DEEP LEARNING APPROACHES TO CLASSIFY AND TRACK AT-RISK FISH SPECIES

Abstract

In marine ecosystems, fish play a crucial role. They are linked to other organisms through the food web and other processes. Because they provide food, humans have had an especially close relationship with them for decades. Fish requires a healthy living environment to survive and grow. Due to poor living conditions, the numbers of several large fish species have decreased. Scientists have long recognized the importance of sufficient fish habitat in maintaining a healthy fish population. As a result, marine biologists and conservationists must detect and track fish species in the real world frequently to determine relative abundance and track population changes. For this purpose, one common method of doing so is to use acoustic transmitters which are surgically implanted in fish that transmit a unique id and sensor data. Energy companies are legislated to not harm species at risk around their fixed infrastructures. A species at risk cannot be tagged using conventional fish tracking technology. Therefore, without harming the species and obtaining the required information, the alternative idea is to use sonars, cameras, etc to collect the data and use deep learning algorithms to analyze the data for fish species detection, classification, and tracking. Accordingly, two deep learning models called YOLOv3 and Mask-RCNN were applied to acoustic images. Even different augmentation techniques such as hue, saturation, and random rotation were applied to achieve 0.73 mean average precision(mAp) using YOLOv3, and about 0.62 mean average precision(mAp) using Mask-RCNN at the intersection over union (IOU) 0.4. These results helped in understanding that deep learning models can be applied along with different augmentation techniques to achieve the best results on acoustic data. For tracking of fish species, YOLOv4 along with the integration of the Norfair tracking algorithm was tested on the Wells dam dataset. The model was able to achieve the maximum Multiple objects tracking accuracy (MOTA) of about 66.9% on 20fps videos. Our findings show that deep learning models can replace human effort in watching hundreds of thousands of videos for fish species detection and classification, and that tracking algorithms and video cameras can also replace fish tagging in some situations.