An Integrated UAV-Based Computer Vision Framework for Non-Invasive Biomass Estimation of Red Tilapia (Oreochromis spp.)

Biomass estimation in aquaculture usually requires capturing and weighing fish. This study tested whether images from an unmanned aerial vehicle (UAV) can supply the mean weight of red tilapia (Oreochromis spp.) without handling the animals. A YOLO26 instance-segmentation cascade located the culture unit, the fish and a floating 15 × 35 cm panel; a per-image homography to the panel gave metric scale; a skeleton and B-spline gave curvilinear length; and a length–weight model gave mean weight, which a known census converted into biomass. Four culture units were flown at 7, 15 and 30 m, three times a day for 8 weeks, and 396 images from Weeks 6–8, unused in model development, formed the external set. Segmentation agreed with semi-automatic references (fish F1 = 0.97), automated and manual counts differed by one fish in 547, and the homography placed 330 control points with a mean error of 0.53 cm. The Week 8 estimate differed from the Week 9 harvest by 9.3% (mean absolute percentage error, MAPE; n = 4 units), the length-scale factor (1.333) behaved as an empirical constant, and baselines without images reached 4.1–6.3%. The framework tracks mean weight without handling fish and is best seen as a complement to periodic weighing.