- Posted by Jimmy
So begins the chapter in Melville’s “Moby Dick” that documents the killing of a Right Whale. Named by whalers as the ‘right’ whale to hunt, these 70-ton mammals had enormous value for their plentiful oil and baleen. However, by the mid-nineteenth century, hunters had decimated their population to the brink of extinction.
The Right Whale still struggles for survival. With only an estimated 400 individuals left, they are the rarest of the large whales. Primarily located along the Eastern coast of the United States, the most significant modern threat has been boats and ocean congestion. Vessel strikes and entanglements in fishing gear account for nearly half of Right Whale deaths.
Countless efforts have been made to prevent these accidents. Believe it or not, the most promising have relied upon machine learning to make sense of oceanic noise. This technology allows vessels to identify (and ultimately avoid) the Right Whale when they come within earshot.
The Right Whale makes a unique call known as an “up-call” to let other whales know of their presence. Machine learning algorithms have been particularly adept at teasing the up-call apart from other environmental noises, especially those that come from boats. One method developed by researchers at Oakland University that uses tree-based classifiers had an 85% success rate in identifying the up-call.
The researchers’ detection scheme used spectrograms for visual and statistical renderings of sound frequencies.
The Oakland University researchers isolated the specific ‘path’ in the spectrogram that represents the up-call. With this path, we can determine features of the up-call, such as duration and maximum frequency. These features were then used in a tree-based classification, a predictive model that could determine whether some future sound was a Right Whale up-call.
Let’s dive in.
Sound data, like in Example (1), needs to be processed and cleared before we can use it further. This includes eliminating the weakest 80% of the sounds and clearing as much secondary noise as possible.
Next, we find paths that connect the points. We can imagine these to be the paths that a particle would take to travel from point 1 to point 2. (For more information on how the researchers calculated which specific path to use: see page 4.)
After eliminating any points that cannot be connected by paths, we repeat the same process. This will zero in on the particular band that corresponds to the up-call. Eventually it will look something like this:
Now that we have isolated the path, we can directly quantify several features of the up-call, including signal duration, minimum frequency, maximum bandwidth etc.
The scores for these features are normalized, and then they are used within a tree classification, a well-developed machine learning approach. After being trained with many Right Whale up-calls, this classification model ‘learns’ its sound features. This means that given some new sound A, the model can determine whether sound A’s features correspond to those features of an up-call. The algorithm will then classify sound A as either (1) up-call or (2) not an up-call.
This is good news for the Right Whale. As we develop better techniques to identify them, we can prevent accidents and help conservation efforts. Similar research using spectrogram correlation has been used to classify bird sounds in order to prevent bird strikes near airplanes. Taken together, these reinforce the fact that machine learning can solve real-world problems.