Logo Lanfrica

danieloluwoleola-pixel/pristine-X-tilapia-detection

Domain:

agriculture

Record type:

model
Creator:
dan
Host:
Africa's first AI-powered Tilapia detection model — PRISTINE X # PRISTINE X — Tilapia Detection Model v1 > Africa's first AI-powered underwater Nile Tilapia detection model, > built for sub-Saharan African farm conditions. ## What This Is A YOLOv8 model fine-tuned on Nile Tilapia (*Oreochromis niloticus*) underwater footage. Built as the first proof of concept for PRISTINE X — a precision aquaculture technology company developing AI monitoring systems for intensive RAS fish farms in sub-Saharan Africa. ## Model Performance (v1) | Metric | Score | |--------|-------| | mAP50 | 82.1% | | Precision | 81.8% | | Recall | 73.4% | | Training images | 44 | | Training time | <1 min (T4 GPU) | | Model size | 6.2MB | ## Generalisation Testing The model was evaluated on two independent video sources it had never seen during training. Results confirm the model generalises beyond its training distribution across different environments and water clarity conditions. | Test | Source | Frames | Detection Rate | Avg Confidence | |------|--------|--------|----------------|----------------| | Training | youtu.be/oN96Xs-JTnk | 44 | 100% | 0.820 | | Test Video 1 | youtu.be/Bqsy7enwtsI | 107 | 100% | 0.693 | | Test Video 2 | youtu.be/1Wy62mLNqWM | 147 | 100% | 0.585 | **Note:** Test Video 2 was filmed in murky RAS tank water — conditions representative of Nigerian intensive fish farm environments. The confidence drop from 0.820 → 0.693 → 0.585 across increasing water turbidity is expected and scientifically consistent. Detection rate remained 100% across all tests. ## The Story Behind This Generic object detection models (COCO-trained YOLOv8) were pointed at Nigerian Tilapia farm footage. Across 7,831 frames — zero fish detected. The model saw donuts, hot dogs, and birds instead. This model was built to fix that. Trained on 44 annotated frames of real Tilapia underwater footage, it correctly detects and labels every fish in real time — proof that the technology works and that the only missing ingredient is a large, properly collected Afri …