AI System for Aquaculture and Fish Farming
We develop AI systems for aquaculture that solve tasks impossible for manual control: real-time monitoring of thousands of fish, disease detection 12–24 hours before clinical symptoms, and feeding optimization based on cage biomass. Over 6 years, we have deployed solutions on 15 farms, achieving an average FCR reduction of 8% and a 20% decrease in mortality. Unlike traditional selective sampling once a month and visual inspection, the AI system provides a continuous picture: every 10 minutes you know the average weight, health status, and behavior of the school. Underwater cameras with illumination operate 24/7, transmitting video streams to an edge server for instant processing.
Key Challenges
- Biomass and growth monitoring: Know how many fish are in the cage and their average weight without harvesting. Traditional selective sampling provides data only once a month.
- Early disease detection: Ichthyophthirius, nodavirus — a 24-hour delay can wipe out the population.
- Feed optimization: Feed accounts for 50–70% of operational costs; overfeeding pollutes the water.
- Water quality: DO, pH, NH3, temperature are stress predictors. Hypoxia forecast 45–60 minutes ahead.
How Stereo Vision Estimates Biomass
The core of biomass estimation is stereo vision. An underwater stereo camera captures video, disparity reconstructs a 3D point cloud. Fish are detected, length and volume measured, and weight calculated using allometric equations.
Stereo Biomass Pipeline
- Stereo calibration (OpenCV: intrinsics, extrinsics, lens distortion)
- Stereo matching: SGBM or Deep Learning-based (PSMNet, CFNet) for turbid water
- Fish detection: YOLOv8 + DeepSORT/ByteTrack for tracking
- 3D segmentation: SAM + depth projection → 3D bbox
- Measurement: Euclidean distance between tail and head in 3D; volume via ellipsoid approximation
- Biomass: weight = a × length^b (calibrated parameters)
For Atlantic salmon, MAE of average weight estimation ≈ 4.1% versus ±8–12% for sampling-based methods. Measuring 500 fish takes 10 minutes of video vs 3 hours of manual sampling. Thus, AI biomass estimation is 2–3 times more accurate and 20 times faster. — According to independent testing by the Norwegian Institute Nofima, AI biomass estimation accuracy reaches 95%.
When turbidity exceeds 2 NTU, we use structured light (laser stripes) and domain adaptation on synthetic data with augmented turbidity. CLAHE preprocessing improves visibility by ~30%.
Why Fish Behavior Is the Best Disease Indicator
Sick fish change behavior before visible symptoms appear: reduced swimming speed, disrupted schooling patterns, staying near the surface. We extract features: average speed (optical flow + Kalman tracker), schooling density, depth distribution, time spent in the top 20% of the volume, synchrony index.
Model: Temporal Fusion Transformer on time series (window 6–24 hours). On a dataset with ichthyophthiriasis, AUC-ROC = 0.88, average warning time before clinical symptoms — 18 hours. Early detection reduces mortality by 15%, saving significant costs per cage per season.
Feeding Management
A surface camera detects uneaten feed (pellet detection via CV). Inputs: biomass, water temperature, FCR history. Output: optimal dose. LightGBM as baseline model, Stable-Baselines3 (RL) for advanced. On a salmon farm (12 cages, 6-month A/B test): AI-driven feeding reduced FCR from 1.32 to 1.19, saving 9.8% feed. This substantially cuts costs, as feed accounts for 50–70% of operational expenses.
Water Quality Monitoring and Event Prediction
IoT sensors and historical data. LSTM predicts DO drop 45–60 minutes before critical levels. Algal bloom predicted by Sentinel-2 satellite imagery (chlorophyll-a) plus local sensors. Yield forecasting builds on historical data and environmental conditions.
Technology Stack
| Task | Tools |
|---|---|
| Biomass estimation | OpenCV, YOLOv8, SAM, ByteTrack |
| Behavior analysis | PyTorch, LSTM/TFT, OpenCV |
| Feed management | LightGBM, Stable-Baselines3 |
| Water quality | InfluxDB, Grafana, LSTM |
| Edge deployment | NVIDIA Jetson Orin, TensorRT |
Comparison: AI vs Traditional Approach
| Parameter | AI System | Traditional Method |
|---|---|---|
| Biomass accuracy | MAE 4.1% | ±8–12% (sampling) |
| Time to measure 500 fish | 10 minutes | 3 hours |
| Disease warning | 18 hours before symptoms | after symptoms |
| FCR optimization | 8–10% savings | manual management |
| Water monitoring | continuous + forecast | once daily |
What Is Included
- Audit of current production: infrastructure assessment, KPI setting.
- Architecture design: sensor, camera, and edge device selection.
- ML pipeline development: data collection, model training, validation.
- Integration with existing systems (ERP, SCADA).
- On-site deployment (NVIDIA Jetson, RTSP cameras, sensors).
- Staff training and documentation.
- Post-release support: monitoring, model retraining.
Estimated Timelines
Biomass monitoring system (stereo camera + estimation): 10–16 weeks. Full platform (biomass + disease + feeding + water): 8–14 months. Cost is calculated individually per farm.
We guarantee stable operation — 99% uptime for edge servers and model accuracy as stated. Our experience: 6+ years in AI for agritech, 15 completed computer vision projects. Get a consultation for your project — we will prepare a commercial proposal within 3 business days. Request an individual assessment for your farm.