Bioelectric Field Measurements: Augmenting Image Recognition for Fisheries Management

Accurate and efficient stock assessment methods of commercially relevant fish species are extremely important toward sustainable fisheries management, which has significant economic implications. Optical data, such as benthic images generated by the Habitat Mapping Camera System (HabCam), have been successfully used for this purpose, but data analysis—involving manual annotations of thousands of sampled images—is time- and labor-intensive. Therefore, the automation of this process using machine learning implementations of image recognition has the potential to significantly improve stock assessment efficiency. However, the prevalence of naturally camouflaged species, such as the summer flounder (Paralichthys dentatus), limits the accuracy of these automated optical methods. Based on numerous previous studies demonstrating the production of measurable electric fields by various marine species, we propose the use of bioelectric field measurements to guide automated object detection algorithms, allowing thorough computational analysis on regions of high electrical activity. In this study, we investigate the feasibility of this approach by training an open-source convolutional neural network-based algorithm—You Only Look Once (YOLO)—on HabCam data for the detection of sea scallops, flatfish, skates, and roundfish. Two test datasets were created, containing raw HabCam images or cropped images that simulate the result of field measurement-assisted preprocessing, and the accuracies of detection with both datasets were compared: Results show improved recall—an increase of more than 7% in flatfish detections—as well as increased precision and intersection over union with the preprocessed test set, indicating a significant benefit from the incorporation of non-optical data. However, preliminary testing with fabricated electric field sensors shows limited sensitivity and ability to attenuate noise when compared against signal magnitudes reported in literature. Therefore, although there is a clear increase in accuracy stemming from the incorporation of non-optical data, bioelectric fields do not appear to be a feasible source of this data. Alternative sources should be explored in the future, as its incorporation shows the potential to significantly improve image recognition algorithms and streamline fisheries management.