AI-Powered Sustainable Fishing with LLM-Enhanced Domain Knowledge Integration

furuno · Other · year n/a · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Sustainable fishing and marine electronics

Furuno wanted to help fishermen identify fish species and size from fish-finder and onboard vessel signals so they could avoid juvenile or restricted catches and fish more sustainably despite severe labeled-data scarcity.

Constraints:

• Japanese waters include more than 300 fish species, but many species have only around 100 data points
• traditional image-recognition training would require hundreds of thousands of examples
• fishing vessels often operate far from land with old, non-internet-connected equipment
• the most valuable knowledge is tacit fisherman intuition that is hard to translate into data
• models need to run practically at the edge on vessels

Architecture

Furuno proposes an ensemble AI system for sustainable fishing. A fish-finder image recognition model distinguishes visual fish-finder patterns, a classifier uses swimming depth and ultrasonic signals, and a fisherman-specific knowledge model weights local seasonal and behavioral expertise. LLMs make it easier for fishermen to input tacit knowledge as text or voice, while satellite communications and broader equipment collection expand inputs from a small set of sources toward more than 20. The system is intended to run on laptop PCs as edge devices aboard vessels and eventually be sold as an add-on subscription service.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  onboard-fish-finder["onboard-fish-finder"] --> image-recognition-model["image-recognition-model"]
  onboard-fish-finder["onboard-fish-finder"] --> classifier-model["classifier-model"]
  onboard-equipment-data["onboard-equipment-data"] --> classifier-model["classifier-model"]
  onboard-equipment-data["onboard-equipment-data"] --> ensemble-decision-layer["ensemble-decision-layer"]
  fisherman-knowledge-input["fisherman-knowledge-input"] --> knowledge-model["knowledge-model"]
  image-recognition-model["image-recognition-model"] --> ensemble-decision-layer["ensemble-decision-layer"]
  classifier-model["classifier-model"] --> ensemble-decision-layer["ensemble-decision-layer"]
  knowledge-model["knowledge-model"] --> ensemble-decision-layer["ensemble-decision-layer"]
  knowledge-model["knowledge-model"] --> fisherman-knowledge-input["fisherman-knowledge-input"]
  ensemble-decision-layer["ensemble-decision-layer"] --> edge-laptop-runtime["edge-laptop-runtime"]

Components

Data flow

A vessel collects fish-finder images, ultrasonic signals, swimming-depth cues, and other onboard equipment data. The image model analyzes fish-finder visual appearance, the classifier analyzes depth and ultrasonic signatures, and the knowledge model applies fisherman-specific context such as seasonal migration or target species rules. LLM-supported knowledge capture lets fishermen enter intuitive expertise that was previously hard to formalize. The ensemble combines these signals to recommend whether fish are likely target species and size, helping fishermen decide when to cast nets or wait.

Why this architecture

Furuno chose an ensemble with fisherman-specific knowledge because pure image recognition cannot handle hundreds of species with tiny labeled datasets. Limiting models to a few local target species reduces noise and data requirements, while the knowledge model captures what experienced fishermen already know but cannot easily encode. LLM input lowers the friction of turning tacit knowledge into usable data, and edge laptop deployment fits vessels that are far from land and historically disconnected.

Models used: image recognition model · classifier model · knowledge model · LLM · ensemble model

Scale

• Furuno developed the world's first fish finder in 1948.
  “we developed the world first fish finder”transcript, stated

  “back in 1948”transcript, stated
• Furuno has more than seven decades of marine electronics history.
  “over the past seven decades our main focus has been on product selling manufacturing and delivering”transcript, stated
• Many species have only about 100 data points, while image recognition normally needs hundreds of thousands of examples.
  “image recognition requires hundred of thousand of example but the reality is that for many fish species we have only 100 of data points”transcript, stated
• Japan has over 300 caught fish species, making full labeled-data collection a century-scale problem.
  “with over 300 species of fish cat in Japan it's estimated that it would take more than a century to GA enough data”transcript, stated
• The team wants to expand information sources from the current two to over 20.
  “we are aiming to significantly broaden our information sources expanding from the current to to over 20”transcript, stated
• The planned add-on service price was about $300 per month per fisherman.
  “like a monthly is like $300 additional service to each fisherman that's their plan”transcript, stated
• Fishing vessels are described as about 20 years behind land-based industries such as agriculture.
  “fishing vessel are often said to R about 20 years behind landbased industri such as agriculture”transcript, stated

Alternatives considered (roads not taken)

Training a broad model across hundreds of fish species

Rejected because: There is not enough labeled data, and collecting enough across all species would take more than a century.

Relying only on fish-finder images

Rejected because: Visual signals alone are insufficient; depth, ultrasonic signals, seasonal patterns, and fisherman knowledge all matter.

Centralized cloud-only inference

Rejected because: Vessels operate offshore with onboard equipment constraints, so laptop PCs are becoming practical edge devices.

Patterns

fine-tuning-adaptation (model adaptation)

The model is adapted to each fisherman's target species and local knowledge instead of attempting a single universal classifier.

data-quality-loop (data strategy)

Fishermen contribute and refine structured knowledge over time, turning scarce labeled data and local expertise into a stronger model.

structured-output-contract (knowledge capture)

LLM-enabled knowledge input turns tacit fisherman knowledge in signal, text, and voice forms into structured model inputs.

Tradeoffs

Limit each model to a few local target species.

Chose Narrow, fisherman-specific segmentation over A universal classifier for hundreds of fish species.

Rationale: The narrow model reduces irrelevant training data and makes learning feasible under severe data scarcity.

Make fishermen owners of their models.

Chose Practitioner-owned knowledge models over Opaque expert-fed AI models.

Rationale: Ownership increases motivation to use, teach, and improve the AI rather than treating it as a black box.

Run the application on edge devices aboard vessels.

Chose Laptop PC edge runtime over Depending only on legacy fish-finder hardware or always-connected cloud systems.

Rationale: Modern laptop PCs are becoming common and powerful enough to run the AI application in ocean environments.

Gotchas (what bit them)

Labeled fish data is far too sparse for conventional image recognition.

Root cause: Many fish species have only around 100 data points, while image recognition generally needs hundreds of thousands of examples.

Resolution: Constrain models to local target species and combine image, classifier, and fisherman knowledge models.

Expert fishing knowledge was hard to convert into machine-usable data.

Root cause: Fishermen's decades of expertise are intricate, intuitive, and historically not captured by equipment data alone.

Resolution: Use LLMs and a knowledge model so fishermen can input tacit knowledge more directly.

Vessels have older, disconnected onboard equipment.

Root cause: Typical Japanese fishing vessels use multiple older machines that are not connected to the internet or equipped with AI.

Resolution: Use newer satellite communications to collect equipment data and laptop PCs as more capable edge devices.

Evaluation

Early evaluation appears based on fixed-net fishing use cases where the AI helps fishermen predict when fish arrive and therefore when to go catch, with future assessment tied to adoption and subscription value.

Metrics: ability to predict when fish arrive at fixed nets · fisherman productivity · model improvement from collected data · subscription viability around $300 per month

Outcomes

• Early fixed-net results were described as promising because fishermen could better know when to go catch.
• Furuno planned to monetize the system as an add-on service for fishermen at about $300 per month.
• The system remained a developing AI/service model rather than a fully reported production deployment.

The source cites more than 300 fish species in Japan, only about 100 data points for many species, a century-scale data-collection problem, expansion from two sources to over 20, and a planned $300 monthly service.