0 avis
Otolith age estimation by machine learning approaches
Archive ouverte
Edité par CCSD -
International audience. Age and growth are primordial essential data in stock assessment and management. However, contracting experts for age estimation using calcified pieces costs several million euros annually. Yet, alternative methods exist for fish ageing using the otolith shape (i.e., otolith shape descriptors or Elliptic Fourier Analysis). The goal of this study is to use a new descriptor of the otolith shape with Mojette Transform as an input of k-Nearest Neighbors (k-NN), Random Forest (RF) and Multi-Layer Perceptron (MLP) classifiers. Mojette Transform is the exact discrete Radon transform used in tomographic reconstruction, image watermarking, or video compression. Its mathematical properties allow reducing the information and having enough redundancy to characterize the object/image according a sufficient numbers of projections from the binarized image. Each projection is the sum of pixel luminance crossed with a specific angle. For otoliths, this projection well reflects the succession of the growth segments. Preliminary experiments were conducted on 8578 plaice (Pleuronectes platessa) samples collected during the surveys CGFS and in the fishing markets from 2010 to 2017 covering the Eastern English Channel. The experts estimated the age from 0 to 8 years old. The calibrated image for the left sagittal otolith was realized for each fish. The image database was labeled by expert interpretation according to international rules. The recognition rate is based on the comparison with the different classifiers label and the expert data. After rescaling (Gray transform centering) and resizing (from 50x50 pixels), RF seemed to be the best classifier according to raw image or Mojette bins with a 51.9% error rate and increasing to 88.9 % error rate according to precision of 1 year. The database was built from all otolith images (n=8578) used for stock assessment without prior filtering or images cleaning, image quality (broken otoliths, dirty otoliths) impacted the results and must be evaluated.These results could be improved by optimizing machine learning parameters and by selecting discriminant projections.
Consulter en ligne
Suggestions
Du même auteur
