Improving underwater localization accuracy with machine learning.

Area of Science:

• Acoustic sensor networks
• Machine learning applications
• Wildlife tracking technology

Background:

• Accurate tracking of fish passage is crucial for hydroelectric dam operations.
• Time-difference-of-arrival (TDOA) acoustic arrays are used for fish tracking.
• Localization errors can limit the effectiveness of TDOA systems.

Purpose of the Study:

• To calibrate localization errors in a TDOA-based acoustic sensor array.
• To improve the accuracy of tracking salmon passage through a hydroelectric dam.
• To enhance the performance of machine learning algorithms in sensor network calibration.

Main Methods:

• Applied machine learning classification and regression algorithms.
• Used approximate maximum likelihood for initial tag localization.
• Employed ensembles of classification trees to filter data points with large localization errors.
• Developed a machine-learned regression model for error calibration.

Main Results:

• Reduced median distance error by 50% for stationary tracks and 34% for mobile tracks.
• Extended sub-meter localization accuracy range from 100 m to 250 m horizontally.
• Significantly decreased median distance errors in the depth direction (e.g., 0.49 m to 0.04 m for stationary tracks).

Conclusions:

• Machine learning effectively calibrates TDOA sensor network errors.
• The developed methods enhance accuracy and range for tracking aquatic animal passage.
• Applicable to other TDOA-based sensor networks in stable environments.