Pre-trained models#

The currently available model checkpoints can be seen by running the command:

echofilter --list-checkpoints

All current checkpoints were trained on data acquired by FORCE.

Model checkpoints#

The architecture used for all current models is a U-Net with a backbone of 6 EfficientNet blocks in each direction (encoding and decoding). There are horizontal skip connections between compression and expansion blocks at the same spatial scale and a latent space of 32 channels throughout the network. The depth dimension of the input is halved (doubled) after each block, whilst the time dimension is halved (doubled) every other block.

For details about how the Echofilter models were trained, and our findings about their empirical performance, please consult our companion paper:

SC Lowe, LP McGarry, J Douglas, J Newport, S Oore, C Whidden, DJ Hasselman (2022). Echofilter: A Deep Learning Segmention Model Improves the Automation, Standardization, and Timeliness for Post-Processing Echosounder Data in Tidal Energy Streams. Front. Mar. Sci., 9, 1–21. doi: 10.3389/fmars.2022.867857.

An overview for of notable model checkpoints available in echofilter are provided below.

echofilter-v1_bifacing_700ep#

Trained on both upfacing stationary and downfacing mobile data.
Overall IoU performance of 99.15% on downfacing mobile and 93.0%–94.9% on upfacing stationary test data.
Default model checkpoint.

echofilter-v1_bifacing_300ep#

Trained on both upfacing stationary and downfacing mobile data.
Overall IoU performance of 99.02% on downfacing mobile and 93.2%–95.0% on upfacing stationary test data.

echofilter-v1_bifacing_100ep#

Trained on both upfacing stationary and downfacing mobile data.
Overall IoU performance of 98.93% on downfacing mobile and 93.5%–94.9% on upfacing stationary test data.
Sample outputs on upfacing stationary data were thoroughly verified via manual inspection by trained analysts.

echofilter-v1_upfacing_600ep#

Trained on upfacing stationary data only.
Overall IoU performance of 92.1%–95.1% on upfacing stationary test data.

echofilter-v1_upfacing_200ep#

Trained on upfacing stationary data only.
Overall IoU performance of 93.3%–95.1% on upfacing stationary test data.
Sample outputs thoroughly were thoroughly verified via manual inspection by trained analysts.

echofilter-v0.5_downfacing_300ep#

Trained on downfacing mobile data only.

Training Datasets#

The machine learning model was trained on upfacing stationary and downfacing mobile data provided by Fundy Ocean Research Centre for Energy (FORCE). The training and evaluation data is available for download. Queries regarding dataset access should be directed to FORCE, info@fundyforce.ca.

Stationary#

data collection

bottom-mounted stationary, autonomous

orientation

uplooking

echosounder

120 kHz Simrad WBAT

locations

FORCE tidal power demonstration site, Minas Passage
- 45°21’47.34”N 64°25’38.94”W
- December 2017 through November 2018
SMEC, Grand Passage
- 44°15’49.80”N 66°20’12.60”W
- December 2019 through January 2020

organization

FORCE

Mobile#