horizon Europe funded ULTFARMS project, research into acoustic detection of seaweed and mussel/oyster aquaculture using multibeam echosounders is conducted. Through the acquisition of multibeam water column data and conversion into point cloud data (containing x, y, z, intensity and beam number), automatic detection of relevant cultures is attempted. The conversion of the raw multibeam data into point cloud data is performed using commercial software packages (Qimera and AutoClean), but the open-source software “Ping” (https://github.com/themachinethatgoesping) is a promising candidate for future applications.

To obtain automatic detection and volume calculation, several steps are conducted using tailor-made Python scripts. First, the point cloud data are filtered based on their intensity values, discarding low-intensity scatterers and retaining aquaculture installations and (unfortunately) some noise. Second, noise and outliers are removed using statistical outlier removal. Both standard deviation of the point cloud data and outlier detection, deleting points with few neighboring points, is used to retain the dense point cloud areas. Thirdly, clustering of the data is introduced based on the intensity values or the proximity of points using unsupervised machine learning methods including K-means clustering (grouping points into predefined clusters based on their proximity to cluster centers), Gaussian Mixture Model (assigning points to clusters by modeling data as a mixture of probability distributions) or Hdbscan (automatically identifying clusters based on the  varying shapes and densities in a dataset). The result is clusters of seaweed or individual volumes of mussel aquaculture installations. Finally, for each cluster, the volume is calculated using weighted voxelization; each voxel is assigned a weight based on the number of points in the voxel. Voxels with a large weight are considered to be entirely consisting of aquaculture species, while those with a low weight are only partly filled and thus only partly considered in the total volume. In some instances, interpolation of datapoints between beam numbers is needed to obtain a sufficient resolution. This depends on the beam spacing and hence the ping frequency and vessel speed.

The scripts are still under development and improvements are still being implemented. Undoubtedly, being able to automatically detect volumes of clusters in aquaculture installations will prove to be a huge cost-reducing step in future aquaculture installations.