Abstract
EchoFlow is a three-stage, containerised workflow that converts raw Kongsberg EK80
echosounder files into human-readable echograms and machine-interpretable attention
maps. The EK80 is the dominant wideband echosounder from Kongsberg, the leading
manufacturer in fisheries acoustics; surveys with these instruments routinely produce terabytes
of multi-frequency data.
1. Conversion – raw .raw pings are decoded and calibrated to volume-back-scattering
strength with PyEcholab (Wall et al., 2018).
2. Pre-processing – echograms are contrast-stretched, down-sampled and tiled as PNG
images.
3. Inference – a Vision Transformer trained with DINO (Caron et al., 2021) yields perpatch attention heat-maps that highlight fish schools, seabed returns, and other salient
structures.
Each stage is encapsulated in its own Docker image and orchestrated with Docker Compose.
Continuous Integration (CI) ensures that a test file always produces at least one attention map
per echogram frequency. The test file can be inspected as artifacts from the CI actions.