Difference between revisions of "Using AI to save endangered whales"
(17 intermediate revisions by the same user not shown) | |||
Line 1: | Line 1: | ||
− | [[File:Poc2.png|left|frameless|585x585px]] | + | ('''Français''': [[L’IA au secours des baleines en voie de disparition]])[[File:Poc2.png|left|frameless|585x585px]] |
Line 13: | Line 13: | ||
== The Challenge == | == The Challenge == | ||
+ | [[File:750x544 NARW SI Cases 0.jpg|thumb|'''<small>Annual North Atlantic Right Whale Serious Injury (SI) Cases of whales last seen alive, 2017-2021, U.S. and Canada</small>''' <ref>https://www.fisheries.noaa.gov/national/marine-life-distress/2017-2021-north-atlantic-right-whale-unusual-mortality-event</ref>.|228x228px]] | ||
+ | The North Atlantic Right Whale (NARW) is one of the most endangered whale species, with only about 366 remaining in the world. In 2017, 12 individuals died in the Gulf of St. Lawrence. The high mortality rate is mainly due to the collision with vessels and the entanglement with fishing gears. | ||
+ | |||
+ | Protection measures, which include vessel speed reduction, fishing closure, and investing in new acoustic technologies, were then put in place by the Department of Fisheries and Oceans (DFO) to prevent the recurrence of such events. | ||
+ | |||
+ | Passive Acoustic Monitoring (PAM) is an observation method where hydrophones are deployed in the ocean to capture sounds from the surrounding environment. Marine mammal acoustic experts periodically, 2 to 4 times a year, collect the recordings from the hydrophones. PAM has become a crucial tool in observing endangered whales. | ||
+ | |||
+ | Currently, acoustic data analysis is performed manually by marine mammal acoustic experts. However, '''manual recognition is tricky, resource-intensive, time-consuming, and very few scientists can perform it on the fly.''' | ||
+ | |||
+ | == The Solution == | ||
+ | [[File:NARW Detector1.png|thumb|258x258px|<small>'''AI-powered near real-time detector of NARW'''</small>]] | ||
+ | The deployed hydrophones can detect whale calls and transmit this information in near real-time, providing continuous round-the-clock information over the season. Automating the process of acoustic data analysis can result in near real-time detection of NARW. The main idea is to "teach<nowiki>''</nowiki> a computer to recognize the sounds of NARW from the acoustic recordings by identifying specific patterns in the data. Such a tool can drastically minimize to time consumed by marine biologists to perform manual acoustic data analysis from 14 days to 4 to 5 hours <ref>https://gcn.com/2020/04/ai-streamlines-acoustic-id-of-beluga-whales/291304/</ref>. | ||
+ | |||
+ | [[File:Narw detector.png|thumb|559x559px|'''<small>A predictive model for detecting NARW upcalls</small>''']] | ||
+ | Supported by the 2020 – 2021 Results Fund, a Proof of Concept (POC) was developed for a predictive model for the automated detection of NARW upcalls from acoustic data. The detection problem is formulated as a binary image classification problem where the image is a spectrogram. A spectrogram is a visual representation of the spectrum of frequencies of a signal as it varies with time and is considered an effective method of displaying marine mammal vocalizations. | ||
+ | |||
+ | The POC is developed using the Open source Ketos toolkit <ref>https://docs.meridian.cs.dal.ca/ketos/</ref>, provided by Dalhousie University, for underwater acoustic analysis. The AI team at DFO has also collaborated with researchers from Dalhousie University to further enhance the accuracy of the predictive model in differentiating between two kinds of whales: Humpback whales and North Atlantic Right whales. | ||
+ | |||
+ | Eventually, the goal is to develop a system that will monitor the sounds from whales' calls from a network of hydrophones every hour to detect whales and '''send a real-time warning to vessels to slow down or change their course when the whales are present.''' | ||
+ | |||
+ | == References == |
Latest revision as of 12:52, 31 December 2021
(Français: L’IA au secours des baleines en voie de disparition)
Computers can learn to recognize the sound of endangered whales. DFO has developed a predictive model to identify North Atlantic Right Whales from underwater acoustic data. The insights gained from the model can be used to develop a warning system for preventing vessels from fatally striking the endangered species.
The Challenge
![](/images/7/70/750x544_NARW_SI_Cases_0.jpg)
The North Atlantic Right Whale (NARW) is one of the most endangered whale species, with only about 366 remaining in the world. In 2017, 12 individuals died in the Gulf of St. Lawrence. The high mortality rate is mainly due to the collision with vessels and the entanglement with fishing gears.
Protection measures, which include vessel speed reduction, fishing closure, and investing in new acoustic technologies, were then put in place by the Department of Fisheries and Oceans (DFO) to prevent the recurrence of such events.
Passive Acoustic Monitoring (PAM) is an observation method where hydrophones are deployed in the ocean to capture sounds from the surrounding environment. Marine mammal acoustic experts periodically, 2 to 4 times a year, collect the recordings from the hydrophones. PAM has become a crucial tool in observing endangered whales.
Currently, acoustic data analysis is performed manually by marine mammal acoustic experts. However, manual recognition is tricky, resource-intensive, time-consuming, and very few scientists can perform it on the fly.
The Solution
The deployed hydrophones can detect whale calls and transmit this information in near real-time, providing continuous round-the-clock information over the season. Automating the process of acoustic data analysis can result in near real-time detection of NARW. The main idea is to "teach'' a computer to recognize the sounds of NARW from the acoustic recordings by identifying specific patterns in the data. Such a tool can drastically minimize to time consumed by marine biologists to perform manual acoustic data analysis from 14 days to 4 to 5 hours [2].
Supported by the 2020 – 2021 Results Fund, a Proof of Concept (POC) was developed for a predictive model for the automated detection of NARW upcalls from acoustic data. The detection problem is formulated as a binary image classification problem where the image is a spectrogram. A spectrogram is a visual representation of the spectrum of frequencies of a signal as it varies with time and is considered an effective method of displaying marine mammal vocalizations.
The POC is developed using the Open source Ketos toolkit [3], provided by Dalhousie University, for underwater acoustic analysis. The AI team at DFO has also collaborated with researchers from Dalhousie University to further enhance the accuracy of the predictive model in differentiating between two kinds of whales: Humpback whales and North Atlantic Right whales.
Eventually, the goal is to develop a system that will monitor the sounds from whales' calls from a network of hydrophones every hour to detect whales and send a real-time warning to vessels to slow down or change their course when the whales are present.