3. AI Algorithms
The acoustic signal of the hydrophone is processed in 1 minute packages, which are immediately processed in-situ to get automatic detection of acoustic sources using AI/ML algorithms.
Table of contents
Table of contents
3.1 AI detections
The artificial intelligence algorithms used in HT are developed and continuously improved by the team. Currently, the HT is able to provide automatic detections of dolphins (several types), whales (two types) and vessels. The algorithm can detect multiple classes per timeframe, and as we can expect an unbalanced presence of our detection classes, plain accuracy is not the most accurate score to present here. The accuracy of the models can be given in terms of various metrics: precision, recall or a combined measured so-called F1-Score. The F1-Score represents a balance of precision; how often a positive prediction is actually expected to be correct (higher precision means fewer false positives), and recall; how many actual positive cases were successfully detected (higher recall means fewer false negatives). It provides a more realistic score than plain accuracy, which does not account for the trade-off between these two factors. The table and picture bellow illustrate the current precision, recall and F1-score of Hydrotwin AI models.
Table 1 - F1-Score of the classes detected by the AI algorithms.
Class
F1-Score
Dolphins
77%
Whales
64%
Vessels
80%
In the dashboard you can visualize the percentage of detections per class every 30 minutes, the default message rate of HT. If the duration of sampling is 12 minutes, and there is 50% activity of vessels, that means the AI algorithm detected that class for 50% of that sampling period (6 minutes)