This project set out to extend Ai.Fish artificial intelligence into an onboard system using low-powered computing devices also known as "Edge" technology. Onboard artificial intelligence supports near real-time analysis of video while the vessel is at sea. Such technology opens the door to information available to fishermen to make different decisions while fishing; real-time fisheries data available for remote transmission to support fishery management, compliance, and enforcement; and opportunities to improve fishing sustainability with data at scale.
Using our annotation library developed in prior work we were able to employ 718,539 gear annotations for fishing floats used in longline fishing in Hawaii to develop a model to detect gear. The number of floats deployed influences the number of hooks deployed and this in turn may impact the number of fish caught during any given haul. Detecting gear may also have other benefits including the ability to reconcile gear deployed with gear retrieved while at sea to reduce gear loss and pollution associated with lost gear. Float counts obtained through real-time analysis were compared with manual counts from human review of video footage after the trips were completed. Results were encouraging with average precision above 85%.
In Alaska, AI was able to successfully identify the key target species: Halibut and Sablefish. We also experience some success detecting Skates as bycatch species. However, at times, Sablefish were confused with Pacific Cod. In Hawaii, all true positive detections among the 222 clips reviewed were bigeye tuna, a key target species for the fishery, however, we again saw confusion between yellowfin and bigeye tunas for some detections. More work is required to enlarge our training dataset with sufficient examples of bycatch species now that we have consistent results for target species.
Edge system configuration
Further investigation of the potential value to fishermen of onboard systems will be beneficial.
Offboarding results computed at sea, monitoring device activity, and potentially installing software updates or fixes during at-sea operation rely on network connectivity via cellular or WiFi signal. Recent uptake of Starlink WiFi services by fishing fleets presents an opportunity but such services remain costly for fishermen and minimizing bandwidth as well as dealing with periods where the system is off to reduce costs are key challenges to feasible onboard analysis. We did successfully deliver a software patch to one Alaska vessel while at sea proving the potential for small software updates supporting real-time improvements.
Specialized algorithms tailored to fishing fleet needs and typical species paired with low-powered computing systems and reliable cameras represent a highly impactful opportunity in sustainable fishing.
