Detecting the North Atlantic Right Whale Using Satellite Imagery

Abstract

The North Atlantic right whale (Eubalaena glacialis) is critically endangered, with only about 370 individuals remaining. Modern conservation efforts rely on accurate knowledge of their location and movements, and, while established surveying methods such as aerial survey flights produce high-quality data for this purpose, they can be costly and are unable to cover large areas. Satellite imaging has been proposed as an additional tool to aid in the detection and monitoring of the whales, allowing for much broader coverage at a lower cost, though at a reduced accuracy. This thesis describes the first attempts at observing the North Atlantic right whale in satellite imagery, and the development of an automated detector model, including a new form of training data.

On April 24th, 2021, concurrent WorldView-3 satellite imagery and aerial photographs were acquired in Cape Cod Bay, Massachusetts. Ideal environmental conditions and an abundance of whales in the area resulted in 39 whale observations in the imagery, which were confirmed by the aerial survey. It was demonstrated that North Atlantic right whales were fairly easily visible in 30 cm and 15 cm satellite imagery, and that they were able to be identified on a species level due to visible markings unique to the right whale.

While right whales are often easily visible in such satellite imagery, visually identifying whales in a large number of images would be a very slow and tedious task. To develop an automated whale detector model, a large number of examples of right whales in satellite imagery are needed to allow the model to "understand" all the different ways a whale can look in such imagery However, at the time only the 39 observations mentioned above were available. Here, aerial photographs were modified to resemble satellite imagery using a deep learning approach called Neural Style Transfer (NST), in which the style of an existing satellite image of a right whale is transferred to the content of an aerial photograph. A unique set of satellite/aerial 'reference pairs' was developed, allowing for direct comparison between actual satellite imagery and the newly developed 'satellite-like' NST images using image similarity metrics. This demonstrated that the NST images were significantly more similar to satellite imagery than unmodified aerial photographs, and allowed for the immediate increase of examples of right whales in 'satellite imagery' from 39 to many thousands.

The NST images were directly compared to other types of training imagery, including unmodified aerial photographs, colour-normalized aerial photographs, and satellite imagery itself, by training a detector model using each of these training data types and comparing detection accuracies between them. It was found that models trained on NST simulated images slightly outperformed those trained on colour normalized photographs, while both significantly outperformed unmodified aerial photographs. Models trained on satellite imagery had an excellent precision but rather poor recall, likely due to the small amount of training data relative to the other datasets. These results indicate that performing some type of preprocessing modification to aerial photographs before training is highly desirable, though the trade-off between a slight increase in accuracy with NST and the significantly lower preprocessing time with colour normalized photographs will be a decision point for the end user.