A novel automated approach for fish biomass estimation in turbid environments through deep learning, object detection, and regression

• Published in: Ecological informatics Vol. 81
• Main Authors: Jansi Rani, S.V., Ioannou, Iacovos, Swetha, R., Dhivya Lakshmi, R.M., Vassiliou, Vasos
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.07.2024
• Subjects: Fish biomass; mYOLOv8; Object detection; Regression; YOLOv8; CNN; CLAHE; YOLOv8; Regression; KNN; mYOLOv8; RMSE; SVM; IoU; MAE; SNR; DNA; Object detection; RCNN; Fish biomass; MSE
• ISSN: 1574-9541
• DOI: 10.1016/j.ecoinf.2024.102663

Estimating fish biomass is crucial in the fisheries sector, where traditional methods often harm fish through manual sampling and anesthetics. A non-invasive approach is introduced using underwater films to estimate fish biomass in turbid conditions. This study presents the “Aquatic WeightNet” dataset, targeting the Genetically Improved Formed Tilapia (GIFT) Tilapia species, and addresses the challenge of unclear images with preprocessing techniques like dehazing and Contrast Limited Adaptive Histogram Equalization (CLAHE). YOLOv8, a leading object detection model modified to accommodate the custom Aquatic WeightNet dataset's varied image sizes with five detection heads, P2 to P6, is employed, achieving a recall of 0.997 and a mean Average Precision (mAP) of 0.899 within the 50–95% Intersection over Union (IoU) range. Fish biomass estimation assesses depth, length, and width using regression models for calculation. A three-phase grid search identifies the most effective models, with the Extra Trees Regressor outperforming depth estimation with mean absolute error (MAE) of 0.63 and coefficient of determination (R2) of 0.87 and the Random Forest Regressor for length and width (MAE of 0.01 and R2 of 0.99). For biomass estimation, the Extra Trees Regressor again performs well (MAE of 0.004 and R2 of 0.99), which is critical for determining optimal feed quantities to enhance aquaculture efficiency. This study emphasizes a non-invasive method to estimate fish biomass, optimizing the effectiveness and ecological sustainability of fish farming in murky waters through advanced detection algorithms and robust regression models.