Effects of image data quality on a convolutional neural network trained in-tank fish detection model for recirculating aquaculture systems

• Published in: Computers and electronics in agriculture Vol. 205; p. 107644
• Main Authors: Ranjan, Rakesh, Sharrer, Kata, Tsukuda, Scott, Good, Christopher
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.02.2023
• Subjects: Artificial intelligence; Machine learning; Precision aquaculture; RAS; Underwater imaging; Precision aquaculture; Underwater imaging; RAS; Artificial intelligence; Machine learning
• ISSN: 0168-1699; 1872-7107
• DOI: 10.1016/j.compag.2023.107644

•A sensing platform was developed for underwater imagery data acquisition in RAS.•Effect of data quality on the fish detection model performance was investigated.•Sensor selection considerably influenced the model performance.•Light condition did not demonstrate the substantial effect.•Optimized fish detection model achieved satisfactory mAP and F1 score. Artificial intelligence can answer fish production-related questions and assist growers with important management decisions in recirculating aquaculture systems (RAS). However, convolutional neural network-aided machine learning approaches are data-intensive, with model accuracy subject to the input image quality. Underwater imagery data acquisition, relatively high fish density, and water turbidity impart major challenges in acquiring high-quality imagery data. This study was conducted to investigate the effects of sensor selection, image quality, data size, imaging conditions, and pre-processing operations on the machine learning model accuracy for fish detection under RAS production conditions. An imaging platform (RASense1.0) was developed with four off-the-shelf sensors customized for underwater image acquisition. Data acquired from the imaging sensors under two light conditions (i.e., Ambient and Supplemental) were arranged in sets of 100 images and annotated as partial and whole fish. The annotated images were augmented and trained using a one-stage YOLOv5 model. There was a substantial improvement in mean average precision (mAP) and F1 score while increasing the size of the image datasets up to 700 images and 80 epochs. Similarly, image augmentation substantially improved model accuracy for smaller dataset models trained with less than 700 images. Beyond this, there was no improvement in mAP (∼86 %). Sensor selection significantly affected model precision, recall, and mAP; however, light conditions did not demonstrate a considerable effect on model accuracy. While comparing the performance of the one-stage YOLOv5 against a two-stage Faster R-CNN, both models performed similarly in terms of mAP scores; however, training time for the former was 6–14 times lower than the latter.