This study (open access) builds on a previously-published proof of concept. The authors are working towards producing a hand-held multi-mode scanner (combining fluorescence, visible, NIR, and short-wave IR spectroscopy) to support species verification of white fish fillets in business-to-business supply (currently reliant, largely, on visual recognition by experienced traders).
The explain that one of the key challenges in using machine learning for fish species identification is managing the large number of classes, as the variety of fish species is extensive. In their previous research, they introduced a novel multi-mode, highly multi-class machine learning framework based on a hierarchy of dispute models. This approach involved training a global model, and then recognizing groups of classes that have feature subspaces too similar for effective single-stage classification. By partitioning the overall space into smaller, distinct subspaces, they trained specialized models that are more tailored to these specific subsets of the dataset. In practice, the global model initially classified a sample to determine the appropriate subspace, while the dispute model then identified the precise species within that subspace.
The objective of this latest study was to apply this approach data acquired with the multi-mode handheld spectroscopy device. Tissue spectra were acquired at 25 positions on 68 fillets from 11 species, in both frozen and thawed states.
They report that feature-level fusion across the four spectroscopy modes enabled higher classification accuracy than any single mode alone. A global machine-learning model classified all species with 85 ± 2.8 %, while specialized dispute models for commonly misclassified species improved performance to 90 % ± 6.1 %. Individual models for thawed and frozen fillets achieved 90 ± 6.0 % and 90 ± 5.4 %, respectively, with dispute models in the thawed dataset increasing accuracy to 93 ± 4.3 %.
They conclude that their results demonstrate that portable multi-mode spectroscopy, combined with machine learning, can provide a fast, non-destructive and reliable tool for on-site fish species identification.
Comments